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"
17 #include "X86InstrBuilder.h"
18 #include "X86ISelLowering.h"
19 #include "X86TargetMachine.h"
20 #include "X86TargetObjectFile.h"
21 #include "Utils/X86ShuffleDecode.h"
22 #include "llvm/CallingConv.h"
23 #include "llvm/Constants.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/GlobalAlias.h"
26 #include "llvm/GlobalVariable.h"
27 #include "llvm/Function.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/Intrinsics.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/CodeGen/IntrinsicLowering.h"
32 #include "llvm/CodeGen/MachineFrameInfo.h"
33 #include "llvm/CodeGen/MachineFunction.h"
34 #include "llvm/CodeGen/MachineInstrBuilder.h"
35 #include "llvm/CodeGen/MachineJumpTableInfo.h"
36 #include "llvm/CodeGen/MachineModuleInfo.h"
37 #include "llvm/CodeGen/MachineRegisterInfo.h"
38 #include "llvm/CodeGen/PseudoSourceValue.h"
39 #include "llvm/MC/MCAsmInfo.h"
40 #include "llvm/MC/MCContext.h"
41 #include "llvm/MC/MCExpr.h"
42 #include "llvm/MC/MCSymbol.h"
43 #include "llvm/ADT/BitVector.h"
44 #include "llvm/ADT/SmallSet.h"
45 #include "llvm/ADT/Statistic.h"
46 #include "llvm/ADT/StringExtras.h"
47 #include "llvm/ADT/VectorExtras.h"
48 #include "llvm/Support/CallSite.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/Dwarf.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/raw_ostream.h"
55 using namespace dwarf;
57 STATISTIC(NumTailCalls, "Number of tail calls");
59 // Forward declarations.
60 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
63 static SDValue Insert128BitVector(SDValue Result,
69 static SDValue Extract128BitVector(SDValue Vec,
74 static SDValue ConcatVectors(SDValue Lower, SDValue Upper, SelectionDAG &DAG);
77 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
78 /// sets things up to match to an AVX VEXTRACTF128 instruction or a
79 /// simple subregister reference. Idx is an index in the 128 bits we
80 /// want. It need not be aligned to a 128-bit bounday. That makes
81 /// lowering EXTRACT_VECTOR_ELT operations easier.
82 static SDValue Extract128BitVector(SDValue Vec,
86 EVT VT = Vec.getValueType();
87 assert(VT.getSizeInBits() == 256 && "Unexpected vector size!");
88 EVT ElVT = VT.getVectorElementType();
89 int Factor = VT.getSizeInBits()/128;
90 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
91 VT.getVectorNumElements()/Factor);
93 // Extract from UNDEF is UNDEF.
94 if (Vec.getOpcode() == ISD::UNDEF)
95 return DAG.getNode(ISD::UNDEF, dl, ResultVT);
97 if (isa<ConstantSDNode>(Idx)) {
98 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
100 // Extract the relevant 128 bits. Generate an EXTRACT_SUBVECTOR
101 // we can match to VEXTRACTF128.
102 unsigned ElemsPerChunk = 128 / ElVT.getSizeInBits();
104 // This is the index of the first element of the 128-bit chunk
106 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / 128)
109 SDValue VecIdx = DAG.getConstant(NormalizedIdxVal, MVT::i32);
110 SDValue Result = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec,
119 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
120 /// sets things up to match to an AVX VINSERTF128 instruction or a
121 /// simple superregister reference. Idx is an index in the 128 bits
122 /// we want. It need not be aligned to a 128-bit bounday. That makes
123 /// lowering INSERT_VECTOR_ELT operations easier.
124 static SDValue Insert128BitVector(SDValue Result,
129 if (isa<ConstantSDNode>(Idx)) {
130 EVT VT = Vec.getValueType();
131 assert(VT.getSizeInBits() == 128 && "Unexpected vector size!");
133 EVT ElVT = VT.getVectorElementType();
134 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
135 EVT ResultVT = Result.getValueType();
137 // Insert the relevant 128 bits.
138 unsigned ElemsPerChunk = 128/ElVT.getSizeInBits();
140 // This is the index of the first element of the 128-bit chunk
142 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/128)
145 SDValue VecIdx = DAG.getConstant(NormalizedIdxVal, MVT::i32);
146 Result = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec,
154 /// Given two vectors, concat them.
155 static SDValue ConcatVectors(SDValue Lower, SDValue Upper, SelectionDAG &DAG) {
156 DebugLoc dl = Lower.getDebugLoc();
158 assert(Lower.getValueType() == Upper.getValueType() && "Mismatched vectors!");
160 EVT VT = EVT::getVectorVT(*DAG.getContext(),
161 Lower.getValueType().getVectorElementType(),
162 Lower.getValueType().getVectorNumElements() * 2);
164 // TODO: Generalize to arbitrary vector length (this assumes 256-bit vectors).
165 assert(VT.getSizeInBits() == 256 && "Unsupported vector concat!");
167 // Insert the upper subvector.
168 SDValue Vec = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT), Upper,
170 // This is half the length of the result
171 // vector. Start inserting the upper 128
173 Lower.getValueType().getVectorNumElements(),
177 // Insert the lower subvector.
178 Vec = Insert128BitVector(Vec, Lower, DAG.getConstant(0, MVT::i32), DAG, dl);
182 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
183 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
184 bool is64Bit = Subtarget->is64Bit();
186 if (Subtarget->isTargetEnvMacho()) {
188 return new X8664_MachoTargetObjectFile();
189 return new TargetLoweringObjectFileMachO();
192 if (Subtarget->isTargetELF())
193 return new TargetLoweringObjectFileELF();
194 if (Subtarget->isTargetCOFF() && !Subtarget->isTargetEnvMacho())
195 return new TargetLoweringObjectFileCOFF();
196 llvm_unreachable("unknown subtarget type");
199 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
200 : TargetLowering(TM, createTLOF(TM)) {
201 Subtarget = &TM.getSubtarget<X86Subtarget>();
202 X86ScalarSSEf64 = Subtarget->hasXMMInt();
203 X86ScalarSSEf32 = Subtarget->hasXMM();
204 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
206 RegInfo = TM.getRegisterInfo();
207 TD = getTargetData();
209 // Set up the TargetLowering object.
210 static MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
212 // X86 is weird, it always uses i8 for shift amounts and setcc results.
213 setBooleanContents(ZeroOrOneBooleanContent);
215 // For 64-bit since we have so many registers use the ILP scheduler, for
216 // 32-bit code use the register pressure specific scheduling.
217 if (Subtarget->is64Bit())
218 setSchedulingPreference(Sched::ILP);
220 setSchedulingPreference(Sched::RegPressure);
221 setStackPointerRegisterToSaveRestore(X86StackPtr);
223 if (Subtarget->isTargetWindows() && !Subtarget->isTargetCygMing()) {
224 // Setup Windows compiler runtime calls.
225 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
226 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
227 setLibcallName(RTLIB::SREM_I64, "_allrem");
228 setLibcallName(RTLIB::UREM_I64, "_aullrem");
229 setLibcallName(RTLIB::MUL_I64, "_allmul");
230 setLibcallName(RTLIB::FPTOUINT_F64_I64, "_ftol2");
231 setLibcallName(RTLIB::FPTOUINT_F32_I64, "_ftol2");
232 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
233 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
234 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
235 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
236 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
237 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::C);
238 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::C);
241 if (Subtarget->isTargetDarwin()) {
242 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
243 setUseUnderscoreSetJmp(false);
244 setUseUnderscoreLongJmp(false);
245 } else if (Subtarget->isTargetMingw()) {
246 // MS runtime is weird: it exports _setjmp, but longjmp!
247 setUseUnderscoreSetJmp(true);
248 setUseUnderscoreLongJmp(false);
250 setUseUnderscoreSetJmp(true);
251 setUseUnderscoreLongJmp(true);
254 // Set up the register classes.
255 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
256 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
257 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
258 if (Subtarget->is64Bit())
259 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
261 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
263 // We don't accept any truncstore of integer registers.
264 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
265 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
266 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
267 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
268 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
269 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
271 // SETOEQ and SETUNE require checking two conditions.
272 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
273 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
274 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
275 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
276 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
277 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
279 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
281 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
282 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
283 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
285 if (Subtarget->is64Bit()) {
286 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
287 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
288 } else if (!UseSoftFloat) {
289 // We have an algorithm for SSE2->double, and we turn this into a
290 // 64-bit FILD followed by conditional FADD for other targets.
291 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
292 // We have an algorithm for SSE2, and we turn this into a 64-bit
293 // FILD for other targets.
294 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
297 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
299 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
300 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
303 // SSE has no i16 to fp conversion, only i32
304 if (X86ScalarSSEf32) {
305 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
306 // f32 and f64 cases are Legal, f80 case is not
307 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
309 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
310 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
313 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
314 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
317 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
318 // are Legal, f80 is custom lowered.
319 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
320 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
322 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
324 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
325 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
327 if (X86ScalarSSEf32) {
328 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
329 // f32 and f64 cases are Legal, f80 case is not
330 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
332 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
333 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
336 // Handle FP_TO_UINT by promoting the destination to a larger signed
338 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
339 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
340 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
342 if (Subtarget->is64Bit()) {
343 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
344 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
345 } else if (!UseSoftFloat) {
346 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
347 // Expand FP_TO_UINT into a select.
348 // FIXME: We would like to use a Custom expander here eventually to do
349 // the optimal thing for SSE vs. the default expansion in the legalizer.
350 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
352 // With SSE3 we can use fisttpll to convert to a signed i64; without
353 // SSE, we're stuck with a fistpll.
354 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
357 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
358 if (!X86ScalarSSEf64) {
359 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
360 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
361 if (Subtarget->is64Bit()) {
362 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
363 // Without SSE, i64->f64 goes through memory.
364 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
368 // Scalar integer divide and remainder are lowered to use operations that
369 // produce two results, to match the available instructions. This exposes
370 // the two-result form to trivial CSE, which is able to combine x/y and x%y
371 // into a single instruction.
373 // Scalar integer multiply-high is also lowered to use two-result
374 // operations, to match the available instructions. However, plain multiply
375 // (low) operations are left as Legal, as there are single-result
376 // instructions for this in x86. Using the two-result multiply instructions
377 // when both high and low results are needed must be arranged by dagcombine.
378 for (unsigned i = 0, e = 4; i != e; ++i) {
380 setOperationAction(ISD::MULHS, VT, Expand);
381 setOperationAction(ISD::MULHU, VT, Expand);
382 setOperationAction(ISD::SDIV, VT, Expand);
383 setOperationAction(ISD::UDIV, VT, Expand);
384 setOperationAction(ISD::SREM, VT, Expand);
385 setOperationAction(ISD::UREM, VT, Expand);
387 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
388 setOperationAction(ISD::ADDC, VT, Custom);
389 setOperationAction(ISD::ADDE, VT, Custom);
390 setOperationAction(ISD::SUBC, VT, Custom);
391 setOperationAction(ISD::SUBE, VT, Custom);
394 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
395 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
396 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
397 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
398 if (Subtarget->is64Bit())
399 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
400 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
401 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
402 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
403 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
404 setOperationAction(ISD::FREM , MVT::f32 , Expand);
405 setOperationAction(ISD::FREM , MVT::f64 , Expand);
406 setOperationAction(ISD::FREM , MVT::f80 , Expand);
407 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
409 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
410 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
411 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
412 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
413 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
414 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
415 if (Subtarget->is64Bit()) {
416 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
417 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
420 if (Subtarget->hasPOPCNT()) {
421 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
423 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
424 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
425 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
426 if (Subtarget->is64Bit())
427 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
430 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
431 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
433 // These should be promoted to a larger select which is supported.
434 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
435 // X86 wants to expand cmov itself.
436 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
437 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
438 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
439 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
440 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
441 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
442 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
443 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
444 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
445 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
446 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
447 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
448 if (Subtarget->is64Bit()) {
449 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
450 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
452 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
455 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
456 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
457 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
458 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
459 if (Subtarget->is64Bit())
460 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
461 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
462 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
463 if (Subtarget->is64Bit()) {
464 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
465 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
466 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
467 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
468 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
470 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
471 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
472 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
473 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
474 if (Subtarget->is64Bit()) {
475 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
476 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
477 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
480 if (Subtarget->hasXMM())
481 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
483 // We may not have a libcall for MEMBARRIER so we should lower this.
484 setOperationAction(ISD::MEMBARRIER , MVT::Other, Custom);
486 // On X86 and X86-64, atomic operations are lowered to locked instructions.
487 // Locked instructions, in turn, have implicit fence semantics (all memory
488 // operations are flushed before issuing the locked instruction, and they
489 // are not buffered), so we can fold away the common pattern of
490 // fence-atomic-fence.
491 setShouldFoldAtomicFences(true);
493 // Expand certain atomics
494 for (unsigned i = 0, e = 4; i != e; ++i) {
496 setOperationAction(ISD::ATOMIC_CMP_SWAP, VT, Custom);
497 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
500 if (!Subtarget->is64Bit()) {
501 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
502 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
503 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
504 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
505 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
506 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
507 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
510 // FIXME - use subtarget debug flags
511 if (!Subtarget->isTargetDarwin() &&
512 !Subtarget->isTargetELF() &&
513 !Subtarget->isTargetCygMing()) {
514 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
517 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
518 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
519 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
520 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
521 if (Subtarget->is64Bit()) {
522 setExceptionPointerRegister(X86::RAX);
523 setExceptionSelectorRegister(X86::RDX);
525 setExceptionPointerRegister(X86::EAX);
526 setExceptionSelectorRegister(X86::EDX);
528 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
529 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
531 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
533 setOperationAction(ISD::TRAP, MVT::Other, Legal);
535 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
536 setOperationAction(ISD::VASTART , MVT::Other, Custom);
537 setOperationAction(ISD::VAEND , MVT::Other, Expand);
538 if (Subtarget->is64Bit()) {
539 setOperationAction(ISD::VAARG , MVT::Other, Custom);
540 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
542 setOperationAction(ISD::VAARG , MVT::Other, Expand);
543 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
546 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
547 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
548 setOperationAction(ISD::DYNAMIC_STACKALLOC,
549 (Subtarget->is64Bit() ? MVT::i64 : MVT::i32),
550 (Subtarget->isTargetCOFF()
551 && !Subtarget->isTargetEnvMacho()
554 if (!UseSoftFloat && X86ScalarSSEf64) {
555 // f32 and f64 use SSE.
556 // Set up the FP register classes.
557 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
558 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
560 // Use ANDPD to simulate FABS.
561 setOperationAction(ISD::FABS , MVT::f64, Custom);
562 setOperationAction(ISD::FABS , MVT::f32, Custom);
564 // Use XORP to simulate FNEG.
565 setOperationAction(ISD::FNEG , MVT::f64, Custom);
566 setOperationAction(ISD::FNEG , MVT::f32, Custom);
568 // Use ANDPD and ORPD to simulate FCOPYSIGN.
569 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
570 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
572 // Lower this to FGETSIGNx86 plus an AND.
573 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
574 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
576 // We don't support sin/cos/fmod
577 setOperationAction(ISD::FSIN , MVT::f64, Expand);
578 setOperationAction(ISD::FCOS , MVT::f64, Expand);
579 setOperationAction(ISD::FSIN , MVT::f32, Expand);
580 setOperationAction(ISD::FCOS , MVT::f32, Expand);
582 // Expand FP immediates into loads from the stack, except for the special
584 addLegalFPImmediate(APFloat(+0.0)); // xorpd
585 addLegalFPImmediate(APFloat(+0.0f)); // xorps
586 } else if (!UseSoftFloat && X86ScalarSSEf32) {
587 // Use SSE for f32, x87 for f64.
588 // Set up the FP register classes.
589 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
590 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
592 // Use ANDPS to simulate FABS.
593 setOperationAction(ISD::FABS , MVT::f32, Custom);
595 // Use XORP to simulate FNEG.
596 setOperationAction(ISD::FNEG , MVT::f32, Custom);
598 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
600 // Use ANDPS and ORPS to simulate FCOPYSIGN.
601 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
602 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
604 // We don't support sin/cos/fmod
605 setOperationAction(ISD::FSIN , MVT::f32, Expand);
606 setOperationAction(ISD::FCOS , MVT::f32, Expand);
608 // Special cases we handle for FP constants.
609 addLegalFPImmediate(APFloat(+0.0f)); // xorps
610 addLegalFPImmediate(APFloat(+0.0)); // FLD0
611 addLegalFPImmediate(APFloat(+1.0)); // FLD1
612 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
613 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
616 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
617 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
619 } else if (!UseSoftFloat) {
620 // f32 and f64 in x87.
621 // Set up the FP register classes.
622 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
623 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
625 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
626 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
627 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
628 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
631 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
632 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
634 addLegalFPImmediate(APFloat(+0.0)); // FLD0
635 addLegalFPImmediate(APFloat(+1.0)); // FLD1
636 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
637 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
638 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
639 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
640 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
641 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
644 // We don't support FMA.
645 setOperationAction(ISD::FMA, MVT::f64, Expand);
646 setOperationAction(ISD::FMA, MVT::f32, Expand);
648 // Long double always uses X87.
650 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
651 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
652 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
654 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
655 addLegalFPImmediate(TmpFlt); // FLD0
657 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
660 APFloat TmpFlt2(+1.0);
661 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
663 addLegalFPImmediate(TmpFlt2); // FLD1
664 TmpFlt2.changeSign();
665 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
669 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
670 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
673 setOperationAction(ISD::FMA, MVT::f80, Expand);
676 // Always use a library call for pow.
677 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
678 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
679 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
681 setOperationAction(ISD::FLOG, MVT::f80, Expand);
682 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
683 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
684 setOperationAction(ISD::FEXP, MVT::f80, Expand);
685 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
687 // First set operation action for all vector types to either promote
688 // (for widening) or expand (for scalarization). Then we will selectively
689 // turn on ones that can be effectively codegen'd.
690 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
691 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
692 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
693 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
694 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
695 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
696 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
697 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
698 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
699 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
700 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
701 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
702 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
703 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
704 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
705 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
706 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
707 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
708 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
709 setOperationAction(ISD::INSERT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
710 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
711 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
712 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
713 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
714 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
715 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
716 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
717 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
718 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
719 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
720 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
721 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
722 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
723 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
724 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
725 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
726 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
727 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
728 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
729 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
730 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
731 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
732 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
733 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
734 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
735 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
736 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
737 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
738 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
739 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
740 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
741 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
742 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
743 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
744 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
745 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
746 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
747 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
748 setTruncStoreAction((MVT::SimpleValueType)VT,
749 (MVT::SimpleValueType)InnerVT, Expand);
750 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
751 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
752 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
755 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
756 // with -msoft-float, disable use of MMX as well.
757 if (!UseSoftFloat && Subtarget->hasMMX()) {
758 addRegisterClass(MVT::x86mmx, X86::VR64RegisterClass);
759 // No operations on x86mmx supported, everything uses intrinsics.
762 // MMX-sized vectors (other than x86mmx) are expected to be expanded
763 // into smaller operations.
764 setOperationAction(ISD::MULHS, MVT::v8i8, Expand);
765 setOperationAction(ISD::MULHS, MVT::v4i16, Expand);
766 setOperationAction(ISD::MULHS, MVT::v2i32, Expand);
767 setOperationAction(ISD::MULHS, MVT::v1i64, Expand);
768 setOperationAction(ISD::AND, MVT::v8i8, Expand);
769 setOperationAction(ISD::AND, MVT::v4i16, Expand);
770 setOperationAction(ISD::AND, MVT::v2i32, Expand);
771 setOperationAction(ISD::AND, MVT::v1i64, Expand);
772 setOperationAction(ISD::OR, MVT::v8i8, Expand);
773 setOperationAction(ISD::OR, MVT::v4i16, Expand);
774 setOperationAction(ISD::OR, MVT::v2i32, Expand);
775 setOperationAction(ISD::OR, MVT::v1i64, Expand);
776 setOperationAction(ISD::XOR, MVT::v8i8, Expand);
777 setOperationAction(ISD::XOR, MVT::v4i16, Expand);
778 setOperationAction(ISD::XOR, MVT::v2i32, Expand);
779 setOperationAction(ISD::XOR, MVT::v1i64, Expand);
780 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Expand);
781 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Expand);
782 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Expand);
783 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Expand);
784 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
785 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
786 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
787 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
788 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
789 setOperationAction(ISD::BITCAST, MVT::v8i8, Expand);
790 setOperationAction(ISD::BITCAST, MVT::v4i16, Expand);
791 setOperationAction(ISD::BITCAST, MVT::v2i32, Expand);
792 setOperationAction(ISD::BITCAST, MVT::v1i64, Expand);
794 if (!UseSoftFloat && Subtarget->hasXMM()) {
795 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
797 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
798 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
799 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
800 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
801 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
802 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
803 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
804 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
805 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
806 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
807 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
808 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
811 if (!UseSoftFloat && Subtarget->hasXMMInt()) {
812 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
814 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
815 // registers cannot be used even for integer operations.
816 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
817 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
818 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
819 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
821 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
822 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
823 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
824 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
825 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
826 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
827 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
828 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
829 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
830 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
831 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
832 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
833 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
834 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
835 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
836 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
838 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
839 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
840 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
841 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
843 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
844 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
845 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
846 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
847 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
849 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
850 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
851 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
852 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
853 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
855 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
856 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
857 EVT VT = (MVT::SimpleValueType)i;
858 // Do not attempt to custom lower non-power-of-2 vectors
859 if (!isPowerOf2_32(VT.getVectorNumElements()))
861 // Do not attempt to custom lower non-128-bit vectors
862 if (!VT.is128BitVector())
864 setOperationAction(ISD::BUILD_VECTOR,
865 VT.getSimpleVT().SimpleTy, Custom);
866 setOperationAction(ISD::VECTOR_SHUFFLE,
867 VT.getSimpleVT().SimpleTy, Custom);
868 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
869 VT.getSimpleVT().SimpleTy, Custom);
872 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
873 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
874 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
875 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
876 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
877 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
879 if (Subtarget->is64Bit()) {
880 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
881 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
884 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
885 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
886 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
889 // Do not attempt to promote non-128-bit vectors
890 if (!VT.is128BitVector())
893 setOperationAction(ISD::AND, SVT, Promote);
894 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
895 setOperationAction(ISD::OR, SVT, Promote);
896 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
897 setOperationAction(ISD::XOR, SVT, Promote);
898 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
899 setOperationAction(ISD::LOAD, SVT, Promote);
900 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
901 setOperationAction(ISD::SELECT, SVT, Promote);
902 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
905 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
907 // Custom lower v2i64 and v2f64 selects.
908 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
909 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
910 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
911 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
913 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
914 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
917 if (Subtarget->hasSSE41()) {
918 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
919 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
920 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
921 setOperationAction(ISD::FRINT, MVT::f32, Legal);
922 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
923 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
924 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
925 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
926 setOperationAction(ISD::FRINT, MVT::f64, Legal);
927 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
929 // FIXME: Do we need to handle scalar-to-vector here?
930 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
932 // Can turn SHL into an integer multiply.
933 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
934 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
936 // i8 and i16 vectors are custom , because the source register and source
937 // source memory operand types are not the same width. f32 vectors are
938 // custom since the immediate controlling the insert encodes additional
940 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
941 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
942 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
943 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
945 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
946 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
947 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
948 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
950 if (Subtarget->is64Bit()) {
951 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
952 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
956 if (Subtarget->hasSSE2()) {
957 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
958 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
959 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
961 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
962 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
963 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
965 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
966 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
969 if (Subtarget->hasSSE42())
970 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
972 if (!UseSoftFloat && Subtarget->hasAVX()) {
973 addRegisterClass(MVT::v32i8, X86::VR256RegisterClass);
974 addRegisterClass(MVT::v16i16, X86::VR256RegisterClass);
975 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
976 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
977 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
978 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
980 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
981 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
982 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
984 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
985 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
986 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
987 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
988 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
989 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
991 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
992 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
993 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
994 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
995 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
996 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
998 // Custom lower several nodes for 256-bit types.
999 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1000 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
1001 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
1004 // Extract subvector is special because the value type
1005 // (result) is 128-bit but the source is 256-bit wide.
1006 if (VT.is128BitVector())
1007 setOperationAction(ISD::EXTRACT_SUBVECTOR, SVT, Custom);
1009 // Do not attempt to custom lower other non-256-bit vectors
1010 if (!VT.is256BitVector())
1013 setOperationAction(ISD::BUILD_VECTOR, SVT, Custom);
1014 setOperationAction(ISD::VECTOR_SHUFFLE, SVT, Custom);
1015 setOperationAction(ISD::INSERT_VECTOR_ELT, SVT, Custom);
1016 setOperationAction(ISD::EXTRACT_VECTOR_ELT, SVT, Custom);
1017 setOperationAction(ISD::SCALAR_TO_VECTOR, SVT, Custom);
1018 setOperationAction(ISD::INSERT_SUBVECTOR, SVT, Custom);
1021 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1022 for (unsigned i = (unsigned)MVT::v32i8; i != (unsigned)MVT::v4i64; ++i) {
1023 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
1026 // Do not attempt to promote non-256-bit vectors
1027 if (!VT.is256BitVector())
1030 setOperationAction(ISD::AND, SVT, Promote);
1031 AddPromotedToType (ISD::AND, SVT, MVT::v4i64);
1032 setOperationAction(ISD::OR, SVT, Promote);
1033 AddPromotedToType (ISD::OR, SVT, MVT::v4i64);
1034 setOperationAction(ISD::XOR, SVT, Promote);
1035 AddPromotedToType (ISD::XOR, SVT, MVT::v4i64);
1036 setOperationAction(ISD::LOAD, SVT, Promote);
1037 AddPromotedToType (ISD::LOAD, SVT, MVT::v4i64);
1038 setOperationAction(ISD::SELECT, SVT, Promote);
1039 AddPromotedToType (ISD::SELECT, SVT, MVT::v4i64);
1043 // SIGN_EXTEND_INREGs are evaluated by the extend type. Handle the expansion
1044 // of this type with custom code.
1045 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1046 VT != (unsigned)MVT::LAST_VECTOR_VALUETYPE; VT++) {
1047 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT, Custom);
1050 // We want to custom lower some of our intrinsics.
1051 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1054 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1055 // handle type legalization for these operations here.
1057 // FIXME: We really should do custom legalization for addition and
1058 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1059 // than generic legalization for 64-bit multiplication-with-overflow, though.
1060 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1061 // Add/Sub/Mul with overflow operations are custom lowered.
1063 setOperationAction(ISD::SADDO, VT, Custom);
1064 setOperationAction(ISD::UADDO, VT, Custom);
1065 setOperationAction(ISD::SSUBO, VT, Custom);
1066 setOperationAction(ISD::USUBO, VT, Custom);
1067 setOperationAction(ISD::SMULO, VT, Custom);
1068 setOperationAction(ISD::UMULO, VT, Custom);
1071 // There are no 8-bit 3-address imul/mul instructions
1072 setOperationAction(ISD::SMULO, MVT::i8, Expand);
1073 setOperationAction(ISD::UMULO, MVT::i8, Expand);
1075 if (!Subtarget->is64Bit()) {
1076 // These libcalls are not available in 32-bit.
1077 setLibcallName(RTLIB::SHL_I128, 0);
1078 setLibcallName(RTLIB::SRL_I128, 0);
1079 setLibcallName(RTLIB::SRA_I128, 0);
1082 // We have target-specific dag combine patterns for the following nodes:
1083 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1084 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1085 setTargetDAGCombine(ISD::BUILD_VECTOR);
1086 setTargetDAGCombine(ISD::SELECT);
1087 setTargetDAGCombine(ISD::SHL);
1088 setTargetDAGCombine(ISD::SRA);
1089 setTargetDAGCombine(ISD::SRL);
1090 setTargetDAGCombine(ISD::OR);
1091 setTargetDAGCombine(ISD::AND);
1092 setTargetDAGCombine(ISD::ADD);
1093 setTargetDAGCombine(ISD::SUB);
1094 setTargetDAGCombine(ISD::STORE);
1095 setTargetDAGCombine(ISD::ZERO_EXTEND);
1096 setTargetDAGCombine(ISD::SINT_TO_FP);
1097 if (Subtarget->is64Bit())
1098 setTargetDAGCombine(ISD::MUL);
1100 computeRegisterProperties();
1102 // On Darwin, -Os means optimize for size without hurting performance,
1103 // do not reduce the limit.
1104 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1105 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1106 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1107 maxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1108 maxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1109 maxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1110 setPrefLoopAlignment(16);
1111 benefitFromCodePlacementOpt = true;
1113 setPrefFunctionAlignment(4);
1117 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1122 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1123 /// the desired ByVal argument alignment.
1124 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1127 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1128 if (VTy->getBitWidth() == 128)
1130 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1131 unsigned EltAlign = 0;
1132 getMaxByValAlign(ATy->getElementType(), EltAlign);
1133 if (EltAlign > MaxAlign)
1134 MaxAlign = EltAlign;
1135 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1136 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1137 unsigned EltAlign = 0;
1138 getMaxByValAlign(STy->getElementType(i), EltAlign);
1139 if (EltAlign > MaxAlign)
1140 MaxAlign = EltAlign;
1148 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1149 /// function arguments in the caller parameter area. For X86, aggregates
1150 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1151 /// are at 4-byte boundaries.
1152 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1153 if (Subtarget->is64Bit()) {
1154 // Max of 8 and alignment of type.
1155 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1162 if (Subtarget->hasXMM())
1163 getMaxByValAlign(Ty, Align);
1167 /// getOptimalMemOpType - Returns the target specific optimal type for load
1168 /// and store operations as a result of memset, memcpy, and memmove
1169 /// lowering. If DstAlign is zero that means it's safe to destination
1170 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1171 /// means there isn't a need to check it against alignment requirement,
1172 /// probably because the source does not need to be loaded. If
1173 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1174 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1175 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1176 /// constant so it does not need to be loaded.
1177 /// It returns EVT::Other if the type should be determined using generic
1178 /// target-independent logic.
1180 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1181 unsigned DstAlign, unsigned SrcAlign,
1182 bool NonScalarIntSafe,
1184 MachineFunction &MF) const {
1185 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1186 // linux. This is because the stack realignment code can't handle certain
1187 // cases like PR2962. This should be removed when PR2962 is fixed.
1188 const Function *F = MF.getFunction();
1189 if (NonScalarIntSafe &&
1190 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1192 (Subtarget->isUnalignedMemAccessFast() ||
1193 ((DstAlign == 0 || DstAlign >= 16) &&
1194 (SrcAlign == 0 || SrcAlign >= 16))) &&
1195 Subtarget->getStackAlignment() >= 16) {
1196 if (Subtarget->hasSSE2())
1198 if (Subtarget->hasSSE1())
1200 } else if (!MemcpyStrSrc && Size >= 8 &&
1201 !Subtarget->is64Bit() &&
1202 Subtarget->getStackAlignment() >= 8 &&
1203 Subtarget->hasXMMInt()) {
1204 // Do not use f64 to lower memcpy if source is string constant. It's
1205 // better to use i32 to avoid the loads.
1209 if (Subtarget->is64Bit() && Size >= 8)
1214 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1215 /// current function. The returned value is a member of the
1216 /// MachineJumpTableInfo::JTEntryKind enum.
1217 unsigned X86TargetLowering::getJumpTableEncoding() const {
1218 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1220 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1221 Subtarget->isPICStyleGOT())
1222 return MachineJumpTableInfo::EK_Custom32;
1224 // Otherwise, use the normal jump table encoding heuristics.
1225 return TargetLowering::getJumpTableEncoding();
1229 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1230 const MachineBasicBlock *MBB,
1231 unsigned uid,MCContext &Ctx) const{
1232 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1233 Subtarget->isPICStyleGOT());
1234 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1236 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1237 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1240 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1242 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1243 SelectionDAG &DAG) const {
1244 if (!Subtarget->is64Bit())
1245 // This doesn't have DebugLoc associated with it, but is not really the
1246 // same as a Register.
1247 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1251 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1252 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1254 const MCExpr *X86TargetLowering::
1255 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1256 MCContext &Ctx) const {
1257 // X86-64 uses RIP relative addressing based on the jump table label.
1258 if (Subtarget->isPICStyleRIPRel())
1259 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1261 // Otherwise, the reference is relative to the PIC base.
1262 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1265 // FIXME: Why this routine is here? Move to RegInfo!
1266 std::pair<const TargetRegisterClass*, uint8_t>
1267 X86TargetLowering::findRepresentativeClass(EVT VT) const{
1268 const TargetRegisterClass *RRC = 0;
1270 switch (VT.getSimpleVT().SimpleTy) {
1272 return TargetLowering::findRepresentativeClass(VT);
1273 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1274 RRC = (Subtarget->is64Bit()
1275 ? X86::GR64RegisterClass : X86::GR32RegisterClass);
1278 RRC = X86::VR64RegisterClass;
1280 case MVT::f32: case MVT::f64:
1281 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1282 case MVT::v4f32: case MVT::v2f64:
1283 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1285 RRC = X86::VR128RegisterClass;
1288 return std::make_pair(RRC, Cost);
1291 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1292 unsigned &Offset) const {
1293 if (!Subtarget->isTargetLinux())
1296 if (Subtarget->is64Bit()) {
1297 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1299 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1312 //===----------------------------------------------------------------------===//
1313 // Return Value Calling Convention Implementation
1314 //===----------------------------------------------------------------------===//
1316 #include "X86GenCallingConv.inc"
1319 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1320 MachineFunction &MF, bool isVarArg,
1321 const SmallVectorImpl<ISD::OutputArg> &Outs,
1322 LLVMContext &Context) const {
1323 SmallVector<CCValAssign, 16> RVLocs;
1324 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1326 return CCInfo.CheckReturn(Outs, RetCC_X86);
1330 X86TargetLowering::LowerReturn(SDValue Chain,
1331 CallingConv::ID CallConv, bool isVarArg,
1332 const SmallVectorImpl<ISD::OutputArg> &Outs,
1333 const SmallVectorImpl<SDValue> &OutVals,
1334 DebugLoc dl, SelectionDAG &DAG) const {
1335 MachineFunction &MF = DAG.getMachineFunction();
1336 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1338 SmallVector<CCValAssign, 16> RVLocs;
1339 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1340 RVLocs, *DAG.getContext());
1341 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1343 // Add the regs to the liveout set for the function.
1344 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1345 for (unsigned i = 0; i != RVLocs.size(); ++i)
1346 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1347 MRI.addLiveOut(RVLocs[i].getLocReg());
1351 SmallVector<SDValue, 6> RetOps;
1352 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1353 // Operand #1 = Bytes To Pop
1354 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1357 // Copy the result values into the output registers.
1358 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1359 CCValAssign &VA = RVLocs[i];
1360 assert(VA.isRegLoc() && "Can only return in registers!");
1361 SDValue ValToCopy = OutVals[i];
1362 EVT ValVT = ValToCopy.getValueType();
1364 // If this is x86-64, and we disabled SSE, we can't return FP values,
1365 // or SSE or MMX vectors.
1366 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1367 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1368 (Subtarget->is64Bit() && !Subtarget->hasXMM())) {
1369 report_fatal_error("SSE register return with SSE disabled");
1371 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1372 // llvm-gcc has never done it right and no one has noticed, so this
1373 // should be OK for now.
1374 if (ValVT == MVT::f64 &&
1375 (Subtarget->is64Bit() && !Subtarget->hasXMMInt()))
1376 report_fatal_error("SSE2 register return with SSE2 disabled");
1378 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1379 // the RET instruction and handled by the FP Stackifier.
1380 if (VA.getLocReg() == X86::ST0 ||
1381 VA.getLocReg() == X86::ST1) {
1382 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1383 // change the value to the FP stack register class.
1384 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1385 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1386 RetOps.push_back(ValToCopy);
1387 // Don't emit a copytoreg.
1391 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1392 // which is returned in RAX / RDX.
1393 if (Subtarget->is64Bit()) {
1394 if (ValVT == MVT::x86mmx) {
1395 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1396 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
1397 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1399 // If we don't have SSE2 available, convert to v4f32 so the generated
1400 // register is legal.
1401 if (!Subtarget->hasSSE2())
1402 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
1407 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1408 Flag = Chain.getValue(1);
1411 // The x86-64 ABI for returning structs by value requires that we copy
1412 // the sret argument into %rax for the return. We saved the argument into
1413 // a virtual register in the entry block, so now we copy the value out
1415 if (Subtarget->is64Bit() &&
1416 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1417 MachineFunction &MF = DAG.getMachineFunction();
1418 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1419 unsigned Reg = FuncInfo->getSRetReturnReg();
1421 "SRetReturnReg should have been set in LowerFormalArguments().");
1422 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1424 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1425 Flag = Chain.getValue(1);
1427 // RAX now acts like a return value.
1428 MRI.addLiveOut(X86::RAX);
1431 RetOps[0] = Chain; // Update chain.
1433 // Add the flag if we have it.
1435 RetOps.push_back(Flag);
1437 return DAG.getNode(X86ISD::RET_FLAG, dl,
1438 MVT::Other, &RetOps[0], RetOps.size());
1441 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N) const {
1442 if (N->getNumValues() != 1)
1444 if (!N->hasNUsesOfValue(1, 0))
1447 SDNode *Copy = *N->use_begin();
1448 if (Copy->getOpcode() != ISD::CopyToReg &&
1449 Copy->getOpcode() != ISD::FP_EXTEND)
1452 bool HasRet = false;
1453 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1455 if (UI->getOpcode() != X86ISD::RET_FLAG)
1464 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
1465 ISD::NodeType ExtendKind) const {
1467 // TODO: Is this also valid on 32-bit?
1468 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
1469 ReturnMVT = MVT::i8;
1471 ReturnMVT = MVT::i32;
1473 EVT MinVT = getRegisterType(Context, ReturnMVT);
1474 return VT.bitsLT(MinVT) ? MinVT : VT;
1477 /// LowerCallResult - Lower the result values of a call into the
1478 /// appropriate copies out of appropriate physical registers.
1481 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1482 CallingConv::ID CallConv, bool isVarArg,
1483 const SmallVectorImpl<ISD::InputArg> &Ins,
1484 DebugLoc dl, SelectionDAG &DAG,
1485 SmallVectorImpl<SDValue> &InVals) const {
1487 // Assign locations to each value returned by this call.
1488 SmallVector<CCValAssign, 16> RVLocs;
1489 bool Is64Bit = Subtarget->is64Bit();
1490 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1491 getTargetMachine(), RVLocs, *DAG.getContext());
1492 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1494 // Copy all of the result registers out of their specified physreg.
1495 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1496 CCValAssign &VA = RVLocs[i];
1497 EVT CopyVT = VA.getValVT();
1499 // If this is x86-64, and we disabled SSE, we can't return FP values
1500 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1501 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasXMM())) {
1502 report_fatal_error("SSE register return with SSE disabled");
1507 // If this is a call to a function that returns an fp value on the floating
1508 // point stack, we must guarantee the the value is popped from the stack, so
1509 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1510 // if the return value is not used. We use the FpPOP_RETVAL instruction
1512 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1513 // If we prefer to use the value in xmm registers, copy it out as f80 and
1514 // use a truncate to move it from fp stack reg to xmm reg.
1515 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1516 SDValue Ops[] = { Chain, InFlag };
1517 Chain = SDValue(DAG.getMachineNode(X86::FpPOP_RETVAL, dl, CopyVT,
1518 MVT::Other, MVT::Glue, Ops, 2), 1);
1519 Val = Chain.getValue(0);
1521 // Round the f80 to the right size, which also moves it to the appropriate
1523 if (CopyVT != VA.getValVT())
1524 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1525 // This truncation won't change the value.
1526 DAG.getIntPtrConstant(1));
1528 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1529 CopyVT, InFlag).getValue(1);
1530 Val = Chain.getValue(0);
1532 InFlag = Chain.getValue(2);
1533 InVals.push_back(Val);
1540 //===----------------------------------------------------------------------===//
1541 // C & StdCall & Fast Calling Convention implementation
1542 //===----------------------------------------------------------------------===//
1543 // StdCall calling convention seems to be standard for many Windows' API
1544 // routines and around. It differs from C calling convention just a little:
1545 // callee should clean up the stack, not caller. Symbols should be also
1546 // decorated in some fancy way :) It doesn't support any vector arguments.
1547 // For info on fast calling convention see Fast Calling Convention (tail call)
1548 // implementation LowerX86_32FastCCCallTo.
1550 /// CallIsStructReturn - Determines whether a call uses struct return
1552 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1556 return Outs[0].Flags.isSRet();
1559 /// ArgsAreStructReturn - Determines whether a function uses struct
1560 /// return semantics.
1562 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1566 return Ins[0].Flags.isSRet();
1569 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1570 /// by "Src" to address "Dst" with size and alignment information specified by
1571 /// the specific parameter attribute. The copy will be passed as a byval
1572 /// function parameter.
1574 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1575 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1577 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1579 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1580 /*isVolatile*/false, /*AlwaysInline=*/true,
1581 MachinePointerInfo(), MachinePointerInfo());
1584 /// IsTailCallConvention - Return true if the calling convention is one that
1585 /// supports tail call optimization.
1586 static bool IsTailCallConvention(CallingConv::ID CC) {
1587 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1590 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1591 if (!CI->isTailCall())
1595 CallingConv::ID CalleeCC = CS.getCallingConv();
1596 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
1602 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1603 /// a tailcall target by changing its ABI.
1604 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1605 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1609 X86TargetLowering::LowerMemArgument(SDValue Chain,
1610 CallingConv::ID CallConv,
1611 const SmallVectorImpl<ISD::InputArg> &Ins,
1612 DebugLoc dl, SelectionDAG &DAG,
1613 const CCValAssign &VA,
1614 MachineFrameInfo *MFI,
1616 // Create the nodes corresponding to a load from this parameter slot.
1617 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1618 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1619 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1622 // If value is passed by pointer we have address passed instead of the value
1624 if (VA.getLocInfo() == CCValAssign::Indirect)
1625 ValVT = VA.getLocVT();
1627 ValVT = VA.getValVT();
1629 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1630 // changed with more analysis.
1631 // In case of tail call optimization mark all arguments mutable. Since they
1632 // could be overwritten by lowering of arguments in case of a tail call.
1633 if (Flags.isByVal()) {
1634 unsigned Bytes = Flags.getByValSize();
1635 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
1636 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
1637 return DAG.getFrameIndex(FI, getPointerTy());
1639 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1640 VA.getLocMemOffset(), isImmutable);
1641 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1642 return DAG.getLoad(ValVT, dl, Chain, FIN,
1643 MachinePointerInfo::getFixedStack(FI),
1649 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1650 CallingConv::ID CallConv,
1652 const SmallVectorImpl<ISD::InputArg> &Ins,
1655 SmallVectorImpl<SDValue> &InVals)
1657 MachineFunction &MF = DAG.getMachineFunction();
1658 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1660 const Function* Fn = MF.getFunction();
1661 if (Fn->hasExternalLinkage() &&
1662 Subtarget->isTargetCygMing() &&
1663 Fn->getName() == "main")
1664 FuncInfo->setForceFramePointer(true);
1666 MachineFrameInfo *MFI = MF.getFrameInfo();
1667 bool Is64Bit = Subtarget->is64Bit();
1668 bool IsWin64 = Subtarget->isTargetWin64();
1670 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1671 "Var args not supported with calling convention fastcc or ghc");
1673 // Assign locations to all of the incoming arguments.
1674 SmallVector<CCValAssign, 16> ArgLocs;
1675 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1676 ArgLocs, *DAG.getContext());
1678 // Allocate shadow area for Win64
1680 CCInfo.AllocateStack(32, 8);
1683 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
1685 unsigned LastVal = ~0U;
1687 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1688 CCValAssign &VA = ArgLocs[i];
1689 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1691 assert(VA.getValNo() != LastVal &&
1692 "Don't support value assigned to multiple locs yet");
1693 LastVal = VA.getValNo();
1695 if (VA.isRegLoc()) {
1696 EVT RegVT = VA.getLocVT();
1697 TargetRegisterClass *RC = NULL;
1698 if (RegVT == MVT::i32)
1699 RC = X86::GR32RegisterClass;
1700 else if (Is64Bit && RegVT == MVT::i64)
1701 RC = X86::GR64RegisterClass;
1702 else if (RegVT == MVT::f32)
1703 RC = X86::FR32RegisterClass;
1704 else if (RegVT == MVT::f64)
1705 RC = X86::FR64RegisterClass;
1706 else if (RegVT.isVector() && RegVT.getSizeInBits() == 256)
1707 RC = X86::VR256RegisterClass;
1708 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1709 RC = X86::VR128RegisterClass;
1710 else if (RegVT == MVT::x86mmx)
1711 RC = X86::VR64RegisterClass;
1713 llvm_unreachable("Unknown argument type!");
1715 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1716 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1718 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1719 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1721 if (VA.getLocInfo() == CCValAssign::SExt)
1722 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1723 DAG.getValueType(VA.getValVT()));
1724 else if (VA.getLocInfo() == CCValAssign::ZExt)
1725 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1726 DAG.getValueType(VA.getValVT()));
1727 else if (VA.getLocInfo() == CCValAssign::BCvt)
1728 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
1730 if (VA.isExtInLoc()) {
1731 // Handle MMX values passed in XMM regs.
1732 if (RegVT.isVector()) {
1733 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(),
1736 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1739 assert(VA.isMemLoc());
1740 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1743 // If value is passed via pointer - do a load.
1744 if (VA.getLocInfo() == CCValAssign::Indirect)
1745 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
1746 MachinePointerInfo(), false, false, 0);
1748 InVals.push_back(ArgValue);
1751 // The x86-64 ABI for returning structs by value requires that we copy
1752 // the sret argument into %rax for the return. Save the argument into
1753 // a virtual register so that we can access it from the return points.
1754 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1755 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1756 unsigned Reg = FuncInfo->getSRetReturnReg();
1758 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1759 FuncInfo->setSRetReturnReg(Reg);
1761 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1762 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1765 unsigned StackSize = CCInfo.getNextStackOffset();
1766 // Align stack specially for tail calls.
1767 if (FuncIsMadeTailCallSafe(CallConv))
1768 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1770 // If the function takes variable number of arguments, make a frame index for
1771 // the start of the first vararg value... for expansion of llvm.va_start.
1773 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1774 CallConv != CallingConv::X86_ThisCall)) {
1775 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1778 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1780 // FIXME: We should really autogenerate these arrays
1781 static const unsigned GPR64ArgRegsWin64[] = {
1782 X86::RCX, X86::RDX, X86::R8, X86::R9
1784 static const unsigned GPR64ArgRegs64Bit[] = {
1785 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1787 static const unsigned XMMArgRegs64Bit[] = {
1788 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1789 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1791 const unsigned *GPR64ArgRegs;
1792 unsigned NumXMMRegs = 0;
1795 // The XMM registers which might contain var arg parameters are shadowed
1796 // in their paired GPR. So we only need to save the GPR to their home
1798 TotalNumIntRegs = 4;
1799 GPR64ArgRegs = GPR64ArgRegsWin64;
1801 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1802 GPR64ArgRegs = GPR64ArgRegs64Bit;
1804 NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs64Bit, TotalNumXMMRegs);
1806 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1809 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1810 assert(!(NumXMMRegs && !Subtarget->hasXMM()) &&
1811 "SSE register cannot be used when SSE is disabled!");
1812 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1813 "SSE register cannot be used when SSE is disabled!");
1814 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasXMM())
1815 // Kernel mode asks for SSE to be disabled, so don't push them
1817 TotalNumXMMRegs = 0;
1820 const TargetFrameLowering &TFI = *getTargetMachine().getFrameLowering();
1821 // Get to the caller-allocated home save location. Add 8 to account
1822 // for the return address.
1823 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
1824 FuncInfo->setRegSaveFrameIndex(
1825 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
1826 // Fixup to set vararg frame on shadow area (4 x i64).
1828 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
1830 // For X86-64, if there are vararg parameters that are passed via
1831 // registers, then we must store them to their spots on the stack so they
1832 // may be loaded by deferencing the result of va_next.
1833 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1834 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1835 FuncInfo->setRegSaveFrameIndex(
1836 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1840 // Store the integer parameter registers.
1841 SmallVector<SDValue, 8> MemOps;
1842 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1844 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1845 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1846 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1847 DAG.getIntPtrConstant(Offset));
1848 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1849 X86::GR64RegisterClass);
1850 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1852 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1853 MachinePointerInfo::getFixedStack(
1854 FuncInfo->getRegSaveFrameIndex(), Offset),
1856 MemOps.push_back(Store);
1860 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1861 // Now store the XMM (fp + vector) parameter registers.
1862 SmallVector<SDValue, 11> SaveXMMOps;
1863 SaveXMMOps.push_back(Chain);
1865 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1866 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1867 SaveXMMOps.push_back(ALVal);
1869 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1870 FuncInfo->getRegSaveFrameIndex()));
1871 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1872 FuncInfo->getVarArgsFPOffset()));
1874 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1875 unsigned VReg = MF.addLiveIn(XMMArgRegs64Bit[NumXMMRegs],
1876 X86::VR128RegisterClass);
1877 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1878 SaveXMMOps.push_back(Val);
1880 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1882 &SaveXMMOps[0], SaveXMMOps.size()));
1885 if (!MemOps.empty())
1886 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1887 &MemOps[0], MemOps.size());
1891 // Some CCs need callee pop.
1892 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, GuaranteedTailCallOpt)) {
1893 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1895 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1896 // If this is an sret function, the return should pop the hidden pointer.
1897 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1898 FuncInfo->setBytesToPopOnReturn(4);
1902 // RegSaveFrameIndex is X86-64 only.
1903 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1904 if (CallConv == CallingConv::X86_FastCall ||
1905 CallConv == CallingConv::X86_ThisCall)
1906 // fastcc functions can't have varargs.
1907 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1914 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1915 SDValue StackPtr, SDValue Arg,
1916 DebugLoc dl, SelectionDAG &DAG,
1917 const CCValAssign &VA,
1918 ISD::ArgFlagsTy Flags) const {
1919 unsigned LocMemOffset = VA.getLocMemOffset();
1920 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1921 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1922 if (Flags.isByVal())
1923 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1925 return DAG.getStore(Chain, dl, Arg, PtrOff,
1926 MachinePointerInfo::getStack(LocMemOffset),
1930 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1931 /// optimization is performed and it is required.
1933 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1934 SDValue &OutRetAddr, SDValue Chain,
1935 bool IsTailCall, bool Is64Bit,
1936 int FPDiff, DebugLoc dl) const {
1937 // Adjust the Return address stack slot.
1938 EVT VT = getPointerTy();
1939 OutRetAddr = getReturnAddressFrameIndex(DAG);
1941 // Load the "old" Return address.
1942 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
1944 return SDValue(OutRetAddr.getNode(), 1);
1947 /// EmitTailCallStoreRetAddr - Emit a store of the return address if tail call
1948 /// optimization is performed and it is required (FPDiff!=0).
1950 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1951 SDValue Chain, SDValue RetAddrFrIdx,
1952 bool Is64Bit, int FPDiff, DebugLoc dl) {
1953 // Store the return address to the appropriate stack slot.
1954 if (!FPDiff) return Chain;
1955 // Calculate the new stack slot for the return address.
1956 int SlotSize = Is64Bit ? 8 : 4;
1957 int NewReturnAddrFI =
1958 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1959 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1960 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1961 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1962 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
1968 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1969 CallingConv::ID CallConv, bool isVarArg,
1971 const SmallVectorImpl<ISD::OutputArg> &Outs,
1972 const SmallVectorImpl<SDValue> &OutVals,
1973 const SmallVectorImpl<ISD::InputArg> &Ins,
1974 DebugLoc dl, SelectionDAG &DAG,
1975 SmallVectorImpl<SDValue> &InVals) const {
1976 MachineFunction &MF = DAG.getMachineFunction();
1977 bool Is64Bit = Subtarget->is64Bit();
1978 bool IsWin64 = Subtarget->isTargetWin64();
1979 bool IsStructRet = CallIsStructReturn(Outs);
1980 bool IsSibcall = false;
1983 // Check if it's really possible to do a tail call.
1984 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1985 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1986 Outs, OutVals, Ins, DAG);
1988 // Sibcalls are automatically detected tailcalls which do not require
1990 if (!GuaranteedTailCallOpt && isTailCall)
1997 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1998 "Var args not supported with calling convention fastcc or ghc");
2000 // Analyze operands of the call, assigning locations to each operand.
2001 SmallVector<CCValAssign, 16> ArgLocs;
2002 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
2003 ArgLocs, *DAG.getContext());
2005 // Allocate shadow area for Win64
2007 CCInfo.AllocateStack(32, 8);
2010 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2012 // Get a count of how many bytes are to be pushed on the stack.
2013 unsigned NumBytes = CCInfo.getNextStackOffset();
2015 // This is a sibcall. The memory operands are available in caller's
2016 // own caller's stack.
2018 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
2019 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2022 if (isTailCall && !IsSibcall) {
2023 // Lower arguments at fp - stackoffset + fpdiff.
2024 unsigned NumBytesCallerPushed =
2025 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
2026 FPDiff = NumBytesCallerPushed - NumBytes;
2028 // Set the delta of movement of the returnaddr stackslot.
2029 // But only set if delta is greater than previous delta.
2030 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
2031 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
2035 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2037 SDValue RetAddrFrIdx;
2038 // Load return address for tail calls.
2039 if (isTailCall && FPDiff)
2040 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2041 Is64Bit, FPDiff, dl);
2043 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2044 SmallVector<SDValue, 8> MemOpChains;
2047 // Walk the register/memloc assignments, inserting copies/loads. In the case
2048 // of tail call optimization arguments are handle later.
2049 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2050 CCValAssign &VA = ArgLocs[i];
2051 EVT RegVT = VA.getLocVT();
2052 SDValue Arg = OutVals[i];
2053 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2054 bool isByVal = Flags.isByVal();
2056 // Promote the value if needed.
2057 switch (VA.getLocInfo()) {
2058 default: llvm_unreachable("Unknown loc info!");
2059 case CCValAssign::Full: break;
2060 case CCValAssign::SExt:
2061 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2063 case CCValAssign::ZExt:
2064 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2066 case CCValAssign::AExt:
2067 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
2068 // Special case: passing MMX values in XMM registers.
2069 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2070 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2071 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2073 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2075 case CCValAssign::BCvt:
2076 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2078 case CCValAssign::Indirect: {
2079 // Store the argument.
2080 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2081 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2082 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2083 MachinePointerInfo::getFixedStack(FI),
2090 if (VA.isRegLoc()) {
2091 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2092 if (isVarArg && IsWin64) {
2093 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2094 // shadow reg if callee is a varargs function.
2095 unsigned ShadowReg = 0;
2096 switch (VA.getLocReg()) {
2097 case X86::XMM0: ShadowReg = X86::RCX; break;
2098 case X86::XMM1: ShadowReg = X86::RDX; break;
2099 case X86::XMM2: ShadowReg = X86::R8; break;
2100 case X86::XMM3: ShadowReg = X86::R9; break;
2103 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2105 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2106 assert(VA.isMemLoc());
2107 if (StackPtr.getNode() == 0)
2108 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
2109 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2110 dl, DAG, VA, Flags));
2114 if (!MemOpChains.empty())
2115 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2116 &MemOpChains[0], MemOpChains.size());
2118 // Build a sequence of copy-to-reg nodes chained together with token chain
2119 // and flag operands which copy the outgoing args into registers.
2121 // Tail call byval lowering might overwrite argument registers so in case of
2122 // tail call optimization the copies to registers are lowered later.
2124 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2125 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2126 RegsToPass[i].second, InFlag);
2127 InFlag = Chain.getValue(1);
2130 if (Subtarget->isPICStyleGOT()) {
2131 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2134 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
2135 DAG.getNode(X86ISD::GlobalBaseReg,
2136 DebugLoc(), getPointerTy()),
2138 InFlag = Chain.getValue(1);
2140 // If we are tail calling and generating PIC/GOT style code load the
2141 // address of the callee into ECX. The value in ecx is used as target of
2142 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2143 // for tail calls on PIC/GOT architectures. Normally we would just put the
2144 // address of GOT into ebx and then call target@PLT. But for tail calls
2145 // ebx would be restored (since ebx is callee saved) before jumping to the
2148 // Note: The actual moving to ECX is done further down.
2149 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2150 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2151 !G->getGlobal()->hasProtectedVisibility())
2152 Callee = LowerGlobalAddress(Callee, DAG);
2153 else if (isa<ExternalSymbolSDNode>(Callee))
2154 Callee = LowerExternalSymbol(Callee, DAG);
2158 if (Is64Bit && isVarArg && !IsWin64) {
2159 // From AMD64 ABI document:
2160 // For calls that may call functions that use varargs or stdargs
2161 // (prototype-less calls or calls to functions containing ellipsis (...) in
2162 // the declaration) %al is used as hidden argument to specify the number
2163 // of SSE registers used. The contents of %al do not need to match exactly
2164 // the number of registers, but must be an ubound on the number of SSE
2165 // registers used and is in the range 0 - 8 inclusive.
2167 // Count the number of XMM registers allocated.
2168 static const unsigned XMMArgRegs[] = {
2169 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2170 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2172 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2173 assert((Subtarget->hasXMM() || !NumXMMRegs)
2174 && "SSE registers cannot be used when SSE is disabled");
2176 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2177 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2178 InFlag = Chain.getValue(1);
2182 // For tail calls lower the arguments to the 'real' stack slot.
2184 // Force all the incoming stack arguments to be loaded from the stack
2185 // before any new outgoing arguments are stored to the stack, because the
2186 // outgoing stack slots may alias the incoming argument stack slots, and
2187 // the alias isn't otherwise explicit. This is slightly more conservative
2188 // than necessary, because it means that each store effectively depends
2189 // on every argument instead of just those arguments it would clobber.
2190 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2192 SmallVector<SDValue, 8> MemOpChains2;
2195 // Do not flag preceding copytoreg stuff together with the following stuff.
2197 if (GuaranteedTailCallOpt) {
2198 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2199 CCValAssign &VA = ArgLocs[i];
2202 assert(VA.isMemLoc());
2203 SDValue Arg = OutVals[i];
2204 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2205 // Create frame index.
2206 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2207 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2208 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2209 FIN = DAG.getFrameIndex(FI, getPointerTy());
2211 if (Flags.isByVal()) {
2212 // Copy relative to framepointer.
2213 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2214 if (StackPtr.getNode() == 0)
2215 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2217 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2219 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2223 // Store relative to framepointer.
2224 MemOpChains2.push_back(
2225 DAG.getStore(ArgChain, dl, Arg, FIN,
2226 MachinePointerInfo::getFixedStack(FI),
2232 if (!MemOpChains2.empty())
2233 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2234 &MemOpChains2[0], MemOpChains2.size());
2236 // Copy arguments to their registers.
2237 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2238 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2239 RegsToPass[i].second, InFlag);
2240 InFlag = Chain.getValue(1);
2244 // Store the return address to the appropriate stack slot.
2245 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2249 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2250 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2251 // In the 64-bit large code model, we have to make all calls
2252 // through a register, since the call instruction's 32-bit
2253 // pc-relative offset may not be large enough to hold the whole
2255 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2256 // If the callee is a GlobalAddress node (quite common, every direct call
2257 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2260 // We should use extra load for direct calls to dllimported functions in
2262 const GlobalValue *GV = G->getGlobal();
2263 if (!GV->hasDLLImportLinkage()) {
2264 unsigned char OpFlags = 0;
2265 bool ExtraLoad = false;
2266 unsigned WrapperKind = ISD::DELETED_NODE;
2268 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2269 // external symbols most go through the PLT in PIC mode. If the symbol
2270 // has hidden or protected visibility, or if it is static or local, then
2271 // we don't need to use the PLT - we can directly call it.
2272 if (Subtarget->isTargetELF() &&
2273 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2274 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2275 OpFlags = X86II::MO_PLT;
2276 } else if (Subtarget->isPICStyleStubAny() &&
2277 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2278 (!Subtarget->getTargetTriple().isMacOSX() ||
2279 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2280 // PC-relative references to external symbols should go through $stub,
2281 // unless we're building with the leopard linker or later, which
2282 // automatically synthesizes these stubs.
2283 OpFlags = X86II::MO_DARWIN_STUB;
2284 } else if (Subtarget->isPICStyleRIPRel() &&
2285 isa<Function>(GV) &&
2286 cast<Function>(GV)->hasFnAttr(Attribute::NonLazyBind)) {
2287 // If the function is marked as non-lazy, generate an indirect call
2288 // which loads from the GOT directly. This avoids runtime overhead
2289 // at the cost of eager binding (and one extra byte of encoding).
2290 OpFlags = X86II::MO_GOTPCREL;
2291 WrapperKind = X86ISD::WrapperRIP;
2295 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2296 G->getOffset(), OpFlags);
2298 // Add a wrapper if needed.
2299 if (WrapperKind != ISD::DELETED_NODE)
2300 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
2301 // Add extra indirection if needed.
2303 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
2304 MachinePointerInfo::getGOT(),
2307 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2308 unsigned char OpFlags = 0;
2310 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
2311 // external symbols should go through the PLT.
2312 if (Subtarget->isTargetELF() &&
2313 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2314 OpFlags = X86II::MO_PLT;
2315 } else if (Subtarget->isPICStyleStubAny() &&
2316 (!Subtarget->getTargetTriple().isMacOSX() ||
2317 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2318 // PC-relative references to external symbols should go through $stub,
2319 // unless we're building with the leopard linker or later, which
2320 // automatically synthesizes these stubs.
2321 OpFlags = X86II::MO_DARWIN_STUB;
2324 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2328 // Returns a chain & a flag for retval copy to use.
2329 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2330 SmallVector<SDValue, 8> Ops;
2332 if (!IsSibcall && isTailCall) {
2333 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2334 DAG.getIntPtrConstant(0, true), InFlag);
2335 InFlag = Chain.getValue(1);
2338 Ops.push_back(Chain);
2339 Ops.push_back(Callee);
2342 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2344 // Add argument registers to the end of the list so that they are known live
2346 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2347 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2348 RegsToPass[i].second.getValueType()));
2350 // Add an implicit use GOT pointer in EBX.
2351 if (!isTailCall && Subtarget->isPICStyleGOT())
2352 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2354 // Add an implicit use of AL for non-Windows x86 64-bit vararg functions.
2355 if (Is64Bit && isVarArg && !IsWin64)
2356 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2358 if (InFlag.getNode())
2359 Ops.push_back(InFlag);
2363 //// If this is the first return lowered for this function, add the regs
2364 //// to the liveout set for the function.
2365 // This isn't right, although it's probably harmless on x86; liveouts
2366 // should be computed from returns not tail calls. Consider a void
2367 // function making a tail call to a function returning int.
2368 return DAG.getNode(X86ISD::TC_RETURN, dl,
2369 NodeTys, &Ops[0], Ops.size());
2372 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2373 InFlag = Chain.getValue(1);
2375 // Create the CALLSEQ_END node.
2376 unsigned NumBytesForCalleeToPush;
2377 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, GuaranteedTailCallOpt))
2378 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2379 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2380 // If this is a call to a struct-return function, the callee
2381 // pops the hidden struct pointer, so we have to push it back.
2382 // This is common for Darwin/X86, Linux & Mingw32 targets.
2383 NumBytesForCalleeToPush = 4;
2385 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2387 // Returns a flag for retval copy to use.
2389 Chain = DAG.getCALLSEQ_END(Chain,
2390 DAG.getIntPtrConstant(NumBytes, true),
2391 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2394 InFlag = Chain.getValue(1);
2397 // Handle result values, copying them out of physregs into vregs that we
2399 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2400 Ins, dl, DAG, InVals);
2404 //===----------------------------------------------------------------------===//
2405 // Fast Calling Convention (tail call) implementation
2406 //===----------------------------------------------------------------------===//
2408 // Like std call, callee cleans arguments, convention except that ECX is
2409 // reserved for storing the tail called function address. Only 2 registers are
2410 // free for argument passing (inreg). Tail call optimization is performed
2412 // * tailcallopt is enabled
2413 // * caller/callee are fastcc
2414 // On X86_64 architecture with GOT-style position independent code only local
2415 // (within module) calls are supported at the moment.
2416 // To keep the stack aligned according to platform abi the function
2417 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2418 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2419 // If a tail called function callee has more arguments than the caller the
2420 // caller needs to make sure that there is room to move the RETADDR to. This is
2421 // achieved by reserving an area the size of the argument delta right after the
2422 // original REtADDR, but before the saved framepointer or the spilled registers
2423 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2435 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2436 /// for a 16 byte align requirement.
2438 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2439 SelectionDAG& DAG) const {
2440 MachineFunction &MF = DAG.getMachineFunction();
2441 const TargetMachine &TM = MF.getTarget();
2442 const TargetFrameLowering &TFI = *TM.getFrameLowering();
2443 unsigned StackAlignment = TFI.getStackAlignment();
2444 uint64_t AlignMask = StackAlignment - 1;
2445 int64_t Offset = StackSize;
2446 uint64_t SlotSize = TD->getPointerSize();
2447 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2448 // Number smaller than 12 so just add the difference.
2449 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2451 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2452 Offset = ((~AlignMask) & Offset) + StackAlignment +
2453 (StackAlignment-SlotSize);
2458 /// MatchingStackOffset - Return true if the given stack call argument is
2459 /// already available in the same position (relatively) of the caller's
2460 /// incoming argument stack.
2462 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2463 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2464 const X86InstrInfo *TII) {
2465 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2467 if (Arg.getOpcode() == ISD::CopyFromReg) {
2468 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2469 if (!TargetRegisterInfo::isVirtualRegister(VR))
2471 MachineInstr *Def = MRI->getVRegDef(VR);
2474 if (!Flags.isByVal()) {
2475 if (!TII->isLoadFromStackSlot(Def, FI))
2478 unsigned Opcode = Def->getOpcode();
2479 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2480 Def->getOperand(1).isFI()) {
2481 FI = Def->getOperand(1).getIndex();
2482 Bytes = Flags.getByValSize();
2486 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2487 if (Flags.isByVal())
2488 // ByVal argument is passed in as a pointer but it's now being
2489 // dereferenced. e.g.
2490 // define @foo(%struct.X* %A) {
2491 // tail call @bar(%struct.X* byval %A)
2494 SDValue Ptr = Ld->getBasePtr();
2495 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2498 FI = FINode->getIndex();
2499 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
2500 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
2501 FI = FINode->getIndex();
2502 Bytes = Flags.getByValSize();
2506 assert(FI != INT_MAX);
2507 if (!MFI->isFixedObjectIndex(FI))
2509 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2512 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2513 /// for tail call optimization. Targets which want to do tail call
2514 /// optimization should implement this function.
2516 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2517 CallingConv::ID CalleeCC,
2519 bool isCalleeStructRet,
2520 bool isCallerStructRet,
2521 const SmallVectorImpl<ISD::OutputArg> &Outs,
2522 const SmallVectorImpl<SDValue> &OutVals,
2523 const SmallVectorImpl<ISD::InputArg> &Ins,
2524 SelectionDAG& DAG) const {
2525 if (!IsTailCallConvention(CalleeCC) &&
2526 CalleeCC != CallingConv::C)
2529 // If -tailcallopt is specified, make fastcc functions tail-callable.
2530 const MachineFunction &MF = DAG.getMachineFunction();
2531 const Function *CallerF = DAG.getMachineFunction().getFunction();
2532 CallingConv::ID CallerCC = CallerF->getCallingConv();
2533 bool CCMatch = CallerCC == CalleeCC;
2535 if (GuaranteedTailCallOpt) {
2536 if (IsTailCallConvention(CalleeCC) && CCMatch)
2541 // Look for obvious safe cases to perform tail call optimization that do not
2542 // require ABI changes. This is what gcc calls sibcall.
2544 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2545 // emit a special epilogue.
2546 if (RegInfo->needsStackRealignment(MF))
2549 // Also avoid sibcall optimization if either caller or callee uses struct
2550 // return semantics.
2551 if (isCalleeStructRet || isCallerStructRet)
2554 // An stdcall caller is expected to clean up its arguments; the callee
2555 // isn't going to do that.
2556 if (!CCMatch && CallerCC==CallingConv::X86_StdCall)
2559 // Do not sibcall optimize vararg calls unless all arguments are passed via
2561 if (isVarArg && !Outs.empty()) {
2563 // Optimizing for varargs on Win64 is unlikely to be safe without
2564 // additional testing.
2565 if (Subtarget->isTargetWin64())
2568 SmallVector<CCValAssign, 16> ArgLocs;
2569 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2570 getTargetMachine(), ArgLocs, *DAG.getContext());
2572 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2573 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
2574 if (!ArgLocs[i].isRegLoc())
2578 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2579 // Therefore if it's not used by the call it is not safe to optimize this into
2581 bool Unused = false;
2582 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2589 SmallVector<CCValAssign, 16> RVLocs;
2590 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(),
2591 getTargetMachine(), RVLocs, *DAG.getContext());
2592 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2593 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2594 CCValAssign &VA = RVLocs[i];
2595 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2600 // If the calling conventions do not match, then we'd better make sure the
2601 // results are returned in the same way as what the caller expects.
2603 SmallVector<CCValAssign, 16> RVLocs1;
2604 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
2605 getTargetMachine(), RVLocs1, *DAG.getContext());
2606 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2608 SmallVector<CCValAssign, 16> RVLocs2;
2609 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
2610 getTargetMachine(), RVLocs2, *DAG.getContext());
2611 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2613 if (RVLocs1.size() != RVLocs2.size())
2615 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2616 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2618 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2620 if (RVLocs1[i].isRegLoc()) {
2621 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2624 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2630 // If the callee takes no arguments then go on to check the results of the
2632 if (!Outs.empty()) {
2633 // Check if stack adjustment is needed. For now, do not do this if any
2634 // argument is passed on the stack.
2635 SmallVector<CCValAssign, 16> ArgLocs;
2636 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2637 getTargetMachine(), ArgLocs, *DAG.getContext());
2639 // Allocate shadow area for Win64
2640 if (Subtarget->isTargetWin64()) {
2641 CCInfo.AllocateStack(32, 8);
2644 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2645 if (CCInfo.getNextStackOffset()) {
2646 MachineFunction &MF = DAG.getMachineFunction();
2647 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2650 // Check if the arguments are already laid out in the right way as
2651 // the caller's fixed stack objects.
2652 MachineFrameInfo *MFI = MF.getFrameInfo();
2653 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2654 const X86InstrInfo *TII =
2655 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2656 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2657 CCValAssign &VA = ArgLocs[i];
2658 SDValue Arg = OutVals[i];
2659 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2660 if (VA.getLocInfo() == CCValAssign::Indirect)
2662 if (!VA.isRegLoc()) {
2663 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2670 // If the tailcall address may be in a register, then make sure it's
2671 // possible to register allocate for it. In 32-bit, the call address can
2672 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2673 // callee-saved registers are restored. These happen to be the same
2674 // registers used to pass 'inreg' arguments so watch out for those.
2675 if (!Subtarget->is64Bit() &&
2676 !isa<GlobalAddressSDNode>(Callee) &&
2677 !isa<ExternalSymbolSDNode>(Callee)) {
2678 unsigned NumInRegs = 0;
2679 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2680 CCValAssign &VA = ArgLocs[i];
2683 unsigned Reg = VA.getLocReg();
2686 case X86::EAX: case X86::EDX: case X86::ECX:
2687 if (++NumInRegs == 3)
2699 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2700 return X86::createFastISel(funcInfo);
2704 //===----------------------------------------------------------------------===//
2705 // Other Lowering Hooks
2706 //===----------------------------------------------------------------------===//
2708 static bool MayFoldLoad(SDValue Op) {
2709 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
2712 static bool MayFoldIntoStore(SDValue Op) {
2713 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
2716 static bool isTargetShuffle(unsigned Opcode) {
2718 default: return false;
2719 case X86ISD::PSHUFD:
2720 case X86ISD::PSHUFHW:
2721 case X86ISD::PSHUFLW:
2722 case X86ISD::SHUFPD:
2723 case X86ISD::PALIGN:
2724 case X86ISD::SHUFPS:
2725 case X86ISD::MOVLHPS:
2726 case X86ISD::MOVLHPD:
2727 case X86ISD::MOVHLPS:
2728 case X86ISD::MOVLPS:
2729 case X86ISD::MOVLPD:
2730 case X86ISD::MOVSHDUP:
2731 case X86ISD::MOVSLDUP:
2732 case X86ISD::MOVDDUP:
2735 case X86ISD::UNPCKLPS:
2736 case X86ISD::UNPCKLPD:
2737 case X86ISD::VUNPCKLPS:
2738 case X86ISD::VUNPCKLPD:
2739 case X86ISD::VUNPCKLPSY:
2740 case X86ISD::VUNPCKLPDY:
2741 case X86ISD::PUNPCKLWD:
2742 case X86ISD::PUNPCKLBW:
2743 case X86ISD::PUNPCKLDQ:
2744 case X86ISD::PUNPCKLQDQ:
2745 case X86ISD::UNPCKHPS:
2746 case X86ISD::UNPCKHPD:
2747 case X86ISD::PUNPCKHWD:
2748 case X86ISD::PUNPCKHBW:
2749 case X86ISD::PUNPCKHDQ:
2750 case X86ISD::PUNPCKHQDQ:
2751 case X86ISD::VPERMIL:
2757 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2758 SDValue V1, SelectionDAG &DAG) {
2760 default: llvm_unreachable("Unknown x86 shuffle node");
2761 case X86ISD::MOVSHDUP:
2762 case X86ISD::MOVSLDUP:
2763 case X86ISD::MOVDDUP:
2764 return DAG.getNode(Opc, dl, VT, V1);
2770 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2771 SDValue V1, unsigned TargetMask, SelectionDAG &DAG) {
2773 default: llvm_unreachable("Unknown x86 shuffle node");
2774 case X86ISD::PSHUFD:
2775 case X86ISD::PSHUFHW:
2776 case X86ISD::PSHUFLW:
2777 case X86ISD::VPERMIL:
2778 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
2784 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2785 SDValue V1, SDValue V2, unsigned TargetMask, SelectionDAG &DAG) {
2787 default: llvm_unreachable("Unknown x86 shuffle node");
2788 case X86ISD::PALIGN:
2789 case X86ISD::SHUFPD:
2790 case X86ISD::SHUFPS:
2791 return DAG.getNode(Opc, dl, VT, V1, V2,
2792 DAG.getConstant(TargetMask, MVT::i8));
2797 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2798 SDValue V1, SDValue V2, SelectionDAG &DAG) {
2800 default: llvm_unreachable("Unknown x86 shuffle node");
2801 case X86ISD::MOVLHPS:
2802 case X86ISD::MOVLHPD:
2803 case X86ISD::MOVHLPS:
2804 case X86ISD::MOVLPS:
2805 case X86ISD::MOVLPD:
2808 case X86ISD::UNPCKLPS:
2809 case X86ISD::UNPCKLPD:
2810 case X86ISD::VUNPCKLPS:
2811 case X86ISD::VUNPCKLPD:
2812 case X86ISD::VUNPCKLPSY:
2813 case X86ISD::VUNPCKLPDY:
2814 case X86ISD::PUNPCKLWD:
2815 case X86ISD::PUNPCKLBW:
2816 case X86ISD::PUNPCKLDQ:
2817 case X86ISD::PUNPCKLQDQ:
2818 case X86ISD::UNPCKHPS:
2819 case X86ISD::UNPCKHPD:
2820 case X86ISD::PUNPCKHWD:
2821 case X86ISD::PUNPCKHBW:
2822 case X86ISD::PUNPCKHDQ:
2823 case X86ISD::PUNPCKHQDQ:
2824 return DAG.getNode(Opc, dl, VT, V1, V2);
2829 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2830 MachineFunction &MF = DAG.getMachineFunction();
2831 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2832 int ReturnAddrIndex = FuncInfo->getRAIndex();
2834 if (ReturnAddrIndex == 0) {
2835 // Set up a frame object for the return address.
2836 uint64_t SlotSize = TD->getPointerSize();
2837 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2839 FuncInfo->setRAIndex(ReturnAddrIndex);
2842 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2846 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2847 bool hasSymbolicDisplacement) {
2848 // Offset should fit into 32 bit immediate field.
2849 if (!isInt<32>(Offset))
2852 // If we don't have a symbolic displacement - we don't have any extra
2854 if (!hasSymbolicDisplacement)
2857 // FIXME: Some tweaks might be needed for medium code model.
2858 if (M != CodeModel::Small && M != CodeModel::Kernel)
2861 // For small code model we assume that latest object is 16MB before end of 31
2862 // bits boundary. We may also accept pretty large negative constants knowing
2863 // that all objects are in the positive half of address space.
2864 if (M == CodeModel::Small && Offset < 16*1024*1024)
2867 // For kernel code model we know that all object resist in the negative half
2868 // of 32bits address space. We may not accept negative offsets, since they may
2869 // be just off and we may accept pretty large positive ones.
2870 if (M == CodeModel::Kernel && Offset > 0)
2876 /// isCalleePop - Determines whether the callee is required to pop its
2877 /// own arguments. Callee pop is necessary to support tail calls.
2878 bool X86::isCalleePop(CallingConv::ID CallingConv,
2879 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
2883 switch (CallingConv) {
2886 case CallingConv::X86_StdCall:
2888 case CallingConv::X86_FastCall:
2890 case CallingConv::X86_ThisCall:
2892 case CallingConv::Fast:
2894 case CallingConv::GHC:
2899 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2900 /// specific condition code, returning the condition code and the LHS/RHS of the
2901 /// comparison to make.
2902 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2903 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2905 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2906 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2907 // X > -1 -> X == 0, jump !sign.
2908 RHS = DAG.getConstant(0, RHS.getValueType());
2909 return X86::COND_NS;
2910 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2911 // X < 0 -> X == 0, jump on sign.
2913 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2915 RHS = DAG.getConstant(0, RHS.getValueType());
2916 return X86::COND_LE;
2920 switch (SetCCOpcode) {
2921 default: llvm_unreachable("Invalid integer condition!");
2922 case ISD::SETEQ: return X86::COND_E;
2923 case ISD::SETGT: return X86::COND_G;
2924 case ISD::SETGE: return X86::COND_GE;
2925 case ISD::SETLT: return X86::COND_L;
2926 case ISD::SETLE: return X86::COND_LE;
2927 case ISD::SETNE: return X86::COND_NE;
2928 case ISD::SETULT: return X86::COND_B;
2929 case ISD::SETUGT: return X86::COND_A;
2930 case ISD::SETULE: return X86::COND_BE;
2931 case ISD::SETUGE: return X86::COND_AE;
2935 // First determine if it is required or is profitable to flip the operands.
2937 // If LHS is a foldable load, but RHS is not, flip the condition.
2938 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
2939 !ISD::isNON_EXTLoad(RHS.getNode())) {
2940 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2941 std::swap(LHS, RHS);
2944 switch (SetCCOpcode) {
2950 std::swap(LHS, RHS);
2954 // On a floating point condition, the flags are set as follows:
2956 // 0 | 0 | 0 | X > Y
2957 // 0 | 0 | 1 | X < Y
2958 // 1 | 0 | 0 | X == Y
2959 // 1 | 1 | 1 | unordered
2960 switch (SetCCOpcode) {
2961 default: llvm_unreachable("Condcode should be pre-legalized away");
2963 case ISD::SETEQ: return X86::COND_E;
2964 case ISD::SETOLT: // flipped
2966 case ISD::SETGT: return X86::COND_A;
2967 case ISD::SETOLE: // flipped
2969 case ISD::SETGE: return X86::COND_AE;
2970 case ISD::SETUGT: // flipped
2972 case ISD::SETLT: return X86::COND_B;
2973 case ISD::SETUGE: // flipped
2975 case ISD::SETLE: return X86::COND_BE;
2977 case ISD::SETNE: return X86::COND_NE;
2978 case ISD::SETUO: return X86::COND_P;
2979 case ISD::SETO: return X86::COND_NP;
2981 case ISD::SETUNE: return X86::COND_INVALID;
2985 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2986 /// code. Current x86 isa includes the following FP cmov instructions:
2987 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2988 static bool hasFPCMov(unsigned X86CC) {
3004 /// isFPImmLegal - Returns true if the target can instruction select the
3005 /// specified FP immediate natively. If false, the legalizer will
3006 /// materialize the FP immediate as a load from a constant pool.
3007 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3008 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3009 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3015 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3016 /// the specified range (L, H].
3017 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3018 return (Val < 0) || (Val >= Low && Val < Hi);
3021 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3022 /// specified value.
3023 static bool isUndefOrEqual(int Val, int CmpVal) {
3024 if (Val < 0 || Val == CmpVal)
3029 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
3030 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
3031 /// the second operand.
3032 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3033 if (VT == MVT::v4f32 || VT == MVT::v4i32 )
3034 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
3035 if (VT == MVT::v2f64 || VT == MVT::v2i64)
3036 return (Mask[0] < 2 && Mask[1] < 2);
3040 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
3041 SmallVector<int, 8> M;
3043 return ::isPSHUFDMask(M, N->getValueType(0));
3046 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
3047 /// is suitable for input to PSHUFHW.
3048 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3049 if (VT != MVT::v8i16)
3052 // Lower quadword copied in order or undef.
3053 for (int i = 0; i != 4; ++i)
3054 if (Mask[i] >= 0 && Mask[i] != i)
3057 // Upper quadword shuffled.
3058 for (int i = 4; i != 8; ++i)
3059 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
3065 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
3066 SmallVector<int, 8> M;
3068 return ::isPSHUFHWMask(M, N->getValueType(0));
3071 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
3072 /// is suitable for input to PSHUFLW.
3073 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3074 if (VT != MVT::v8i16)
3077 // Upper quadword copied in order.
3078 for (int i = 4; i != 8; ++i)
3079 if (Mask[i] >= 0 && Mask[i] != i)
3082 // Lower quadword shuffled.
3083 for (int i = 0; i != 4; ++i)
3090 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
3091 SmallVector<int, 8> M;
3093 return ::isPSHUFLWMask(M, N->getValueType(0));
3096 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
3097 /// is suitable for input to PALIGNR.
3098 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
3100 int i, e = VT.getVectorNumElements();
3102 // Do not handle v2i64 / v2f64 shuffles with palignr.
3103 if (e < 4 || !hasSSSE3)
3106 for (i = 0; i != e; ++i)
3110 // All undef, not a palignr.
3114 // Determine if it's ok to perform a palignr with only the LHS, since we
3115 // don't have access to the actual shuffle elements to see if RHS is undef.
3116 bool Unary = Mask[i] < (int)e;
3117 bool NeedsUnary = false;
3119 int s = Mask[i] - i;
3121 // Check the rest of the elements to see if they are consecutive.
3122 for (++i; i != e; ++i) {
3127 Unary = Unary && (m < (int)e);
3128 NeedsUnary = NeedsUnary || (m < s);
3130 if (NeedsUnary && !Unary)
3132 if (Unary && m != ((s+i) & (e-1)))
3134 if (!Unary && m != (s+i))
3140 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
3141 SmallVector<int, 8> M;
3143 return ::isPALIGNRMask(M, N->getValueType(0), true);
3146 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
3147 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
3148 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3149 int NumElems = VT.getVectorNumElements();
3150 if (NumElems != 2 && NumElems != 4)
3153 int Half = NumElems / 2;
3154 for (int i = 0; i < Half; ++i)
3155 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3157 for (int i = Half; i < NumElems; ++i)
3158 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3164 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
3165 SmallVector<int, 8> M;
3167 return ::isSHUFPMask(M, N->getValueType(0));
3170 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
3171 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
3172 /// half elements to come from vector 1 (which would equal the dest.) and
3173 /// the upper half to come from vector 2.
3174 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3175 int NumElems = VT.getVectorNumElements();
3177 if (NumElems != 2 && NumElems != 4)
3180 int Half = NumElems / 2;
3181 for (int i = 0; i < Half; ++i)
3182 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3184 for (int i = Half; i < NumElems; ++i)
3185 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3190 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
3191 SmallVector<int, 8> M;
3193 return isCommutedSHUFPMask(M, N->getValueType(0));
3196 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3197 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3198 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
3199 if (N->getValueType(0).getVectorNumElements() != 4)
3202 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3203 return isUndefOrEqual(N->getMaskElt(0), 6) &&
3204 isUndefOrEqual(N->getMaskElt(1), 7) &&
3205 isUndefOrEqual(N->getMaskElt(2), 2) &&
3206 isUndefOrEqual(N->getMaskElt(3), 3);
3209 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3210 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3212 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
3213 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3218 return isUndefOrEqual(N->getMaskElt(0), 2) &&
3219 isUndefOrEqual(N->getMaskElt(1), 3) &&
3220 isUndefOrEqual(N->getMaskElt(2), 2) &&
3221 isUndefOrEqual(N->getMaskElt(3), 3);
3224 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3225 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3226 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
3227 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3229 if (NumElems != 2 && NumElems != 4)
3232 for (unsigned i = 0; i < NumElems/2; ++i)
3233 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
3236 for (unsigned i = NumElems/2; i < NumElems; ++i)
3237 if (!isUndefOrEqual(N->getMaskElt(i), i))
3243 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3244 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3245 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
3246 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3248 if ((NumElems != 2 && NumElems != 4)
3249 || N->getValueType(0).getSizeInBits() > 128)
3252 for (unsigned i = 0; i < NumElems/2; ++i)
3253 if (!isUndefOrEqual(N->getMaskElt(i), i))
3256 for (unsigned i = 0; i < NumElems/2; ++i)
3257 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
3263 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3264 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3265 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3266 bool V2IsSplat = false) {
3267 int NumElts = VT.getVectorNumElements();
3268 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3271 // Handle vector lengths > 128 bits. Define a "section" as a set of
3272 // 128 bits. AVX defines UNPCK* to operate independently on 128-bit
3274 unsigned NumSections = VT.getSizeInBits() / 128;
3275 if (NumSections == 0 ) NumSections = 1; // Handle MMX
3276 unsigned NumSectionElts = NumElts / NumSections;
3279 unsigned End = NumSectionElts;
3280 for (unsigned s = 0; s < NumSections; ++s) {
3281 for (unsigned i = Start, j = s * NumSectionElts;
3285 int BitI1 = Mask[i+1];
3286 if (!isUndefOrEqual(BitI, j))
3289 if (!isUndefOrEqual(BitI1, NumElts))
3292 if (!isUndefOrEqual(BitI1, j + NumElts))
3296 // Process the next 128 bits.
3297 Start += NumSectionElts;
3298 End += NumSectionElts;
3304 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3305 SmallVector<int, 8> M;
3307 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
3310 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3311 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3312 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3313 bool V2IsSplat = false) {
3314 int NumElts = VT.getVectorNumElements();
3315 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3318 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3320 int BitI1 = Mask[i+1];
3321 if (!isUndefOrEqual(BitI, j + NumElts/2))
3324 if (isUndefOrEqual(BitI1, NumElts))
3327 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
3334 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3335 SmallVector<int, 8> M;
3337 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3340 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3341 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3343 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3344 int NumElems = VT.getVectorNumElements();
3345 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3348 // Handle vector lengths > 128 bits. Define a "section" as a set of
3349 // 128 bits. AVX defines UNPCK* to operate independently on 128-bit
3351 unsigned NumSections = VT.getSizeInBits() / 128;
3352 if (NumSections == 0 ) NumSections = 1; // Handle MMX
3353 unsigned NumSectionElts = NumElems / NumSections;
3355 for (unsigned s = 0; s < NumSections; ++s) {
3356 for (unsigned i = s * NumSectionElts, j = s * NumSectionElts;
3357 i != NumSectionElts * (s + 1);
3360 int BitI1 = Mask[i+1];
3362 if (!isUndefOrEqual(BitI, j))
3364 if (!isUndefOrEqual(BitI1, j))
3372 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3373 SmallVector<int, 8> M;
3375 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3378 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3379 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3381 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3382 int NumElems = VT.getVectorNumElements();
3383 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3386 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3388 int BitI1 = Mask[i+1];
3389 if (!isUndefOrEqual(BitI, j))
3391 if (!isUndefOrEqual(BitI1, j))
3397 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3398 SmallVector<int, 8> M;
3400 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3403 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3404 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3405 /// MOVSD, and MOVD, i.e. setting the lowest element.
3406 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3407 if (VT.getVectorElementType().getSizeInBits() < 32)
3410 int NumElts = VT.getVectorNumElements();
3412 if (!isUndefOrEqual(Mask[0], NumElts))
3415 for (int i = 1; i < NumElts; ++i)
3416 if (!isUndefOrEqual(Mask[i], i))
3422 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3423 SmallVector<int, 8> M;
3425 return ::isMOVLMask(M, N->getValueType(0));
3428 /// isVPERMILMask - Return true if the specified VECTOR_SHUFFLE operand
3429 /// specifies a shuffle of elements that is suitable for input to VPERMIL*.
3430 static bool isVPERMILMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3431 unsigned NumElts = VT.getVectorNumElements();
3432 unsigned NumLanes = VT.getSizeInBits()/128;
3434 // Match any permutation of 128-bit vector with 32/64-bit types
3435 if (NumLanes == 1) {
3436 if (NumElts == 4 || NumElts == 2)
3441 // Only match 256-bit with 32/64-bit types
3442 if (NumElts != 8 && NumElts != 4)
3445 // The mask on the high lane should be the same as the low. Actually,
3446 // they can differ if any of the corresponding index in a lane is undef.
3447 int LaneSize = NumElts/NumLanes;
3448 for (int i = 0; i < LaneSize; ++i) {
3449 int HighElt = i+LaneSize;
3450 if (Mask[i] < 0 || Mask[HighElt] < 0)
3453 if (Mask[HighElt]-Mask[i] != LaneSize)
3460 /// getShuffleVPERMILImmediateediate - Return the appropriate immediate to shuffle
3461 /// the specified VECTOR_MASK mask with VPERMIL* instructions.
3462 static unsigned getShuffleVPERMILImmediate(SDNode *N) {
3463 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3464 EVT VT = SVOp->getValueType(0);
3466 int NumElts = VT.getVectorNumElements();
3467 int NumLanes = VT.getSizeInBits()/128;
3470 for (int i = 0; i < NumElts/NumLanes /* lane size */; ++i)
3471 Mask |= SVOp->getMaskElt(i) << (i*2);
3476 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3477 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3478 /// element of vector 2 and the other elements to come from vector 1 in order.
3479 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3480 bool V2IsSplat = false, bool V2IsUndef = false) {
3481 int NumOps = VT.getVectorNumElements();
3482 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3485 if (!isUndefOrEqual(Mask[0], 0))
3488 for (int i = 1; i < NumOps; ++i)
3489 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3490 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3491 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3497 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3498 bool V2IsUndef = false) {
3499 SmallVector<int, 8> M;
3501 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3504 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3505 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3506 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3507 if (N->getValueType(0).getVectorNumElements() != 4)
3510 // Expect 1, 1, 3, 3
3511 for (unsigned i = 0; i < 2; ++i) {
3512 int Elt = N->getMaskElt(i);
3513 if (Elt >= 0 && Elt != 1)
3518 for (unsigned i = 2; i < 4; ++i) {
3519 int Elt = N->getMaskElt(i);
3520 if (Elt >= 0 && Elt != 3)
3525 // Don't use movshdup if it can be done with a shufps.
3526 // FIXME: verify that matching u, u, 3, 3 is what we want.
3530 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3531 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3532 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3533 if (N->getValueType(0).getVectorNumElements() != 4)
3536 // Expect 0, 0, 2, 2
3537 for (unsigned i = 0; i < 2; ++i)
3538 if (N->getMaskElt(i) > 0)
3542 for (unsigned i = 2; i < 4; ++i) {
3543 int Elt = N->getMaskElt(i);
3544 if (Elt >= 0 && Elt != 2)
3549 // Don't use movsldup if it can be done with a shufps.
3553 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3554 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3555 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3556 int e = N->getValueType(0).getVectorNumElements() / 2;
3558 for (int i = 0; i < e; ++i)
3559 if (!isUndefOrEqual(N->getMaskElt(i), i))
3561 for (int i = 0; i < e; ++i)
3562 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3567 /// isVEXTRACTF128Index - Return true if the specified
3568 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
3569 /// suitable for input to VEXTRACTF128.
3570 bool X86::isVEXTRACTF128Index(SDNode *N) {
3571 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3574 // The index should be aligned on a 128-bit boundary.
3576 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3578 unsigned VL = N->getValueType(0).getVectorNumElements();
3579 unsigned VBits = N->getValueType(0).getSizeInBits();
3580 unsigned ElSize = VBits / VL;
3581 bool Result = (Index * ElSize) % 128 == 0;
3586 /// isVINSERTF128Index - Return true if the specified INSERT_SUBVECTOR
3587 /// operand specifies a subvector insert that is suitable for input to
3589 bool X86::isVINSERTF128Index(SDNode *N) {
3590 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3593 // The index should be aligned on a 128-bit boundary.
3595 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3597 unsigned VL = N->getValueType(0).getVectorNumElements();
3598 unsigned VBits = N->getValueType(0).getSizeInBits();
3599 unsigned ElSize = VBits / VL;
3600 bool Result = (Index * ElSize) % 128 == 0;
3605 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3606 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3607 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3608 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3609 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3611 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3613 for (int i = 0; i < NumOperands; ++i) {
3614 int Val = SVOp->getMaskElt(NumOperands-i-1);
3615 if (Val < 0) Val = 0;
3616 if (Val >= NumOperands) Val -= NumOperands;
3618 if (i != NumOperands - 1)
3624 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3625 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3626 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3627 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3629 // 8 nodes, but we only care about the last 4.
3630 for (unsigned i = 7; i >= 4; --i) {
3631 int Val = SVOp->getMaskElt(i);
3640 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3641 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3642 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3643 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3645 // 8 nodes, but we only care about the first 4.
3646 for (int i = 3; i >= 0; --i) {
3647 int Val = SVOp->getMaskElt(i);
3656 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3657 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3658 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3659 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3660 EVT VVT = N->getValueType(0);
3661 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3665 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3666 Val = SVOp->getMaskElt(i);
3670 return (Val - i) * EltSize;
3673 /// getExtractVEXTRACTF128Immediate - Return the appropriate immediate
3674 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
3676 unsigned X86::getExtractVEXTRACTF128Immediate(SDNode *N) {
3677 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3678 llvm_unreachable("Illegal extract subvector for VEXTRACTF128");
3681 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3683 EVT VecVT = N->getOperand(0).getValueType();
3684 EVT ElVT = VecVT.getVectorElementType();
3686 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
3687 return Index / NumElemsPerChunk;
3690 /// getInsertVINSERTF128Immediate - Return the appropriate immediate
3691 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
3693 unsigned X86::getInsertVINSERTF128Immediate(SDNode *N) {
3694 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3695 llvm_unreachable("Illegal insert subvector for VINSERTF128");
3698 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3700 EVT VecVT = N->getValueType(0);
3701 EVT ElVT = VecVT.getVectorElementType();
3703 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
3704 return Index / NumElemsPerChunk;
3707 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3709 bool X86::isZeroNode(SDValue Elt) {
3710 return ((isa<ConstantSDNode>(Elt) &&
3711 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3712 (isa<ConstantFPSDNode>(Elt) &&
3713 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3716 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3717 /// their permute mask.
3718 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3719 SelectionDAG &DAG) {
3720 EVT VT = SVOp->getValueType(0);
3721 unsigned NumElems = VT.getVectorNumElements();
3722 SmallVector<int, 8> MaskVec;
3724 for (unsigned i = 0; i != NumElems; ++i) {
3725 int idx = SVOp->getMaskElt(i);
3727 MaskVec.push_back(idx);
3728 else if (idx < (int)NumElems)
3729 MaskVec.push_back(idx + NumElems);
3731 MaskVec.push_back(idx - NumElems);
3733 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3734 SVOp->getOperand(0), &MaskVec[0]);
3737 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3738 /// the two vector operands have swapped position.
3739 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3740 unsigned NumElems = VT.getVectorNumElements();
3741 for (unsigned i = 0; i != NumElems; ++i) {
3745 else if (idx < (int)NumElems)
3746 Mask[i] = idx + NumElems;
3748 Mask[i] = idx - NumElems;
3752 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3753 /// match movhlps. The lower half elements should come from upper half of
3754 /// V1 (and in order), and the upper half elements should come from the upper
3755 /// half of V2 (and in order).
3756 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3757 if (Op->getValueType(0).getVectorNumElements() != 4)
3759 for (unsigned i = 0, e = 2; i != e; ++i)
3760 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3762 for (unsigned i = 2; i != 4; ++i)
3763 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3768 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3769 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3771 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3772 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3774 N = N->getOperand(0).getNode();
3775 if (!ISD::isNON_EXTLoad(N))
3778 *LD = cast<LoadSDNode>(N);
3782 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3783 /// match movlp{s|d}. The lower half elements should come from lower half of
3784 /// V1 (and in order), and the upper half elements should come from the upper
3785 /// half of V2 (and in order). And since V1 will become the source of the
3786 /// MOVLP, it must be either a vector load or a scalar load to vector.
3787 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3788 ShuffleVectorSDNode *Op) {
3789 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3791 // Is V2 is a vector load, don't do this transformation. We will try to use
3792 // load folding shufps op.
3793 if (ISD::isNON_EXTLoad(V2))
3796 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3798 if (NumElems != 2 && NumElems != 4)
3800 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3801 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3803 for (unsigned i = NumElems/2; i != NumElems; ++i)
3804 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3809 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3811 static bool isSplatVector(SDNode *N) {
3812 if (N->getOpcode() != ISD::BUILD_VECTOR)
3815 SDValue SplatValue = N->getOperand(0);
3816 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3817 if (N->getOperand(i) != SplatValue)
3822 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3823 /// to an zero vector.
3824 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3825 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3826 SDValue V1 = N->getOperand(0);
3827 SDValue V2 = N->getOperand(1);
3828 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3829 for (unsigned i = 0; i != NumElems; ++i) {
3830 int Idx = N->getMaskElt(i);
3831 if (Idx >= (int)NumElems) {
3832 unsigned Opc = V2.getOpcode();
3833 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3835 if (Opc != ISD::BUILD_VECTOR ||
3836 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3838 } else if (Idx >= 0) {
3839 unsigned Opc = V1.getOpcode();
3840 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3842 if (Opc != ISD::BUILD_VECTOR ||
3843 !X86::isZeroNode(V1.getOperand(Idx)))
3850 /// getZeroVector - Returns a vector of specified type with all zero elements.
3852 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3854 assert(VT.isVector() && "Expected a vector type");
3856 // Always build SSE zero vectors as <4 x i32> bitcasted
3857 // to their dest type. This ensures they get CSE'd.
3859 if (VT.getSizeInBits() == 128) { // SSE
3860 if (HasSSE2) { // SSE2
3861 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3862 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3864 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3865 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3867 } else if (VT.getSizeInBits() == 256) { // AVX
3868 // 256-bit logic and arithmetic instructions in AVX are
3869 // all floating-point, no support for integer ops. Default
3870 // to emitting fp zeroed vectors then.
3871 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3872 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3873 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
3875 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
3878 /// getOnesVector - Returns a vector of specified type with all bits set.
3879 /// Always build ones vectors as <4 x i32> or <8 x i32> bitcasted to
3880 /// their original type, ensuring they get CSE'd.
3881 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3882 assert(VT.isVector() && "Expected a vector type");
3883 assert((VT.is128BitVector() || VT.is256BitVector())
3884 && "Expected a 128-bit or 256-bit vector type");
3886 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3889 if (VT.is256BitVector()) {
3890 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3891 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops, 8);
3893 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3894 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
3897 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3898 /// that point to V2 points to its first element.
3899 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3900 EVT VT = SVOp->getValueType(0);
3901 unsigned NumElems = VT.getVectorNumElements();
3903 bool Changed = false;
3904 SmallVector<int, 8> MaskVec;
3905 SVOp->getMask(MaskVec);
3907 for (unsigned i = 0; i != NumElems; ++i) {
3908 if (MaskVec[i] > (int)NumElems) {
3909 MaskVec[i] = NumElems;
3914 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3915 SVOp->getOperand(1), &MaskVec[0]);
3916 return SDValue(SVOp, 0);
3919 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3920 /// operation of specified width.
3921 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3923 unsigned NumElems = VT.getVectorNumElements();
3924 SmallVector<int, 8> Mask;
3925 Mask.push_back(NumElems);
3926 for (unsigned i = 1; i != NumElems; ++i)
3928 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3931 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3932 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3934 unsigned NumElems = VT.getVectorNumElements();
3935 SmallVector<int, 8> Mask;
3936 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3938 Mask.push_back(i + NumElems);
3940 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3943 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
3944 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3946 unsigned NumElems = VT.getVectorNumElements();
3947 unsigned Half = NumElems/2;
3948 SmallVector<int, 8> Mask;
3949 for (unsigned i = 0; i != Half; ++i) {
3950 Mask.push_back(i + Half);
3951 Mask.push_back(i + NumElems + Half);
3953 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3956 // PromoteSplatv8v16 - All i16 and i8 vector types can't be used directly by
3957 // a generic shuffle instruction because the target has no such instructions.
3958 // Generate shuffles which repeat i16 and i8 several times until they can be
3959 // represented by v4f32 and then be manipulated by target suported shuffles.
3960 static SDValue PromoteSplatv8v16(SDValue V, SelectionDAG &DAG, int &EltNo) {
3961 EVT VT = V.getValueType();
3962 int NumElems = VT.getVectorNumElements();
3963 DebugLoc dl = V.getDebugLoc();
3965 while (NumElems > 4) {
3966 if (EltNo < NumElems/2) {
3967 V = getUnpackl(DAG, dl, VT, V, V);
3969 V = getUnpackh(DAG, dl, VT, V, V);
3970 EltNo -= NumElems/2;
3977 /// getLegalSplat - Generate a legal splat with supported x86 shuffles
3978 static SDValue getLegalSplat(SelectionDAG &DAG, SDValue V, int EltNo) {
3979 EVT VT = V.getValueType();
3980 DebugLoc dl = V.getDebugLoc();
3981 assert((VT.getSizeInBits() == 128 || VT.getSizeInBits() == 256)
3982 && "Vector size not supported");
3984 bool Is128 = VT.getSizeInBits() == 128;
3985 EVT NVT = Is128 ? MVT::v4f32 : MVT::v8f32;
3986 V = DAG.getNode(ISD::BITCAST, dl, NVT, V);
3989 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3990 V = DAG.getVectorShuffle(NVT, dl, V, DAG.getUNDEF(NVT), &SplatMask[0]);
3992 // The second half of indicies refer to the higher part, which is a
3993 // duplication of the lower one. This makes this shuffle a perfect match
3994 // for the VPERM instruction.
3995 int SplatMask[8] = { EltNo, EltNo, EltNo, EltNo,
3996 EltNo+4, EltNo+4, EltNo+4, EltNo+4 };
3997 V = DAG.getVectorShuffle(NVT, dl, V, DAG.getUNDEF(NVT), &SplatMask[0]);
4000 return DAG.getNode(ISD::BITCAST, dl, VT, V);
4003 /// PromoteSplat - Promote a splat of v4i32, v8i16 or v16i8 to v4f32 and
4004 /// v8i32, v16i16 or v32i8 to v8f32.
4005 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
4006 EVT SrcVT = SV->getValueType(0);
4007 SDValue V1 = SV->getOperand(0);
4008 DebugLoc dl = SV->getDebugLoc();
4010 int EltNo = SV->getSplatIndex();
4011 int NumElems = SrcVT.getVectorNumElements();
4012 unsigned Size = SrcVT.getSizeInBits();
4014 // Extract the 128-bit part containing the splat element and update
4015 // the splat element index when it refers to the higher register.
4017 unsigned Idx = (EltNo > NumElems/2) ? NumElems/2 : 0;
4018 V1 = Extract128BitVector(V1, DAG.getConstant(Idx, MVT::i32), DAG, dl);
4020 EltNo -= NumElems/2;
4023 // Make this 128-bit vector duplicate i8 and i16 elements
4025 V1 = PromoteSplatv8v16(V1, DAG, EltNo);
4027 // Recreate the 256-bit vector and place the same 128-bit vector
4028 // into the low and high part. This is necessary because we want
4029 // to use VPERM to shuffle the v8f32 vector, and VPERM only shuffles
4030 // inside each separate v4f32 lane.
4032 SDValue InsV = Insert128BitVector(DAG.getUNDEF(SrcVT), V1,
4033 DAG.getConstant(0, MVT::i32), DAG, dl);
4034 V1 = Insert128BitVector(InsV, V1,
4035 DAG.getConstant(NumElems/2, MVT::i32), DAG, dl);
4038 return getLegalSplat(DAG, V1, EltNo);
4041 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
4042 /// vector of zero or undef vector. This produces a shuffle where the low
4043 /// element of V2 is swizzled into the zero/undef vector, landing at element
4044 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4045 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4046 bool isZero, bool HasSSE2,
4047 SelectionDAG &DAG) {
4048 EVT VT = V2.getValueType();
4050 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
4051 unsigned NumElems = VT.getVectorNumElements();
4052 SmallVector<int, 16> MaskVec;
4053 for (unsigned i = 0; i != NumElems; ++i)
4054 // If this is the insertion idx, put the low elt of V2 here.
4055 MaskVec.push_back(i == Idx ? NumElems : i);
4056 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
4059 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4060 /// element of the result of the vector shuffle.
4061 static SDValue getShuffleScalarElt(SDNode *N, int Index, SelectionDAG &DAG,
4064 return SDValue(); // Limit search depth.
4066 SDValue V = SDValue(N, 0);
4067 EVT VT = V.getValueType();
4068 unsigned Opcode = V.getOpcode();
4070 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4071 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4072 Index = SV->getMaskElt(Index);
4075 return DAG.getUNDEF(VT.getVectorElementType());
4077 int NumElems = VT.getVectorNumElements();
4078 SDValue NewV = (Index < NumElems) ? SV->getOperand(0) : SV->getOperand(1);
4079 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG, Depth+1);
4082 // Recurse into target specific vector shuffles to find scalars.
4083 if (isTargetShuffle(Opcode)) {
4084 int NumElems = VT.getVectorNumElements();
4085 SmallVector<unsigned, 16> ShuffleMask;
4089 case X86ISD::SHUFPS:
4090 case X86ISD::SHUFPD:
4091 ImmN = N->getOperand(N->getNumOperands()-1);
4092 DecodeSHUFPSMask(NumElems,
4093 cast<ConstantSDNode>(ImmN)->getZExtValue(),
4096 case X86ISD::PUNPCKHBW:
4097 case X86ISD::PUNPCKHWD:
4098 case X86ISD::PUNPCKHDQ:
4099 case X86ISD::PUNPCKHQDQ:
4100 DecodePUNPCKHMask(NumElems, ShuffleMask);
4102 case X86ISD::UNPCKHPS:
4103 case X86ISD::UNPCKHPD:
4104 DecodeUNPCKHPMask(NumElems, ShuffleMask);
4106 case X86ISD::PUNPCKLBW:
4107 case X86ISD::PUNPCKLWD:
4108 case X86ISD::PUNPCKLDQ:
4109 case X86ISD::PUNPCKLQDQ:
4110 DecodePUNPCKLMask(VT, ShuffleMask);
4112 case X86ISD::UNPCKLPS:
4113 case X86ISD::UNPCKLPD:
4114 case X86ISD::VUNPCKLPS:
4115 case X86ISD::VUNPCKLPD:
4116 case X86ISD::VUNPCKLPSY:
4117 case X86ISD::VUNPCKLPDY:
4118 DecodeUNPCKLPMask(VT, ShuffleMask);
4120 case X86ISD::MOVHLPS:
4121 DecodeMOVHLPSMask(NumElems, ShuffleMask);
4123 case X86ISD::MOVLHPS:
4124 DecodeMOVLHPSMask(NumElems, ShuffleMask);
4126 case X86ISD::PSHUFD:
4127 ImmN = N->getOperand(N->getNumOperands()-1);
4128 DecodePSHUFMask(NumElems,
4129 cast<ConstantSDNode>(ImmN)->getZExtValue(),
4132 case X86ISD::PSHUFHW:
4133 ImmN = N->getOperand(N->getNumOperands()-1);
4134 DecodePSHUFHWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
4137 case X86ISD::PSHUFLW:
4138 ImmN = N->getOperand(N->getNumOperands()-1);
4139 DecodePSHUFLWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
4143 case X86ISD::MOVSD: {
4144 // The index 0 always comes from the first element of the second source,
4145 // this is why MOVSS and MOVSD are used in the first place. The other
4146 // elements come from the other positions of the first source vector.
4147 unsigned OpNum = (Index == 0) ? 1 : 0;
4148 return getShuffleScalarElt(V.getOperand(OpNum).getNode(), Index, DAG,
4151 case X86ISD::VPERMIL:
4152 ImmN = N->getOperand(N->getNumOperands()-1);
4153 DecodeVPERMILMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4156 assert("not implemented for target shuffle node");
4160 Index = ShuffleMask[Index];
4162 return DAG.getUNDEF(VT.getVectorElementType());
4164 SDValue NewV = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
4165 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG,
4169 // Actual nodes that may contain scalar elements
4170 if (Opcode == ISD::BITCAST) {
4171 V = V.getOperand(0);
4172 EVT SrcVT = V.getValueType();
4173 unsigned NumElems = VT.getVectorNumElements();
4175 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4179 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4180 return (Index == 0) ? V.getOperand(0)
4181 : DAG.getUNDEF(VT.getVectorElementType());
4183 if (V.getOpcode() == ISD::BUILD_VECTOR)
4184 return V.getOperand(Index);
4189 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
4190 /// shuffle operation which come from a consecutively from a zero. The
4191 /// search can start in two different directions, from left or right.
4193 unsigned getNumOfConsecutiveZeros(SDNode *N, int NumElems,
4194 bool ZerosFromLeft, SelectionDAG &DAG) {
4197 while (i < NumElems) {
4198 unsigned Index = ZerosFromLeft ? i : NumElems-i-1;
4199 SDValue Elt = getShuffleScalarElt(N, Index, DAG, 0);
4200 if (!(Elt.getNode() &&
4201 (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt))))
4209 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies from MaskI to
4210 /// MaskE correspond consecutively to elements from one of the vector operands,
4211 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
4213 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp, int MaskI, int MaskE,
4214 int OpIdx, int NumElems, unsigned &OpNum) {
4215 bool SeenV1 = false;
4216 bool SeenV2 = false;
4218 for (int i = MaskI; i <= MaskE; ++i, ++OpIdx) {
4219 int Idx = SVOp->getMaskElt(i);
4220 // Ignore undef indicies
4229 // Only accept consecutive elements from the same vector
4230 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
4234 OpNum = SeenV1 ? 0 : 1;
4238 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
4239 /// logical left shift of a vector.
4240 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4241 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4242 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4243 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4244 false /* check zeros from right */, DAG);
4250 // Considering the elements in the mask that are not consecutive zeros,
4251 // check if they consecutively come from only one of the source vectors.
4253 // V1 = {X, A, B, C} 0
4255 // vector_shuffle V1, V2 <1, 2, 3, X>
4257 if (!isShuffleMaskConsecutive(SVOp,
4258 0, // Mask Start Index
4259 NumElems-NumZeros-1, // Mask End Index
4260 NumZeros, // Where to start looking in the src vector
4261 NumElems, // Number of elements in vector
4262 OpSrc)) // Which source operand ?
4267 ShVal = SVOp->getOperand(OpSrc);
4271 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
4272 /// logical left shift of a vector.
4273 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4274 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4275 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4276 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4277 true /* check zeros from left */, DAG);
4283 // Considering the elements in the mask that are not consecutive zeros,
4284 // check if they consecutively come from only one of the source vectors.
4286 // 0 { A, B, X, X } = V2
4288 // vector_shuffle V1, V2 <X, X, 4, 5>
4290 if (!isShuffleMaskConsecutive(SVOp,
4291 NumZeros, // Mask Start Index
4292 NumElems-1, // Mask End Index
4293 0, // Where to start looking in the src vector
4294 NumElems, // Number of elements in vector
4295 OpSrc)) // Which source operand ?
4300 ShVal = SVOp->getOperand(OpSrc);
4304 /// isVectorShift - Returns true if the shuffle can be implemented as a
4305 /// logical left or right shift of a vector.
4306 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4307 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4308 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
4309 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
4315 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
4317 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4318 unsigned NumNonZero, unsigned NumZero,
4320 const TargetLowering &TLI) {
4324 DebugLoc dl = Op.getDebugLoc();
4327 for (unsigned i = 0; i < 16; ++i) {
4328 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4329 if (ThisIsNonZero && First) {
4331 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4333 V = DAG.getUNDEF(MVT::v8i16);
4338 SDValue ThisElt(0, 0), LastElt(0, 0);
4339 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4340 if (LastIsNonZero) {
4341 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4342 MVT::i16, Op.getOperand(i-1));
4344 if (ThisIsNonZero) {
4345 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4346 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4347 ThisElt, DAG.getConstant(8, MVT::i8));
4349 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4353 if (ThisElt.getNode())
4354 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4355 DAG.getIntPtrConstant(i/2));
4359 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
4362 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
4364 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4365 unsigned NumNonZero, unsigned NumZero,
4367 const TargetLowering &TLI) {
4371 DebugLoc dl = Op.getDebugLoc();
4374 for (unsigned i = 0; i < 8; ++i) {
4375 bool isNonZero = (NonZeros & (1 << i)) != 0;
4379 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4381 V = DAG.getUNDEF(MVT::v8i16);
4384 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4385 MVT::v8i16, V, Op.getOperand(i),
4386 DAG.getIntPtrConstant(i));
4393 /// getVShift - Return a vector logical shift node.
4395 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4396 unsigned NumBits, SelectionDAG &DAG,
4397 const TargetLowering &TLI, DebugLoc dl) {
4398 EVT ShVT = MVT::v2i64;
4399 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
4400 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
4401 return DAG.getNode(ISD::BITCAST, dl, VT,
4402 DAG.getNode(Opc, dl, ShVT, SrcOp,
4403 DAG.getConstant(NumBits,
4404 TLI.getShiftAmountTy(SrcOp.getValueType()))));
4408 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
4409 SelectionDAG &DAG) const {
4411 // Check if the scalar load can be widened into a vector load. And if
4412 // the address is "base + cst" see if the cst can be "absorbed" into
4413 // the shuffle mask.
4414 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4415 SDValue Ptr = LD->getBasePtr();
4416 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4418 EVT PVT = LD->getValueType(0);
4419 if (PVT != MVT::i32 && PVT != MVT::f32)
4424 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4425 FI = FINode->getIndex();
4427 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4428 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4429 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4430 Offset = Ptr.getConstantOperandVal(1);
4431 Ptr = Ptr.getOperand(0);
4436 SDValue Chain = LD->getChain();
4437 // Make sure the stack object alignment is at least 16.
4438 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4439 if (DAG.InferPtrAlignment(Ptr) < 16) {
4440 if (MFI->isFixedObjectIndex(FI)) {
4441 // Can't change the alignment. FIXME: It's possible to compute
4442 // the exact stack offset and reference FI + adjust offset instead.
4443 // If someone *really* cares about this. That's the way to implement it.
4446 MFI->setObjectAlignment(FI, 16);
4450 // (Offset % 16) must be multiple of 4. Then address is then
4451 // Ptr + (Offset & ~15).
4454 if ((Offset % 16) & 3)
4456 int64_t StartOffset = Offset & ~15;
4458 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
4459 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4461 int EltNo = (Offset - StartOffset) >> 2;
4462 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
4463 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
4464 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,
4465 LD->getPointerInfo().getWithOffset(StartOffset),
4467 // Canonicalize it to a v4i32 shuffle.
4468 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1);
4469 return DAG.getNode(ISD::BITCAST, dl, VT,
4470 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
4471 DAG.getUNDEF(MVT::v4i32),&Mask[0]));
4477 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
4478 /// vector of type 'VT', see if the elements can be replaced by a single large
4479 /// load which has the same value as a build_vector whose operands are 'elts'.
4481 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4483 /// FIXME: we'd also like to handle the case where the last elements are zero
4484 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4485 /// There's even a handy isZeroNode for that purpose.
4486 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
4487 DebugLoc &DL, SelectionDAG &DAG) {
4488 EVT EltVT = VT.getVectorElementType();
4489 unsigned NumElems = Elts.size();
4491 LoadSDNode *LDBase = NULL;
4492 unsigned LastLoadedElt = -1U;
4494 // For each element in the initializer, see if we've found a load or an undef.
4495 // If we don't find an initial load element, or later load elements are
4496 // non-consecutive, bail out.
4497 for (unsigned i = 0; i < NumElems; ++i) {
4498 SDValue Elt = Elts[i];
4500 if (!Elt.getNode() ||
4501 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4504 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4506 LDBase = cast<LoadSDNode>(Elt.getNode());
4510 if (Elt.getOpcode() == ISD::UNDEF)
4513 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4514 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
4519 // If we have found an entire vector of loads and undefs, then return a large
4520 // load of the entire vector width starting at the base pointer. If we found
4521 // consecutive loads for the low half, generate a vzext_load node.
4522 if (LastLoadedElt == NumElems - 1) {
4523 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
4524 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4525 LDBase->getPointerInfo(),
4526 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
4527 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4528 LDBase->getPointerInfo(),
4529 LDBase->isVolatile(), LDBase->isNonTemporal(),
4530 LDBase->getAlignment());
4531 } else if (NumElems == 4 && LastLoadedElt == 1) {
4532 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4533 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4534 SDValue ResNode = DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys,
4536 LDBase->getMemOperand());
4537 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
4543 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
4544 DebugLoc dl = Op.getDebugLoc();
4546 EVT VT = Op.getValueType();
4547 EVT ExtVT = VT.getVectorElementType();
4549 unsigned NumElems = Op.getNumOperands();
4551 // For AVX-length vectors, build the individual 128-bit pieces and
4552 // use shuffles to put them in place.
4553 if (VT.getSizeInBits() > 256 &&
4554 Subtarget->hasAVX() &&
4555 !ISD::isBuildVectorAllZeros(Op.getNode())) {
4556 SmallVector<SDValue, 8> V;
4558 for (unsigned i = 0; i < NumElems; ++i) {
4559 V[i] = Op.getOperand(i);
4562 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
4564 // Build the lower subvector.
4565 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[0], NumElems/2);
4566 // Build the upper subvector.
4567 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[NumElems / 2],
4570 return ConcatVectors(Lower, Upper, DAG);
4574 // - pxor (SSE2), xorps (SSE1), vpxor (128 AVX), xorp[s|d] (256 AVX)
4576 // - pcmpeqd (SSE2 and 128 AVX), fallback to constant pools (256 AVX)
4577 if (ISD::isBuildVectorAllZeros(Op.getNode()) ||
4578 ISD::isBuildVectorAllOnes(Op.getNode())) {
4579 // Canonicalize this to <4 x i32> or <8 x 32> (SSE) to
4580 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
4581 // eliminated on x86-32 hosts.
4582 if (Op.getValueType() == MVT::v4i32 ||
4583 Op.getValueType() == MVT::v8i32)
4586 if (ISD::isBuildVectorAllOnes(Op.getNode()))
4587 return getOnesVector(Op.getValueType(), DAG, dl);
4588 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
4591 unsigned EVTBits = ExtVT.getSizeInBits();
4593 unsigned NumZero = 0;
4594 unsigned NumNonZero = 0;
4595 unsigned NonZeros = 0;
4596 bool IsAllConstants = true;
4597 SmallSet<SDValue, 8> Values;
4598 for (unsigned i = 0; i < NumElems; ++i) {
4599 SDValue Elt = Op.getOperand(i);
4600 if (Elt.getOpcode() == ISD::UNDEF)
4603 if (Elt.getOpcode() != ISD::Constant &&
4604 Elt.getOpcode() != ISD::ConstantFP)
4605 IsAllConstants = false;
4606 if (X86::isZeroNode(Elt))
4609 NonZeros |= (1 << i);
4614 // All undef vector. Return an UNDEF. All zero vectors were handled above.
4615 if (NumNonZero == 0)
4616 return DAG.getUNDEF(VT);
4618 // Special case for single non-zero, non-undef, element.
4619 if (NumNonZero == 1) {
4620 unsigned Idx = CountTrailingZeros_32(NonZeros);
4621 SDValue Item = Op.getOperand(Idx);
4623 // If this is an insertion of an i64 value on x86-32, and if the top bits of
4624 // the value are obviously zero, truncate the value to i32 and do the
4625 // insertion that way. Only do this if the value is non-constant or if the
4626 // value is a constant being inserted into element 0. It is cheaper to do
4627 // a constant pool load than it is to do a movd + shuffle.
4628 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
4629 (!IsAllConstants || Idx == 0)) {
4630 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
4632 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
4633 EVT VecVT = MVT::v4i32;
4634 unsigned VecElts = 4;
4636 // Truncate the value (which may itself be a constant) to i32, and
4637 // convert it to a vector with movd (S2V+shuffle to zero extend).
4638 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
4639 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
4640 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4641 Subtarget->hasSSE2(), DAG);
4643 // Now we have our 32-bit value zero extended in the low element of
4644 // a vector. If Idx != 0, swizzle it into place.
4646 SmallVector<int, 4> Mask;
4647 Mask.push_back(Idx);
4648 for (unsigned i = 1; i != VecElts; ++i)
4650 Item = DAG.getVectorShuffle(VecVT, dl, Item,
4651 DAG.getUNDEF(Item.getValueType()),
4654 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Item);
4658 // If we have a constant or non-constant insertion into the low element of
4659 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
4660 // the rest of the elements. This will be matched as movd/movq/movss/movsd
4661 // depending on what the source datatype is.
4664 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4665 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
4666 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
4667 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4668 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
4669 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
4671 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
4672 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
4673 assert(VT.getSizeInBits() == 128 && "Expected an SSE value type!");
4674 EVT MiddleVT = MVT::v4i32;
4675 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
4676 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4677 Subtarget->hasSSE2(), DAG);
4678 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
4682 // Is it a vector logical left shift?
4683 if (NumElems == 2 && Idx == 1 &&
4684 X86::isZeroNode(Op.getOperand(0)) &&
4685 !X86::isZeroNode(Op.getOperand(1))) {
4686 unsigned NumBits = VT.getSizeInBits();
4687 return getVShift(true, VT,
4688 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4689 VT, Op.getOperand(1)),
4690 NumBits/2, DAG, *this, dl);
4693 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
4696 // Otherwise, if this is a vector with i32 or f32 elements, and the element
4697 // is a non-constant being inserted into an element other than the low one,
4698 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
4699 // movd/movss) to move this into the low element, then shuffle it into
4701 if (EVTBits == 32) {
4702 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4704 // Turn it into a shuffle of zero and zero-extended scalar to vector.
4705 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
4706 Subtarget->hasSSE2(), DAG);
4707 SmallVector<int, 8> MaskVec;
4708 for (unsigned i = 0; i < NumElems; i++)
4709 MaskVec.push_back(i == Idx ? 0 : 1);
4710 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
4714 // Splat is obviously ok. Let legalizer expand it to a shuffle.
4715 if (Values.size() == 1) {
4716 if (EVTBits == 32) {
4717 // Instead of a shuffle like this:
4718 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
4719 // Check if it's possible to issue this instead.
4720 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
4721 unsigned Idx = CountTrailingZeros_32(NonZeros);
4722 SDValue Item = Op.getOperand(Idx);
4723 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
4724 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
4729 // A vector full of immediates; various special cases are already
4730 // handled, so this is best done with a single constant-pool load.
4734 // Let legalizer expand 2-wide build_vectors.
4735 if (EVTBits == 64) {
4736 if (NumNonZero == 1) {
4737 // One half is zero or undef.
4738 unsigned Idx = CountTrailingZeros_32(NonZeros);
4739 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
4740 Op.getOperand(Idx));
4741 return getShuffleVectorZeroOrUndef(V2, Idx, true,
4742 Subtarget->hasSSE2(), DAG);
4747 // If element VT is < 32 bits, convert it to inserts into a zero vector.
4748 if (EVTBits == 8 && NumElems == 16) {
4749 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
4751 if (V.getNode()) return V;
4754 if (EVTBits == 16 && NumElems == 8) {
4755 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
4757 if (V.getNode()) return V;
4760 // If element VT is == 32 bits, turn it into a number of shuffles.
4761 SmallVector<SDValue, 8> V;
4763 if (NumElems == 4 && NumZero > 0) {
4764 for (unsigned i = 0; i < 4; ++i) {
4765 bool isZero = !(NonZeros & (1 << i));
4767 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4769 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4772 for (unsigned i = 0; i < 2; ++i) {
4773 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
4776 V[i] = V[i*2]; // Must be a zero vector.
4779 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
4782 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
4785 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
4790 SmallVector<int, 8> MaskVec;
4791 bool Reverse = (NonZeros & 0x3) == 2;
4792 for (unsigned i = 0; i < 2; ++i)
4793 MaskVec.push_back(Reverse ? 1-i : i);
4794 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4795 for (unsigned i = 0; i < 2; ++i)
4796 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4797 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4800 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4801 // Check for a build vector of consecutive loads.
4802 for (unsigned i = 0; i < NumElems; ++i)
4803 V[i] = Op.getOperand(i);
4805 // Check for elements which are consecutive loads.
4806 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4810 // For SSE 4.1, use insertps to put the high elements into the low element.
4811 if (getSubtarget()->hasSSE41()) {
4813 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
4814 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
4816 Result = DAG.getUNDEF(VT);
4818 for (unsigned i = 1; i < NumElems; ++i) {
4819 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
4820 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
4821 Op.getOperand(i), DAG.getIntPtrConstant(i));
4826 // Otherwise, expand into a number of unpckl*, start by extending each of
4827 // our (non-undef) elements to the full vector width with the element in the
4828 // bottom slot of the vector (which generates no code for SSE).
4829 for (unsigned i = 0; i < NumElems; ++i) {
4830 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4831 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4833 V[i] = DAG.getUNDEF(VT);
4836 // Next, we iteratively mix elements, e.g. for v4f32:
4837 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4838 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4839 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4840 unsigned EltStride = NumElems >> 1;
4841 while (EltStride != 0) {
4842 for (unsigned i = 0; i < EltStride; ++i) {
4843 // If V[i+EltStride] is undef and this is the first round of mixing,
4844 // then it is safe to just drop this shuffle: V[i] is already in the
4845 // right place, the one element (since it's the first round) being
4846 // inserted as undef can be dropped. This isn't safe for successive
4847 // rounds because they will permute elements within both vectors.
4848 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
4849 EltStride == NumElems/2)
4852 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
4862 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4863 // We support concatenate two MMX registers and place them in a MMX
4864 // register. This is better than doing a stack convert.
4865 DebugLoc dl = Op.getDebugLoc();
4866 EVT ResVT = Op.getValueType();
4867 assert(Op.getNumOperands() == 2);
4868 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4869 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4871 SDValue InVec = DAG.getNode(ISD::BITCAST,dl, MVT::v1i64, Op.getOperand(0));
4872 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4873 InVec = Op.getOperand(1);
4874 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4875 unsigned NumElts = ResVT.getVectorNumElements();
4876 VecOp = DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
4877 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4878 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4880 InVec = DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, InVec);
4881 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4882 Mask[0] = 0; Mask[1] = 2;
4883 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4885 return DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
4888 // v8i16 shuffles - Prefer shuffles in the following order:
4889 // 1. [all] pshuflw, pshufhw, optional move
4890 // 2. [ssse3] 1 x pshufb
4891 // 3. [ssse3] 2 x pshufb + 1 x por
4892 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4894 X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
4895 SelectionDAG &DAG) const {
4896 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4897 SDValue V1 = SVOp->getOperand(0);
4898 SDValue V2 = SVOp->getOperand(1);
4899 DebugLoc dl = SVOp->getDebugLoc();
4900 SmallVector<int, 8> MaskVals;
4902 // Determine if more than 1 of the words in each of the low and high quadwords
4903 // of the result come from the same quadword of one of the two inputs. Undef
4904 // mask values count as coming from any quadword, for better codegen.
4905 SmallVector<unsigned, 4> LoQuad(4);
4906 SmallVector<unsigned, 4> HiQuad(4);
4907 BitVector InputQuads(4);
4908 for (unsigned i = 0; i < 8; ++i) {
4909 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4910 int EltIdx = SVOp->getMaskElt(i);
4911 MaskVals.push_back(EltIdx);
4920 InputQuads.set(EltIdx / 4);
4923 int BestLoQuad = -1;
4924 unsigned MaxQuad = 1;
4925 for (unsigned i = 0; i < 4; ++i) {
4926 if (LoQuad[i] > MaxQuad) {
4928 MaxQuad = LoQuad[i];
4932 int BestHiQuad = -1;
4934 for (unsigned i = 0; i < 4; ++i) {
4935 if (HiQuad[i] > MaxQuad) {
4937 MaxQuad = HiQuad[i];
4941 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4942 // of the two input vectors, shuffle them into one input vector so only a
4943 // single pshufb instruction is necessary. If There are more than 2 input
4944 // quads, disable the next transformation since it does not help SSSE3.
4945 bool V1Used = InputQuads[0] || InputQuads[1];
4946 bool V2Used = InputQuads[2] || InputQuads[3];
4947 if (Subtarget->hasSSSE3()) {
4948 if (InputQuads.count() == 2 && V1Used && V2Used) {
4949 BestLoQuad = InputQuads.find_first();
4950 BestHiQuad = InputQuads.find_next(BestLoQuad);
4952 if (InputQuads.count() > 2) {
4958 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4959 // the shuffle mask. If a quad is scored as -1, that means that it contains
4960 // words from all 4 input quadwords.
4962 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4963 SmallVector<int, 8> MaskV;
4964 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4965 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4966 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4967 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
4968 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
4969 NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
4971 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4972 // source words for the shuffle, to aid later transformations.
4973 bool AllWordsInNewV = true;
4974 bool InOrder[2] = { true, true };
4975 for (unsigned i = 0; i != 8; ++i) {
4976 int idx = MaskVals[i];
4978 InOrder[i/4] = false;
4979 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4981 AllWordsInNewV = false;
4985 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4986 if (AllWordsInNewV) {
4987 for (int i = 0; i != 8; ++i) {
4988 int idx = MaskVals[i];
4991 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4992 if ((idx != i) && idx < 4)
4994 if ((idx != i) && idx > 3)
5003 // If we've eliminated the use of V2, and the new mask is a pshuflw or
5004 // pshufhw, that's as cheap as it gets. Return the new shuffle.
5005 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
5006 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
5007 unsigned TargetMask = 0;
5008 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
5009 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
5010 TargetMask = pshufhw ? X86::getShufflePSHUFHWImmediate(NewV.getNode()):
5011 X86::getShufflePSHUFLWImmediate(NewV.getNode());
5012 V1 = NewV.getOperand(0);
5013 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
5017 // If we have SSSE3, and all words of the result are from 1 input vector,
5018 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
5019 // is present, fall back to case 4.
5020 if (Subtarget->hasSSSE3()) {
5021 SmallVector<SDValue,16> pshufbMask;
5023 // If we have elements from both input vectors, set the high bit of the
5024 // shuffle mask element to zero out elements that come from V2 in the V1
5025 // mask, and elements that come from V1 in the V2 mask, so that the two
5026 // results can be OR'd together.
5027 bool TwoInputs = V1Used && V2Used;
5028 for (unsigned i = 0; i != 8; ++i) {
5029 int EltIdx = MaskVals[i] * 2;
5030 if (TwoInputs && (EltIdx >= 16)) {
5031 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5032 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5035 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
5036 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
5038 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V1);
5039 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
5040 DAG.getNode(ISD::BUILD_VECTOR, dl,
5041 MVT::v16i8, &pshufbMask[0], 16));
5043 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5045 // Calculate the shuffle mask for the second input, shuffle it, and
5046 // OR it with the first shuffled input.
5048 for (unsigned i = 0; i != 8; ++i) {
5049 int EltIdx = MaskVals[i] * 2;
5051 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5052 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5055 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
5056 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
5058 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V2);
5059 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
5060 DAG.getNode(ISD::BUILD_VECTOR, dl,
5061 MVT::v16i8, &pshufbMask[0], 16));
5062 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
5063 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5066 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
5067 // and update MaskVals with new element order.
5068 BitVector InOrder(8);
5069 if (BestLoQuad >= 0) {
5070 SmallVector<int, 8> MaskV;
5071 for (int i = 0; i != 4; ++i) {
5072 int idx = MaskVals[i];
5074 MaskV.push_back(-1);
5076 } else if ((idx / 4) == BestLoQuad) {
5077 MaskV.push_back(idx & 3);
5080 MaskV.push_back(-1);
5083 for (unsigned i = 4; i != 8; ++i)
5085 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
5088 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
5089 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
5091 X86::getShufflePSHUFLWImmediate(NewV.getNode()),
5095 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
5096 // and update MaskVals with the new element order.
5097 if (BestHiQuad >= 0) {
5098 SmallVector<int, 8> MaskV;
5099 for (unsigned i = 0; i != 4; ++i)
5101 for (unsigned i = 4; i != 8; ++i) {
5102 int idx = MaskVals[i];
5104 MaskV.push_back(-1);
5106 } else if ((idx / 4) == BestHiQuad) {
5107 MaskV.push_back((idx & 3) + 4);
5110 MaskV.push_back(-1);
5113 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
5116 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
5117 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
5119 X86::getShufflePSHUFHWImmediate(NewV.getNode()),
5123 // In case BestHi & BestLo were both -1, which means each quadword has a word
5124 // from each of the four input quadwords, calculate the InOrder bitvector now
5125 // before falling through to the insert/extract cleanup.
5126 if (BestLoQuad == -1 && BestHiQuad == -1) {
5128 for (int i = 0; i != 8; ++i)
5129 if (MaskVals[i] < 0 || MaskVals[i] == i)
5133 // The other elements are put in the right place using pextrw and pinsrw.
5134 for (unsigned i = 0; i != 8; ++i) {
5137 int EltIdx = MaskVals[i];
5140 SDValue ExtOp = (EltIdx < 8)
5141 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
5142 DAG.getIntPtrConstant(EltIdx))
5143 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
5144 DAG.getIntPtrConstant(EltIdx - 8));
5145 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
5146 DAG.getIntPtrConstant(i));
5151 // v16i8 shuffles - Prefer shuffles in the following order:
5152 // 1. [ssse3] 1 x pshufb
5153 // 2. [ssse3] 2 x pshufb + 1 x por
5154 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
5156 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
5158 const X86TargetLowering &TLI) {
5159 SDValue V1 = SVOp->getOperand(0);
5160 SDValue V2 = SVOp->getOperand(1);
5161 DebugLoc dl = SVOp->getDebugLoc();
5162 SmallVector<int, 16> MaskVals;
5163 SVOp->getMask(MaskVals);
5165 // If we have SSSE3, case 1 is generated when all result bytes come from
5166 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
5167 // present, fall back to case 3.
5168 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
5171 for (unsigned i = 0; i < 16; ++i) {
5172 int EltIdx = MaskVals[i];
5181 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
5182 if (TLI.getSubtarget()->hasSSSE3()) {
5183 SmallVector<SDValue,16> pshufbMask;
5185 // If all result elements are from one input vector, then only translate
5186 // undef mask values to 0x80 (zero out result) in the pshufb mask.
5188 // Otherwise, we have elements from both input vectors, and must zero out
5189 // elements that come from V2 in the first mask, and V1 in the second mask
5190 // so that we can OR them together.
5191 bool TwoInputs = !(V1Only || V2Only);
5192 for (unsigned i = 0; i != 16; ++i) {
5193 int EltIdx = MaskVals[i];
5194 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
5195 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5198 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
5200 // If all the elements are from V2, assign it to V1 and return after
5201 // building the first pshufb.
5204 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
5205 DAG.getNode(ISD::BUILD_VECTOR, dl,
5206 MVT::v16i8, &pshufbMask[0], 16));
5210 // Calculate the shuffle mask for the second input, shuffle it, and
5211 // OR it with the first shuffled input.
5213 for (unsigned i = 0; i != 16; ++i) {
5214 int EltIdx = MaskVals[i];
5216 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5219 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
5221 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
5222 DAG.getNode(ISD::BUILD_VECTOR, dl,
5223 MVT::v16i8, &pshufbMask[0], 16));
5224 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
5227 // No SSSE3 - Calculate in place words and then fix all out of place words
5228 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
5229 // the 16 different words that comprise the two doublequadword input vectors.
5230 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5231 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
5232 SDValue NewV = V2Only ? V2 : V1;
5233 for (int i = 0; i != 8; ++i) {
5234 int Elt0 = MaskVals[i*2];
5235 int Elt1 = MaskVals[i*2+1];
5237 // This word of the result is all undef, skip it.
5238 if (Elt0 < 0 && Elt1 < 0)
5241 // This word of the result is already in the correct place, skip it.
5242 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
5244 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
5247 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
5248 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
5251 // If Elt0 and Elt1 are defined, are consecutive, and can be load
5252 // using a single extract together, load it and store it.
5253 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
5254 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
5255 DAG.getIntPtrConstant(Elt1 / 2));
5256 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5257 DAG.getIntPtrConstant(i));
5261 // If Elt1 is defined, extract it from the appropriate source. If the
5262 // source byte is not also odd, shift the extracted word left 8 bits
5263 // otherwise clear the bottom 8 bits if we need to do an or.
5265 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
5266 DAG.getIntPtrConstant(Elt1 / 2));
5267 if ((Elt1 & 1) == 0)
5268 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
5270 TLI.getShiftAmountTy(InsElt.getValueType())));
5272 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
5273 DAG.getConstant(0xFF00, MVT::i16));
5275 // If Elt0 is defined, extract it from the appropriate source. If the
5276 // source byte is not also even, shift the extracted word right 8 bits. If
5277 // Elt1 was also defined, OR the extracted values together before
5278 // inserting them in the result.
5280 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
5281 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
5282 if ((Elt0 & 1) != 0)
5283 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
5285 TLI.getShiftAmountTy(InsElt0.getValueType())));
5287 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
5288 DAG.getConstant(0x00FF, MVT::i16));
5289 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
5292 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5293 DAG.getIntPtrConstant(i));
5295 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
5298 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
5299 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
5300 /// done when every pair / quad of shuffle mask elements point to elements in
5301 /// the right sequence. e.g.
5302 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
5304 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
5305 SelectionDAG &DAG, DebugLoc dl) {
5306 EVT VT = SVOp->getValueType(0);
5307 SDValue V1 = SVOp->getOperand(0);
5308 SDValue V2 = SVOp->getOperand(1);
5309 unsigned NumElems = VT.getVectorNumElements();
5310 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
5312 switch (VT.getSimpleVT().SimpleTy) {
5313 default: assert(false && "Unexpected!");
5314 case MVT::v4f32: NewVT = MVT::v2f64; break;
5315 case MVT::v4i32: NewVT = MVT::v2i64; break;
5316 case MVT::v8i16: NewVT = MVT::v4i32; break;
5317 case MVT::v16i8: NewVT = MVT::v4i32; break;
5320 int Scale = NumElems / NewWidth;
5321 SmallVector<int, 8> MaskVec;
5322 for (unsigned i = 0; i < NumElems; i += Scale) {
5324 for (int j = 0; j < Scale; ++j) {
5325 int EltIdx = SVOp->getMaskElt(i+j);
5329 StartIdx = EltIdx - (EltIdx % Scale);
5330 if (EltIdx != StartIdx + j)
5334 MaskVec.push_back(-1);
5336 MaskVec.push_back(StartIdx / Scale);
5339 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
5340 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
5341 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
5344 /// getVZextMovL - Return a zero-extending vector move low node.
5346 static SDValue getVZextMovL(EVT VT, EVT OpVT,
5347 SDValue SrcOp, SelectionDAG &DAG,
5348 const X86Subtarget *Subtarget, DebugLoc dl) {
5349 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
5350 LoadSDNode *LD = NULL;
5351 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
5352 LD = dyn_cast<LoadSDNode>(SrcOp);
5354 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
5356 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
5357 if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
5358 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
5359 SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
5360 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
5362 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
5363 return DAG.getNode(ISD::BITCAST, dl, VT,
5364 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5365 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5373 return DAG.getNode(ISD::BITCAST, dl, VT,
5374 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5375 DAG.getNode(ISD::BITCAST, dl,
5379 /// LowerVECTOR_SHUFFLE_256 - Handle all 256-bit wide vectors shuffles
5380 /// which could not be matched by any known target speficic shuffle
5382 LowerVECTOR_SHUFFLE_256(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
5386 /// LowerVECTOR_SHUFFLE_128v4 - Handle all 128-bit wide vectors with
5387 /// 4 elements, and match them with several different shuffle types.
5389 LowerVECTOR_SHUFFLE_128v4(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
5390 SDValue V1 = SVOp->getOperand(0);
5391 SDValue V2 = SVOp->getOperand(1);
5392 DebugLoc dl = SVOp->getDebugLoc();
5393 EVT VT = SVOp->getValueType(0);
5395 assert(VT.getSizeInBits() == 128 && "Unsupported vector size");
5397 SmallVector<std::pair<int, int>, 8> Locs;
5399 SmallVector<int, 8> Mask1(4U, -1);
5400 SmallVector<int, 8> PermMask;
5401 SVOp->getMask(PermMask);
5405 for (unsigned i = 0; i != 4; ++i) {
5406 int Idx = PermMask[i];
5408 Locs[i] = std::make_pair(-1, -1);
5410 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
5412 Locs[i] = std::make_pair(0, NumLo);
5416 Locs[i] = std::make_pair(1, NumHi);
5418 Mask1[2+NumHi] = Idx;
5424 if (NumLo <= 2 && NumHi <= 2) {
5425 // If no more than two elements come from either vector. This can be
5426 // implemented with two shuffles. First shuffle gather the elements.
5427 // The second shuffle, which takes the first shuffle as both of its
5428 // vector operands, put the elements into the right order.
5429 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5431 SmallVector<int, 8> Mask2(4U, -1);
5433 for (unsigned i = 0; i != 4; ++i) {
5434 if (Locs[i].first == -1)
5437 unsigned Idx = (i < 2) ? 0 : 4;
5438 Idx += Locs[i].first * 2 + Locs[i].second;
5443 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
5444 } else if (NumLo == 3 || NumHi == 3) {
5445 // Otherwise, we must have three elements from one vector, call it X, and
5446 // one element from the other, call it Y. First, use a shufps to build an
5447 // intermediate vector with the one element from Y and the element from X
5448 // that will be in the same half in the final destination (the indexes don't
5449 // matter). Then, use a shufps to build the final vector, taking the half
5450 // containing the element from Y from the intermediate, and the other half
5453 // Normalize it so the 3 elements come from V1.
5454 CommuteVectorShuffleMask(PermMask, VT);
5458 // Find the element from V2.
5460 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
5461 int Val = PermMask[HiIndex];
5468 Mask1[0] = PermMask[HiIndex];
5470 Mask1[2] = PermMask[HiIndex^1];
5472 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5475 Mask1[0] = PermMask[0];
5476 Mask1[1] = PermMask[1];
5477 Mask1[2] = HiIndex & 1 ? 6 : 4;
5478 Mask1[3] = HiIndex & 1 ? 4 : 6;
5479 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5481 Mask1[0] = HiIndex & 1 ? 2 : 0;
5482 Mask1[1] = HiIndex & 1 ? 0 : 2;
5483 Mask1[2] = PermMask[2];
5484 Mask1[3] = PermMask[3];
5489 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
5493 // Break it into (shuffle shuffle_hi, shuffle_lo).
5496 SmallVector<int,8> LoMask(4U, -1);
5497 SmallVector<int,8> HiMask(4U, -1);
5499 SmallVector<int,8> *MaskPtr = &LoMask;
5500 unsigned MaskIdx = 0;
5503 for (unsigned i = 0; i != 4; ++i) {
5510 int Idx = PermMask[i];
5512 Locs[i] = std::make_pair(-1, -1);
5513 } else if (Idx < 4) {
5514 Locs[i] = std::make_pair(MaskIdx, LoIdx);
5515 (*MaskPtr)[LoIdx] = Idx;
5518 Locs[i] = std::make_pair(MaskIdx, HiIdx);
5519 (*MaskPtr)[HiIdx] = Idx;
5524 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
5525 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
5526 SmallVector<int, 8> MaskOps;
5527 for (unsigned i = 0; i != 4; ++i) {
5528 if (Locs[i].first == -1) {
5529 MaskOps.push_back(-1);
5531 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
5532 MaskOps.push_back(Idx);
5535 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
5538 static bool MayFoldVectorLoad(SDValue V) {
5539 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
5540 V = V.getOperand(0);
5541 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5542 V = V.getOperand(0);
5548 // FIXME: the version above should always be used. Since there's
5549 // a bug where several vector shuffles can't be folded because the
5550 // DAG is not updated during lowering and a node claims to have two
5551 // uses while it only has one, use this version, and let isel match
5552 // another instruction if the load really happens to have more than
5553 // one use. Remove this version after this bug get fixed.
5554 // rdar://8434668, PR8156
5555 static bool RelaxedMayFoldVectorLoad(SDValue V) {
5556 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
5557 V = V.getOperand(0);
5558 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5559 V = V.getOperand(0);
5560 if (ISD::isNormalLoad(V.getNode()))
5565 /// CanFoldShuffleIntoVExtract - Check if the current shuffle is used by
5566 /// a vector extract, and if both can be later optimized into a single load.
5567 /// This is done in visitEXTRACT_VECTOR_ELT and the conditions are checked
5568 /// here because otherwise a target specific shuffle node is going to be
5569 /// emitted for this shuffle, and the optimization not done.
5570 /// FIXME: This is probably not the best approach, but fix the problem
5571 /// until the right path is decided.
5573 bool CanXFormVExtractWithShuffleIntoLoad(SDValue V, SelectionDAG &DAG,
5574 const TargetLowering &TLI) {
5575 EVT VT = V.getValueType();
5576 ShuffleVectorSDNode *SVOp = dyn_cast<ShuffleVectorSDNode>(V);
5578 // Be sure that the vector shuffle is present in a pattern like this:
5579 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), c) -> (f32 load $addr)
5583 SDNode *N = *V.getNode()->use_begin();
5584 if (N->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5587 SDValue EltNo = N->getOperand(1);
5588 if (!isa<ConstantSDNode>(EltNo))
5591 // If the bit convert changed the number of elements, it is unsafe
5592 // to examine the mask.
5593 bool HasShuffleIntoBitcast = false;
5594 if (V.getOpcode() == ISD::BITCAST) {
5595 EVT SrcVT = V.getOperand(0).getValueType();
5596 if (SrcVT.getVectorNumElements() != VT.getVectorNumElements())
5598 V = V.getOperand(0);
5599 HasShuffleIntoBitcast = true;
5602 // Select the input vector, guarding against out of range extract vector.
5603 unsigned NumElems = VT.getVectorNumElements();
5604 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5605 int Idx = (Elt > NumElems) ? -1 : SVOp->getMaskElt(Elt);
5606 V = (Idx < (int)NumElems) ? V.getOperand(0) : V.getOperand(1);
5608 // Skip one more bit_convert if necessary
5609 if (V.getOpcode() == ISD::BITCAST)
5610 V = V.getOperand(0);
5612 if (ISD::isNormalLoad(V.getNode())) {
5613 // Is the original load suitable?
5614 LoadSDNode *LN0 = cast<LoadSDNode>(V);
5616 // FIXME: avoid the multi-use bug that is preventing lots of
5617 // of foldings to be detected, this is still wrong of course, but
5618 // give the temporary desired behavior, and if it happens that
5619 // the load has real more uses, during isel it will not fold, and
5620 // will generate poor code.
5621 if (!LN0 || LN0->isVolatile()) // || !LN0->hasOneUse()
5624 if (!HasShuffleIntoBitcast)
5627 // If there's a bitcast before the shuffle, check if the load type and
5628 // alignment is valid.
5629 unsigned Align = LN0->getAlignment();
5631 TLI.getTargetData()->getABITypeAlignment(
5632 VT.getTypeForEVT(*DAG.getContext()));
5634 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
5642 SDValue getMOVDDup(SDValue &Op, DebugLoc &dl, SDValue V1, SelectionDAG &DAG) {
5643 EVT VT = Op.getValueType();
5645 // Canonizalize to v2f64.
5646 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
5647 return DAG.getNode(ISD::BITCAST, dl, VT,
5648 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
5653 SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
5655 SDValue V1 = Op.getOperand(0);
5656 SDValue V2 = Op.getOperand(1);
5657 EVT VT = Op.getValueType();
5659 assert(VT != MVT::v2i64 && "unsupported shuffle type");
5661 if (HasSSE2 && VT == MVT::v2f64)
5662 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
5665 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V2, DAG);
5669 SDValue getMOVHighToLow(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG) {
5670 SDValue V1 = Op.getOperand(0);
5671 SDValue V2 = Op.getOperand(1);
5672 EVT VT = Op.getValueType();
5674 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
5675 "unsupported shuffle type");
5677 if (V2.getOpcode() == ISD::UNDEF)
5681 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
5685 SDValue getMOVLP(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
5686 SDValue V1 = Op.getOperand(0);
5687 SDValue V2 = Op.getOperand(1);
5688 EVT VT = Op.getValueType();
5689 unsigned NumElems = VT.getVectorNumElements();
5691 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
5692 // operand of these instructions is only memory, so check if there's a
5693 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
5695 bool CanFoldLoad = false;
5697 // Trivial case, when V2 comes from a load.
5698 if (MayFoldVectorLoad(V2))
5701 // When V1 is a load, it can be folded later into a store in isel, example:
5702 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
5704 // (MOVLPSmr addr:$src1, VR128:$src2)
5705 // So, recognize this potential and also use MOVLPS or MOVLPD
5706 if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
5709 // Both of them can't be memory operations though.
5710 if (MayFoldVectorLoad(V1) && MayFoldVectorLoad(V2))
5711 CanFoldLoad = false;
5714 if (HasSSE2 && NumElems == 2)
5715 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
5718 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
5721 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5722 // movl and movlp will both match v2i64, but v2i64 is never matched by
5723 // movl earlier because we make it strict to avoid messing with the movlp load
5724 // folding logic (see the code above getMOVLP call). Match it here then,
5725 // this is horrible, but will stay like this until we move all shuffle
5726 // matching to x86 specific nodes. Note that for the 1st condition all
5727 // types are matched with movsd.
5728 if ((HasSSE2 && NumElems == 2) || !X86::isMOVLMask(SVOp))
5729 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5731 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5734 assert(VT != MVT::v4i32 && "unsupported shuffle type");
5736 // Invert the operand order and use SHUFPS to match it.
5737 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V2, V1,
5738 X86::getShuffleSHUFImmediate(SVOp), DAG);
5741 static inline unsigned getUNPCKLOpcode(EVT VT, const X86Subtarget *Subtarget) {
5742 switch(VT.getSimpleVT().SimpleTy) {
5743 case MVT::v4i32: return X86ISD::PUNPCKLDQ;
5744 case MVT::v2i64: return X86ISD::PUNPCKLQDQ;
5746 return Subtarget->hasAVX() ? X86ISD::VUNPCKLPS : X86ISD::UNPCKLPS;
5748 return Subtarget->hasAVX() ? X86ISD::VUNPCKLPD : X86ISD::UNPCKLPD;
5749 case MVT::v8f32: return X86ISD::VUNPCKLPSY;
5750 case MVT::v4f64: return X86ISD::VUNPCKLPDY;
5751 case MVT::v16i8: return X86ISD::PUNPCKLBW;
5752 case MVT::v8i16: return X86ISD::PUNPCKLWD;
5754 llvm_unreachable("Unknown type for unpckl");
5759 static inline unsigned getUNPCKHOpcode(EVT VT) {
5760 switch(VT.getSimpleVT().SimpleTy) {
5761 case MVT::v4i32: return X86ISD::PUNPCKHDQ;
5762 case MVT::v2i64: return X86ISD::PUNPCKHQDQ;
5763 case MVT::v4f32: return X86ISD::UNPCKHPS;
5764 case MVT::v2f64: return X86ISD::UNPCKHPD;
5765 case MVT::v16i8: return X86ISD::PUNPCKHBW;
5766 case MVT::v8i16: return X86ISD::PUNPCKHWD;
5768 llvm_unreachable("Unknown type for unpckh");
5774 SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG,
5775 const TargetLowering &TLI,
5776 const X86Subtarget *Subtarget) {
5777 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5778 EVT VT = Op.getValueType();
5779 DebugLoc dl = Op.getDebugLoc();
5780 SDValue V1 = Op.getOperand(0);
5781 SDValue V2 = Op.getOperand(1);
5783 if (isZeroShuffle(SVOp))
5784 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
5786 // Handle splat operations
5787 if (SVOp->isSplat()) {
5788 unsigned NumElem = VT.getVectorNumElements();
5789 // Special case, this is the only place now where it's allowed to return
5790 // a vector_shuffle operation without using a target specific node, because
5791 // *hopefully* it will be optimized away by the dag combiner. FIXME: should
5792 // this be moved to DAGCombine instead?
5793 if (NumElem <= 4 && CanXFormVExtractWithShuffleIntoLoad(Op, DAG, TLI))
5796 // Handle splats by matching through known masks
5797 if ((VT.is128BitVector() && NumElem <= 4) ||
5798 (VT.is256BitVector() && NumElem <= 8))
5801 // All i16 and i8 vector types can't be used directly by a generic shuffle
5802 // instruction because the target has no such instruction. Generate shuffles
5803 // which repeat i16 and i8 several times until they fit in i32, and then can
5804 // be manipulated by target suported shuffles. After the insertion of the
5805 // necessary shuffles, the result is bitcasted back to v4f32 or v8f32.
5806 return PromoteSplat(SVOp, DAG);
5809 // If the shuffle can be profitably rewritten as a narrower shuffle, then
5811 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
5812 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5813 if (NewOp.getNode())
5814 return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
5815 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
5816 // FIXME: Figure out a cleaner way to do this.
5817 // Try to make use of movq to zero out the top part.
5818 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
5819 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5820 if (NewOp.getNode()) {
5821 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
5822 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
5823 DAG, Subtarget, dl);
5825 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
5826 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5827 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
5828 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
5829 DAG, Subtarget, dl);
5836 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
5837 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5838 SDValue V1 = Op.getOperand(0);
5839 SDValue V2 = Op.getOperand(1);
5840 EVT VT = Op.getValueType();
5841 DebugLoc dl = Op.getDebugLoc();
5842 unsigned NumElems = VT.getVectorNumElements();
5843 bool isMMX = VT.getSizeInBits() == 64;
5844 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
5845 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
5846 bool V1IsSplat = false;
5847 bool V2IsSplat = false;
5848 bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
5849 bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
5850 bool HasSSSE3 = Subtarget->hasSSSE3() || Subtarget->hasAVX();
5851 MachineFunction &MF = DAG.getMachineFunction();
5852 bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
5854 // Shuffle operations on MMX not supported.
5858 // Vector shuffle lowering takes 3 steps:
5860 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
5861 // narrowing and commutation of operands should be handled.
5862 // 2) Matching of shuffles with known shuffle masks to x86 target specific
5864 // 3) Rewriting of unmatched masks into new generic shuffle operations,
5865 // so the shuffle can be broken into other shuffles and the legalizer can
5866 // try the lowering again.
5868 // The general ideia is that no vector_shuffle operation should be left to
5869 // be matched during isel, all of them must be converted to a target specific
5872 // Normalize the input vectors. Here splats, zeroed vectors, profitable
5873 // narrowing and commutation of operands should be handled. The actual code
5874 // doesn't include all of those, work in progress...
5875 SDValue NewOp = NormalizeVectorShuffle(Op, DAG, *this, Subtarget);
5876 if (NewOp.getNode())
5879 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
5880 // unpckh_undef). Only use pshufd if speed is more important than size.
5881 if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp))
5882 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5883 return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()), dl, VT, V1, V1, DAG);
5884 if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp))
5885 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5886 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5888 if (X86::isMOVDDUPMask(SVOp) && HasSSE3 && V2IsUndef &&
5889 RelaxedMayFoldVectorLoad(V1))
5890 return getMOVDDup(Op, dl, V1, DAG);
5892 if (X86::isMOVHLPS_v_undef_Mask(SVOp))
5893 return getMOVHighToLow(Op, dl, DAG);
5895 // Use to match splats
5896 if (HasSSE2 && X86::isUNPCKHMask(SVOp) && V2IsUndef &&
5897 (VT == MVT::v2f64 || VT == MVT::v2i64))
5898 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5900 if (X86::isPSHUFDMask(SVOp)) {
5901 // The actual implementation will match the mask in the if above and then
5902 // during isel it can match several different instructions, not only pshufd
5903 // as its name says, sad but true, emulate the behavior for now...
5904 if (X86::isMOVDDUPMask(SVOp) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
5905 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
5907 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5909 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
5910 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
5912 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5913 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V1,
5916 if (VT == MVT::v4f32)
5917 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V1,
5921 // Check if this can be converted into a logical shift.
5922 bool isLeft = false;
5925 bool isShift = getSubtarget()->hasSSE2() &&
5926 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
5927 if (isShift && ShVal.hasOneUse()) {
5928 // If the shifted value has multiple uses, it may be cheaper to use
5929 // v_set0 + movlhps or movhlps, etc.
5930 EVT EltVT = VT.getVectorElementType();
5931 ShAmt *= EltVT.getSizeInBits();
5932 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5935 if (X86::isMOVLMask(SVOp)) {
5938 if (ISD::isBuildVectorAllZeros(V1.getNode()))
5939 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
5940 if (!X86::isMOVLPMask(SVOp)) {
5941 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5942 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5944 if (VT == MVT::v4i32 || VT == MVT::v4f32)
5945 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5949 // FIXME: fold these into legal mask.
5950 if (X86::isMOVLHPSMask(SVOp) && !X86::isUNPCKLMask(SVOp))
5951 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
5953 if (X86::isMOVHLPSMask(SVOp))
5954 return getMOVHighToLow(Op, dl, DAG);
5956 if (X86::isMOVSHDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5957 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
5959 if (X86::isMOVSLDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5960 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
5962 if (X86::isMOVLPMask(SVOp))
5963 return getMOVLP(Op, dl, DAG, HasSSE2);
5965 if (ShouldXformToMOVHLPS(SVOp) ||
5966 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
5967 return CommuteVectorShuffle(SVOp, DAG);
5970 // No better options. Use a vshl / vsrl.
5971 EVT EltVT = VT.getVectorElementType();
5972 ShAmt *= EltVT.getSizeInBits();
5973 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5976 bool Commuted = false;
5977 // FIXME: This should also accept a bitcast of a splat? Be careful, not
5978 // 1,1,1,1 -> v8i16 though.
5979 V1IsSplat = isSplatVector(V1.getNode());
5980 V2IsSplat = isSplatVector(V2.getNode());
5982 // Canonicalize the splat or undef, if present, to be on the RHS.
5983 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
5984 Op = CommuteVectorShuffle(SVOp, DAG);
5985 SVOp = cast<ShuffleVectorSDNode>(Op);
5986 V1 = SVOp->getOperand(0);
5987 V2 = SVOp->getOperand(1);
5988 std::swap(V1IsSplat, V2IsSplat);
5989 std::swap(V1IsUndef, V2IsUndef);
5993 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
5994 // Shuffling low element of v1 into undef, just return v1.
5997 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
5998 // the instruction selector will not match, so get a canonical MOVL with
5999 // swapped operands to undo the commute.
6000 return getMOVL(DAG, dl, VT, V2, V1);
6003 if (X86::isUNPCKLMask(SVOp))
6004 return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()),
6005 dl, VT, V1, V2, DAG);
6007 if (X86::isUNPCKHMask(SVOp))
6008 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V2, DAG);
6011 // Normalize mask so all entries that point to V2 points to its first
6012 // element then try to match unpck{h|l} again. If match, return a
6013 // new vector_shuffle with the corrected mask.
6014 SDValue NewMask = NormalizeMask(SVOp, DAG);
6015 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
6016 if (NSVOp != SVOp) {
6017 if (X86::isUNPCKLMask(NSVOp, true)) {
6019 } else if (X86::isUNPCKHMask(NSVOp, true)) {
6026 // Commute is back and try unpck* again.
6027 // FIXME: this seems wrong.
6028 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
6029 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
6031 if (X86::isUNPCKLMask(NewSVOp))
6032 return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()),
6033 dl, VT, V2, V1, DAG);
6035 if (X86::isUNPCKHMask(NewSVOp))
6036 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V2, V1, DAG);
6039 // Normalize the node to match x86 shuffle ops if needed
6040 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
6041 return CommuteVectorShuffle(SVOp, DAG);
6043 // The checks below are all present in isShuffleMaskLegal, but they are
6044 // inlined here right now to enable us to directly emit target specific
6045 // nodes, and remove one by one until they don't return Op anymore.
6046 SmallVector<int, 16> M;
6049 if (isPALIGNRMask(M, VT, HasSSSE3))
6050 return getTargetShuffleNode(X86ISD::PALIGN, dl, VT, V1, V2,
6051 X86::getShufflePALIGNRImmediate(SVOp),
6054 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
6055 SVOp->getSplatIndex() == 0 && V2IsUndef) {
6056 if (VT == MVT::v2f64) {
6057 X86ISD::NodeType Opcode =
6058 getSubtarget()->hasAVX() ? X86ISD::VUNPCKLPD : X86ISD::UNPCKLPD;
6059 return getTargetShuffleNode(Opcode, dl, VT, V1, V1, DAG);
6061 if (VT == MVT::v2i64)
6062 return getTargetShuffleNode(X86ISD::PUNPCKLQDQ, dl, VT, V1, V1, DAG);
6065 if (isPSHUFHWMask(M, VT))
6066 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
6067 X86::getShufflePSHUFHWImmediate(SVOp),
6070 if (isPSHUFLWMask(M, VT))
6071 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
6072 X86::getShufflePSHUFLWImmediate(SVOp),
6075 if (isSHUFPMask(M, VT)) {
6076 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
6077 if (VT == MVT::v4f32 || VT == MVT::v4i32)
6078 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V2,
6080 if (VT == MVT::v2f64 || VT == MVT::v2i64)
6081 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V2,
6085 if (X86::isUNPCKL_v_undef_Mask(SVOp))
6086 if (VT != MVT::v2i64 && VT != MVT::v2f64)
6087 return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()),
6088 dl, VT, V1, V1, DAG);
6089 if (X86::isUNPCKH_v_undef_Mask(SVOp))
6090 if (VT != MVT::v2i64 && VT != MVT::v2f64)
6091 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
6093 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
6094 if (VT == MVT::v8i16) {
6095 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
6096 if (NewOp.getNode())
6100 if (VT == MVT::v16i8) {
6101 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
6102 if (NewOp.getNode())
6106 // Handle all 128-bit wide vectors with 4 elements, and match them with
6107 // several different shuffle types.
6108 if (NumElems == 4 && VT.getSizeInBits() == 128)
6109 return LowerVECTOR_SHUFFLE_128v4(SVOp, DAG);
6111 //===--------------------------------------------------------------------===//
6112 // Custom lower or generate target specific nodes for 256-bit shuffles.
6114 // Handle VPERMIL permutations
6115 if (isVPERMILMask(M, VT)) {
6116 unsigned TargetMask = getShuffleVPERMILImmediate(SVOp);
6117 if (VT == MVT::v8f32)
6118 return getTargetShuffleNode(X86ISD::VPERMIL, dl, VT, V1, TargetMask, DAG);
6121 // Handle general 256-bit shuffles
6122 if (VT.is256BitVector())
6123 return LowerVECTOR_SHUFFLE_256(SVOp, DAG);
6129 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
6130 SelectionDAG &DAG) const {
6131 EVT VT = Op.getValueType();
6132 DebugLoc dl = Op.getDebugLoc();
6133 if (VT.getSizeInBits() == 8) {
6134 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
6135 Op.getOperand(0), Op.getOperand(1));
6136 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
6137 DAG.getValueType(VT));
6138 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6139 } else if (VT.getSizeInBits() == 16) {
6140 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6141 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
6143 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
6144 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6145 DAG.getNode(ISD::BITCAST, dl,
6149 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
6150 Op.getOperand(0), Op.getOperand(1));
6151 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
6152 DAG.getValueType(VT));
6153 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6154 } else if (VT == MVT::f32) {
6155 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
6156 // the result back to FR32 register. It's only worth matching if the
6157 // result has a single use which is a store or a bitcast to i32. And in
6158 // the case of a store, it's not worth it if the index is a constant 0,
6159 // because a MOVSSmr can be used instead, which is smaller and faster.
6160 if (!Op.hasOneUse())
6162 SDNode *User = *Op.getNode()->use_begin();
6163 if ((User->getOpcode() != ISD::STORE ||
6164 (isa<ConstantSDNode>(Op.getOperand(1)) &&
6165 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
6166 (User->getOpcode() != ISD::BITCAST ||
6167 User->getValueType(0) != MVT::i32))
6169 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6170 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
6173 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
6174 } else if (VT == MVT::i32) {
6175 // ExtractPS works with constant index.
6176 if (isa<ConstantSDNode>(Op.getOperand(1)))
6184 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
6185 SelectionDAG &DAG) const {
6186 if (!isa<ConstantSDNode>(Op.getOperand(1)))
6189 SDValue Vec = Op.getOperand(0);
6190 EVT VecVT = Vec.getValueType();
6192 // If this is a 256-bit vector result, first extract the 128-bit
6193 // vector and then extract from the 128-bit vector.
6194 if (VecVT.getSizeInBits() > 128) {
6195 DebugLoc dl = Op.getNode()->getDebugLoc();
6196 unsigned NumElems = VecVT.getVectorNumElements();
6197 SDValue Idx = Op.getOperand(1);
6199 if (!isa<ConstantSDNode>(Idx))
6202 unsigned ExtractNumElems = NumElems / (VecVT.getSizeInBits() / 128);
6203 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
6205 // Get the 128-bit vector.
6206 bool Upper = IdxVal >= ExtractNumElems;
6207 Vec = Extract128BitVector(Vec, Idx, DAG, dl);
6210 SDValue ScaledIdx = Idx;
6212 ScaledIdx = DAG.getNode(ISD::SUB, dl, Idx.getValueType(), Idx,
6213 DAG.getConstant(ExtractNumElems,
6214 Idx.getValueType()));
6215 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
6219 assert(Vec.getValueSizeInBits() <= 128 && "Unexpected vector length");
6221 if (Subtarget->hasSSE41()) {
6222 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
6227 EVT VT = Op.getValueType();
6228 DebugLoc dl = Op.getDebugLoc();
6229 // TODO: handle v16i8.
6230 if (VT.getSizeInBits() == 16) {
6231 SDValue Vec = Op.getOperand(0);
6232 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6234 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
6235 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6236 DAG.getNode(ISD::BITCAST, dl,
6239 // Transform it so it match pextrw which produces a 32-bit result.
6240 EVT EltVT = MVT::i32;
6241 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
6242 Op.getOperand(0), Op.getOperand(1));
6243 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
6244 DAG.getValueType(VT));
6245 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6246 } else if (VT.getSizeInBits() == 32) {
6247 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6251 // SHUFPS the element to the lowest double word, then movss.
6252 int Mask[4] = { Idx, -1, -1, -1 };
6253 EVT VVT = Op.getOperand(0).getValueType();
6254 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
6255 DAG.getUNDEF(VVT), Mask);
6256 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
6257 DAG.getIntPtrConstant(0));
6258 } else if (VT.getSizeInBits() == 64) {
6259 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
6260 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
6261 // to match extract_elt for f64.
6262 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6266 // UNPCKHPD the element to the lowest double word, then movsd.
6267 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
6268 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
6269 int Mask[2] = { 1, -1 };
6270 EVT VVT = Op.getOperand(0).getValueType();
6271 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
6272 DAG.getUNDEF(VVT), Mask);
6273 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
6274 DAG.getIntPtrConstant(0));
6281 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
6282 SelectionDAG &DAG) const {
6283 EVT VT = Op.getValueType();
6284 EVT EltVT = VT.getVectorElementType();
6285 DebugLoc dl = Op.getDebugLoc();
6287 SDValue N0 = Op.getOperand(0);
6288 SDValue N1 = Op.getOperand(1);
6289 SDValue N2 = Op.getOperand(2);
6291 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
6292 isa<ConstantSDNode>(N2)) {
6294 if (VT == MVT::v8i16)
6295 Opc = X86ISD::PINSRW;
6296 else if (VT == MVT::v16i8)
6297 Opc = X86ISD::PINSRB;
6299 Opc = X86ISD::PINSRB;
6301 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
6303 if (N1.getValueType() != MVT::i32)
6304 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
6305 if (N2.getValueType() != MVT::i32)
6306 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
6307 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
6308 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
6309 // Bits [7:6] of the constant are the source select. This will always be
6310 // zero here. The DAG Combiner may combine an extract_elt index into these
6311 // bits. For example (insert (extract, 3), 2) could be matched by putting
6312 // the '3' into bits [7:6] of X86ISD::INSERTPS.
6313 // Bits [5:4] of the constant are the destination select. This is the
6314 // value of the incoming immediate.
6315 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
6316 // combine either bitwise AND or insert of float 0.0 to set these bits.
6317 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
6318 // Create this as a scalar to vector..
6319 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
6320 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
6321 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
6322 // PINSR* works with constant index.
6329 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
6330 EVT VT = Op.getValueType();
6331 EVT EltVT = VT.getVectorElementType();
6333 DebugLoc dl = Op.getDebugLoc();
6334 SDValue N0 = Op.getOperand(0);
6335 SDValue N1 = Op.getOperand(1);
6336 SDValue N2 = Op.getOperand(2);
6338 // If this is a 256-bit vector result, first insert into a 128-bit
6339 // vector and then insert into the 256-bit vector.
6340 if (VT.getSizeInBits() > 128) {
6341 if (!isa<ConstantSDNode>(N2))
6344 // Get the 128-bit vector.
6345 unsigned NumElems = VT.getVectorNumElements();
6346 unsigned IdxVal = cast<ConstantSDNode>(N2)->getZExtValue();
6347 bool Upper = IdxVal >= NumElems / 2;
6349 SDValue SubN0 = Extract128BitVector(N0, N2, DAG, dl);
6352 SDValue ScaledN2 = N2;
6354 ScaledN2 = DAG.getNode(ISD::SUB, dl, N2.getValueType(), N2,
6355 DAG.getConstant(NumElems /
6356 (VT.getSizeInBits() / 128),
6357 N2.getValueType()));
6358 Op = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, SubN0.getValueType(), SubN0,
6361 // Insert the 128-bit vector
6362 // FIXME: Why UNDEF?
6363 return Insert128BitVector(N0, Op, N2, DAG, dl);
6366 if (Subtarget->hasSSE41())
6367 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
6369 if (EltVT == MVT::i8)
6372 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
6373 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
6374 // as its second argument.
6375 if (N1.getValueType() != MVT::i32)
6376 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
6377 if (N2.getValueType() != MVT::i32)
6378 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
6379 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
6385 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6386 LLVMContext *Context = DAG.getContext();
6387 DebugLoc dl = Op.getDebugLoc();
6388 EVT OpVT = Op.getValueType();
6390 // If this is a 256-bit vector result, first insert into a 128-bit
6391 // vector and then insert into the 256-bit vector.
6392 if (OpVT.getSizeInBits() > 128) {
6393 // Insert into a 128-bit vector.
6394 EVT VT128 = EVT::getVectorVT(*Context,
6395 OpVT.getVectorElementType(),
6396 OpVT.getVectorNumElements() / 2);
6398 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
6400 // Insert the 128-bit vector.
6401 return Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, OpVT), Op,
6402 DAG.getConstant(0, MVT::i32),
6406 if (Op.getValueType() == MVT::v1i64 &&
6407 Op.getOperand(0).getValueType() == MVT::i64)
6408 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
6410 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
6411 assert(Op.getValueType().getSimpleVT().getSizeInBits() == 128 &&
6412 "Expected an SSE type!");
6413 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(),
6414 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
6417 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
6418 // a simple subregister reference or explicit instructions to grab
6419 // upper bits of a vector.
6421 X86TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
6422 if (Subtarget->hasAVX()) {
6423 DebugLoc dl = Op.getNode()->getDebugLoc();
6424 SDValue Vec = Op.getNode()->getOperand(0);
6425 SDValue Idx = Op.getNode()->getOperand(1);
6427 if (Op.getNode()->getValueType(0).getSizeInBits() == 128
6428 && Vec.getNode()->getValueType(0).getSizeInBits() == 256) {
6429 return Extract128BitVector(Vec, Idx, DAG, dl);
6435 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
6436 // simple superregister reference or explicit instructions to insert
6437 // the upper bits of a vector.
6439 X86TargetLowering::LowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
6440 if (Subtarget->hasAVX()) {
6441 DebugLoc dl = Op.getNode()->getDebugLoc();
6442 SDValue Vec = Op.getNode()->getOperand(0);
6443 SDValue SubVec = Op.getNode()->getOperand(1);
6444 SDValue Idx = Op.getNode()->getOperand(2);
6446 if (Op.getNode()->getValueType(0).getSizeInBits() == 256
6447 && SubVec.getNode()->getValueType(0).getSizeInBits() == 128) {
6448 return Insert128BitVector(Vec, SubVec, Idx, DAG, dl);
6454 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
6455 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
6456 // one of the above mentioned nodes. It has to be wrapped because otherwise
6457 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
6458 // be used to form addressing mode. These wrapped nodes will be selected
6461 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
6462 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
6464 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6466 unsigned char OpFlag = 0;
6467 unsigned WrapperKind = X86ISD::Wrapper;
6468 CodeModel::Model M = getTargetMachine().getCodeModel();
6470 if (Subtarget->isPICStyleRIPRel() &&
6471 (M == CodeModel::Small || M == CodeModel::Kernel))
6472 WrapperKind = X86ISD::WrapperRIP;
6473 else if (Subtarget->isPICStyleGOT())
6474 OpFlag = X86II::MO_GOTOFF;
6475 else if (Subtarget->isPICStyleStubPIC())
6476 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6478 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
6480 CP->getOffset(), OpFlag);
6481 DebugLoc DL = CP->getDebugLoc();
6482 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6483 // With PIC, the address is actually $g + Offset.
6485 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6486 DAG.getNode(X86ISD::GlobalBaseReg,
6487 DebugLoc(), getPointerTy()),
6494 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
6495 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
6497 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6499 unsigned char OpFlag = 0;
6500 unsigned WrapperKind = X86ISD::Wrapper;
6501 CodeModel::Model M = getTargetMachine().getCodeModel();
6503 if (Subtarget->isPICStyleRIPRel() &&
6504 (M == CodeModel::Small || M == CodeModel::Kernel))
6505 WrapperKind = X86ISD::WrapperRIP;
6506 else if (Subtarget->isPICStyleGOT())
6507 OpFlag = X86II::MO_GOTOFF;
6508 else if (Subtarget->isPICStyleStubPIC())
6509 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6511 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
6513 DebugLoc DL = JT->getDebugLoc();
6514 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6516 // With PIC, the address is actually $g + Offset.
6518 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6519 DAG.getNode(X86ISD::GlobalBaseReg,
6520 DebugLoc(), getPointerTy()),
6527 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
6528 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
6530 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6532 unsigned char OpFlag = 0;
6533 unsigned WrapperKind = X86ISD::Wrapper;
6534 CodeModel::Model M = getTargetMachine().getCodeModel();
6536 if (Subtarget->isPICStyleRIPRel() &&
6537 (M == CodeModel::Small || M == CodeModel::Kernel))
6538 WrapperKind = X86ISD::WrapperRIP;
6539 else if (Subtarget->isPICStyleGOT())
6540 OpFlag = X86II::MO_GOTOFF;
6541 else if (Subtarget->isPICStyleStubPIC())
6542 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6544 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
6546 DebugLoc DL = Op.getDebugLoc();
6547 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6550 // With PIC, the address is actually $g + Offset.
6551 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
6552 !Subtarget->is64Bit()) {
6553 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6554 DAG.getNode(X86ISD::GlobalBaseReg,
6555 DebugLoc(), getPointerTy()),
6563 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
6564 // Create the TargetBlockAddressAddress node.
6565 unsigned char OpFlags =
6566 Subtarget->ClassifyBlockAddressReference();
6567 CodeModel::Model M = getTargetMachine().getCodeModel();
6568 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
6569 DebugLoc dl = Op.getDebugLoc();
6570 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
6571 /*isTarget=*/true, OpFlags);
6573 if (Subtarget->isPICStyleRIPRel() &&
6574 (M == CodeModel::Small || M == CodeModel::Kernel))
6575 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6577 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6579 // With PIC, the address is actually $g + Offset.
6580 if (isGlobalRelativeToPICBase(OpFlags)) {
6581 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6582 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6590 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
6592 SelectionDAG &DAG) const {
6593 // Create the TargetGlobalAddress node, folding in the constant
6594 // offset if it is legal.
6595 unsigned char OpFlags =
6596 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
6597 CodeModel::Model M = getTargetMachine().getCodeModel();
6599 if (OpFlags == X86II::MO_NO_FLAG &&
6600 X86::isOffsetSuitableForCodeModel(Offset, M)) {
6601 // A direct static reference to a global.
6602 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
6605 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
6608 if (Subtarget->isPICStyleRIPRel() &&
6609 (M == CodeModel::Small || M == CodeModel::Kernel))
6610 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6612 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6614 // With PIC, the address is actually $g + Offset.
6615 if (isGlobalRelativeToPICBase(OpFlags)) {
6616 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6617 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6621 // For globals that require a load from a stub to get the address, emit the
6623 if (isGlobalStubReference(OpFlags))
6624 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
6625 MachinePointerInfo::getGOT(), false, false, 0);
6627 // If there was a non-zero offset that we didn't fold, create an explicit
6630 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
6631 DAG.getConstant(Offset, getPointerTy()));
6637 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
6638 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
6639 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
6640 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
6644 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
6645 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
6646 unsigned char OperandFlags) {
6647 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6648 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6649 DebugLoc dl = GA->getDebugLoc();
6650 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6651 GA->getValueType(0),
6655 SDValue Ops[] = { Chain, TGA, *InFlag };
6656 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
6658 SDValue Ops[] = { Chain, TGA };
6659 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
6662 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
6663 MFI->setAdjustsStack(true);
6665 SDValue Flag = Chain.getValue(1);
6666 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
6669 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
6671 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6674 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
6675 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
6676 DAG.getNode(X86ISD::GlobalBaseReg,
6677 DebugLoc(), PtrVT), InFlag);
6678 InFlag = Chain.getValue(1);
6680 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
6683 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
6685 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6687 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
6688 X86::RAX, X86II::MO_TLSGD);
6691 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
6692 // "local exec" model.
6693 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6694 const EVT PtrVT, TLSModel::Model model,
6696 DebugLoc dl = GA->getDebugLoc();
6698 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
6699 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
6700 is64Bit ? 257 : 256));
6702 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
6703 DAG.getIntPtrConstant(0),
6704 MachinePointerInfo(Ptr), false, false, 0);
6706 unsigned char OperandFlags = 0;
6707 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
6709 unsigned WrapperKind = X86ISD::Wrapper;
6710 if (model == TLSModel::LocalExec) {
6711 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
6712 } else if (is64Bit) {
6713 assert(model == TLSModel::InitialExec);
6714 OperandFlags = X86II::MO_GOTTPOFF;
6715 WrapperKind = X86ISD::WrapperRIP;
6717 assert(model == TLSModel::InitialExec);
6718 OperandFlags = X86II::MO_INDNTPOFF;
6721 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
6723 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6724 GA->getValueType(0),
6725 GA->getOffset(), OperandFlags);
6726 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
6728 if (model == TLSModel::InitialExec)
6729 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
6730 MachinePointerInfo::getGOT(), false, false, 0);
6732 // The address of the thread local variable is the add of the thread
6733 // pointer with the offset of the variable.
6734 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
6738 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
6740 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
6741 const GlobalValue *GV = GA->getGlobal();
6743 if (Subtarget->isTargetELF()) {
6744 // TODO: implement the "local dynamic" model
6745 // TODO: implement the "initial exec"model for pic executables
6747 // If GV is an alias then use the aliasee for determining
6748 // thread-localness.
6749 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
6750 GV = GA->resolveAliasedGlobal(false);
6752 TLSModel::Model model
6753 = getTLSModel(GV, getTargetMachine().getRelocationModel());
6756 case TLSModel::GeneralDynamic:
6757 case TLSModel::LocalDynamic: // not implemented
6758 if (Subtarget->is64Bit())
6759 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
6760 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
6762 case TLSModel::InitialExec:
6763 case TLSModel::LocalExec:
6764 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
6765 Subtarget->is64Bit());
6767 } else if (Subtarget->isTargetDarwin()) {
6768 // Darwin only has one model of TLS. Lower to that.
6769 unsigned char OpFlag = 0;
6770 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
6771 X86ISD::WrapperRIP : X86ISD::Wrapper;
6773 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6775 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
6776 !Subtarget->is64Bit();
6778 OpFlag = X86II::MO_TLVP_PIC_BASE;
6780 OpFlag = X86II::MO_TLVP;
6781 DebugLoc DL = Op.getDebugLoc();
6782 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
6783 GA->getValueType(0),
6784 GA->getOffset(), OpFlag);
6785 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6787 // With PIC32, the address is actually $g + Offset.
6789 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6790 DAG.getNode(X86ISD::GlobalBaseReg,
6791 DebugLoc(), getPointerTy()),
6794 // Lowering the machine isd will make sure everything is in the right
6796 SDValue Chain = DAG.getEntryNode();
6797 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6798 SDValue Args[] = { Chain, Offset };
6799 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args, 2);
6801 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
6802 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6803 MFI->setAdjustsStack(true);
6805 // And our return value (tls address) is in the standard call return value
6807 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
6808 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
6812 "TLS not implemented for this target.");
6814 llvm_unreachable("Unreachable");
6819 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values and
6820 /// take a 2 x i32 value to shift plus a shift amount.
6821 SDValue X86TargetLowering::LowerShiftParts(SDValue Op, SelectionDAG &DAG) const {
6822 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
6823 EVT VT = Op.getValueType();
6824 unsigned VTBits = VT.getSizeInBits();
6825 DebugLoc dl = Op.getDebugLoc();
6826 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
6827 SDValue ShOpLo = Op.getOperand(0);
6828 SDValue ShOpHi = Op.getOperand(1);
6829 SDValue ShAmt = Op.getOperand(2);
6830 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
6831 DAG.getConstant(VTBits - 1, MVT::i8))
6832 : DAG.getConstant(0, VT);
6835 if (Op.getOpcode() == ISD::SHL_PARTS) {
6836 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
6837 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
6839 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
6840 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
6843 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
6844 DAG.getConstant(VTBits, MVT::i8));
6845 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6846 AndNode, DAG.getConstant(0, MVT::i8));
6849 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6850 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
6851 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
6853 if (Op.getOpcode() == ISD::SHL_PARTS) {
6854 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6855 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6857 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6858 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6861 SDValue Ops[2] = { Lo, Hi };
6862 return DAG.getMergeValues(Ops, 2, dl);
6865 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
6866 SelectionDAG &DAG) const {
6867 EVT SrcVT = Op.getOperand(0).getValueType();
6869 if (SrcVT.isVector())
6872 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
6873 "Unknown SINT_TO_FP to lower!");
6875 // These are really Legal; return the operand so the caller accepts it as
6877 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
6879 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
6880 Subtarget->is64Bit()) {
6884 DebugLoc dl = Op.getDebugLoc();
6885 unsigned Size = SrcVT.getSizeInBits()/8;
6886 MachineFunction &MF = DAG.getMachineFunction();
6887 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
6888 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6889 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6891 MachinePointerInfo::getFixedStack(SSFI),
6893 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
6896 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
6898 SelectionDAG &DAG) const {
6900 DebugLoc DL = Op.getDebugLoc();
6902 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
6904 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
6906 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
6908 unsigned ByteSize = SrcVT.getSizeInBits()/8;
6910 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
6911 MachineMemOperand *MMO;
6913 int SSFI = FI->getIndex();
6915 DAG.getMachineFunction()
6916 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6917 MachineMemOperand::MOLoad, ByteSize, ByteSize);
6919 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
6920 StackSlot = StackSlot.getOperand(1);
6922 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
6923 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
6925 Tys, Ops, array_lengthof(Ops),
6929 Chain = Result.getValue(1);
6930 SDValue InFlag = Result.getValue(2);
6932 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
6933 // shouldn't be necessary except that RFP cannot be live across
6934 // multiple blocks. When stackifier is fixed, they can be uncoupled.
6935 MachineFunction &MF = DAG.getMachineFunction();
6936 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
6937 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
6938 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6939 Tys = DAG.getVTList(MVT::Other);
6941 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
6943 MachineMemOperand *MMO =
6944 DAG.getMachineFunction()
6945 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6946 MachineMemOperand::MOStore, SSFISize, SSFISize);
6948 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
6949 Ops, array_lengthof(Ops),
6950 Op.getValueType(), MMO);
6951 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
6952 MachinePointerInfo::getFixedStack(SSFI),
6959 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
6960 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
6961 SelectionDAG &DAG) const {
6962 // This algorithm is not obvious. Here it is in C code, more or less:
6964 double uint64_to_double( uint32_t hi, uint32_t lo ) {
6965 static const __m128i exp = { 0x4330000045300000ULL, 0 };
6966 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
6968 // Copy ints to xmm registers.
6969 __m128i xh = _mm_cvtsi32_si128( hi );
6970 __m128i xl = _mm_cvtsi32_si128( lo );
6972 // Combine into low half of a single xmm register.
6973 __m128i x = _mm_unpacklo_epi32( xh, xl );
6977 // Merge in appropriate exponents to give the integer bits the right
6979 x = _mm_unpacklo_epi32( x, exp );
6981 // Subtract away the biases to deal with the IEEE-754 double precision
6983 d = _mm_sub_pd( (__m128d) x, bias );
6985 // All conversions up to here are exact. The correctly rounded result is
6986 // calculated using the current rounding mode using the following
6988 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
6989 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
6990 // store doesn't really need to be here (except
6991 // maybe to zero the other double)
6996 DebugLoc dl = Op.getDebugLoc();
6997 LLVMContext *Context = DAG.getContext();
6999 // Build some magic constants.
7000 std::vector<Constant*> CV0;
7001 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
7002 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
7003 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
7004 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
7005 Constant *C0 = ConstantVector::get(CV0);
7006 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
7008 std::vector<Constant*> CV1;
7010 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
7012 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
7013 Constant *C1 = ConstantVector::get(CV1);
7014 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
7016 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
7017 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7019 DAG.getIntPtrConstant(1)));
7020 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
7021 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7023 DAG.getIntPtrConstant(0)));
7024 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
7025 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
7026 MachinePointerInfo::getConstantPool(),
7028 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
7029 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck2);
7030 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
7031 MachinePointerInfo::getConstantPool(),
7033 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
7035 // Add the halves; easiest way is to swap them into another reg first.
7036 int ShufMask[2] = { 1, -1 };
7037 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
7038 DAG.getUNDEF(MVT::v2f64), ShufMask);
7039 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
7040 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
7041 DAG.getIntPtrConstant(0));
7044 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
7045 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
7046 SelectionDAG &DAG) const {
7047 DebugLoc dl = Op.getDebugLoc();
7048 // FP constant to bias correct the final result.
7049 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
7052 // Load the 32-bit value into an XMM register.
7053 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
7054 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7056 DAG.getIntPtrConstant(0)));
7058 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7059 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
7060 DAG.getIntPtrConstant(0));
7062 // Or the load with the bias.
7063 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
7064 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7065 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
7067 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7068 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
7069 MVT::v2f64, Bias)));
7070 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7071 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
7072 DAG.getIntPtrConstant(0));
7074 // Subtract the bias.
7075 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
7077 // Handle final rounding.
7078 EVT DestVT = Op.getValueType();
7080 if (DestVT.bitsLT(MVT::f64)) {
7081 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
7082 DAG.getIntPtrConstant(0));
7083 } else if (DestVT.bitsGT(MVT::f64)) {
7084 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
7087 // Handle final rounding.
7091 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
7092 SelectionDAG &DAG) const {
7093 SDValue N0 = Op.getOperand(0);
7094 DebugLoc dl = Op.getDebugLoc();
7096 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
7097 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
7098 // the optimization here.
7099 if (DAG.SignBitIsZero(N0))
7100 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
7102 EVT SrcVT = N0.getValueType();
7103 EVT DstVT = Op.getValueType();
7104 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
7105 return LowerUINT_TO_FP_i64(Op, DAG);
7106 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
7107 return LowerUINT_TO_FP_i32(Op, DAG);
7109 // Make a 64-bit buffer, and use it to build an FILD.
7110 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
7111 if (SrcVT == MVT::i32) {
7112 SDValue WordOff = DAG.getConstant(4, getPointerTy());
7113 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
7114 getPointerTy(), StackSlot, WordOff);
7115 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7116 StackSlot, MachinePointerInfo(),
7118 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
7119 OffsetSlot, MachinePointerInfo(),
7121 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
7125 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
7126 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7127 StackSlot, MachinePointerInfo(),
7129 // For i64 source, we need to add the appropriate power of 2 if the input
7130 // was negative. This is the same as the optimization in
7131 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
7132 // we must be careful to do the computation in x87 extended precision, not
7133 // in SSE. (The generic code can't know it's OK to do this, or how to.)
7134 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
7135 MachineMemOperand *MMO =
7136 DAG.getMachineFunction()
7137 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7138 MachineMemOperand::MOLoad, 8, 8);
7140 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
7141 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
7142 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops, 3,
7145 APInt FF(32, 0x5F800000ULL);
7147 // Check whether the sign bit is set.
7148 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
7149 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
7152 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
7153 SDValue FudgePtr = DAG.getConstantPool(
7154 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
7157 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
7158 SDValue Zero = DAG.getIntPtrConstant(0);
7159 SDValue Four = DAG.getIntPtrConstant(4);
7160 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
7162 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
7164 // Load the value out, extending it from f32 to f80.
7165 // FIXME: Avoid the extend by constructing the right constant pool?
7166 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
7167 FudgePtr, MachinePointerInfo::getConstantPool(),
7168 MVT::f32, false, false, 4);
7169 // Extend everything to 80 bits to force it to be done on x87.
7170 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
7171 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
7174 std::pair<SDValue,SDValue> X86TargetLowering::
7175 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
7176 DebugLoc DL = Op.getDebugLoc();
7178 EVT DstTy = Op.getValueType();
7181 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
7185 assert(DstTy.getSimpleVT() <= MVT::i64 &&
7186 DstTy.getSimpleVT() >= MVT::i16 &&
7187 "Unknown FP_TO_SINT to lower!");
7189 // These are really Legal.
7190 if (DstTy == MVT::i32 &&
7191 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
7192 return std::make_pair(SDValue(), SDValue());
7193 if (Subtarget->is64Bit() &&
7194 DstTy == MVT::i64 &&
7195 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
7196 return std::make_pair(SDValue(), SDValue());
7198 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
7200 MachineFunction &MF = DAG.getMachineFunction();
7201 unsigned MemSize = DstTy.getSizeInBits()/8;
7202 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
7203 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7208 switch (DstTy.getSimpleVT().SimpleTy) {
7209 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
7210 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
7211 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
7212 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
7215 SDValue Chain = DAG.getEntryNode();
7216 SDValue Value = Op.getOperand(0);
7217 EVT TheVT = Op.getOperand(0).getValueType();
7218 if (isScalarFPTypeInSSEReg(TheVT)) {
7219 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
7220 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
7221 MachinePointerInfo::getFixedStack(SSFI),
7223 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
7225 Chain, StackSlot, DAG.getValueType(TheVT)
7228 MachineMemOperand *MMO =
7229 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7230 MachineMemOperand::MOLoad, MemSize, MemSize);
7231 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, 3,
7233 Chain = Value.getValue(1);
7234 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
7235 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7238 MachineMemOperand *MMO =
7239 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7240 MachineMemOperand::MOStore, MemSize, MemSize);
7242 // Build the FP_TO_INT*_IN_MEM
7243 SDValue Ops[] = { Chain, Value, StackSlot };
7244 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
7245 Ops, 3, DstTy, MMO);
7247 return std::make_pair(FIST, StackSlot);
7250 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
7251 SelectionDAG &DAG) const {
7252 if (Op.getValueType().isVector())
7255 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
7256 SDValue FIST = Vals.first, StackSlot = Vals.second;
7257 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
7258 if (FIST.getNode() == 0) return Op;
7261 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
7262 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
7265 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
7266 SelectionDAG &DAG) const {
7267 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
7268 SDValue FIST = Vals.first, StackSlot = Vals.second;
7269 assert(FIST.getNode() && "Unexpected failure");
7272 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
7273 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
7276 SDValue X86TargetLowering::LowerFABS(SDValue Op,
7277 SelectionDAG &DAG) const {
7278 LLVMContext *Context = DAG.getContext();
7279 DebugLoc dl = Op.getDebugLoc();
7280 EVT VT = Op.getValueType();
7283 EltVT = VT.getVectorElementType();
7284 std::vector<Constant*> CV;
7285 if (EltVT == MVT::f64) {
7286 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
7290 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
7296 Constant *C = ConstantVector::get(CV);
7297 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7298 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7299 MachinePointerInfo::getConstantPool(),
7301 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
7304 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
7305 LLVMContext *Context = DAG.getContext();
7306 DebugLoc dl = Op.getDebugLoc();
7307 EVT VT = Op.getValueType();
7310 EltVT = VT.getVectorElementType();
7311 std::vector<Constant*> CV;
7312 if (EltVT == MVT::f64) {
7313 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
7317 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
7323 Constant *C = ConstantVector::get(CV);
7324 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7325 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7326 MachinePointerInfo::getConstantPool(),
7328 if (VT.isVector()) {
7329 return DAG.getNode(ISD::BITCAST, dl, VT,
7330 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
7331 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7333 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, Mask)));
7335 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
7339 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
7340 LLVMContext *Context = DAG.getContext();
7341 SDValue Op0 = Op.getOperand(0);
7342 SDValue Op1 = Op.getOperand(1);
7343 DebugLoc dl = Op.getDebugLoc();
7344 EVT VT = Op.getValueType();
7345 EVT SrcVT = Op1.getValueType();
7347 // If second operand is smaller, extend it first.
7348 if (SrcVT.bitsLT(VT)) {
7349 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
7352 // And if it is bigger, shrink it first.
7353 if (SrcVT.bitsGT(VT)) {
7354 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
7358 // At this point the operands and the result should have the same
7359 // type, and that won't be f80 since that is not custom lowered.
7361 // First get the sign bit of second operand.
7362 std::vector<Constant*> CV;
7363 if (SrcVT == MVT::f64) {
7364 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
7365 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
7367 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
7368 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7369 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7370 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7372 Constant *C = ConstantVector::get(CV);
7373 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7374 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
7375 MachinePointerInfo::getConstantPool(),
7377 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
7379 // Shift sign bit right or left if the two operands have different types.
7380 if (SrcVT.bitsGT(VT)) {
7381 // Op0 is MVT::f32, Op1 is MVT::f64.
7382 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
7383 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
7384 DAG.getConstant(32, MVT::i32));
7385 SignBit = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, SignBit);
7386 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
7387 DAG.getIntPtrConstant(0));
7390 // Clear first operand sign bit.
7392 if (VT == MVT::f64) {
7393 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
7394 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
7396 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
7397 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7398 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7399 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7401 C = ConstantVector::get(CV);
7402 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7403 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7404 MachinePointerInfo::getConstantPool(),
7406 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
7408 // Or the value with the sign bit.
7409 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
7412 SDValue X86TargetLowering::LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) const {
7413 SDValue N0 = Op.getOperand(0);
7414 DebugLoc dl = Op.getDebugLoc();
7415 EVT VT = Op.getValueType();
7417 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
7418 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
7419 DAG.getConstant(1, VT));
7420 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT));
7423 /// Emit nodes that will be selected as "test Op0,Op0", or something
7425 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
7426 SelectionDAG &DAG) const {
7427 DebugLoc dl = Op.getDebugLoc();
7429 // CF and OF aren't always set the way we want. Determine which
7430 // of these we need.
7431 bool NeedCF = false;
7432 bool NeedOF = false;
7435 case X86::COND_A: case X86::COND_AE:
7436 case X86::COND_B: case X86::COND_BE:
7439 case X86::COND_G: case X86::COND_GE:
7440 case X86::COND_L: case X86::COND_LE:
7441 case X86::COND_O: case X86::COND_NO:
7446 // See if we can use the EFLAGS value from the operand instead of
7447 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
7448 // we prove that the arithmetic won't overflow, we can't use OF or CF.
7449 if (Op.getResNo() != 0 || NeedOF || NeedCF)
7450 // Emit a CMP with 0, which is the TEST pattern.
7451 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
7452 DAG.getConstant(0, Op.getValueType()));
7454 unsigned Opcode = 0;
7455 unsigned NumOperands = 0;
7456 switch (Op.getNode()->getOpcode()) {
7458 // Due to an isel shortcoming, be conservative if this add is likely to be
7459 // selected as part of a load-modify-store instruction. When the root node
7460 // in a match is a store, isel doesn't know how to remap non-chain non-flag
7461 // uses of other nodes in the match, such as the ADD in this case. This
7462 // leads to the ADD being left around and reselected, with the result being
7463 // two adds in the output. Alas, even if none our users are stores, that
7464 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
7465 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
7466 // climbing the DAG back to the root, and it doesn't seem to be worth the
7468 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7469 UE = Op.getNode()->use_end(); UI != UE; ++UI)
7470 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
7473 if (ConstantSDNode *C =
7474 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
7475 // An add of one will be selected as an INC.
7476 if (C->getAPIntValue() == 1) {
7477 Opcode = X86ISD::INC;
7482 // An add of negative one (subtract of one) will be selected as a DEC.
7483 if (C->getAPIntValue().isAllOnesValue()) {
7484 Opcode = X86ISD::DEC;
7490 // Otherwise use a regular EFLAGS-setting add.
7491 Opcode = X86ISD::ADD;
7495 // If the primary and result isn't used, don't bother using X86ISD::AND,
7496 // because a TEST instruction will be better.
7497 bool NonFlagUse = false;
7498 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7499 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
7501 unsigned UOpNo = UI.getOperandNo();
7502 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
7503 // Look pass truncate.
7504 UOpNo = User->use_begin().getOperandNo();
7505 User = *User->use_begin();
7508 if (User->getOpcode() != ISD::BRCOND &&
7509 User->getOpcode() != ISD::SETCC &&
7510 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
7523 // Due to the ISEL shortcoming noted above, be conservative if this op is
7524 // likely to be selected as part of a load-modify-store instruction.
7525 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7526 UE = Op.getNode()->use_end(); UI != UE; ++UI)
7527 if (UI->getOpcode() == ISD::STORE)
7530 // Otherwise use a regular EFLAGS-setting instruction.
7531 switch (Op.getNode()->getOpcode()) {
7532 default: llvm_unreachable("unexpected operator!");
7533 case ISD::SUB: Opcode = X86ISD::SUB; break;
7534 case ISD::OR: Opcode = X86ISD::OR; break;
7535 case ISD::XOR: Opcode = X86ISD::XOR; break;
7536 case ISD::AND: Opcode = X86ISD::AND; break;
7548 return SDValue(Op.getNode(), 1);
7555 // Emit a CMP with 0, which is the TEST pattern.
7556 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
7557 DAG.getConstant(0, Op.getValueType()));
7559 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
7560 SmallVector<SDValue, 4> Ops;
7561 for (unsigned i = 0; i != NumOperands; ++i)
7562 Ops.push_back(Op.getOperand(i));
7564 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
7565 DAG.ReplaceAllUsesWith(Op, New);
7566 return SDValue(New.getNode(), 1);
7569 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
7571 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
7572 SelectionDAG &DAG) const {
7573 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
7574 if (C->getAPIntValue() == 0)
7575 return EmitTest(Op0, X86CC, DAG);
7577 DebugLoc dl = Op0.getDebugLoc();
7578 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
7581 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
7582 /// if it's possible.
7583 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
7584 DebugLoc dl, SelectionDAG &DAG) const {
7585 SDValue Op0 = And.getOperand(0);
7586 SDValue Op1 = And.getOperand(1);
7587 if (Op0.getOpcode() == ISD::TRUNCATE)
7588 Op0 = Op0.getOperand(0);
7589 if (Op1.getOpcode() == ISD::TRUNCATE)
7590 Op1 = Op1.getOperand(0);
7593 if (Op1.getOpcode() == ISD::SHL)
7594 std::swap(Op0, Op1);
7595 if (Op0.getOpcode() == ISD::SHL) {
7596 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
7597 if (And00C->getZExtValue() == 1) {
7598 // If we looked past a truncate, check that it's only truncating away
7600 unsigned BitWidth = Op0.getValueSizeInBits();
7601 unsigned AndBitWidth = And.getValueSizeInBits();
7602 if (BitWidth > AndBitWidth) {
7603 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
7604 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
7605 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
7609 RHS = Op0.getOperand(1);
7611 } else if (Op1.getOpcode() == ISD::Constant) {
7612 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
7613 SDValue AndLHS = Op0;
7614 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
7615 LHS = AndLHS.getOperand(0);
7616 RHS = AndLHS.getOperand(1);
7620 if (LHS.getNode()) {
7621 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
7622 // instruction. Since the shift amount is in-range-or-undefined, we know
7623 // that doing a bittest on the i32 value is ok. We extend to i32 because
7624 // the encoding for the i16 version is larger than the i32 version.
7625 // Also promote i16 to i32 for performance / code size reason.
7626 if (LHS.getValueType() == MVT::i8 ||
7627 LHS.getValueType() == MVT::i16)
7628 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
7630 // If the operand types disagree, extend the shift amount to match. Since
7631 // BT ignores high bits (like shifts) we can use anyextend.
7632 if (LHS.getValueType() != RHS.getValueType())
7633 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
7635 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
7636 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
7637 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7638 DAG.getConstant(Cond, MVT::i8), BT);
7644 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
7645 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
7646 SDValue Op0 = Op.getOperand(0);
7647 SDValue Op1 = Op.getOperand(1);
7648 DebugLoc dl = Op.getDebugLoc();
7649 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
7651 // Optimize to BT if possible.
7652 // Lower (X & (1 << N)) == 0 to BT(X, N).
7653 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
7654 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
7655 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
7656 Op1.getOpcode() == ISD::Constant &&
7657 cast<ConstantSDNode>(Op1)->isNullValue() &&
7658 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7659 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
7660 if (NewSetCC.getNode())
7664 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
7666 if (Op1.getOpcode() == ISD::Constant &&
7667 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
7668 cast<ConstantSDNode>(Op1)->isNullValue()) &&
7669 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7671 // If the input is a setcc, then reuse the input setcc or use a new one with
7672 // the inverted condition.
7673 if (Op0.getOpcode() == X86ISD::SETCC) {
7674 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
7675 bool Invert = (CC == ISD::SETNE) ^
7676 cast<ConstantSDNode>(Op1)->isNullValue();
7677 if (!Invert) return Op0;
7679 CCode = X86::GetOppositeBranchCondition(CCode);
7680 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7681 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
7685 bool isFP = Op1.getValueType().isFloatingPoint();
7686 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
7687 if (X86CC == X86::COND_INVALID)
7690 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
7691 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7692 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
7695 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
7697 SDValue Op0 = Op.getOperand(0);
7698 SDValue Op1 = Op.getOperand(1);
7699 SDValue CC = Op.getOperand(2);
7700 EVT VT = Op.getValueType();
7701 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
7702 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
7703 DebugLoc dl = Op.getDebugLoc();
7707 EVT VT0 = Op0.getValueType();
7708 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
7709 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
7712 switch (SetCCOpcode) {
7715 case ISD::SETEQ: SSECC = 0; break;
7717 case ISD::SETGT: Swap = true; // Fallthrough
7719 case ISD::SETOLT: SSECC = 1; break;
7721 case ISD::SETGE: Swap = true; // Fallthrough
7723 case ISD::SETOLE: SSECC = 2; break;
7724 case ISD::SETUO: SSECC = 3; break;
7726 case ISD::SETNE: SSECC = 4; break;
7727 case ISD::SETULE: Swap = true;
7728 case ISD::SETUGE: SSECC = 5; break;
7729 case ISD::SETULT: Swap = true;
7730 case ISD::SETUGT: SSECC = 6; break;
7731 case ISD::SETO: SSECC = 7; break;
7734 std::swap(Op0, Op1);
7736 // In the two special cases we can't handle, emit two comparisons.
7738 if (SetCCOpcode == ISD::SETUEQ) {
7740 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
7741 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
7742 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
7744 else if (SetCCOpcode == ISD::SETONE) {
7746 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
7747 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
7748 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
7750 llvm_unreachable("Illegal FP comparison");
7752 // Handle all other FP comparisons here.
7753 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
7756 // We are handling one of the integer comparisons here. Since SSE only has
7757 // GT and EQ comparisons for integer, swapping operands and multiple
7758 // operations may be required for some comparisons.
7759 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
7760 bool Swap = false, Invert = false, FlipSigns = false;
7762 switch (VT.getSimpleVT().SimpleTy) {
7764 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
7765 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
7766 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
7767 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
7770 switch (SetCCOpcode) {
7772 case ISD::SETNE: Invert = true;
7773 case ISD::SETEQ: Opc = EQOpc; break;
7774 case ISD::SETLT: Swap = true;
7775 case ISD::SETGT: Opc = GTOpc; break;
7776 case ISD::SETGE: Swap = true;
7777 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
7778 case ISD::SETULT: Swap = true;
7779 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
7780 case ISD::SETUGE: Swap = true;
7781 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
7784 std::swap(Op0, Op1);
7786 // Since SSE has no unsigned integer comparisons, we need to flip the sign
7787 // bits of the inputs before performing those operations.
7789 EVT EltVT = VT.getVectorElementType();
7790 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
7792 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
7793 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
7795 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
7796 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
7799 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
7801 // If the logical-not of the result is required, perform that now.
7803 Result = DAG.getNOT(dl, Result, VT);
7808 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
7809 static bool isX86LogicalCmp(SDValue Op) {
7810 unsigned Opc = Op.getNode()->getOpcode();
7811 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
7813 if (Op.getResNo() == 1 &&
7814 (Opc == X86ISD::ADD ||
7815 Opc == X86ISD::SUB ||
7816 Opc == X86ISD::ADC ||
7817 Opc == X86ISD::SBB ||
7818 Opc == X86ISD::SMUL ||
7819 Opc == X86ISD::UMUL ||
7820 Opc == X86ISD::INC ||
7821 Opc == X86ISD::DEC ||
7822 Opc == X86ISD::OR ||
7823 Opc == X86ISD::XOR ||
7824 Opc == X86ISD::AND))
7827 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
7833 static bool isZero(SDValue V) {
7834 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
7835 return C && C->isNullValue();
7838 static bool isAllOnes(SDValue V) {
7839 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
7840 return C && C->isAllOnesValue();
7843 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
7844 bool addTest = true;
7845 SDValue Cond = Op.getOperand(0);
7846 SDValue Op1 = Op.getOperand(1);
7847 SDValue Op2 = Op.getOperand(2);
7848 DebugLoc DL = Op.getDebugLoc();
7851 if (Cond.getOpcode() == ISD::SETCC) {
7852 SDValue NewCond = LowerSETCC(Cond, DAG);
7853 if (NewCond.getNode())
7857 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
7858 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
7859 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
7860 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
7861 if (Cond.getOpcode() == X86ISD::SETCC &&
7862 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
7863 isZero(Cond.getOperand(1).getOperand(1))) {
7864 SDValue Cmp = Cond.getOperand(1);
7866 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
7868 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
7869 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
7870 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
7872 SDValue CmpOp0 = Cmp.getOperand(0);
7873 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
7874 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
7876 SDValue Res = // Res = 0 or -1.
7877 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
7878 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
7880 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
7881 Res = DAG.getNOT(DL, Res, Res.getValueType());
7883 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
7884 if (N2C == 0 || !N2C->isNullValue())
7885 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
7890 // Look past (and (setcc_carry (cmp ...)), 1).
7891 if (Cond.getOpcode() == ISD::AND &&
7892 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7893 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7894 if (C && C->getAPIntValue() == 1)
7895 Cond = Cond.getOperand(0);
7898 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7899 // setting operand in place of the X86ISD::SETCC.
7900 if (Cond.getOpcode() == X86ISD::SETCC ||
7901 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7902 CC = Cond.getOperand(0);
7904 SDValue Cmp = Cond.getOperand(1);
7905 unsigned Opc = Cmp.getOpcode();
7906 EVT VT = Op.getValueType();
7908 bool IllegalFPCMov = false;
7909 if (VT.isFloatingPoint() && !VT.isVector() &&
7910 !isScalarFPTypeInSSEReg(VT)) // FPStack?
7911 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
7913 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
7914 Opc == X86ISD::BT) { // FIXME
7921 // Look pass the truncate.
7922 if (Cond.getOpcode() == ISD::TRUNCATE)
7923 Cond = Cond.getOperand(0);
7925 // We know the result of AND is compared against zero. Try to match
7927 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7928 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
7929 if (NewSetCC.getNode()) {
7930 CC = NewSetCC.getOperand(0);
7931 Cond = NewSetCC.getOperand(1);
7938 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7939 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7942 // a < b ? -1 : 0 -> RES = ~setcc_carry
7943 // a < b ? 0 : -1 -> RES = setcc_carry
7944 // a >= b ? -1 : 0 -> RES = setcc_carry
7945 // a >= b ? 0 : -1 -> RES = ~setcc_carry
7946 if (Cond.getOpcode() == X86ISD::CMP) {
7947 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
7949 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
7950 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
7951 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
7952 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
7953 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
7954 return DAG.getNOT(DL, Res, Res.getValueType());
7959 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
7960 // condition is true.
7961 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
7962 SDValue Ops[] = { Op2, Op1, CC, Cond };
7963 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops, array_lengthof(Ops));
7966 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
7967 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
7968 // from the AND / OR.
7969 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
7970 Opc = Op.getOpcode();
7971 if (Opc != ISD::OR && Opc != ISD::AND)
7973 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7974 Op.getOperand(0).hasOneUse() &&
7975 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
7976 Op.getOperand(1).hasOneUse());
7979 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
7980 // 1 and that the SETCC node has a single use.
7981 static bool isXor1OfSetCC(SDValue Op) {
7982 if (Op.getOpcode() != ISD::XOR)
7984 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7985 if (N1C && N1C->getAPIntValue() == 1) {
7986 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7987 Op.getOperand(0).hasOneUse();
7992 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
7993 bool addTest = true;
7994 SDValue Chain = Op.getOperand(0);
7995 SDValue Cond = Op.getOperand(1);
7996 SDValue Dest = Op.getOperand(2);
7997 DebugLoc dl = Op.getDebugLoc();
8000 if (Cond.getOpcode() == ISD::SETCC) {
8001 SDValue NewCond = LowerSETCC(Cond, DAG);
8002 if (NewCond.getNode())
8006 // FIXME: LowerXALUO doesn't handle these!!
8007 else if (Cond.getOpcode() == X86ISD::ADD ||
8008 Cond.getOpcode() == X86ISD::SUB ||
8009 Cond.getOpcode() == X86ISD::SMUL ||
8010 Cond.getOpcode() == X86ISD::UMUL)
8011 Cond = LowerXALUO(Cond, DAG);
8014 // Look pass (and (setcc_carry (cmp ...)), 1).
8015 if (Cond.getOpcode() == ISD::AND &&
8016 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
8017 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
8018 if (C && C->getAPIntValue() == 1)
8019 Cond = Cond.getOperand(0);
8022 // If condition flag is set by a X86ISD::CMP, then use it as the condition
8023 // setting operand in place of the X86ISD::SETCC.
8024 if (Cond.getOpcode() == X86ISD::SETCC ||
8025 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
8026 CC = Cond.getOperand(0);
8028 SDValue Cmp = Cond.getOperand(1);
8029 unsigned Opc = Cmp.getOpcode();
8030 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
8031 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
8035 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
8039 // These can only come from an arithmetic instruction with overflow,
8040 // e.g. SADDO, UADDO.
8041 Cond = Cond.getNode()->getOperand(1);
8048 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
8049 SDValue Cmp = Cond.getOperand(0).getOperand(1);
8050 if (CondOpc == ISD::OR) {
8051 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
8052 // two branches instead of an explicit OR instruction with a
8054 if (Cmp == Cond.getOperand(1).getOperand(1) &&
8055 isX86LogicalCmp(Cmp)) {
8056 CC = Cond.getOperand(0).getOperand(0);
8057 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8058 Chain, Dest, CC, Cmp);
8059 CC = Cond.getOperand(1).getOperand(0);
8063 } else { // ISD::AND
8064 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
8065 // two branches instead of an explicit AND instruction with a
8066 // separate test. However, we only do this if this block doesn't
8067 // have a fall-through edge, because this requires an explicit
8068 // jmp when the condition is false.
8069 if (Cmp == Cond.getOperand(1).getOperand(1) &&
8070 isX86LogicalCmp(Cmp) &&
8071 Op.getNode()->hasOneUse()) {
8072 X86::CondCode CCode =
8073 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
8074 CCode = X86::GetOppositeBranchCondition(CCode);
8075 CC = DAG.getConstant(CCode, MVT::i8);
8076 SDNode *User = *Op.getNode()->use_begin();
8077 // Look for an unconditional branch following this conditional branch.
8078 // We need this because we need to reverse the successors in order
8079 // to implement FCMP_OEQ.
8080 if (User->getOpcode() == ISD::BR) {
8081 SDValue FalseBB = User->getOperand(1);
8083 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
8084 assert(NewBR == User);
8088 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8089 Chain, Dest, CC, Cmp);
8090 X86::CondCode CCode =
8091 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
8092 CCode = X86::GetOppositeBranchCondition(CCode);
8093 CC = DAG.getConstant(CCode, MVT::i8);
8099 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
8100 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
8101 // It should be transformed during dag combiner except when the condition
8102 // is set by a arithmetics with overflow node.
8103 X86::CondCode CCode =
8104 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
8105 CCode = X86::GetOppositeBranchCondition(CCode);
8106 CC = DAG.getConstant(CCode, MVT::i8);
8107 Cond = Cond.getOperand(0).getOperand(1);
8113 // Look pass the truncate.
8114 if (Cond.getOpcode() == ISD::TRUNCATE)
8115 Cond = Cond.getOperand(0);
8117 // We know the result of AND is compared against zero. Try to match
8119 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
8120 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
8121 if (NewSetCC.getNode()) {
8122 CC = NewSetCC.getOperand(0);
8123 Cond = NewSetCC.getOperand(1);
8130 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
8131 Cond = EmitTest(Cond, X86::COND_NE, DAG);
8133 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8134 Chain, Dest, CC, Cond);
8138 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
8139 // Calls to _alloca is needed to probe the stack when allocating more than 4k
8140 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
8141 // that the guard pages used by the OS virtual memory manager are allocated in
8142 // correct sequence.
8144 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
8145 SelectionDAG &DAG) const {
8146 assert((Subtarget->isTargetCygMing() || Subtarget->isTargetWindows()) &&
8147 "This should be used only on Windows targets");
8148 assert(!Subtarget->isTargetEnvMacho());
8149 DebugLoc dl = Op.getDebugLoc();
8152 SDValue Chain = Op.getOperand(0);
8153 SDValue Size = Op.getOperand(1);
8154 // FIXME: Ensure alignment here
8158 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
8159 unsigned Reg = (Subtarget->is64Bit() ? X86::RAX : X86::EAX);
8161 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
8162 Flag = Chain.getValue(1);
8164 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
8166 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
8167 Flag = Chain.getValue(1);
8169 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
8171 SDValue Ops1[2] = { Chain.getValue(0), Chain };
8172 return DAG.getMergeValues(Ops1, 2, dl);
8175 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
8176 MachineFunction &MF = DAG.getMachineFunction();
8177 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
8179 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
8180 DebugLoc DL = Op.getDebugLoc();
8182 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
8183 // vastart just stores the address of the VarArgsFrameIndex slot into the
8184 // memory location argument.
8185 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
8187 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
8188 MachinePointerInfo(SV), false, false, 0);
8192 // gp_offset (0 - 6 * 8)
8193 // fp_offset (48 - 48 + 8 * 16)
8194 // overflow_arg_area (point to parameters coming in memory).
8196 SmallVector<SDValue, 8> MemOps;
8197 SDValue FIN = Op.getOperand(1);
8199 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
8200 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
8202 FIN, MachinePointerInfo(SV), false, false, 0);
8203 MemOps.push_back(Store);
8206 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8207 FIN, DAG.getIntPtrConstant(4));
8208 Store = DAG.getStore(Op.getOperand(0), DL,
8209 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
8211 FIN, MachinePointerInfo(SV, 4), false, false, 0);
8212 MemOps.push_back(Store);
8214 // Store ptr to overflow_arg_area
8215 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8216 FIN, DAG.getIntPtrConstant(4));
8217 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
8219 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
8220 MachinePointerInfo(SV, 8),
8222 MemOps.push_back(Store);
8224 // Store ptr to reg_save_area.
8225 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8226 FIN, DAG.getIntPtrConstant(8));
8227 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
8229 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
8230 MachinePointerInfo(SV, 16), false, false, 0);
8231 MemOps.push_back(Store);
8232 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
8233 &MemOps[0], MemOps.size());
8236 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
8237 assert(Subtarget->is64Bit() &&
8238 "LowerVAARG only handles 64-bit va_arg!");
8239 assert((Subtarget->isTargetLinux() ||
8240 Subtarget->isTargetDarwin()) &&
8241 "Unhandled target in LowerVAARG");
8242 assert(Op.getNode()->getNumOperands() == 4);
8243 SDValue Chain = Op.getOperand(0);
8244 SDValue SrcPtr = Op.getOperand(1);
8245 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
8246 unsigned Align = Op.getConstantOperandVal(3);
8247 DebugLoc dl = Op.getDebugLoc();
8249 EVT ArgVT = Op.getNode()->getValueType(0);
8250 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
8251 uint32_t ArgSize = getTargetData()->getTypeAllocSize(ArgTy);
8254 // Decide which area this value should be read from.
8255 // TODO: Implement the AMD64 ABI in its entirety. This simple
8256 // selection mechanism works only for the basic types.
8257 if (ArgVT == MVT::f80) {
8258 llvm_unreachable("va_arg for f80 not yet implemented");
8259 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
8260 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
8261 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
8262 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
8264 llvm_unreachable("Unhandled argument type in LowerVAARG");
8268 // Sanity Check: Make sure using fp_offset makes sense.
8269 assert(!UseSoftFloat &&
8270 !(DAG.getMachineFunction()
8271 .getFunction()->hasFnAttr(Attribute::NoImplicitFloat)) &&
8272 Subtarget->hasXMM());
8275 // Insert VAARG_64 node into the DAG
8276 // VAARG_64 returns two values: Variable Argument Address, Chain
8277 SmallVector<SDValue, 11> InstOps;
8278 InstOps.push_back(Chain);
8279 InstOps.push_back(SrcPtr);
8280 InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
8281 InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
8282 InstOps.push_back(DAG.getConstant(Align, MVT::i32));
8283 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
8284 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
8285 VTs, &InstOps[0], InstOps.size(),
8287 MachinePointerInfo(SV),
8292 Chain = VAARG.getValue(1);
8294 // Load the next argument and return it
8295 return DAG.getLoad(ArgVT, dl,
8298 MachinePointerInfo(),
8302 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
8303 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
8304 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
8305 SDValue Chain = Op.getOperand(0);
8306 SDValue DstPtr = Op.getOperand(1);
8307 SDValue SrcPtr = Op.getOperand(2);
8308 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
8309 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
8310 DebugLoc DL = Op.getDebugLoc();
8312 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
8313 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
8315 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
8319 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
8320 DebugLoc dl = Op.getDebugLoc();
8321 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8323 default: return SDValue(); // Don't custom lower most intrinsics.
8324 // Comparison intrinsics.
8325 case Intrinsic::x86_sse_comieq_ss:
8326 case Intrinsic::x86_sse_comilt_ss:
8327 case Intrinsic::x86_sse_comile_ss:
8328 case Intrinsic::x86_sse_comigt_ss:
8329 case Intrinsic::x86_sse_comige_ss:
8330 case Intrinsic::x86_sse_comineq_ss:
8331 case Intrinsic::x86_sse_ucomieq_ss:
8332 case Intrinsic::x86_sse_ucomilt_ss:
8333 case Intrinsic::x86_sse_ucomile_ss:
8334 case Intrinsic::x86_sse_ucomigt_ss:
8335 case Intrinsic::x86_sse_ucomige_ss:
8336 case Intrinsic::x86_sse_ucomineq_ss:
8337 case Intrinsic::x86_sse2_comieq_sd:
8338 case Intrinsic::x86_sse2_comilt_sd:
8339 case Intrinsic::x86_sse2_comile_sd:
8340 case Intrinsic::x86_sse2_comigt_sd:
8341 case Intrinsic::x86_sse2_comige_sd:
8342 case Intrinsic::x86_sse2_comineq_sd:
8343 case Intrinsic::x86_sse2_ucomieq_sd:
8344 case Intrinsic::x86_sse2_ucomilt_sd:
8345 case Intrinsic::x86_sse2_ucomile_sd:
8346 case Intrinsic::x86_sse2_ucomigt_sd:
8347 case Intrinsic::x86_sse2_ucomige_sd:
8348 case Intrinsic::x86_sse2_ucomineq_sd: {
8350 ISD::CondCode CC = ISD::SETCC_INVALID;
8353 case Intrinsic::x86_sse_comieq_ss:
8354 case Intrinsic::x86_sse2_comieq_sd:
8358 case Intrinsic::x86_sse_comilt_ss:
8359 case Intrinsic::x86_sse2_comilt_sd:
8363 case Intrinsic::x86_sse_comile_ss:
8364 case Intrinsic::x86_sse2_comile_sd:
8368 case Intrinsic::x86_sse_comigt_ss:
8369 case Intrinsic::x86_sse2_comigt_sd:
8373 case Intrinsic::x86_sse_comige_ss:
8374 case Intrinsic::x86_sse2_comige_sd:
8378 case Intrinsic::x86_sse_comineq_ss:
8379 case Intrinsic::x86_sse2_comineq_sd:
8383 case Intrinsic::x86_sse_ucomieq_ss:
8384 case Intrinsic::x86_sse2_ucomieq_sd:
8385 Opc = X86ISD::UCOMI;
8388 case Intrinsic::x86_sse_ucomilt_ss:
8389 case Intrinsic::x86_sse2_ucomilt_sd:
8390 Opc = X86ISD::UCOMI;
8393 case Intrinsic::x86_sse_ucomile_ss:
8394 case Intrinsic::x86_sse2_ucomile_sd:
8395 Opc = X86ISD::UCOMI;
8398 case Intrinsic::x86_sse_ucomigt_ss:
8399 case Intrinsic::x86_sse2_ucomigt_sd:
8400 Opc = X86ISD::UCOMI;
8403 case Intrinsic::x86_sse_ucomige_ss:
8404 case Intrinsic::x86_sse2_ucomige_sd:
8405 Opc = X86ISD::UCOMI;
8408 case Intrinsic::x86_sse_ucomineq_ss:
8409 case Intrinsic::x86_sse2_ucomineq_sd:
8410 Opc = X86ISD::UCOMI;
8415 SDValue LHS = Op.getOperand(1);
8416 SDValue RHS = Op.getOperand(2);
8417 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
8418 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
8419 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
8420 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
8421 DAG.getConstant(X86CC, MVT::i8), Cond);
8422 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
8424 // ptest and testp intrinsics. The intrinsic these come from are designed to
8425 // return an integer value, not just an instruction so lower it to the ptest
8426 // or testp pattern and a setcc for the result.
8427 case Intrinsic::x86_sse41_ptestz:
8428 case Intrinsic::x86_sse41_ptestc:
8429 case Intrinsic::x86_sse41_ptestnzc:
8430 case Intrinsic::x86_avx_ptestz_256:
8431 case Intrinsic::x86_avx_ptestc_256:
8432 case Intrinsic::x86_avx_ptestnzc_256:
8433 case Intrinsic::x86_avx_vtestz_ps:
8434 case Intrinsic::x86_avx_vtestc_ps:
8435 case Intrinsic::x86_avx_vtestnzc_ps:
8436 case Intrinsic::x86_avx_vtestz_pd:
8437 case Intrinsic::x86_avx_vtestc_pd:
8438 case Intrinsic::x86_avx_vtestnzc_pd:
8439 case Intrinsic::x86_avx_vtestz_ps_256:
8440 case Intrinsic::x86_avx_vtestc_ps_256:
8441 case Intrinsic::x86_avx_vtestnzc_ps_256:
8442 case Intrinsic::x86_avx_vtestz_pd_256:
8443 case Intrinsic::x86_avx_vtestc_pd_256:
8444 case Intrinsic::x86_avx_vtestnzc_pd_256: {
8445 bool IsTestPacked = false;
8448 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
8449 case Intrinsic::x86_avx_vtestz_ps:
8450 case Intrinsic::x86_avx_vtestz_pd:
8451 case Intrinsic::x86_avx_vtestz_ps_256:
8452 case Intrinsic::x86_avx_vtestz_pd_256:
8453 IsTestPacked = true; // Fallthrough
8454 case Intrinsic::x86_sse41_ptestz:
8455 case Intrinsic::x86_avx_ptestz_256:
8457 X86CC = X86::COND_E;
8459 case Intrinsic::x86_avx_vtestc_ps:
8460 case Intrinsic::x86_avx_vtestc_pd:
8461 case Intrinsic::x86_avx_vtestc_ps_256:
8462 case Intrinsic::x86_avx_vtestc_pd_256:
8463 IsTestPacked = true; // Fallthrough
8464 case Intrinsic::x86_sse41_ptestc:
8465 case Intrinsic::x86_avx_ptestc_256:
8467 X86CC = X86::COND_B;
8469 case Intrinsic::x86_avx_vtestnzc_ps:
8470 case Intrinsic::x86_avx_vtestnzc_pd:
8471 case Intrinsic::x86_avx_vtestnzc_ps_256:
8472 case Intrinsic::x86_avx_vtestnzc_pd_256:
8473 IsTestPacked = true; // Fallthrough
8474 case Intrinsic::x86_sse41_ptestnzc:
8475 case Intrinsic::x86_avx_ptestnzc_256:
8477 X86CC = X86::COND_A;
8481 SDValue LHS = Op.getOperand(1);
8482 SDValue RHS = Op.getOperand(2);
8483 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
8484 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
8485 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
8486 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
8487 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
8490 // Fix vector shift instructions where the last operand is a non-immediate
8492 case Intrinsic::x86_sse2_pslli_w:
8493 case Intrinsic::x86_sse2_pslli_d:
8494 case Intrinsic::x86_sse2_pslli_q:
8495 case Intrinsic::x86_sse2_psrli_w:
8496 case Intrinsic::x86_sse2_psrli_d:
8497 case Intrinsic::x86_sse2_psrli_q:
8498 case Intrinsic::x86_sse2_psrai_w:
8499 case Intrinsic::x86_sse2_psrai_d:
8500 case Intrinsic::x86_mmx_pslli_w:
8501 case Intrinsic::x86_mmx_pslli_d:
8502 case Intrinsic::x86_mmx_pslli_q:
8503 case Intrinsic::x86_mmx_psrli_w:
8504 case Intrinsic::x86_mmx_psrli_d:
8505 case Intrinsic::x86_mmx_psrli_q:
8506 case Intrinsic::x86_mmx_psrai_w:
8507 case Intrinsic::x86_mmx_psrai_d: {
8508 SDValue ShAmt = Op.getOperand(2);
8509 if (isa<ConstantSDNode>(ShAmt))
8512 unsigned NewIntNo = 0;
8513 EVT ShAmtVT = MVT::v4i32;
8515 case Intrinsic::x86_sse2_pslli_w:
8516 NewIntNo = Intrinsic::x86_sse2_psll_w;
8518 case Intrinsic::x86_sse2_pslli_d:
8519 NewIntNo = Intrinsic::x86_sse2_psll_d;
8521 case Intrinsic::x86_sse2_pslli_q:
8522 NewIntNo = Intrinsic::x86_sse2_psll_q;
8524 case Intrinsic::x86_sse2_psrli_w:
8525 NewIntNo = Intrinsic::x86_sse2_psrl_w;
8527 case Intrinsic::x86_sse2_psrli_d:
8528 NewIntNo = Intrinsic::x86_sse2_psrl_d;
8530 case Intrinsic::x86_sse2_psrli_q:
8531 NewIntNo = Intrinsic::x86_sse2_psrl_q;
8533 case Intrinsic::x86_sse2_psrai_w:
8534 NewIntNo = Intrinsic::x86_sse2_psra_w;
8536 case Intrinsic::x86_sse2_psrai_d:
8537 NewIntNo = Intrinsic::x86_sse2_psra_d;
8540 ShAmtVT = MVT::v2i32;
8542 case Intrinsic::x86_mmx_pslli_w:
8543 NewIntNo = Intrinsic::x86_mmx_psll_w;
8545 case Intrinsic::x86_mmx_pslli_d:
8546 NewIntNo = Intrinsic::x86_mmx_psll_d;
8548 case Intrinsic::x86_mmx_pslli_q:
8549 NewIntNo = Intrinsic::x86_mmx_psll_q;
8551 case Intrinsic::x86_mmx_psrli_w:
8552 NewIntNo = Intrinsic::x86_mmx_psrl_w;
8554 case Intrinsic::x86_mmx_psrli_d:
8555 NewIntNo = Intrinsic::x86_mmx_psrl_d;
8557 case Intrinsic::x86_mmx_psrli_q:
8558 NewIntNo = Intrinsic::x86_mmx_psrl_q;
8560 case Intrinsic::x86_mmx_psrai_w:
8561 NewIntNo = Intrinsic::x86_mmx_psra_w;
8563 case Intrinsic::x86_mmx_psrai_d:
8564 NewIntNo = Intrinsic::x86_mmx_psra_d;
8566 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
8572 // The vector shift intrinsics with scalars uses 32b shift amounts but
8573 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
8577 ShOps[1] = DAG.getConstant(0, MVT::i32);
8578 if (ShAmtVT == MVT::v4i32) {
8579 ShOps[2] = DAG.getUNDEF(MVT::i32);
8580 ShOps[3] = DAG.getUNDEF(MVT::i32);
8581 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
8583 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
8584 // FIXME this must be lowered to get rid of the invalid type.
8587 EVT VT = Op.getValueType();
8588 ShAmt = DAG.getNode(ISD::BITCAST, dl, VT, ShAmt);
8589 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8590 DAG.getConstant(NewIntNo, MVT::i32),
8591 Op.getOperand(1), ShAmt);
8596 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
8597 SelectionDAG &DAG) const {
8598 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8599 MFI->setReturnAddressIsTaken(true);
8601 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8602 DebugLoc dl = Op.getDebugLoc();
8605 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
8607 DAG.getConstant(TD->getPointerSize(),
8608 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
8609 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
8610 DAG.getNode(ISD::ADD, dl, getPointerTy(),
8612 MachinePointerInfo(), false, false, 0);
8615 // Just load the return address.
8616 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
8617 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
8618 RetAddrFI, MachinePointerInfo(), false, false, 0);
8621 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
8622 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8623 MFI->setFrameAddressIsTaken(true);
8625 EVT VT = Op.getValueType();
8626 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
8627 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8628 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
8629 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
8631 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
8632 MachinePointerInfo(),
8637 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
8638 SelectionDAG &DAG) const {
8639 return DAG.getIntPtrConstant(2*TD->getPointerSize());
8642 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
8643 MachineFunction &MF = DAG.getMachineFunction();
8644 SDValue Chain = Op.getOperand(0);
8645 SDValue Offset = Op.getOperand(1);
8646 SDValue Handler = Op.getOperand(2);
8647 DebugLoc dl = Op.getDebugLoc();
8649 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
8650 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
8652 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
8654 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
8655 DAG.getIntPtrConstant(TD->getPointerSize()));
8656 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
8657 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
8659 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
8660 MF.getRegInfo().addLiveOut(StoreAddrReg);
8662 return DAG.getNode(X86ISD::EH_RETURN, dl,
8664 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
8667 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
8668 SelectionDAG &DAG) const {
8669 SDValue Root = Op.getOperand(0);
8670 SDValue Trmp = Op.getOperand(1); // trampoline
8671 SDValue FPtr = Op.getOperand(2); // nested function
8672 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
8673 DebugLoc dl = Op.getDebugLoc();
8675 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
8677 if (Subtarget->is64Bit()) {
8678 SDValue OutChains[6];
8680 // Large code-model.
8681 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
8682 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
8684 const unsigned char N86R10 = X86_MC::getX86RegNum(X86::R10);
8685 const unsigned char N86R11 = X86_MC::getX86RegNum(X86::R11);
8687 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
8689 // Load the pointer to the nested function into R11.
8690 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
8691 SDValue Addr = Trmp;
8692 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8693 Addr, MachinePointerInfo(TrmpAddr),
8696 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8697 DAG.getConstant(2, MVT::i64));
8698 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
8699 MachinePointerInfo(TrmpAddr, 2),
8702 // Load the 'nest' parameter value into R10.
8703 // R10 is specified in X86CallingConv.td
8704 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
8705 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8706 DAG.getConstant(10, MVT::i64));
8707 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8708 Addr, MachinePointerInfo(TrmpAddr, 10),
8711 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8712 DAG.getConstant(12, MVT::i64));
8713 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
8714 MachinePointerInfo(TrmpAddr, 12),
8717 // Jump to the nested function.
8718 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
8719 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8720 DAG.getConstant(20, MVT::i64));
8721 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8722 Addr, MachinePointerInfo(TrmpAddr, 20),
8725 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
8726 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8727 DAG.getConstant(22, MVT::i64));
8728 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
8729 MachinePointerInfo(TrmpAddr, 22),
8733 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
8734 return DAG.getMergeValues(Ops, 2, dl);
8736 const Function *Func =
8737 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
8738 CallingConv::ID CC = Func->getCallingConv();
8743 llvm_unreachable("Unsupported calling convention");
8744 case CallingConv::C:
8745 case CallingConv::X86_StdCall: {
8746 // Pass 'nest' parameter in ECX.
8747 // Must be kept in sync with X86CallingConv.td
8750 // Check that ECX wasn't needed by an 'inreg' parameter.
8751 FunctionType *FTy = Func->getFunctionType();
8752 const AttrListPtr &Attrs = Func->getAttributes();
8754 if (!Attrs.isEmpty() && !Func->isVarArg()) {
8755 unsigned InRegCount = 0;
8758 for (FunctionType::param_iterator I = FTy->param_begin(),
8759 E = FTy->param_end(); I != E; ++I, ++Idx)
8760 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
8761 // FIXME: should only count parameters that are lowered to integers.
8762 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
8764 if (InRegCount > 2) {
8765 report_fatal_error("Nest register in use - reduce number of inreg"
8771 case CallingConv::X86_FastCall:
8772 case CallingConv::X86_ThisCall:
8773 case CallingConv::Fast:
8774 // Pass 'nest' parameter in EAX.
8775 // Must be kept in sync with X86CallingConv.td
8780 SDValue OutChains[4];
8783 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8784 DAG.getConstant(10, MVT::i32));
8785 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
8787 // This is storing the opcode for MOV32ri.
8788 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
8789 const unsigned char N86Reg = X86_MC::getX86RegNum(NestReg);
8790 OutChains[0] = DAG.getStore(Root, dl,
8791 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
8792 Trmp, MachinePointerInfo(TrmpAddr),
8795 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8796 DAG.getConstant(1, MVT::i32));
8797 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
8798 MachinePointerInfo(TrmpAddr, 1),
8801 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
8802 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8803 DAG.getConstant(5, MVT::i32));
8804 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
8805 MachinePointerInfo(TrmpAddr, 5),
8808 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8809 DAG.getConstant(6, MVT::i32));
8810 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
8811 MachinePointerInfo(TrmpAddr, 6),
8815 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
8816 return DAG.getMergeValues(Ops, 2, dl);
8820 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
8821 SelectionDAG &DAG) const {
8823 The rounding mode is in bits 11:10 of FPSR, and has the following
8830 FLT_ROUNDS, on the other hand, expects the following:
8837 To perform the conversion, we do:
8838 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
8841 MachineFunction &MF = DAG.getMachineFunction();
8842 const TargetMachine &TM = MF.getTarget();
8843 const TargetFrameLowering &TFI = *TM.getFrameLowering();
8844 unsigned StackAlignment = TFI.getStackAlignment();
8845 EVT VT = Op.getValueType();
8846 DebugLoc DL = Op.getDebugLoc();
8848 // Save FP Control Word to stack slot
8849 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
8850 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8853 MachineMemOperand *MMO =
8854 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8855 MachineMemOperand::MOStore, 2, 2);
8857 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
8858 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
8859 DAG.getVTList(MVT::Other),
8860 Ops, 2, MVT::i16, MMO);
8862 // Load FP Control Word from stack slot
8863 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
8864 MachinePointerInfo(), false, false, 0);
8866 // Transform as necessary
8868 DAG.getNode(ISD::SRL, DL, MVT::i16,
8869 DAG.getNode(ISD::AND, DL, MVT::i16,
8870 CWD, DAG.getConstant(0x800, MVT::i16)),
8871 DAG.getConstant(11, MVT::i8));
8873 DAG.getNode(ISD::SRL, DL, MVT::i16,
8874 DAG.getNode(ISD::AND, DL, MVT::i16,
8875 CWD, DAG.getConstant(0x400, MVT::i16)),
8876 DAG.getConstant(9, MVT::i8));
8879 DAG.getNode(ISD::AND, DL, MVT::i16,
8880 DAG.getNode(ISD::ADD, DL, MVT::i16,
8881 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
8882 DAG.getConstant(1, MVT::i16)),
8883 DAG.getConstant(3, MVT::i16));
8886 return DAG.getNode((VT.getSizeInBits() < 16 ?
8887 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
8890 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
8891 EVT VT = Op.getValueType();
8893 unsigned NumBits = VT.getSizeInBits();
8894 DebugLoc dl = Op.getDebugLoc();
8896 Op = Op.getOperand(0);
8897 if (VT == MVT::i8) {
8898 // Zero extend to i32 since there is not an i8 bsr.
8900 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8903 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
8904 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8905 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
8907 // If src is zero (i.e. bsr sets ZF), returns NumBits.
8910 DAG.getConstant(NumBits+NumBits-1, OpVT),
8911 DAG.getConstant(X86::COND_E, MVT::i8),
8914 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8916 // Finally xor with NumBits-1.
8917 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
8920 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8924 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
8925 EVT VT = Op.getValueType();
8927 unsigned NumBits = VT.getSizeInBits();
8928 DebugLoc dl = Op.getDebugLoc();
8930 Op = Op.getOperand(0);
8931 if (VT == MVT::i8) {
8933 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8936 // Issue a bsf (scan bits forward) which also sets EFLAGS.
8937 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8938 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
8940 // If src is zero (i.e. bsf sets ZF), returns NumBits.
8943 DAG.getConstant(NumBits, OpVT),
8944 DAG.getConstant(X86::COND_E, MVT::i8),
8947 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8950 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8954 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
8955 EVT VT = Op.getValueType();
8956 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
8957 DebugLoc dl = Op.getDebugLoc();
8959 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
8960 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
8961 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
8962 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
8963 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
8965 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
8966 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
8967 // return AloBlo + AloBhi + AhiBlo;
8969 SDValue A = Op.getOperand(0);
8970 SDValue B = Op.getOperand(1);
8972 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8973 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8974 A, DAG.getConstant(32, MVT::i32));
8975 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8976 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8977 B, DAG.getConstant(32, MVT::i32));
8978 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8979 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8981 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8982 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8984 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8985 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8987 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8988 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8989 AloBhi, DAG.getConstant(32, MVT::i32));
8990 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8991 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8992 AhiBlo, DAG.getConstant(32, MVT::i32));
8993 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
8994 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
8998 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
9000 EVT VT = Op.getValueType();
9001 DebugLoc dl = Op.getDebugLoc();
9002 SDValue R = Op.getOperand(0);
9003 SDValue Amt = Op.getOperand(1);
9005 LLVMContext *Context = DAG.getContext();
9008 if (!Subtarget->hasSSE2()) return SDValue();
9010 // Optimize shl/srl/sra with constant shift amount.
9011 if (isSplatVector(Amt.getNode())) {
9012 SDValue SclrAmt = Amt->getOperand(0);
9013 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(SclrAmt)) {
9014 uint64_t ShiftAmt = C->getZExtValue();
9016 if (VT == MVT::v2i64 && Op.getOpcode() == ISD::SHL)
9017 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9018 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9019 R, DAG.getConstant(ShiftAmt, MVT::i32));
9021 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SHL)
9022 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9023 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9024 R, DAG.getConstant(ShiftAmt, MVT::i32));
9026 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SHL)
9027 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9028 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9029 R, DAG.getConstant(ShiftAmt, MVT::i32));
9031 if (VT == MVT::v2i64 && Op.getOpcode() == ISD::SRL)
9032 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9033 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9034 R, DAG.getConstant(ShiftAmt, MVT::i32));
9036 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SRL)
9037 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9038 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9039 R, DAG.getConstant(ShiftAmt, MVT::i32));
9041 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SRL)
9042 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9043 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9044 R, DAG.getConstant(ShiftAmt, MVT::i32));
9046 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SRA)
9047 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9048 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9049 R, DAG.getConstant(ShiftAmt, MVT::i32));
9051 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SRA)
9052 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9053 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9054 R, DAG.getConstant(ShiftAmt, MVT::i32));
9058 // Lower SHL with variable shift amount.
9059 // Cannot lower SHL without SSE2 or later.
9060 if (!Subtarget->hasSSE2()) return SDValue();
9062 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
9063 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9064 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9065 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
9067 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
9069 std::vector<Constant*> CV(4, CI);
9070 Constant *C = ConstantVector::get(CV);
9071 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9072 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9073 MachinePointerInfo::getConstantPool(),
9076 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
9077 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
9078 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
9079 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
9081 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
9083 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9084 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9085 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
9087 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
9088 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
9090 std::vector<Constant*> CVM1(16, CM1);
9091 std::vector<Constant*> CVM2(16, CM2);
9092 Constant *C = ConstantVector::get(CVM1);
9093 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9094 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9095 MachinePointerInfo::getConstantPool(),
9098 // r = pblendv(r, psllw(r & (char16)15, 4), a);
9099 M = DAG.getNode(ISD::AND, dl, VT, R, M);
9100 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9101 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
9102 DAG.getConstant(4, MVT::i32));
9103 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
9105 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
9107 C = ConstantVector::get(CVM2);
9108 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9109 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9110 MachinePointerInfo::getConstantPool(),
9113 // r = pblendv(r, psllw(r & (char16)63, 2), a);
9114 M = DAG.getNode(ISD::AND, dl, VT, R, M);
9115 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9116 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
9117 DAG.getConstant(2, MVT::i32));
9118 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
9120 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
9122 // return pblendv(r, r+r, a);
9123 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT,
9124 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
9130 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
9131 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
9132 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
9133 // looks for this combo and may remove the "setcc" instruction if the "setcc"
9134 // has only one use.
9135 SDNode *N = Op.getNode();
9136 SDValue LHS = N->getOperand(0);
9137 SDValue RHS = N->getOperand(1);
9138 unsigned BaseOp = 0;
9140 DebugLoc DL = Op.getDebugLoc();
9141 switch (Op.getOpcode()) {
9142 default: llvm_unreachable("Unknown ovf instruction!");
9144 // A subtract of one will be selected as a INC. Note that INC doesn't
9145 // set CF, so we can't do this for UADDO.
9146 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
9148 BaseOp = X86ISD::INC;
9152 BaseOp = X86ISD::ADD;
9156 BaseOp = X86ISD::ADD;
9160 // A subtract of one will be selected as a DEC. Note that DEC doesn't
9161 // set CF, so we can't do this for USUBO.
9162 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
9164 BaseOp = X86ISD::DEC;
9168 BaseOp = X86ISD::SUB;
9172 BaseOp = X86ISD::SUB;
9176 BaseOp = X86ISD::SMUL;
9179 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
9180 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
9182 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
9185 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9186 DAG.getConstant(X86::COND_O, MVT::i32),
9187 SDValue(Sum.getNode(), 2));
9189 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
9194 // Also sets EFLAGS.
9195 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
9196 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
9199 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
9200 DAG.getConstant(Cond, MVT::i32),
9201 SDValue(Sum.getNode(), 1));
9203 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
9207 SDValue X86TargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const{
9208 DebugLoc dl = Op.getDebugLoc();
9209 SDNode* Node = Op.getNode();
9210 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
9211 EVT VT = Node->getValueType(0);
9213 if (Subtarget->hasSSE2() && VT.isVector()) {
9214 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
9215 ExtraVT.getScalarType().getSizeInBits();
9216 SDValue ShAmt = DAG.getConstant(BitsDiff, MVT::i32);
9218 unsigned SHLIntrinsicsID = 0;
9219 unsigned SRAIntrinsicsID = 0;
9220 switch (VT.getSimpleVT().SimpleTy) {
9224 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_q;
9225 SRAIntrinsicsID = 0;
9229 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_d;
9230 SRAIntrinsicsID = Intrinsic::x86_sse2_psrai_d;
9234 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_w;
9235 SRAIntrinsicsID = Intrinsic::x86_sse2_psrai_w;
9240 SDValue Tmp1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9241 DAG.getConstant(SHLIntrinsicsID, MVT::i32),
9242 Node->getOperand(0), ShAmt);
9244 // In case of 1 bit sext, no need to shr
9245 if (ExtraVT.getScalarType().getSizeInBits() == 1) return Tmp1;
9247 if (SRAIntrinsicsID) {
9248 Tmp1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9249 DAG.getConstant(SRAIntrinsicsID, MVT::i32),
9259 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
9260 DebugLoc dl = Op.getDebugLoc();
9262 // Go ahead and emit the fence on x86-64 even if we asked for no-sse2.
9263 // There isn't any reason to disable it if the target processor supports it.
9264 if (!Subtarget->hasSSE2() && !Subtarget->is64Bit()) {
9265 SDValue Chain = Op.getOperand(0);
9266 SDValue Zero = DAG.getConstant(0, MVT::i32);
9268 DAG.getRegister(X86::ESP, MVT::i32), // Base
9269 DAG.getTargetConstant(1, MVT::i8), // Scale
9270 DAG.getRegister(0, MVT::i32), // Index
9271 DAG.getTargetConstant(0, MVT::i32), // Disp
9272 DAG.getRegister(0, MVT::i32), // Segment.
9277 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
9278 array_lengthof(Ops));
9279 return SDValue(Res, 0);
9282 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
9284 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
9286 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
9287 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
9288 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
9289 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
9291 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
9292 if (!Op1 && !Op2 && !Op3 && Op4)
9293 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
9295 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
9296 if (Op1 && !Op2 && !Op3 && !Op4)
9297 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
9299 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
9301 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
9304 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
9305 EVT T = Op.getValueType();
9306 DebugLoc DL = Op.getDebugLoc();
9309 switch(T.getSimpleVT().SimpleTy) {
9311 assert(false && "Invalid value type!");
9312 case MVT::i8: Reg = X86::AL; size = 1; break;
9313 case MVT::i16: Reg = X86::AX; size = 2; break;
9314 case MVT::i32: Reg = X86::EAX; size = 4; break;
9316 assert(Subtarget->is64Bit() && "Node not type legal!");
9317 Reg = X86::RAX; size = 8;
9320 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
9321 Op.getOperand(2), SDValue());
9322 SDValue Ops[] = { cpIn.getValue(0),
9325 DAG.getTargetConstant(size, MVT::i8),
9327 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9328 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
9329 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
9332 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
9336 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
9337 SelectionDAG &DAG) const {
9338 assert(Subtarget->is64Bit() && "Result not type legalized?");
9339 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9340 SDValue TheChain = Op.getOperand(0);
9341 DebugLoc dl = Op.getDebugLoc();
9342 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
9343 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
9344 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
9346 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
9347 DAG.getConstant(32, MVT::i8));
9349 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
9352 return DAG.getMergeValues(Ops, 2, dl);
9355 SDValue X86TargetLowering::LowerBITCAST(SDValue Op,
9356 SelectionDAG &DAG) const {
9357 EVT SrcVT = Op.getOperand(0).getValueType();
9358 EVT DstVT = Op.getValueType();
9359 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
9360 Subtarget->hasMMX() && "Unexpected custom BITCAST");
9361 assert((DstVT == MVT::i64 ||
9362 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
9363 "Unexpected custom BITCAST");
9364 // i64 <=> MMX conversions are Legal.
9365 if (SrcVT==MVT::i64 && DstVT.isVector())
9367 if (DstVT==MVT::i64 && SrcVT.isVector())
9369 // MMX <=> MMX conversions are Legal.
9370 if (SrcVT.isVector() && DstVT.isVector())
9372 // All other conversions need to be expanded.
9376 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
9377 SDNode *Node = Op.getNode();
9378 DebugLoc dl = Node->getDebugLoc();
9379 EVT T = Node->getValueType(0);
9380 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
9381 DAG.getConstant(0, T), Node->getOperand(2));
9382 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
9383 cast<AtomicSDNode>(Node)->getMemoryVT(),
9384 Node->getOperand(0),
9385 Node->getOperand(1), negOp,
9386 cast<AtomicSDNode>(Node)->getSrcValue(),
9387 cast<AtomicSDNode>(Node)->getAlignment());
9390 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
9391 EVT VT = Op.getNode()->getValueType(0);
9393 // Let legalize expand this if it isn't a legal type yet.
9394 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
9397 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
9400 bool ExtraOp = false;
9401 switch (Op.getOpcode()) {
9402 default: assert(0 && "Invalid code");
9403 case ISD::ADDC: Opc = X86ISD::ADD; break;
9404 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
9405 case ISD::SUBC: Opc = X86ISD::SUB; break;
9406 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
9410 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
9412 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
9413 Op.getOperand(1), Op.getOperand(2));
9416 /// LowerOperation - Provide custom lowering hooks for some operations.
9418 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
9419 switch (Op.getOpcode()) {
9420 default: llvm_unreachable("Should not custom lower this!");
9421 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op,DAG);
9422 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
9423 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
9424 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
9425 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
9426 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
9427 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
9428 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
9429 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
9430 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
9431 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, DAG);
9432 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
9433 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
9434 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
9435 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
9436 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
9437 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
9438 case ISD::SHL_PARTS:
9439 case ISD::SRA_PARTS:
9440 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
9441 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
9442 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
9443 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
9444 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
9445 case ISD::FABS: return LowerFABS(Op, DAG);
9446 case ISD::FNEG: return LowerFNEG(Op, DAG);
9447 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
9448 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
9449 case ISD::SETCC: return LowerSETCC(Op, DAG);
9450 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
9451 case ISD::SELECT: return LowerSELECT(Op, DAG);
9452 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
9453 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
9454 case ISD::VASTART: return LowerVASTART(Op, DAG);
9455 case ISD::VAARG: return LowerVAARG(Op, DAG);
9456 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
9457 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
9458 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
9459 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
9460 case ISD::FRAME_TO_ARGS_OFFSET:
9461 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
9462 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
9463 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
9464 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
9465 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
9466 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
9467 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
9468 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
9471 case ISD::SHL: return LowerShift(Op, DAG);
9477 case ISD::UMULO: return LowerXALUO(Op, DAG);
9478 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
9479 case ISD::BITCAST: return LowerBITCAST(Op, DAG);
9483 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
9487 void X86TargetLowering::
9488 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
9489 SelectionDAG &DAG, unsigned NewOp) const {
9490 EVT T = Node->getValueType(0);
9491 DebugLoc dl = Node->getDebugLoc();
9492 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
9494 SDValue Chain = Node->getOperand(0);
9495 SDValue In1 = Node->getOperand(1);
9496 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
9497 Node->getOperand(2), DAG.getIntPtrConstant(0));
9498 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
9499 Node->getOperand(2), DAG.getIntPtrConstant(1));
9500 SDValue Ops[] = { Chain, In1, In2L, In2H };
9501 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
9503 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
9504 cast<MemSDNode>(Node)->getMemOperand());
9505 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
9506 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
9507 Results.push_back(Result.getValue(2));
9510 /// ReplaceNodeResults - Replace a node with an illegal result type
9511 /// with a new node built out of custom code.
9512 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
9513 SmallVectorImpl<SDValue>&Results,
9514 SelectionDAG &DAG) const {
9515 DebugLoc dl = N->getDebugLoc();
9516 switch (N->getOpcode()) {
9518 assert(false && "Do not know how to custom type legalize this operation!");
9520 case ISD::SIGN_EXTEND_INREG:
9525 // We don't want to expand or promote these.
9527 case ISD::FP_TO_SINT: {
9528 std::pair<SDValue,SDValue> Vals =
9529 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
9530 SDValue FIST = Vals.first, StackSlot = Vals.second;
9531 if (FIST.getNode() != 0) {
9532 EVT VT = N->getValueType(0);
9533 // Return a load from the stack slot.
9534 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
9535 MachinePointerInfo(), false, false, 0));
9539 case ISD::READCYCLECOUNTER: {
9540 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9541 SDValue TheChain = N->getOperand(0);
9542 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
9543 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
9545 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
9547 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
9548 SDValue Ops[] = { eax, edx };
9549 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
9550 Results.push_back(edx.getValue(1));
9553 case ISD::ATOMIC_CMP_SWAP: {
9554 EVT T = N->getValueType(0);
9555 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
9556 SDValue cpInL, cpInH;
9557 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
9558 DAG.getConstant(0, MVT::i32));
9559 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
9560 DAG.getConstant(1, MVT::i32));
9561 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
9562 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
9564 SDValue swapInL, swapInH;
9565 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
9566 DAG.getConstant(0, MVT::i32));
9567 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
9568 DAG.getConstant(1, MVT::i32));
9569 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
9571 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
9572 swapInL.getValue(1));
9573 SDValue Ops[] = { swapInH.getValue(0),
9575 swapInH.getValue(1) };
9576 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9577 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
9578 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG8_DAG, dl, Tys,
9580 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
9581 MVT::i32, Result.getValue(1));
9582 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
9583 MVT::i32, cpOutL.getValue(2));
9584 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
9585 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
9586 Results.push_back(cpOutH.getValue(1));
9589 case ISD::ATOMIC_LOAD_ADD:
9590 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
9592 case ISD::ATOMIC_LOAD_AND:
9593 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
9595 case ISD::ATOMIC_LOAD_NAND:
9596 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
9598 case ISD::ATOMIC_LOAD_OR:
9599 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
9601 case ISD::ATOMIC_LOAD_SUB:
9602 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
9604 case ISD::ATOMIC_LOAD_XOR:
9605 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
9607 case ISD::ATOMIC_SWAP:
9608 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
9613 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
9615 default: return NULL;
9616 case X86ISD::BSF: return "X86ISD::BSF";
9617 case X86ISD::BSR: return "X86ISD::BSR";
9618 case X86ISD::SHLD: return "X86ISD::SHLD";
9619 case X86ISD::SHRD: return "X86ISD::SHRD";
9620 case X86ISD::FAND: return "X86ISD::FAND";
9621 case X86ISD::FOR: return "X86ISD::FOR";
9622 case X86ISD::FXOR: return "X86ISD::FXOR";
9623 case X86ISD::FSRL: return "X86ISD::FSRL";
9624 case X86ISD::FILD: return "X86ISD::FILD";
9625 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
9626 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
9627 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
9628 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
9629 case X86ISD::FLD: return "X86ISD::FLD";
9630 case X86ISD::FST: return "X86ISD::FST";
9631 case X86ISD::CALL: return "X86ISD::CALL";
9632 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
9633 case X86ISD::BT: return "X86ISD::BT";
9634 case X86ISD::CMP: return "X86ISD::CMP";
9635 case X86ISD::COMI: return "X86ISD::COMI";
9636 case X86ISD::UCOMI: return "X86ISD::UCOMI";
9637 case X86ISD::SETCC: return "X86ISD::SETCC";
9638 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
9639 case X86ISD::FSETCCsd: return "X86ISD::FSETCCsd";
9640 case X86ISD::FSETCCss: return "X86ISD::FSETCCss";
9641 case X86ISD::CMOV: return "X86ISD::CMOV";
9642 case X86ISD::BRCOND: return "X86ISD::BRCOND";
9643 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
9644 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
9645 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
9646 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
9647 case X86ISD::Wrapper: return "X86ISD::Wrapper";
9648 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
9649 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
9650 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
9651 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
9652 case X86ISD::PINSRB: return "X86ISD::PINSRB";
9653 case X86ISD::PINSRW: return "X86ISD::PINSRW";
9654 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
9655 case X86ISD::ANDNP: return "X86ISD::ANDNP";
9656 case X86ISD::PSIGNB: return "X86ISD::PSIGNB";
9657 case X86ISD::PSIGNW: return "X86ISD::PSIGNW";
9658 case X86ISD::PSIGND: return "X86ISD::PSIGND";
9659 case X86ISD::PBLENDVB: return "X86ISD::PBLENDVB";
9660 case X86ISD::FMAX: return "X86ISD::FMAX";
9661 case X86ISD::FMIN: return "X86ISD::FMIN";
9662 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
9663 case X86ISD::FRCP: return "X86ISD::FRCP";
9664 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
9665 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
9666 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
9667 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
9668 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
9669 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
9670 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
9671 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
9672 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
9673 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
9674 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
9675 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
9676 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
9677 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
9678 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
9679 case X86ISD::VSHL: return "X86ISD::VSHL";
9680 case X86ISD::VSRL: return "X86ISD::VSRL";
9681 case X86ISD::CMPPD: return "X86ISD::CMPPD";
9682 case X86ISD::CMPPS: return "X86ISD::CMPPS";
9683 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
9684 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
9685 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
9686 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
9687 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
9688 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
9689 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
9690 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
9691 case X86ISD::ADD: return "X86ISD::ADD";
9692 case X86ISD::SUB: return "X86ISD::SUB";
9693 case X86ISD::ADC: return "X86ISD::ADC";
9694 case X86ISD::SBB: return "X86ISD::SBB";
9695 case X86ISD::SMUL: return "X86ISD::SMUL";
9696 case X86ISD::UMUL: return "X86ISD::UMUL";
9697 case X86ISD::INC: return "X86ISD::INC";
9698 case X86ISD::DEC: return "X86ISD::DEC";
9699 case X86ISD::OR: return "X86ISD::OR";
9700 case X86ISD::XOR: return "X86ISD::XOR";
9701 case X86ISD::AND: return "X86ISD::AND";
9702 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
9703 case X86ISD::PTEST: return "X86ISD::PTEST";
9704 case X86ISD::TESTP: return "X86ISD::TESTP";
9705 case X86ISD::PALIGN: return "X86ISD::PALIGN";
9706 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
9707 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
9708 case X86ISD::PSHUFHW_LD: return "X86ISD::PSHUFHW_LD";
9709 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
9710 case X86ISD::PSHUFLW_LD: return "X86ISD::PSHUFLW_LD";
9711 case X86ISD::SHUFPS: return "X86ISD::SHUFPS";
9712 case X86ISD::SHUFPD: return "X86ISD::SHUFPD";
9713 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
9714 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
9715 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
9716 case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
9717 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
9718 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
9719 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
9720 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
9721 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
9722 case X86ISD::MOVSHDUP_LD: return "X86ISD::MOVSHDUP_LD";
9723 case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
9724 case X86ISD::MOVSD: return "X86ISD::MOVSD";
9725 case X86ISD::MOVSS: return "X86ISD::MOVSS";
9726 case X86ISD::UNPCKLPS: return "X86ISD::UNPCKLPS";
9727 case X86ISD::UNPCKLPD: return "X86ISD::UNPCKLPD";
9728 case X86ISD::VUNPCKLPS: return "X86ISD::VUNPCKLPS";
9729 case X86ISD::VUNPCKLPD: return "X86ISD::VUNPCKLPD";
9730 case X86ISD::VUNPCKLPSY: return "X86ISD::VUNPCKLPSY";
9731 case X86ISD::VUNPCKLPDY: return "X86ISD::VUNPCKLPDY";
9732 case X86ISD::UNPCKHPS: return "X86ISD::UNPCKHPS";
9733 case X86ISD::UNPCKHPD: return "X86ISD::UNPCKHPD";
9734 case X86ISD::PUNPCKLBW: return "X86ISD::PUNPCKLBW";
9735 case X86ISD::PUNPCKLWD: return "X86ISD::PUNPCKLWD";
9736 case X86ISD::PUNPCKLDQ: return "X86ISD::PUNPCKLDQ";
9737 case X86ISD::PUNPCKLQDQ: return "X86ISD::PUNPCKLQDQ";
9738 case X86ISD::PUNPCKHBW: return "X86ISD::PUNPCKHBW";
9739 case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
9740 case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
9741 case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
9742 case X86ISD::VPERMIL: return "X86ISD::VPERMIL";
9743 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
9744 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
9745 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
9749 // isLegalAddressingMode - Return true if the addressing mode represented
9750 // by AM is legal for this target, for a load/store of the specified type.
9751 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
9753 // X86 supports extremely general addressing modes.
9754 CodeModel::Model M = getTargetMachine().getCodeModel();
9755 Reloc::Model R = getTargetMachine().getRelocationModel();
9757 // X86 allows a sign-extended 32-bit immediate field as a displacement.
9758 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
9763 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
9765 // If a reference to this global requires an extra load, we can't fold it.
9766 if (isGlobalStubReference(GVFlags))
9769 // If BaseGV requires a register for the PIC base, we cannot also have a
9770 // BaseReg specified.
9771 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
9774 // If lower 4G is not available, then we must use rip-relative addressing.
9775 if ((M != CodeModel::Small || R != Reloc::Static) &&
9776 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
9786 // These scales always work.
9791 // These scales are formed with basereg+scalereg. Only accept if there is
9796 default: // Other stuff never works.
9804 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
9805 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
9807 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
9808 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
9809 if (NumBits1 <= NumBits2)
9814 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
9815 if (!VT1.isInteger() || !VT2.isInteger())
9817 unsigned NumBits1 = VT1.getSizeInBits();
9818 unsigned NumBits2 = VT2.getSizeInBits();
9819 if (NumBits1 <= NumBits2)
9824 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
9825 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
9826 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
9829 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
9830 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
9831 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
9834 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
9835 // i16 instructions are longer (0x66 prefix) and potentially slower.
9836 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
9839 /// isShuffleMaskLegal - Targets can use this to indicate that they only
9840 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
9841 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
9842 /// are assumed to be legal.
9844 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
9846 // Very little shuffling can be done for 64-bit vectors right now.
9847 if (VT.getSizeInBits() == 64)
9848 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
9850 // FIXME: pshufb, blends, shifts.
9851 return (VT.getVectorNumElements() == 2 ||
9852 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
9853 isMOVLMask(M, VT) ||
9854 isSHUFPMask(M, VT) ||
9855 isPSHUFDMask(M, VT) ||
9856 isPSHUFHWMask(M, VT) ||
9857 isPSHUFLWMask(M, VT) ||
9858 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
9859 isUNPCKLMask(M, VT) ||
9860 isUNPCKHMask(M, VT) ||
9861 isUNPCKL_v_undef_Mask(M, VT) ||
9862 isUNPCKH_v_undef_Mask(M, VT));
9866 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
9868 unsigned NumElts = VT.getVectorNumElements();
9869 // FIXME: This collection of masks seems suspect.
9872 if (NumElts == 4 && VT.getSizeInBits() == 128) {
9873 return (isMOVLMask(Mask, VT) ||
9874 isCommutedMOVLMask(Mask, VT, true) ||
9875 isSHUFPMask(Mask, VT) ||
9876 isCommutedSHUFPMask(Mask, VT));
9881 //===----------------------------------------------------------------------===//
9882 // X86 Scheduler Hooks
9883 //===----------------------------------------------------------------------===//
9885 // private utility function
9887 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
9888 MachineBasicBlock *MBB,
9895 TargetRegisterClass *RC,
9896 bool invSrc) const {
9897 // For the atomic bitwise operator, we generate
9900 // ld t1 = [bitinstr.addr]
9901 // op t2 = t1, [bitinstr.val]
9903 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
9905 // fallthrough -->nextMBB
9906 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9907 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9908 MachineFunction::iterator MBBIter = MBB;
9911 /// First build the CFG
9912 MachineFunction *F = MBB->getParent();
9913 MachineBasicBlock *thisMBB = MBB;
9914 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9915 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9916 F->insert(MBBIter, newMBB);
9917 F->insert(MBBIter, nextMBB);
9919 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9920 nextMBB->splice(nextMBB->begin(), thisMBB,
9921 llvm::next(MachineBasicBlock::iterator(bInstr)),
9923 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9925 // Update thisMBB to fall through to newMBB
9926 thisMBB->addSuccessor(newMBB);
9928 // newMBB jumps to itself and fall through to nextMBB
9929 newMBB->addSuccessor(nextMBB);
9930 newMBB->addSuccessor(newMBB);
9932 // Insert instructions into newMBB based on incoming instruction
9933 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
9934 "unexpected number of operands");
9935 DebugLoc dl = bInstr->getDebugLoc();
9936 MachineOperand& destOper = bInstr->getOperand(0);
9937 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9938 int numArgs = bInstr->getNumOperands() - 1;
9939 for (int i=0; i < numArgs; ++i)
9940 argOpers[i] = &bInstr->getOperand(i+1);
9942 // x86 address has 4 operands: base, index, scale, and displacement
9943 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9944 int valArgIndx = lastAddrIndx + 1;
9946 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
9947 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
9948 for (int i=0; i <= lastAddrIndx; ++i)
9949 (*MIB).addOperand(*argOpers[i]);
9951 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
9953 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
9958 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
9959 assert((argOpers[valArgIndx]->isReg() ||
9960 argOpers[valArgIndx]->isImm()) &&
9962 if (argOpers[valArgIndx]->isReg())
9963 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
9965 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
9967 (*MIB).addOperand(*argOpers[valArgIndx]);
9969 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
9972 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
9973 for (int i=0; i <= lastAddrIndx; ++i)
9974 (*MIB).addOperand(*argOpers[i]);
9976 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9977 (*MIB).setMemRefs(bInstr->memoperands_begin(),
9978 bInstr->memoperands_end());
9980 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
9984 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9986 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
9990 // private utility function: 64 bit atomics on 32 bit host.
9992 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
9993 MachineBasicBlock *MBB,
9998 bool invSrc) const {
9999 // For the atomic bitwise operator, we generate
10000 // thisMBB (instructions are in pairs, except cmpxchg8b)
10001 // ld t1,t2 = [bitinstr.addr]
10003 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
10004 // op t5, t6 <- out1, out2, [bitinstr.val]
10005 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
10006 // mov ECX, EBX <- t5, t6
10007 // mov EAX, EDX <- t1, t2
10008 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
10009 // mov t3, t4 <- EAX, EDX
10011 // result in out1, out2
10012 // fallthrough -->nextMBB
10014 const TargetRegisterClass *RC = X86::GR32RegisterClass;
10015 const unsigned LoadOpc = X86::MOV32rm;
10016 const unsigned NotOpc = X86::NOT32r;
10017 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10018 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10019 MachineFunction::iterator MBBIter = MBB;
10022 /// First build the CFG
10023 MachineFunction *F = MBB->getParent();
10024 MachineBasicBlock *thisMBB = MBB;
10025 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
10026 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
10027 F->insert(MBBIter, newMBB);
10028 F->insert(MBBIter, nextMBB);
10030 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
10031 nextMBB->splice(nextMBB->begin(), thisMBB,
10032 llvm::next(MachineBasicBlock::iterator(bInstr)),
10034 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10036 // Update thisMBB to fall through to newMBB
10037 thisMBB->addSuccessor(newMBB);
10039 // newMBB jumps to itself and fall through to nextMBB
10040 newMBB->addSuccessor(nextMBB);
10041 newMBB->addSuccessor(newMBB);
10043 DebugLoc dl = bInstr->getDebugLoc();
10044 // Insert instructions into newMBB based on incoming instruction
10045 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
10046 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
10047 "unexpected number of operands");
10048 MachineOperand& dest1Oper = bInstr->getOperand(0);
10049 MachineOperand& dest2Oper = bInstr->getOperand(1);
10050 MachineOperand* argOpers[2 + X86::AddrNumOperands];
10051 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
10052 argOpers[i] = &bInstr->getOperand(i+2);
10054 // We use some of the operands multiple times, so conservatively just
10055 // clear any kill flags that might be present.
10056 if (argOpers[i]->isReg() && argOpers[i]->isUse())
10057 argOpers[i]->setIsKill(false);
10060 // x86 address has 5 operands: base, index, scale, displacement, and segment.
10061 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
10063 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
10064 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
10065 for (int i=0; i <= lastAddrIndx; ++i)
10066 (*MIB).addOperand(*argOpers[i]);
10067 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
10068 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
10069 // add 4 to displacement.
10070 for (int i=0; i <= lastAddrIndx-2; ++i)
10071 (*MIB).addOperand(*argOpers[i]);
10072 MachineOperand newOp3 = *(argOpers[3]);
10073 if (newOp3.isImm())
10074 newOp3.setImm(newOp3.getImm()+4);
10076 newOp3.setOffset(newOp3.getOffset()+4);
10077 (*MIB).addOperand(newOp3);
10078 (*MIB).addOperand(*argOpers[lastAddrIndx]);
10080 // t3/4 are defined later, at the bottom of the loop
10081 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
10082 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
10083 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
10084 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
10085 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
10086 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
10088 // The subsequent operations should be using the destination registers of
10089 //the PHI instructions.
10091 t1 = F->getRegInfo().createVirtualRegister(RC);
10092 t2 = F->getRegInfo().createVirtualRegister(RC);
10093 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
10094 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
10096 t1 = dest1Oper.getReg();
10097 t2 = dest2Oper.getReg();
10100 int valArgIndx = lastAddrIndx + 1;
10101 assert((argOpers[valArgIndx]->isReg() ||
10102 argOpers[valArgIndx]->isImm()) &&
10103 "invalid operand");
10104 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
10105 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
10106 if (argOpers[valArgIndx]->isReg())
10107 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
10109 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
10110 if (regOpcL != X86::MOV32rr)
10112 (*MIB).addOperand(*argOpers[valArgIndx]);
10113 assert(argOpers[valArgIndx + 1]->isReg() ==
10114 argOpers[valArgIndx]->isReg());
10115 assert(argOpers[valArgIndx + 1]->isImm() ==
10116 argOpers[valArgIndx]->isImm());
10117 if (argOpers[valArgIndx + 1]->isReg())
10118 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
10120 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
10121 if (regOpcH != X86::MOV32rr)
10123 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
10125 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
10127 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
10130 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
10132 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
10135 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
10136 for (int i=0; i <= lastAddrIndx; ++i)
10137 (*MIB).addOperand(*argOpers[i]);
10139 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
10140 (*MIB).setMemRefs(bInstr->memoperands_begin(),
10141 bInstr->memoperands_end());
10143 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
10144 MIB.addReg(X86::EAX);
10145 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
10146 MIB.addReg(X86::EDX);
10149 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
10151 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
10155 // private utility function
10156 MachineBasicBlock *
10157 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
10158 MachineBasicBlock *MBB,
10159 unsigned cmovOpc) const {
10160 // For the atomic min/max operator, we generate
10163 // ld t1 = [min/max.addr]
10164 // mov t2 = [min/max.val]
10166 // cmov[cond] t2 = t1
10168 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
10170 // fallthrough -->nextMBB
10172 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10173 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10174 MachineFunction::iterator MBBIter = MBB;
10177 /// First build the CFG
10178 MachineFunction *F = MBB->getParent();
10179 MachineBasicBlock *thisMBB = MBB;
10180 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
10181 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
10182 F->insert(MBBIter, newMBB);
10183 F->insert(MBBIter, nextMBB);
10185 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
10186 nextMBB->splice(nextMBB->begin(), thisMBB,
10187 llvm::next(MachineBasicBlock::iterator(mInstr)),
10189 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10191 // Update thisMBB to fall through to newMBB
10192 thisMBB->addSuccessor(newMBB);
10194 // newMBB jumps to newMBB and fall through to nextMBB
10195 newMBB->addSuccessor(nextMBB);
10196 newMBB->addSuccessor(newMBB);
10198 DebugLoc dl = mInstr->getDebugLoc();
10199 // Insert instructions into newMBB based on incoming instruction
10200 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
10201 "unexpected number of operands");
10202 MachineOperand& destOper = mInstr->getOperand(0);
10203 MachineOperand* argOpers[2 + X86::AddrNumOperands];
10204 int numArgs = mInstr->getNumOperands() - 1;
10205 for (int i=0; i < numArgs; ++i)
10206 argOpers[i] = &mInstr->getOperand(i+1);
10208 // x86 address has 4 operands: base, index, scale, and displacement
10209 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
10210 int valArgIndx = lastAddrIndx + 1;
10212 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
10213 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
10214 for (int i=0; i <= lastAddrIndx; ++i)
10215 (*MIB).addOperand(*argOpers[i]);
10217 // We only support register and immediate values
10218 assert((argOpers[valArgIndx]->isReg() ||
10219 argOpers[valArgIndx]->isImm()) &&
10220 "invalid operand");
10222 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
10223 if (argOpers[valArgIndx]->isReg())
10224 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
10226 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
10227 (*MIB).addOperand(*argOpers[valArgIndx]);
10229 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
10232 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
10237 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
10238 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
10242 // Cmp and exchange if none has modified the memory location
10243 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
10244 for (int i=0; i <= lastAddrIndx; ++i)
10245 (*MIB).addOperand(*argOpers[i]);
10247 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
10248 (*MIB).setMemRefs(mInstr->memoperands_begin(),
10249 mInstr->memoperands_end());
10251 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
10252 MIB.addReg(X86::EAX);
10255 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
10257 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
10261 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
10262 // or XMM0_V32I8 in AVX all of this code can be replaced with that
10263 // in the .td file.
10264 MachineBasicBlock *
10265 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
10266 unsigned numArgs, bool memArg) const {
10267 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
10268 "Target must have SSE4.2 or AVX features enabled");
10270 DebugLoc dl = MI->getDebugLoc();
10271 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10273 if (!Subtarget->hasAVX()) {
10275 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
10277 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
10280 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
10282 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
10285 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
10286 for (unsigned i = 0; i < numArgs; ++i) {
10287 MachineOperand &Op = MI->getOperand(i+1);
10288 if (!(Op.isReg() && Op.isImplicit()))
10289 MIB.addOperand(Op);
10291 BuildMI(*BB, MI, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
10292 .addReg(X86::XMM0);
10294 MI->eraseFromParent();
10298 MachineBasicBlock *
10299 X86TargetLowering::EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB) const {
10300 DebugLoc dl = MI->getDebugLoc();
10301 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10303 // Address into RAX/EAX, other two args into ECX, EDX.
10304 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
10305 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
10306 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
10307 for (int i = 0; i < X86::AddrNumOperands; ++i)
10308 MIB.addOperand(MI->getOperand(i));
10310 unsigned ValOps = X86::AddrNumOperands;
10311 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
10312 .addReg(MI->getOperand(ValOps).getReg());
10313 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
10314 .addReg(MI->getOperand(ValOps+1).getReg());
10316 // The instruction doesn't actually take any operands though.
10317 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
10319 MI->eraseFromParent(); // The pseudo is gone now.
10323 MachineBasicBlock *
10324 X86TargetLowering::EmitMwait(MachineInstr *MI, MachineBasicBlock *BB) const {
10325 DebugLoc dl = MI->getDebugLoc();
10326 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10328 // First arg in ECX, the second in EAX.
10329 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
10330 .addReg(MI->getOperand(0).getReg());
10331 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EAX)
10332 .addReg(MI->getOperand(1).getReg());
10334 // The instruction doesn't actually take any operands though.
10335 BuildMI(*BB, MI, dl, TII->get(X86::MWAITrr));
10337 MI->eraseFromParent(); // The pseudo is gone now.
10341 MachineBasicBlock *
10342 X86TargetLowering::EmitVAARG64WithCustomInserter(
10344 MachineBasicBlock *MBB) const {
10345 // Emit va_arg instruction on X86-64.
10347 // Operands to this pseudo-instruction:
10348 // 0 ) Output : destination address (reg)
10349 // 1-5) Input : va_list address (addr, i64mem)
10350 // 6 ) ArgSize : Size (in bytes) of vararg type
10351 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
10352 // 8 ) Align : Alignment of type
10353 // 9 ) EFLAGS (implicit-def)
10355 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
10356 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
10358 unsigned DestReg = MI->getOperand(0).getReg();
10359 MachineOperand &Base = MI->getOperand(1);
10360 MachineOperand &Scale = MI->getOperand(2);
10361 MachineOperand &Index = MI->getOperand(3);
10362 MachineOperand &Disp = MI->getOperand(4);
10363 MachineOperand &Segment = MI->getOperand(5);
10364 unsigned ArgSize = MI->getOperand(6).getImm();
10365 unsigned ArgMode = MI->getOperand(7).getImm();
10366 unsigned Align = MI->getOperand(8).getImm();
10368 // Memory Reference
10369 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
10370 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
10371 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
10373 // Machine Information
10374 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10375 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
10376 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
10377 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
10378 DebugLoc DL = MI->getDebugLoc();
10380 // struct va_list {
10383 // i64 overflow_area (address)
10384 // i64 reg_save_area (address)
10386 // sizeof(va_list) = 24
10387 // alignment(va_list) = 8
10389 unsigned TotalNumIntRegs = 6;
10390 unsigned TotalNumXMMRegs = 8;
10391 bool UseGPOffset = (ArgMode == 1);
10392 bool UseFPOffset = (ArgMode == 2);
10393 unsigned MaxOffset = TotalNumIntRegs * 8 +
10394 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
10396 /* Align ArgSize to a multiple of 8 */
10397 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
10398 bool NeedsAlign = (Align > 8);
10400 MachineBasicBlock *thisMBB = MBB;
10401 MachineBasicBlock *overflowMBB;
10402 MachineBasicBlock *offsetMBB;
10403 MachineBasicBlock *endMBB;
10405 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
10406 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
10407 unsigned OffsetReg = 0;
10409 if (!UseGPOffset && !UseFPOffset) {
10410 // If we only pull from the overflow region, we don't create a branch.
10411 // We don't need to alter control flow.
10412 OffsetDestReg = 0; // unused
10413 OverflowDestReg = DestReg;
10416 overflowMBB = thisMBB;
10419 // First emit code to check if gp_offset (or fp_offset) is below the bound.
10420 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
10421 // If not, pull from overflow_area. (branch to overflowMBB)
10426 // offsetMBB overflowMBB
10431 // Registers for the PHI in endMBB
10432 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
10433 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
10435 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10436 MachineFunction *MF = MBB->getParent();
10437 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10438 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10439 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10441 MachineFunction::iterator MBBIter = MBB;
10444 // Insert the new basic blocks
10445 MF->insert(MBBIter, offsetMBB);
10446 MF->insert(MBBIter, overflowMBB);
10447 MF->insert(MBBIter, endMBB);
10449 // Transfer the remainder of MBB and its successor edges to endMBB.
10450 endMBB->splice(endMBB->begin(), thisMBB,
10451 llvm::next(MachineBasicBlock::iterator(MI)),
10453 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10455 // Make offsetMBB and overflowMBB successors of thisMBB
10456 thisMBB->addSuccessor(offsetMBB);
10457 thisMBB->addSuccessor(overflowMBB);
10459 // endMBB is a successor of both offsetMBB and overflowMBB
10460 offsetMBB->addSuccessor(endMBB);
10461 overflowMBB->addSuccessor(endMBB);
10463 // Load the offset value into a register
10464 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
10465 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
10469 .addDisp(Disp, UseFPOffset ? 4 : 0)
10470 .addOperand(Segment)
10471 .setMemRefs(MMOBegin, MMOEnd);
10473 // Check if there is enough room left to pull this argument.
10474 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
10476 .addImm(MaxOffset + 8 - ArgSizeA8);
10478 // Branch to "overflowMBB" if offset >= max
10479 // Fall through to "offsetMBB" otherwise
10480 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
10481 .addMBB(overflowMBB);
10484 // In offsetMBB, emit code to use the reg_save_area.
10486 assert(OffsetReg != 0);
10488 // Read the reg_save_area address.
10489 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
10490 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
10495 .addOperand(Segment)
10496 .setMemRefs(MMOBegin, MMOEnd);
10498 // Zero-extend the offset
10499 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
10500 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
10503 .addImm(X86::sub_32bit);
10505 // Add the offset to the reg_save_area to get the final address.
10506 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
10507 .addReg(OffsetReg64)
10508 .addReg(RegSaveReg);
10510 // Compute the offset for the next argument
10511 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
10512 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
10514 .addImm(UseFPOffset ? 16 : 8);
10516 // Store it back into the va_list.
10517 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
10521 .addDisp(Disp, UseFPOffset ? 4 : 0)
10522 .addOperand(Segment)
10523 .addReg(NextOffsetReg)
10524 .setMemRefs(MMOBegin, MMOEnd);
10527 BuildMI(offsetMBB, DL, TII->get(X86::JMP_4))
10532 // Emit code to use overflow area
10535 // Load the overflow_area address into a register.
10536 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
10537 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
10542 .addOperand(Segment)
10543 .setMemRefs(MMOBegin, MMOEnd);
10545 // If we need to align it, do so. Otherwise, just copy the address
10546 // to OverflowDestReg.
10548 // Align the overflow address
10549 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
10550 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
10552 // aligned_addr = (addr + (align-1)) & ~(align-1)
10553 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
10554 .addReg(OverflowAddrReg)
10557 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
10559 .addImm(~(uint64_t)(Align-1));
10561 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
10562 .addReg(OverflowAddrReg);
10565 // Compute the next overflow address after this argument.
10566 // (the overflow address should be kept 8-byte aligned)
10567 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
10568 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
10569 .addReg(OverflowDestReg)
10570 .addImm(ArgSizeA8);
10572 // Store the new overflow address.
10573 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
10578 .addOperand(Segment)
10579 .addReg(NextAddrReg)
10580 .setMemRefs(MMOBegin, MMOEnd);
10582 // If we branched, emit the PHI to the front of endMBB.
10584 BuildMI(*endMBB, endMBB->begin(), DL,
10585 TII->get(X86::PHI), DestReg)
10586 .addReg(OffsetDestReg).addMBB(offsetMBB)
10587 .addReg(OverflowDestReg).addMBB(overflowMBB);
10590 // Erase the pseudo instruction
10591 MI->eraseFromParent();
10596 MachineBasicBlock *
10597 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
10599 MachineBasicBlock *MBB) const {
10600 // Emit code to save XMM registers to the stack. The ABI says that the
10601 // number of registers to save is given in %al, so it's theoretically
10602 // possible to do an indirect jump trick to avoid saving all of them,
10603 // however this code takes a simpler approach and just executes all
10604 // of the stores if %al is non-zero. It's less code, and it's probably
10605 // easier on the hardware branch predictor, and stores aren't all that
10606 // expensive anyway.
10608 // Create the new basic blocks. One block contains all the XMM stores,
10609 // and one block is the final destination regardless of whether any
10610 // stores were performed.
10611 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10612 MachineFunction *F = MBB->getParent();
10613 MachineFunction::iterator MBBIter = MBB;
10615 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
10616 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
10617 F->insert(MBBIter, XMMSaveMBB);
10618 F->insert(MBBIter, EndMBB);
10620 // Transfer the remainder of MBB and its successor edges to EndMBB.
10621 EndMBB->splice(EndMBB->begin(), MBB,
10622 llvm::next(MachineBasicBlock::iterator(MI)),
10624 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
10626 // The original block will now fall through to the XMM save block.
10627 MBB->addSuccessor(XMMSaveMBB);
10628 // The XMMSaveMBB will fall through to the end block.
10629 XMMSaveMBB->addSuccessor(EndMBB);
10631 // Now add the instructions.
10632 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10633 DebugLoc DL = MI->getDebugLoc();
10635 unsigned CountReg = MI->getOperand(0).getReg();
10636 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
10637 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
10639 if (!Subtarget->isTargetWin64()) {
10640 // If %al is 0, branch around the XMM save block.
10641 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
10642 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
10643 MBB->addSuccessor(EndMBB);
10646 // In the XMM save block, save all the XMM argument registers.
10647 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
10648 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
10649 MachineMemOperand *MMO =
10650 F->getMachineMemOperand(
10651 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
10652 MachineMemOperand::MOStore,
10653 /*Size=*/16, /*Align=*/16);
10654 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
10655 .addFrameIndex(RegSaveFrameIndex)
10656 .addImm(/*Scale=*/1)
10657 .addReg(/*IndexReg=*/0)
10658 .addImm(/*Disp=*/Offset)
10659 .addReg(/*Segment=*/0)
10660 .addReg(MI->getOperand(i).getReg())
10661 .addMemOperand(MMO);
10664 MI->eraseFromParent(); // The pseudo instruction is gone now.
10669 MachineBasicBlock *
10670 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
10671 MachineBasicBlock *BB) const {
10672 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10673 DebugLoc DL = MI->getDebugLoc();
10675 // To "insert" a SELECT_CC instruction, we actually have to insert the
10676 // diamond control-flow pattern. The incoming instruction knows the
10677 // destination vreg to set, the condition code register to branch on, the
10678 // true/false values to select between, and a branch opcode to use.
10679 const BasicBlock *LLVM_BB = BB->getBasicBlock();
10680 MachineFunction::iterator It = BB;
10686 // cmpTY ccX, r1, r2
10688 // fallthrough --> copy0MBB
10689 MachineBasicBlock *thisMBB = BB;
10690 MachineFunction *F = BB->getParent();
10691 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
10692 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
10693 F->insert(It, copy0MBB);
10694 F->insert(It, sinkMBB);
10696 // If the EFLAGS register isn't dead in the terminator, then claim that it's
10697 // live into the sink and copy blocks.
10698 const MachineFunction *MF = BB->getParent();
10699 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
10700 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
10702 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
10703 const MachineOperand &MO = MI->getOperand(I);
10704 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
10705 unsigned Reg = MO.getReg();
10706 if (Reg != X86::EFLAGS) continue;
10707 copy0MBB->addLiveIn(Reg);
10708 sinkMBB->addLiveIn(Reg);
10711 // Transfer the remainder of BB and its successor edges to sinkMBB.
10712 sinkMBB->splice(sinkMBB->begin(), BB,
10713 llvm::next(MachineBasicBlock::iterator(MI)),
10715 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
10717 // Add the true and fallthrough blocks as its successors.
10718 BB->addSuccessor(copy0MBB);
10719 BB->addSuccessor(sinkMBB);
10721 // Create the conditional branch instruction.
10723 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
10724 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
10727 // %FalseValue = ...
10728 // # fallthrough to sinkMBB
10729 copy0MBB->addSuccessor(sinkMBB);
10732 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
10734 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
10735 TII->get(X86::PHI), MI->getOperand(0).getReg())
10736 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
10737 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
10739 MI->eraseFromParent(); // The pseudo instruction is gone now.
10743 MachineBasicBlock *
10744 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
10745 MachineBasicBlock *BB) const {
10746 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10747 DebugLoc DL = MI->getDebugLoc();
10749 assert(!Subtarget->isTargetEnvMacho());
10751 // The lowering is pretty easy: we're just emitting the call to _alloca. The
10752 // non-trivial part is impdef of ESP.
10754 if (Subtarget->isTargetWin64()) {
10755 if (Subtarget->isTargetCygMing()) {
10756 // ___chkstk(Mingw64):
10757 // Clobbers R10, R11, RAX and EFLAGS.
10759 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
10760 .addExternalSymbol("___chkstk")
10761 .addReg(X86::RAX, RegState::Implicit)
10762 .addReg(X86::RSP, RegState::Implicit)
10763 .addReg(X86::RAX, RegState::Define | RegState::Implicit)
10764 .addReg(X86::RSP, RegState::Define | RegState::Implicit)
10765 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
10767 // __chkstk(MSVCRT): does not update stack pointer.
10768 // Clobbers R10, R11 and EFLAGS.
10769 // FIXME: RAX(allocated size) might be reused and not killed.
10770 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
10771 .addExternalSymbol("__chkstk")
10772 .addReg(X86::RAX, RegState::Implicit)
10773 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
10774 // RAX has the offset to subtracted from RSP.
10775 BuildMI(*BB, MI, DL, TII->get(X86::SUB64rr), X86::RSP)
10780 const char *StackProbeSymbol =
10781 Subtarget->isTargetWindows() ? "_chkstk" : "_alloca";
10783 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
10784 .addExternalSymbol(StackProbeSymbol)
10785 .addReg(X86::EAX, RegState::Implicit)
10786 .addReg(X86::ESP, RegState::Implicit)
10787 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
10788 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
10789 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
10792 MI->eraseFromParent(); // The pseudo instruction is gone now.
10796 MachineBasicBlock *
10797 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
10798 MachineBasicBlock *BB) const {
10799 // This is pretty easy. We're taking the value that we received from
10800 // our load from the relocation, sticking it in either RDI (x86-64)
10801 // or EAX and doing an indirect call. The return value will then
10802 // be in the normal return register.
10803 const X86InstrInfo *TII
10804 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
10805 DebugLoc DL = MI->getDebugLoc();
10806 MachineFunction *F = BB->getParent();
10808 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
10809 assert(MI->getOperand(3).isGlobal() && "This should be a global");
10811 if (Subtarget->is64Bit()) {
10812 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10813 TII->get(X86::MOV64rm), X86::RDI)
10815 .addImm(0).addReg(0)
10816 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10817 MI->getOperand(3).getTargetFlags())
10819 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
10820 addDirectMem(MIB, X86::RDI);
10821 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
10822 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10823 TII->get(X86::MOV32rm), X86::EAX)
10825 .addImm(0).addReg(0)
10826 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10827 MI->getOperand(3).getTargetFlags())
10829 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
10830 addDirectMem(MIB, X86::EAX);
10832 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10833 TII->get(X86::MOV32rm), X86::EAX)
10834 .addReg(TII->getGlobalBaseReg(F))
10835 .addImm(0).addReg(0)
10836 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10837 MI->getOperand(3).getTargetFlags())
10839 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
10840 addDirectMem(MIB, X86::EAX);
10843 MI->eraseFromParent(); // The pseudo instruction is gone now.
10847 MachineBasicBlock *
10848 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
10849 MachineBasicBlock *BB) const {
10850 switch (MI->getOpcode()) {
10851 default: assert(false && "Unexpected instr type to insert");
10852 case X86::TAILJMPd64:
10853 case X86::TAILJMPr64:
10854 case X86::TAILJMPm64:
10855 assert(!"TAILJMP64 would not be touched here.");
10856 case X86::TCRETURNdi64:
10857 case X86::TCRETURNri64:
10858 case X86::TCRETURNmi64:
10859 // Defs of TCRETURNxx64 has Win64's callee-saved registers, as subset.
10860 // On AMD64, additional defs should be added before register allocation.
10861 if (!Subtarget->isTargetWin64()) {
10862 MI->addRegisterDefined(X86::RSI);
10863 MI->addRegisterDefined(X86::RDI);
10864 MI->addRegisterDefined(X86::XMM6);
10865 MI->addRegisterDefined(X86::XMM7);
10866 MI->addRegisterDefined(X86::XMM8);
10867 MI->addRegisterDefined(X86::XMM9);
10868 MI->addRegisterDefined(X86::XMM10);
10869 MI->addRegisterDefined(X86::XMM11);
10870 MI->addRegisterDefined(X86::XMM12);
10871 MI->addRegisterDefined(X86::XMM13);
10872 MI->addRegisterDefined(X86::XMM14);
10873 MI->addRegisterDefined(X86::XMM15);
10876 case X86::WIN_ALLOCA:
10877 return EmitLoweredWinAlloca(MI, BB);
10878 case X86::TLSCall_32:
10879 case X86::TLSCall_64:
10880 return EmitLoweredTLSCall(MI, BB);
10881 case X86::CMOV_GR8:
10882 case X86::CMOV_FR32:
10883 case X86::CMOV_FR64:
10884 case X86::CMOV_V4F32:
10885 case X86::CMOV_V2F64:
10886 case X86::CMOV_V2I64:
10887 case X86::CMOV_GR16:
10888 case X86::CMOV_GR32:
10889 case X86::CMOV_RFP32:
10890 case X86::CMOV_RFP64:
10891 case X86::CMOV_RFP80:
10892 return EmitLoweredSelect(MI, BB);
10894 case X86::FP32_TO_INT16_IN_MEM:
10895 case X86::FP32_TO_INT32_IN_MEM:
10896 case X86::FP32_TO_INT64_IN_MEM:
10897 case X86::FP64_TO_INT16_IN_MEM:
10898 case X86::FP64_TO_INT32_IN_MEM:
10899 case X86::FP64_TO_INT64_IN_MEM:
10900 case X86::FP80_TO_INT16_IN_MEM:
10901 case X86::FP80_TO_INT32_IN_MEM:
10902 case X86::FP80_TO_INT64_IN_MEM: {
10903 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10904 DebugLoc DL = MI->getDebugLoc();
10906 // Change the floating point control register to use "round towards zero"
10907 // mode when truncating to an integer value.
10908 MachineFunction *F = BB->getParent();
10909 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
10910 addFrameReference(BuildMI(*BB, MI, DL,
10911 TII->get(X86::FNSTCW16m)), CWFrameIdx);
10913 // Load the old value of the high byte of the control word...
10915 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
10916 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
10919 // Set the high part to be round to zero...
10920 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
10923 // Reload the modified control word now...
10924 addFrameReference(BuildMI(*BB, MI, DL,
10925 TII->get(X86::FLDCW16m)), CWFrameIdx);
10927 // Restore the memory image of control word to original value
10928 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
10931 // Get the X86 opcode to use.
10933 switch (MI->getOpcode()) {
10934 default: llvm_unreachable("illegal opcode!");
10935 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
10936 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
10937 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
10938 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
10939 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
10940 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
10941 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
10942 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
10943 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
10947 MachineOperand &Op = MI->getOperand(0);
10949 AM.BaseType = X86AddressMode::RegBase;
10950 AM.Base.Reg = Op.getReg();
10952 AM.BaseType = X86AddressMode::FrameIndexBase;
10953 AM.Base.FrameIndex = Op.getIndex();
10955 Op = MI->getOperand(1);
10957 AM.Scale = Op.getImm();
10958 Op = MI->getOperand(2);
10960 AM.IndexReg = Op.getImm();
10961 Op = MI->getOperand(3);
10962 if (Op.isGlobal()) {
10963 AM.GV = Op.getGlobal();
10965 AM.Disp = Op.getImm();
10967 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
10968 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
10970 // Reload the original control word now.
10971 addFrameReference(BuildMI(*BB, MI, DL,
10972 TII->get(X86::FLDCW16m)), CWFrameIdx);
10974 MI->eraseFromParent(); // The pseudo instruction is gone now.
10977 // String/text processing lowering.
10978 case X86::PCMPISTRM128REG:
10979 case X86::VPCMPISTRM128REG:
10980 return EmitPCMP(MI, BB, 3, false /* in-mem */);
10981 case X86::PCMPISTRM128MEM:
10982 case X86::VPCMPISTRM128MEM:
10983 return EmitPCMP(MI, BB, 3, true /* in-mem */);
10984 case X86::PCMPESTRM128REG:
10985 case X86::VPCMPESTRM128REG:
10986 return EmitPCMP(MI, BB, 5, false /* in mem */);
10987 case X86::PCMPESTRM128MEM:
10988 case X86::VPCMPESTRM128MEM:
10989 return EmitPCMP(MI, BB, 5, true /* in mem */);
10991 // Thread synchronization.
10993 return EmitMonitor(MI, BB);
10995 return EmitMwait(MI, BB);
10997 // Atomic Lowering.
10998 case X86::ATOMAND32:
10999 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
11000 X86::AND32ri, X86::MOV32rm,
11002 X86::NOT32r, X86::EAX,
11003 X86::GR32RegisterClass);
11004 case X86::ATOMOR32:
11005 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
11006 X86::OR32ri, X86::MOV32rm,
11008 X86::NOT32r, X86::EAX,
11009 X86::GR32RegisterClass);
11010 case X86::ATOMXOR32:
11011 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
11012 X86::XOR32ri, X86::MOV32rm,
11014 X86::NOT32r, X86::EAX,
11015 X86::GR32RegisterClass);
11016 case X86::ATOMNAND32:
11017 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
11018 X86::AND32ri, X86::MOV32rm,
11020 X86::NOT32r, X86::EAX,
11021 X86::GR32RegisterClass, true);
11022 case X86::ATOMMIN32:
11023 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
11024 case X86::ATOMMAX32:
11025 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
11026 case X86::ATOMUMIN32:
11027 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
11028 case X86::ATOMUMAX32:
11029 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
11031 case X86::ATOMAND16:
11032 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
11033 X86::AND16ri, X86::MOV16rm,
11035 X86::NOT16r, X86::AX,
11036 X86::GR16RegisterClass);
11037 case X86::ATOMOR16:
11038 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
11039 X86::OR16ri, X86::MOV16rm,
11041 X86::NOT16r, X86::AX,
11042 X86::GR16RegisterClass);
11043 case X86::ATOMXOR16:
11044 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
11045 X86::XOR16ri, X86::MOV16rm,
11047 X86::NOT16r, X86::AX,
11048 X86::GR16RegisterClass);
11049 case X86::ATOMNAND16:
11050 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
11051 X86::AND16ri, X86::MOV16rm,
11053 X86::NOT16r, X86::AX,
11054 X86::GR16RegisterClass, true);
11055 case X86::ATOMMIN16:
11056 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
11057 case X86::ATOMMAX16:
11058 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
11059 case X86::ATOMUMIN16:
11060 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
11061 case X86::ATOMUMAX16:
11062 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
11064 case X86::ATOMAND8:
11065 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
11066 X86::AND8ri, X86::MOV8rm,
11068 X86::NOT8r, X86::AL,
11069 X86::GR8RegisterClass);
11071 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
11072 X86::OR8ri, X86::MOV8rm,
11074 X86::NOT8r, X86::AL,
11075 X86::GR8RegisterClass);
11076 case X86::ATOMXOR8:
11077 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
11078 X86::XOR8ri, X86::MOV8rm,
11080 X86::NOT8r, X86::AL,
11081 X86::GR8RegisterClass);
11082 case X86::ATOMNAND8:
11083 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
11084 X86::AND8ri, X86::MOV8rm,
11086 X86::NOT8r, X86::AL,
11087 X86::GR8RegisterClass, true);
11088 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
11089 // This group is for 64-bit host.
11090 case X86::ATOMAND64:
11091 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
11092 X86::AND64ri32, X86::MOV64rm,
11094 X86::NOT64r, X86::RAX,
11095 X86::GR64RegisterClass);
11096 case X86::ATOMOR64:
11097 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
11098 X86::OR64ri32, X86::MOV64rm,
11100 X86::NOT64r, X86::RAX,
11101 X86::GR64RegisterClass);
11102 case X86::ATOMXOR64:
11103 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
11104 X86::XOR64ri32, X86::MOV64rm,
11106 X86::NOT64r, X86::RAX,
11107 X86::GR64RegisterClass);
11108 case X86::ATOMNAND64:
11109 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
11110 X86::AND64ri32, X86::MOV64rm,
11112 X86::NOT64r, X86::RAX,
11113 X86::GR64RegisterClass, true);
11114 case X86::ATOMMIN64:
11115 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
11116 case X86::ATOMMAX64:
11117 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
11118 case X86::ATOMUMIN64:
11119 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
11120 case X86::ATOMUMAX64:
11121 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
11123 // This group does 64-bit operations on a 32-bit host.
11124 case X86::ATOMAND6432:
11125 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11126 X86::AND32rr, X86::AND32rr,
11127 X86::AND32ri, X86::AND32ri,
11129 case X86::ATOMOR6432:
11130 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11131 X86::OR32rr, X86::OR32rr,
11132 X86::OR32ri, X86::OR32ri,
11134 case X86::ATOMXOR6432:
11135 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11136 X86::XOR32rr, X86::XOR32rr,
11137 X86::XOR32ri, X86::XOR32ri,
11139 case X86::ATOMNAND6432:
11140 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11141 X86::AND32rr, X86::AND32rr,
11142 X86::AND32ri, X86::AND32ri,
11144 case X86::ATOMADD6432:
11145 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11146 X86::ADD32rr, X86::ADC32rr,
11147 X86::ADD32ri, X86::ADC32ri,
11149 case X86::ATOMSUB6432:
11150 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11151 X86::SUB32rr, X86::SBB32rr,
11152 X86::SUB32ri, X86::SBB32ri,
11154 case X86::ATOMSWAP6432:
11155 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11156 X86::MOV32rr, X86::MOV32rr,
11157 X86::MOV32ri, X86::MOV32ri,
11159 case X86::VASTART_SAVE_XMM_REGS:
11160 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
11162 case X86::VAARG_64:
11163 return EmitVAARG64WithCustomInserter(MI, BB);
11167 //===----------------------------------------------------------------------===//
11168 // X86 Optimization Hooks
11169 //===----------------------------------------------------------------------===//
11171 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
11175 const SelectionDAG &DAG,
11176 unsigned Depth) const {
11177 unsigned Opc = Op.getOpcode();
11178 assert((Opc >= ISD::BUILTIN_OP_END ||
11179 Opc == ISD::INTRINSIC_WO_CHAIN ||
11180 Opc == ISD::INTRINSIC_W_CHAIN ||
11181 Opc == ISD::INTRINSIC_VOID) &&
11182 "Should use MaskedValueIsZero if you don't know whether Op"
11183 " is a target node!");
11185 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
11199 // These nodes' second result is a boolean.
11200 if (Op.getResNo() == 0)
11203 case X86ISD::SETCC:
11204 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
11205 Mask.getBitWidth() - 1);
11210 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
11211 unsigned Depth) const {
11212 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
11213 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
11214 return Op.getValueType().getScalarType().getSizeInBits();
11220 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
11221 /// node is a GlobalAddress + offset.
11222 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
11223 const GlobalValue* &GA,
11224 int64_t &Offset) const {
11225 if (N->getOpcode() == X86ISD::Wrapper) {
11226 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
11227 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
11228 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
11232 return TargetLowering::isGAPlusOffset(N, GA, Offset);
11235 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
11236 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
11237 /// if the load addresses are consecutive, non-overlapping, and in the right
11239 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
11240 TargetLowering::DAGCombinerInfo &DCI) {
11241 DebugLoc dl = N->getDebugLoc();
11242 EVT VT = N->getValueType(0);
11244 if (VT.getSizeInBits() != 128)
11247 // Don't create instructions with illegal types after legalize types has run.
11248 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11249 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
11252 SmallVector<SDValue, 16> Elts;
11253 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
11254 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
11256 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
11259 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
11260 /// generation and convert it from being a bunch of shuffles and extracts
11261 /// to a simple store and scalar loads to extract the elements.
11262 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
11263 const TargetLowering &TLI) {
11264 SDValue InputVector = N->getOperand(0);
11266 // Only operate on vectors of 4 elements, where the alternative shuffling
11267 // gets to be more expensive.
11268 if (InputVector.getValueType() != MVT::v4i32)
11271 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
11272 // single use which is a sign-extend or zero-extend, and all elements are
11274 SmallVector<SDNode *, 4> Uses;
11275 unsigned ExtractedElements = 0;
11276 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
11277 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
11278 if (UI.getUse().getResNo() != InputVector.getResNo())
11281 SDNode *Extract = *UI;
11282 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
11285 if (Extract->getValueType(0) != MVT::i32)
11287 if (!Extract->hasOneUse())
11289 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
11290 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
11292 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
11295 // Record which element was extracted.
11296 ExtractedElements |=
11297 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
11299 Uses.push_back(Extract);
11302 // If not all the elements were used, this may not be worthwhile.
11303 if (ExtractedElements != 15)
11306 // Ok, we've now decided to do the transformation.
11307 DebugLoc dl = InputVector.getDebugLoc();
11309 // Store the value to a temporary stack slot.
11310 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
11311 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
11312 MachinePointerInfo(), false, false, 0);
11314 // Replace each use (extract) with a load of the appropriate element.
11315 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
11316 UE = Uses.end(); UI != UE; ++UI) {
11317 SDNode *Extract = *UI;
11319 // cOMpute the element's address.
11320 SDValue Idx = Extract->getOperand(1);
11322 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
11323 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
11324 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
11326 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
11327 StackPtr, OffsetVal);
11329 // Load the scalar.
11330 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
11331 ScalarAddr, MachinePointerInfo(),
11334 // Replace the exact with the load.
11335 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
11338 // The replacement was made in place; don't return anything.
11342 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
11343 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
11344 const X86Subtarget *Subtarget) {
11345 DebugLoc DL = N->getDebugLoc();
11346 SDValue Cond = N->getOperand(0);
11347 // Get the LHS/RHS of the select.
11348 SDValue LHS = N->getOperand(1);
11349 SDValue RHS = N->getOperand(2);
11351 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
11352 // instructions match the semantics of the common C idiom x<y?x:y but not
11353 // x<=y?x:y, because of how they handle negative zero (which can be
11354 // ignored in unsafe-math mode).
11355 if (Subtarget->hasSSE2() &&
11356 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
11357 Cond.getOpcode() == ISD::SETCC) {
11358 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
11360 unsigned Opcode = 0;
11361 // Check for x CC y ? x : y.
11362 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
11363 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
11367 // Converting this to a min would handle NaNs incorrectly, and swapping
11368 // the operands would cause it to handle comparisons between positive
11369 // and negative zero incorrectly.
11370 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
11371 if (!UnsafeFPMath &&
11372 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
11374 std::swap(LHS, RHS);
11376 Opcode = X86ISD::FMIN;
11379 // Converting this to a min would handle comparisons between positive
11380 // and negative zero incorrectly.
11381 if (!UnsafeFPMath &&
11382 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
11384 Opcode = X86ISD::FMIN;
11387 // Converting this to a min would handle both negative zeros and NaNs
11388 // incorrectly, but we can swap the operands to fix both.
11389 std::swap(LHS, RHS);
11393 Opcode = X86ISD::FMIN;
11397 // Converting this to a max would handle comparisons between positive
11398 // and negative zero incorrectly.
11399 if (!UnsafeFPMath &&
11400 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
11402 Opcode = X86ISD::FMAX;
11405 // Converting this to a max would handle NaNs incorrectly, and swapping
11406 // the operands would cause it to handle comparisons between positive
11407 // and negative zero incorrectly.
11408 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
11409 if (!UnsafeFPMath &&
11410 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
11412 std::swap(LHS, RHS);
11414 Opcode = X86ISD::FMAX;
11417 // Converting this to a max would handle both negative zeros and NaNs
11418 // incorrectly, but we can swap the operands to fix both.
11419 std::swap(LHS, RHS);
11423 Opcode = X86ISD::FMAX;
11426 // Check for x CC y ? y : x -- a min/max with reversed arms.
11427 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
11428 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
11432 // Converting this to a min would handle comparisons between positive
11433 // and negative zero incorrectly, and swapping the operands would
11434 // cause it to handle NaNs incorrectly.
11435 if (!UnsafeFPMath &&
11436 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
11437 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11439 std::swap(LHS, RHS);
11441 Opcode = X86ISD::FMIN;
11444 // Converting this to a min would handle NaNs incorrectly.
11445 if (!UnsafeFPMath &&
11446 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
11448 Opcode = X86ISD::FMIN;
11451 // Converting this to a min would handle both negative zeros and NaNs
11452 // incorrectly, but we can swap the operands to fix both.
11453 std::swap(LHS, RHS);
11457 Opcode = X86ISD::FMIN;
11461 // Converting this to a max would handle NaNs incorrectly.
11462 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11464 Opcode = X86ISD::FMAX;
11467 // Converting this to a max would handle comparisons between positive
11468 // and negative zero incorrectly, and swapping the operands would
11469 // cause it to handle NaNs incorrectly.
11470 if (!UnsafeFPMath &&
11471 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
11472 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11474 std::swap(LHS, RHS);
11476 Opcode = X86ISD::FMAX;
11479 // Converting this to a max would handle both negative zeros and NaNs
11480 // incorrectly, but we can swap the operands to fix both.
11481 std::swap(LHS, RHS);
11485 Opcode = X86ISD::FMAX;
11491 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
11494 // If this is a select between two integer constants, try to do some
11496 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
11497 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
11498 // Don't do this for crazy integer types.
11499 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
11500 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
11501 // so that TrueC (the true value) is larger than FalseC.
11502 bool NeedsCondInvert = false;
11504 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
11505 // Efficiently invertible.
11506 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
11507 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
11508 isa<ConstantSDNode>(Cond.getOperand(1))))) {
11509 NeedsCondInvert = true;
11510 std::swap(TrueC, FalseC);
11513 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
11514 if (FalseC->getAPIntValue() == 0 &&
11515 TrueC->getAPIntValue().isPowerOf2()) {
11516 if (NeedsCondInvert) // Invert the condition if needed.
11517 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11518 DAG.getConstant(1, Cond.getValueType()));
11520 // Zero extend the condition if needed.
11521 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
11523 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
11524 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
11525 DAG.getConstant(ShAmt, MVT::i8));
11528 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
11529 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
11530 if (NeedsCondInvert) // Invert the condition if needed.
11531 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11532 DAG.getConstant(1, Cond.getValueType()));
11534 // Zero extend the condition if needed.
11535 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
11536 FalseC->getValueType(0), Cond);
11537 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11538 SDValue(FalseC, 0));
11541 // Optimize cases that will turn into an LEA instruction. This requires
11542 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
11543 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
11544 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
11545 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
11547 bool isFastMultiplier = false;
11549 switch ((unsigned char)Diff) {
11551 case 1: // result = add base, cond
11552 case 2: // result = lea base( , cond*2)
11553 case 3: // result = lea base(cond, cond*2)
11554 case 4: // result = lea base( , cond*4)
11555 case 5: // result = lea base(cond, cond*4)
11556 case 8: // result = lea base( , cond*8)
11557 case 9: // result = lea base(cond, cond*8)
11558 isFastMultiplier = true;
11563 if (isFastMultiplier) {
11564 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
11565 if (NeedsCondInvert) // Invert the condition if needed.
11566 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11567 DAG.getConstant(1, Cond.getValueType()));
11569 // Zero extend the condition if needed.
11570 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
11572 // Scale the condition by the difference.
11574 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
11575 DAG.getConstant(Diff, Cond.getValueType()));
11577 // Add the base if non-zero.
11578 if (FalseC->getAPIntValue() != 0)
11579 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11580 SDValue(FalseC, 0));
11590 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
11591 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
11592 TargetLowering::DAGCombinerInfo &DCI) {
11593 DebugLoc DL = N->getDebugLoc();
11595 // If the flag operand isn't dead, don't touch this CMOV.
11596 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
11599 SDValue FalseOp = N->getOperand(0);
11600 SDValue TrueOp = N->getOperand(1);
11601 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
11602 SDValue Cond = N->getOperand(3);
11603 if (CC == X86::COND_E || CC == X86::COND_NE) {
11604 switch (Cond.getOpcode()) {
11608 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
11609 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
11610 return (CC == X86::COND_E) ? FalseOp : TrueOp;
11614 // If this is a select between two integer constants, try to do some
11615 // optimizations. Note that the operands are ordered the opposite of SELECT
11617 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
11618 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
11619 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
11620 // larger than FalseC (the false value).
11621 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
11622 CC = X86::GetOppositeBranchCondition(CC);
11623 std::swap(TrueC, FalseC);
11626 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
11627 // This is efficient for any integer data type (including i8/i16) and
11629 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
11630 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11631 DAG.getConstant(CC, MVT::i8), Cond);
11633 // Zero extend the condition if needed.
11634 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
11636 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
11637 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
11638 DAG.getConstant(ShAmt, MVT::i8));
11639 if (N->getNumValues() == 2) // Dead flag value?
11640 return DCI.CombineTo(N, Cond, SDValue());
11644 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
11645 // for any integer data type, including i8/i16.
11646 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
11647 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11648 DAG.getConstant(CC, MVT::i8), Cond);
11650 // Zero extend the condition if needed.
11651 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
11652 FalseC->getValueType(0), Cond);
11653 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11654 SDValue(FalseC, 0));
11656 if (N->getNumValues() == 2) // Dead flag value?
11657 return DCI.CombineTo(N, Cond, SDValue());
11661 // Optimize cases that will turn into an LEA instruction. This requires
11662 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
11663 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
11664 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
11665 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
11667 bool isFastMultiplier = false;
11669 switch ((unsigned char)Diff) {
11671 case 1: // result = add base, cond
11672 case 2: // result = lea base( , cond*2)
11673 case 3: // result = lea base(cond, cond*2)
11674 case 4: // result = lea base( , cond*4)
11675 case 5: // result = lea base(cond, cond*4)
11676 case 8: // result = lea base( , cond*8)
11677 case 9: // result = lea base(cond, cond*8)
11678 isFastMultiplier = true;
11683 if (isFastMultiplier) {
11684 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
11685 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11686 DAG.getConstant(CC, MVT::i8), Cond);
11687 // Zero extend the condition if needed.
11688 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
11690 // Scale the condition by the difference.
11692 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
11693 DAG.getConstant(Diff, Cond.getValueType()));
11695 // Add the base if non-zero.
11696 if (FalseC->getAPIntValue() != 0)
11697 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11698 SDValue(FalseC, 0));
11699 if (N->getNumValues() == 2) // Dead flag value?
11700 return DCI.CombineTo(N, Cond, SDValue());
11710 /// PerformMulCombine - Optimize a single multiply with constant into two
11711 /// in order to implement it with two cheaper instructions, e.g.
11712 /// LEA + SHL, LEA + LEA.
11713 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
11714 TargetLowering::DAGCombinerInfo &DCI) {
11715 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
11718 EVT VT = N->getValueType(0);
11719 if (VT != MVT::i64)
11722 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
11725 uint64_t MulAmt = C->getZExtValue();
11726 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
11729 uint64_t MulAmt1 = 0;
11730 uint64_t MulAmt2 = 0;
11731 if ((MulAmt % 9) == 0) {
11733 MulAmt2 = MulAmt / 9;
11734 } else if ((MulAmt % 5) == 0) {
11736 MulAmt2 = MulAmt / 5;
11737 } else if ((MulAmt % 3) == 0) {
11739 MulAmt2 = MulAmt / 3;
11742 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
11743 DebugLoc DL = N->getDebugLoc();
11745 if (isPowerOf2_64(MulAmt2) &&
11746 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
11747 // If second multiplifer is pow2, issue it first. We want the multiply by
11748 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
11750 std::swap(MulAmt1, MulAmt2);
11753 if (isPowerOf2_64(MulAmt1))
11754 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
11755 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
11757 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
11758 DAG.getConstant(MulAmt1, VT));
11760 if (isPowerOf2_64(MulAmt2))
11761 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
11762 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
11764 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
11765 DAG.getConstant(MulAmt2, VT));
11767 // Do not add new nodes to DAG combiner worklist.
11768 DCI.CombineTo(N, NewMul, false);
11773 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
11774 SDValue N0 = N->getOperand(0);
11775 SDValue N1 = N->getOperand(1);
11776 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
11777 EVT VT = N0.getValueType();
11779 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
11780 // since the result of setcc_c is all zero's or all ones.
11781 if (N1C && N0.getOpcode() == ISD::AND &&
11782 N0.getOperand(1).getOpcode() == ISD::Constant) {
11783 SDValue N00 = N0.getOperand(0);
11784 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
11785 ((N00.getOpcode() == ISD::ANY_EXTEND ||
11786 N00.getOpcode() == ISD::ZERO_EXTEND) &&
11787 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
11788 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
11789 APInt ShAmt = N1C->getAPIntValue();
11790 Mask = Mask.shl(ShAmt);
11792 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
11793 N00, DAG.getConstant(Mask, VT));
11800 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
11802 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
11803 const X86Subtarget *Subtarget) {
11804 EVT VT = N->getValueType(0);
11805 if (!VT.isVector() && VT.isInteger() &&
11806 N->getOpcode() == ISD::SHL)
11807 return PerformSHLCombine(N, DAG);
11809 // On X86 with SSE2 support, we can transform this to a vector shift if
11810 // all elements are shifted by the same amount. We can't do this in legalize
11811 // because the a constant vector is typically transformed to a constant pool
11812 // so we have no knowledge of the shift amount.
11813 if (!Subtarget->hasSSE2())
11816 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
11819 SDValue ShAmtOp = N->getOperand(1);
11820 EVT EltVT = VT.getVectorElementType();
11821 DebugLoc DL = N->getDebugLoc();
11822 SDValue BaseShAmt = SDValue();
11823 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
11824 unsigned NumElts = VT.getVectorNumElements();
11826 for (; i != NumElts; ++i) {
11827 SDValue Arg = ShAmtOp.getOperand(i);
11828 if (Arg.getOpcode() == ISD::UNDEF) continue;
11832 for (; i != NumElts; ++i) {
11833 SDValue Arg = ShAmtOp.getOperand(i);
11834 if (Arg.getOpcode() == ISD::UNDEF) continue;
11835 if (Arg != BaseShAmt) {
11839 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
11840 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
11841 SDValue InVec = ShAmtOp.getOperand(0);
11842 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
11843 unsigned NumElts = InVec.getValueType().getVectorNumElements();
11845 for (; i != NumElts; ++i) {
11846 SDValue Arg = InVec.getOperand(i);
11847 if (Arg.getOpcode() == ISD::UNDEF) continue;
11851 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
11852 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
11853 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
11854 if (C->getZExtValue() == SplatIdx)
11855 BaseShAmt = InVec.getOperand(1);
11858 if (BaseShAmt.getNode() == 0)
11859 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
11860 DAG.getIntPtrConstant(0));
11864 // The shift amount is an i32.
11865 if (EltVT.bitsGT(MVT::i32))
11866 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
11867 else if (EltVT.bitsLT(MVT::i32))
11868 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
11870 // The shift amount is identical so we can do a vector shift.
11871 SDValue ValOp = N->getOperand(0);
11872 switch (N->getOpcode()) {
11874 llvm_unreachable("Unknown shift opcode!");
11877 if (VT == MVT::v2i64)
11878 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11879 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
11881 if (VT == MVT::v4i32)
11882 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11883 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
11885 if (VT == MVT::v8i16)
11886 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11887 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
11891 if (VT == MVT::v4i32)
11892 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11893 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
11895 if (VT == MVT::v8i16)
11896 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11897 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
11901 if (VT == MVT::v2i64)
11902 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11903 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
11905 if (VT == MVT::v4i32)
11906 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11907 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
11909 if (VT == MVT::v8i16)
11910 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11911 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
11919 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
11920 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
11921 // and friends. Likewise for OR -> CMPNEQSS.
11922 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
11923 TargetLowering::DAGCombinerInfo &DCI,
11924 const X86Subtarget *Subtarget) {
11927 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
11928 // we're requiring SSE2 for both.
11929 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
11930 SDValue N0 = N->getOperand(0);
11931 SDValue N1 = N->getOperand(1);
11932 SDValue CMP0 = N0->getOperand(1);
11933 SDValue CMP1 = N1->getOperand(1);
11934 DebugLoc DL = N->getDebugLoc();
11936 // The SETCCs should both refer to the same CMP.
11937 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
11940 SDValue CMP00 = CMP0->getOperand(0);
11941 SDValue CMP01 = CMP0->getOperand(1);
11942 EVT VT = CMP00.getValueType();
11944 if (VT == MVT::f32 || VT == MVT::f64) {
11945 bool ExpectingFlags = false;
11946 // Check for any users that want flags:
11947 for (SDNode::use_iterator UI = N->use_begin(),
11949 !ExpectingFlags && UI != UE; ++UI)
11950 switch (UI->getOpcode()) {
11955 ExpectingFlags = true;
11957 case ISD::CopyToReg:
11958 case ISD::SIGN_EXTEND:
11959 case ISD::ZERO_EXTEND:
11960 case ISD::ANY_EXTEND:
11964 if (!ExpectingFlags) {
11965 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
11966 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
11968 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
11969 X86::CondCode tmp = cc0;
11974 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
11975 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
11976 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
11977 X86ISD::NodeType NTOperator = is64BitFP ?
11978 X86ISD::FSETCCsd : X86ISD::FSETCCss;
11979 // FIXME: need symbolic constants for these magic numbers.
11980 // See X86ATTInstPrinter.cpp:printSSECC().
11981 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
11982 SDValue OnesOrZeroesF = DAG.getNode(NTOperator, DL, MVT::f32, CMP00, CMP01,
11983 DAG.getConstant(x86cc, MVT::i8));
11984 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, MVT::i32,
11986 SDValue ANDed = DAG.getNode(ISD::AND, DL, MVT::i32, OnesOrZeroesI,
11987 DAG.getConstant(1, MVT::i32));
11988 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
11989 return OneBitOfTruth;
11997 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
11998 TargetLowering::DAGCombinerInfo &DCI,
11999 const X86Subtarget *Subtarget) {
12000 if (DCI.isBeforeLegalizeOps())
12003 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
12007 // Want to form ANDNP nodes:
12008 // 1) In the hopes of then easily combining them with OR and AND nodes
12009 // to form PBLEND/PSIGN.
12010 // 2) To match ANDN packed intrinsics
12011 EVT VT = N->getValueType(0);
12012 if (VT != MVT::v2i64 && VT != MVT::v4i64)
12015 SDValue N0 = N->getOperand(0);
12016 SDValue N1 = N->getOperand(1);
12017 DebugLoc DL = N->getDebugLoc();
12019 // Check LHS for vnot
12020 if (N0.getOpcode() == ISD::XOR &&
12021 ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
12022 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
12024 // Check RHS for vnot
12025 if (N1.getOpcode() == ISD::XOR &&
12026 ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
12027 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
12032 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
12033 TargetLowering::DAGCombinerInfo &DCI,
12034 const X86Subtarget *Subtarget) {
12035 if (DCI.isBeforeLegalizeOps())
12038 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
12042 EVT VT = N->getValueType(0);
12043 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64 && VT != MVT::v2i64)
12046 SDValue N0 = N->getOperand(0);
12047 SDValue N1 = N->getOperand(1);
12049 // look for psign/blend
12050 if (Subtarget->hasSSSE3()) {
12051 if (VT == MVT::v2i64) {
12052 // Canonicalize pandn to RHS
12053 if (N0.getOpcode() == X86ISD::ANDNP)
12055 // or (and (m, x), (pandn m, y))
12056 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
12057 SDValue Mask = N1.getOperand(0);
12058 SDValue X = N1.getOperand(1);
12060 if (N0.getOperand(0) == Mask)
12061 Y = N0.getOperand(1);
12062 if (N0.getOperand(1) == Mask)
12063 Y = N0.getOperand(0);
12065 // Check to see if the mask appeared in both the AND and ANDNP and
12069 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
12070 if (Mask.getOpcode() != ISD::BITCAST ||
12071 X.getOpcode() != ISD::BITCAST ||
12072 Y.getOpcode() != ISD::BITCAST)
12075 // Look through mask bitcast.
12076 Mask = Mask.getOperand(0);
12077 EVT MaskVT = Mask.getValueType();
12079 // Validate that the Mask operand is a vector sra node. The sra node
12080 // will be an intrinsic.
12081 if (Mask.getOpcode() != ISD::INTRINSIC_WO_CHAIN)
12084 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
12085 // there is no psrai.b
12086 switch (cast<ConstantSDNode>(Mask.getOperand(0))->getZExtValue()) {
12087 case Intrinsic::x86_sse2_psrai_w:
12088 case Intrinsic::x86_sse2_psrai_d:
12090 default: return SDValue();
12093 // Check that the SRA is all signbits.
12094 SDValue SraC = Mask.getOperand(2);
12095 unsigned SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
12096 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
12097 if ((SraAmt + 1) != EltBits)
12100 DebugLoc DL = N->getDebugLoc();
12102 // Now we know we at least have a plendvb with the mask val. See if
12103 // we can form a psignb/w/d.
12104 // psign = x.type == y.type == mask.type && y = sub(0, x);
12105 X = X.getOperand(0);
12106 Y = Y.getOperand(0);
12107 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
12108 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
12109 X.getValueType() == MaskVT && X.getValueType() == Y.getValueType()){
12112 case 8: Opc = X86ISD::PSIGNB; break;
12113 case 16: Opc = X86ISD::PSIGNW; break;
12114 case 32: Opc = X86ISD::PSIGND; break;
12118 SDValue Sign = DAG.getNode(Opc, DL, MaskVT, X, Mask.getOperand(1));
12119 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Sign);
12122 // PBLENDVB only available on SSE 4.1
12123 if (!Subtarget->hasSSE41())
12126 X = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, X);
12127 Y = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Y);
12128 Mask = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Mask);
12129 Mask = DAG.getNode(X86ISD::PBLENDVB, DL, MVT::v16i8, X, Y, Mask);
12130 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Mask);
12135 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
12136 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
12138 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
12140 if (!N0.hasOneUse() || !N1.hasOneUse())
12143 SDValue ShAmt0 = N0.getOperand(1);
12144 if (ShAmt0.getValueType() != MVT::i8)
12146 SDValue ShAmt1 = N1.getOperand(1);
12147 if (ShAmt1.getValueType() != MVT::i8)
12149 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
12150 ShAmt0 = ShAmt0.getOperand(0);
12151 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
12152 ShAmt1 = ShAmt1.getOperand(0);
12154 DebugLoc DL = N->getDebugLoc();
12155 unsigned Opc = X86ISD::SHLD;
12156 SDValue Op0 = N0.getOperand(0);
12157 SDValue Op1 = N1.getOperand(0);
12158 if (ShAmt0.getOpcode() == ISD::SUB) {
12159 Opc = X86ISD::SHRD;
12160 std::swap(Op0, Op1);
12161 std::swap(ShAmt0, ShAmt1);
12164 unsigned Bits = VT.getSizeInBits();
12165 if (ShAmt1.getOpcode() == ISD::SUB) {
12166 SDValue Sum = ShAmt1.getOperand(0);
12167 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
12168 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
12169 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
12170 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
12171 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
12172 return DAG.getNode(Opc, DL, VT,
12174 DAG.getNode(ISD::TRUNCATE, DL,
12177 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
12178 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
12180 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
12181 return DAG.getNode(Opc, DL, VT,
12182 N0.getOperand(0), N1.getOperand(0),
12183 DAG.getNode(ISD::TRUNCATE, DL,
12190 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
12191 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
12192 const X86Subtarget *Subtarget) {
12193 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
12194 // the FP state in cases where an emms may be missing.
12195 // A preferable solution to the general problem is to figure out the right
12196 // places to insert EMMS. This qualifies as a quick hack.
12198 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
12199 StoreSDNode *St = cast<StoreSDNode>(N);
12200 EVT VT = St->getValue().getValueType();
12201 if (VT.getSizeInBits() != 64)
12204 const Function *F = DAG.getMachineFunction().getFunction();
12205 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
12206 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
12207 && Subtarget->hasSSE2();
12208 if ((VT.isVector() ||
12209 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
12210 isa<LoadSDNode>(St->getValue()) &&
12211 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
12212 St->getChain().hasOneUse() && !St->isVolatile()) {
12213 SDNode* LdVal = St->getValue().getNode();
12214 LoadSDNode *Ld = 0;
12215 int TokenFactorIndex = -1;
12216 SmallVector<SDValue, 8> Ops;
12217 SDNode* ChainVal = St->getChain().getNode();
12218 // Must be a store of a load. We currently handle two cases: the load
12219 // is a direct child, and it's under an intervening TokenFactor. It is
12220 // possible to dig deeper under nested TokenFactors.
12221 if (ChainVal == LdVal)
12222 Ld = cast<LoadSDNode>(St->getChain());
12223 else if (St->getValue().hasOneUse() &&
12224 ChainVal->getOpcode() == ISD::TokenFactor) {
12225 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
12226 if (ChainVal->getOperand(i).getNode() == LdVal) {
12227 TokenFactorIndex = i;
12228 Ld = cast<LoadSDNode>(St->getValue());
12230 Ops.push_back(ChainVal->getOperand(i));
12234 if (!Ld || !ISD::isNormalLoad(Ld))
12237 // If this is not the MMX case, i.e. we are just turning i64 load/store
12238 // into f64 load/store, avoid the transformation if there are multiple
12239 // uses of the loaded value.
12240 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
12243 DebugLoc LdDL = Ld->getDebugLoc();
12244 DebugLoc StDL = N->getDebugLoc();
12245 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
12246 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
12248 if (Subtarget->is64Bit() || F64IsLegal) {
12249 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
12250 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
12251 Ld->getPointerInfo(), Ld->isVolatile(),
12252 Ld->isNonTemporal(), Ld->getAlignment());
12253 SDValue NewChain = NewLd.getValue(1);
12254 if (TokenFactorIndex != -1) {
12255 Ops.push_back(NewChain);
12256 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
12259 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
12260 St->getPointerInfo(),
12261 St->isVolatile(), St->isNonTemporal(),
12262 St->getAlignment());
12265 // Otherwise, lower to two pairs of 32-bit loads / stores.
12266 SDValue LoAddr = Ld->getBasePtr();
12267 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
12268 DAG.getConstant(4, MVT::i32));
12270 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
12271 Ld->getPointerInfo(),
12272 Ld->isVolatile(), Ld->isNonTemporal(),
12273 Ld->getAlignment());
12274 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
12275 Ld->getPointerInfo().getWithOffset(4),
12276 Ld->isVolatile(), Ld->isNonTemporal(),
12277 MinAlign(Ld->getAlignment(), 4));
12279 SDValue NewChain = LoLd.getValue(1);
12280 if (TokenFactorIndex != -1) {
12281 Ops.push_back(LoLd);
12282 Ops.push_back(HiLd);
12283 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
12287 LoAddr = St->getBasePtr();
12288 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
12289 DAG.getConstant(4, MVT::i32));
12291 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
12292 St->getPointerInfo(),
12293 St->isVolatile(), St->isNonTemporal(),
12294 St->getAlignment());
12295 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
12296 St->getPointerInfo().getWithOffset(4),
12298 St->isNonTemporal(),
12299 MinAlign(St->getAlignment(), 4));
12300 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
12305 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
12306 /// X86ISD::FXOR nodes.
12307 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
12308 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
12309 // F[X]OR(0.0, x) -> x
12310 // F[X]OR(x, 0.0) -> x
12311 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
12312 if (C->getValueAPF().isPosZero())
12313 return N->getOperand(1);
12314 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
12315 if (C->getValueAPF().isPosZero())
12316 return N->getOperand(0);
12320 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
12321 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
12322 // FAND(0.0, x) -> 0.0
12323 // FAND(x, 0.0) -> 0.0
12324 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
12325 if (C->getValueAPF().isPosZero())
12326 return N->getOperand(0);
12327 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
12328 if (C->getValueAPF().isPosZero())
12329 return N->getOperand(1);
12333 static SDValue PerformBTCombine(SDNode *N,
12335 TargetLowering::DAGCombinerInfo &DCI) {
12336 // BT ignores high bits in the bit index operand.
12337 SDValue Op1 = N->getOperand(1);
12338 if (Op1.hasOneUse()) {
12339 unsigned BitWidth = Op1.getValueSizeInBits();
12340 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
12341 APInt KnownZero, KnownOne;
12342 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
12343 !DCI.isBeforeLegalizeOps());
12344 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12345 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
12346 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
12347 DCI.CommitTargetLoweringOpt(TLO);
12352 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
12353 SDValue Op = N->getOperand(0);
12354 if (Op.getOpcode() == ISD::BITCAST)
12355 Op = Op.getOperand(0);
12356 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
12357 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
12358 VT.getVectorElementType().getSizeInBits() ==
12359 OpVT.getVectorElementType().getSizeInBits()) {
12360 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
12365 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
12366 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
12367 // (and (i32 x86isd::setcc_carry), 1)
12368 // This eliminates the zext. This transformation is necessary because
12369 // ISD::SETCC is always legalized to i8.
12370 DebugLoc dl = N->getDebugLoc();
12371 SDValue N0 = N->getOperand(0);
12372 EVT VT = N->getValueType(0);
12373 if (N0.getOpcode() == ISD::AND &&
12375 N0.getOperand(0).hasOneUse()) {
12376 SDValue N00 = N0.getOperand(0);
12377 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
12379 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
12380 if (!C || C->getZExtValue() != 1)
12382 return DAG.getNode(ISD::AND, dl, VT,
12383 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
12384 N00.getOperand(0), N00.getOperand(1)),
12385 DAG.getConstant(1, VT));
12391 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
12392 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG) {
12393 unsigned X86CC = N->getConstantOperandVal(0);
12394 SDValue EFLAG = N->getOperand(1);
12395 DebugLoc DL = N->getDebugLoc();
12397 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
12398 // a zext and produces an all-ones bit which is more useful than 0/1 in some
12400 if (X86CC == X86::COND_B)
12401 return DAG.getNode(ISD::AND, DL, MVT::i8,
12402 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
12403 DAG.getConstant(X86CC, MVT::i8), EFLAG),
12404 DAG.getConstant(1, MVT::i8));
12409 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
12410 const X86TargetLowering *XTLI) {
12411 SDValue Op0 = N->getOperand(0);
12412 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
12413 // a 32-bit target where SSE doesn't support i64->FP operations.
12414 if (Op0.getOpcode() == ISD::LOAD) {
12415 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
12416 EVT VT = Ld->getValueType(0);
12417 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
12418 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
12419 !XTLI->getSubtarget()->is64Bit() &&
12420 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
12421 SDValue FILDChain = XTLI->BuildFILD(SDValue(N, 0), Ld->getValueType(0),
12422 Ld->getChain(), Op0, DAG);
12423 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
12430 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
12431 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
12432 X86TargetLowering::DAGCombinerInfo &DCI) {
12433 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
12434 // the result is either zero or one (depending on the input carry bit).
12435 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
12436 if (X86::isZeroNode(N->getOperand(0)) &&
12437 X86::isZeroNode(N->getOperand(1)) &&
12438 // We don't have a good way to replace an EFLAGS use, so only do this when
12440 SDValue(N, 1).use_empty()) {
12441 DebugLoc DL = N->getDebugLoc();
12442 EVT VT = N->getValueType(0);
12443 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
12444 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
12445 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
12446 DAG.getConstant(X86::COND_B,MVT::i8),
12448 DAG.getConstant(1, VT));
12449 return DCI.CombineTo(N, Res1, CarryOut);
12455 // fold (add Y, (sete X, 0)) -> adc 0, Y
12456 // (add Y, (setne X, 0)) -> sbb -1, Y
12457 // (sub (sete X, 0), Y) -> sbb 0, Y
12458 // (sub (setne X, 0), Y) -> adc -1, Y
12459 static SDValue OptimizeConditonalInDecrement(SDNode *N, SelectionDAG &DAG) {
12460 DebugLoc DL = N->getDebugLoc();
12462 // Look through ZExts.
12463 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
12464 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
12467 SDValue SetCC = Ext.getOperand(0);
12468 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
12471 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
12472 if (CC != X86::COND_E && CC != X86::COND_NE)
12475 SDValue Cmp = SetCC.getOperand(1);
12476 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
12477 !X86::isZeroNode(Cmp.getOperand(1)) ||
12478 !Cmp.getOperand(0).getValueType().isInteger())
12481 SDValue CmpOp0 = Cmp.getOperand(0);
12482 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
12483 DAG.getConstant(1, CmpOp0.getValueType()));
12485 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
12486 if (CC == X86::COND_NE)
12487 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
12488 DL, OtherVal.getValueType(), OtherVal,
12489 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
12490 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
12491 DL, OtherVal.getValueType(), OtherVal,
12492 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
12495 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
12496 DAGCombinerInfo &DCI) const {
12497 SelectionDAG &DAG = DCI.DAG;
12498 switch (N->getOpcode()) {
12500 case ISD::EXTRACT_VECTOR_ELT:
12501 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
12502 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
12503 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
12505 case ISD::SUB: return OptimizeConditonalInDecrement(N, DAG);
12506 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
12507 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
12510 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
12511 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
12512 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
12513 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
12514 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, this);
12516 case X86ISD::FOR: return PerformFORCombine(N, DAG);
12517 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
12518 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
12519 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
12520 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
12521 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG);
12522 case X86ISD::SHUFPS: // Handle all target specific shuffles
12523 case X86ISD::SHUFPD:
12524 case X86ISD::PALIGN:
12525 case X86ISD::PUNPCKHBW:
12526 case X86ISD::PUNPCKHWD:
12527 case X86ISD::PUNPCKHDQ:
12528 case X86ISD::PUNPCKHQDQ:
12529 case X86ISD::UNPCKHPS:
12530 case X86ISD::UNPCKHPD:
12531 case X86ISD::PUNPCKLBW:
12532 case X86ISD::PUNPCKLWD:
12533 case X86ISD::PUNPCKLDQ:
12534 case X86ISD::PUNPCKLQDQ:
12535 case X86ISD::UNPCKLPS:
12536 case X86ISD::UNPCKLPD:
12537 case X86ISD::VUNPCKLPS:
12538 case X86ISD::VUNPCKLPD:
12539 case X86ISD::VUNPCKLPSY:
12540 case X86ISD::VUNPCKLPDY:
12541 case X86ISD::MOVHLPS:
12542 case X86ISD::MOVLHPS:
12543 case X86ISD::PSHUFD:
12544 case X86ISD::PSHUFHW:
12545 case X86ISD::PSHUFLW:
12546 case X86ISD::MOVSS:
12547 case X86ISD::MOVSD:
12548 case X86ISD::VPERMIL:
12549 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI);
12555 /// isTypeDesirableForOp - Return true if the target has native support for
12556 /// the specified value type and it is 'desirable' to use the type for the
12557 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
12558 /// instruction encodings are longer and some i16 instructions are slow.
12559 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
12560 if (!isTypeLegal(VT))
12562 if (VT != MVT::i16)
12569 case ISD::SIGN_EXTEND:
12570 case ISD::ZERO_EXTEND:
12571 case ISD::ANY_EXTEND:
12584 /// IsDesirableToPromoteOp - This method query the target whether it is
12585 /// beneficial for dag combiner to promote the specified node. If true, it
12586 /// should return the desired promotion type by reference.
12587 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
12588 EVT VT = Op.getValueType();
12589 if (VT != MVT::i16)
12592 bool Promote = false;
12593 bool Commute = false;
12594 switch (Op.getOpcode()) {
12597 LoadSDNode *LD = cast<LoadSDNode>(Op);
12598 // If the non-extending load has a single use and it's not live out, then it
12599 // might be folded.
12600 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
12601 Op.hasOneUse()*/) {
12602 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12603 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
12604 // The only case where we'd want to promote LOAD (rather then it being
12605 // promoted as an operand is when it's only use is liveout.
12606 if (UI->getOpcode() != ISD::CopyToReg)
12613 case ISD::SIGN_EXTEND:
12614 case ISD::ZERO_EXTEND:
12615 case ISD::ANY_EXTEND:
12620 SDValue N0 = Op.getOperand(0);
12621 // Look out for (store (shl (load), x)).
12622 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
12635 SDValue N0 = Op.getOperand(0);
12636 SDValue N1 = Op.getOperand(1);
12637 if (!Commute && MayFoldLoad(N1))
12639 // Avoid disabling potential load folding opportunities.
12640 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
12642 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
12652 //===----------------------------------------------------------------------===//
12653 // X86 Inline Assembly Support
12654 //===----------------------------------------------------------------------===//
12656 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
12657 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
12659 std::string AsmStr = IA->getAsmString();
12661 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
12662 SmallVector<StringRef, 4> AsmPieces;
12663 SplitString(AsmStr, AsmPieces, ";\n");
12665 switch (AsmPieces.size()) {
12666 default: return false;
12668 AsmStr = AsmPieces[0];
12670 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
12672 // FIXME: this should verify that we are targeting a 486 or better. If not,
12673 // we will turn this bswap into something that will be lowered to logical ops
12674 // instead of emitting the bswap asm. For now, we don't support 486 or lower
12675 // so don't worry about this.
12677 if (AsmPieces.size() == 2 &&
12678 (AsmPieces[0] == "bswap" ||
12679 AsmPieces[0] == "bswapq" ||
12680 AsmPieces[0] == "bswapl") &&
12681 (AsmPieces[1] == "$0" ||
12682 AsmPieces[1] == "${0:q}")) {
12683 // No need to check constraints, nothing other than the equivalent of
12684 // "=r,0" would be valid here.
12685 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12686 if (!Ty || Ty->getBitWidth() % 16 != 0)
12688 return IntrinsicLowering::LowerToByteSwap(CI);
12690 // rorw $$8, ${0:w} --> llvm.bswap.i16
12691 if (CI->getType()->isIntegerTy(16) &&
12692 AsmPieces.size() == 3 &&
12693 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
12694 AsmPieces[1] == "$$8," &&
12695 AsmPieces[2] == "${0:w}" &&
12696 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
12698 const std::string &ConstraintsStr = IA->getConstraintString();
12699 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
12700 std::sort(AsmPieces.begin(), AsmPieces.end());
12701 if (AsmPieces.size() == 4 &&
12702 AsmPieces[0] == "~{cc}" &&
12703 AsmPieces[1] == "~{dirflag}" &&
12704 AsmPieces[2] == "~{flags}" &&
12705 AsmPieces[3] == "~{fpsr}") {
12706 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12707 if (!Ty || Ty->getBitWidth() % 16 != 0)
12709 return IntrinsicLowering::LowerToByteSwap(CI);
12714 if (CI->getType()->isIntegerTy(32) &&
12715 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
12716 SmallVector<StringRef, 4> Words;
12717 SplitString(AsmPieces[0], Words, " \t,");
12718 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
12719 Words[2] == "${0:w}") {
12721 SplitString(AsmPieces[1], Words, " \t,");
12722 if (Words.size() == 3 && Words[0] == "rorl" && Words[1] == "$$16" &&
12723 Words[2] == "$0") {
12725 SplitString(AsmPieces[2], Words, " \t,");
12726 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
12727 Words[2] == "${0:w}") {
12729 const std::string &ConstraintsStr = IA->getConstraintString();
12730 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
12731 std::sort(AsmPieces.begin(), AsmPieces.end());
12732 if (AsmPieces.size() == 4 &&
12733 AsmPieces[0] == "~{cc}" &&
12734 AsmPieces[1] == "~{dirflag}" &&
12735 AsmPieces[2] == "~{flags}" &&
12736 AsmPieces[3] == "~{fpsr}") {
12737 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12738 if (!Ty || Ty->getBitWidth() % 16 != 0)
12740 return IntrinsicLowering::LowerToByteSwap(CI);
12747 if (CI->getType()->isIntegerTy(64)) {
12748 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
12749 if (Constraints.size() >= 2 &&
12750 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
12751 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
12752 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
12753 SmallVector<StringRef, 4> Words;
12754 SplitString(AsmPieces[0], Words, " \t");
12755 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
12757 SplitString(AsmPieces[1], Words, " \t");
12758 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
12760 SplitString(AsmPieces[2], Words, " \t,");
12761 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
12762 Words[2] == "%edx") {
12763 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12764 if (!Ty || Ty->getBitWidth() % 16 != 0)
12766 return IntrinsicLowering::LowerToByteSwap(CI);
12779 /// getConstraintType - Given a constraint letter, return the type of
12780 /// constraint it is for this target.
12781 X86TargetLowering::ConstraintType
12782 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
12783 if (Constraint.size() == 1) {
12784 switch (Constraint[0]) {
12795 return C_RegisterClass;
12819 return TargetLowering::getConstraintType(Constraint);
12822 /// Examine constraint type and operand type and determine a weight value.
12823 /// This object must already have been set up with the operand type
12824 /// and the current alternative constraint selected.
12825 TargetLowering::ConstraintWeight
12826 X86TargetLowering::getSingleConstraintMatchWeight(
12827 AsmOperandInfo &info, const char *constraint) const {
12828 ConstraintWeight weight = CW_Invalid;
12829 Value *CallOperandVal = info.CallOperandVal;
12830 // If we don't have a value, we can't do a match,
12831 // but allow it at the lowest weight.
12832 if (CallOperandVal == NULL)
12834 Type *type = CallOperandVal->getType();
12835 // Look at the constraint type.
12836 switch (*constraint) {
12838 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
12849 if (CallOperandVal->getType()->isIntegerTy())
12850 weight = CW_SpecificReg;
12855 if (type->isFloatingPointTy())
12856 weight = CW_SpecificReg;
12859 if (type->isX86_MMXTy() && Subtarget->hasMMX())
12860 weight = CW_SpecificReg;
12864 if ((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasXMM())
12865 weight = CW_Register;
12868 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
12869 if (C->getZExtValue() <= 31)
12870 weight = CW_Constant;
12874 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12875 if (C->getZExtValue() <= 63)
12876 weight = CW_Constant;
12880 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12881 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
12882 weight = CW_Constant;
12886 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12887 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
12888 weight = CW_Constant;
12892 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12893 if (C->getZExtValue() <= 3)
12894 weight = CW_Constant;
12898 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12899 if (C->getZExtValue() <= 0xff)
12900 weight = CW_Constant;
12905 if (dyn_cast<ConstantFP>(CallOperandVal)) {
12906 weight = CW_Constant;
12910 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12911 if ((C->getSExtValue() >= -0x80000000LL) &&
12912 (C->getSExtValue() <= 0x7fffffffLL))
12913 weight = CW_Constant;
12917 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12918 if (C->getZExtValue() <= 0xffffffff)
12919 weight = CW_Constant;
12926 /// LowerXConstraint - try to replace an X constraint, which matches anything,
12927 /// with another that has more specific requirements based on the type of the
12928 /// corresponding operand.
12929 const char *X86TargetLowering::
12930 LowerXConstraint(EVT ConstraintVT) const {
12931 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
12932 // 'f' like normal targets.
12933 if (ConstraintVT.isFloatingPoint()) {
12934 if (Subtarget->hasXMMInt())
12936 if (Subtarget->hasXMM())
12940 return TargetLowering::LowerXConstraint(ConstraintVT);
12943 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
12944 /// vector. If it is invalid, don't add anything to Ops.
12945 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
12946 std::string &Constraint,
12947 std::vector<SDValue>&Ops,
12948 SelectionDAG &DAG) const {
12949 SDValue Result(0, 0);
12951 // Only support length 1 constraints for now.
12952 if (Constraint.length() > 1) return;
12954 char ConstraintLetter = Constraint[0];
12955 switch (ConstraintLetter) {
12958 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12959 if (C->getZExtValue() <= 31) {
12960 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12966 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12967 if (C->getZExtValue() <= 63) {
12968 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12974 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12975 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
12976 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12982 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12983 if (C->getZExtValue() <= 255) {
12984 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12990 // 32-bit signed value
12991 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12992 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
12993 C->getSExtValue())) {
12994 // Widen to 64 bits here to get it sign extended.
12995 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
12998 // FIXME gcc accepts some relocatable values here too, but only in certain
12999 // memory models; it's complicated.
13004 // 32-bit unsigned value
13005 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13006 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
13007 C->getZExtValue())) {
13008 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
13012 // FIXME gcc accepts some relocatable values here too, but only in certain
13013 // memory models; it's complicated.
13017 // Literal immediates are always ok.
13018 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
13019 // Widen to 64 bits here to get it sign extended.
13020 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
13024 // In any sort of PIC mode addresses need to be computed at runtime by
13025 // adding in a register or some sort of table lookup. These can't
13026 // be used as immediates.
13027 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
13030 // If we are in non-pic codegen mode, we allow the address of a global (with
13031 // an optional displacement) to be used with 'i'.
13032 GlobalAddressSDNode *GA = 0;
13033 int64_t Offset = 0;
13035 // Match either (GA), (GA+C), (GA+C1+C2), etc.
13037 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
13038 Offset += GA->getOffset();
13040 } else if (Op.getOpcode() == ISD::ADD) {
13041 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
13042 Offset += C->getZExtValue();
13043 Op = Op.getOperand(0);
13046 } else if (Op.getOpcode() == ISD::SUB) {
13047 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
13048 Offset += -C->getZExtValue();
13049 Op = Op.getOperand(0);
13054 // Otherwise, this isn't something we can handle, reject it.
13058 const GlobalValue *GV = GA->getGlobal();
13059 // If we require an extra load to get this address, as in PIC mode, we
13060 // can't accept it.
13061 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
13062 getTargetMachine())))
13065 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
13066 GA->getValueType(0), Offset);
13071 if (Result.getNode()) {
13072 Ops.push_back(Result);
13075 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
13078 std::pair<unsigned, const TargetRegisterClass*>
13079 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
13081 // First, see if this is a constraint that directly corresponds to an LLVM
13083 if (Constraint.size() == 1) {
13084 // GCC Constraint Letters
13085 switch (Constraint[0]) {
13087 // TODO: Slight differences here in allocation order and leaving
13088 // RIP in the class. Do they matter any more here than they do
13089 // in the normal allocation?
13090 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
13091 if (Subtarget->is64Bit()) {
13092 if (VT == MVT::i32 || VT == MVT::f32)
13093 return std::make_pair(0U, X86::GR32RegisterClass);
13094 else if (VT == MVT::i16)
13095 return std::make_pair(0U, X86::GR16RegisterClass);
13096 else if (VT == MVT::i8 || VT == MVT::i1)
13097 return std::make_pair(0U, X86::GR8RegisterClass);
13098 else if (VT == MVT::i64 || VT == MVT::f64)
13099 return std::make_pair(0U, X86::GR64RegisterClass);
13102 // 32-bit fallthrough
13103 case 'Q': // Q_REGS
13104 if (VT == MVT::i32 || VT == MVT::f32)
13105 return std::make_pair(0U, X86::GR32_ABCDRegisterClass);
13106 else if (VT == MVT::i16)
13107 return std::make_pair(0U, X86::GR16_ABCDRegisterClass);
13108 else if (VT == MVT::i8 || VT == MVT::i1)
13109 return std::make_pair(0U, X86::GR8_ABCD_LRegisterClass);
13110 else if (VT == MVT::i64)
13111 return std::make_pair(0U, X86::GR64_ABCDRegisterClass);
13113 case 'r': // GENERAL_REGS
13114 case 'l': // INDEX_REGS
13115 if (VT == MVT::i8 || VT == MVT::i1)
13116 return std::make_pair(0U, X86::GR8RegisterClass);
13117 if (VT == MVT::i16)
13118 return std::make_pair(0U, X86::GR16RegisterClass);
13119 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
13120 return std::make_pair(0U, X86::GR32RegisterClass);
13121 return std::make_pair(0U, X86::GR64RegisterClass);
13122 case 'R': // LEGACY_REGS
13123 if (VT == MVT::i8 || VT == MVT::i1)
13124 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
13125 if (VT == MVT::i16)
13126 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
13127 if (VT == MVT::i32 || !Subtarget->is64Bit())
13128 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
13129 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
13130 case 'f': // FP Stack registers.
13131 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
13132 // value to the correct fpstack register class.
13133 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
13134 return std::make_pair(0U, X86::RFP32RegisterClass);
13135 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
13136 return std::make_pair(0U, X86::RFP64RegisterClass);
13137 return std::make_pair(0U, X86::RFP80RegisterClass);
13138 case 'y': // MMX_REGS if MMX allowed.
13139 if (!Subtarget->hasMMX()) break;
13140 return std::make_pair(0U, X86::VR64RegisterClass);
13141 case 'Y': // SSE_REGS if SSE2 allowed
13142 if (!Subtarget->hasXMMInt()) break;
13144 case 'x': // SSE_REGS if SSE1 allowed
13145 if (!Subtarget->hasXMM()) break;
13147 switch (VT.getSimpleVT().SimpleTy) {
13149 // Scalar SSE types.
13152 return std::make_pair(0U, X86::FR32RegisterClass);
13155 return std::make_pair(0U, X86::FR64RegisterClass);
13163 return std::make_pair(0U, X86::VR128RegisterClass);
13169 // Use the default implementation in TargetLowering to convert the register
13170 // constraint into a member of a register class.
13171 std::pair<unsigned, const TargetRegisterClass*> Res;
13172 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
13174 // Not found as a standard register?
13175 if (Res.second == 0) {
13176 // Map st(0) -> st(7) -> ST0
13177 if (Constraint.size() == 7 && Constraint[0] == '{' &&
13178 tolower(Constraint[1]) == 's' &&
13179 tolower(Constraint[2]) == 't' &&
13180 Constraint[3] == '(' &&
13181 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
13182 Constraint[5] == ')' &&
13183 Constraint[6] == '}') {
13185 Res.first = X86::ST0+Constraint[4]-'0';
13186 Res.second = X86::RFP80RegisterClass;
13190 // GCC allows "st(0)" to be called just plain "st".
13191 if (StringRef("{st}").equals_lower(Constraint)) {
13192 Res.first = X86::ST0;
13193 Res.second = X86::RFP80RegisterClass;
13198 if (StringRef("{flags}").equals_lower(Constraint)) {
13199 Res.first = X86::EFLAGS;
13200 Res.second = X86::CCRRegisterClass;
13204 // 'A' means EAX + EDX.
13205 if (Constraint == "A") {
13206 Res.first = X86::EAX;
13207 Res.second = X86::GR32_ADRegisterClass;
13213 // Otherwise, check to see if this is a register class of the wrong value
13214 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
13215 // turn into {ax},{dx}.
13216 if (Res.second->hasType(VT))
13217 return Res; // Correct type already, nothing to do.
13219 // All of the single-register GCC register classes map their values onto
13220 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
13221 // really want an 8-bit or 32-bit register, map to the appropriate register
13222 // class and return the appropriate register.
13223 if (Res.second == X86::GR16RegisterClass) {
13224 if (VT == MVT::i8) {
13225 unsigned DestReg = 0;
13226 switch (Res.first) {
13228 case X86::AX: DestReg = X86::AL; break;
13229 case X86::DX: DestReg = X86::DL; break;
13230 case X86::CX: DestReg = X86::CL; break;
13231 case X86::BX: DestReg = X86::BL; break;
13234 Res.first = DestReg;
13235 Res.second = X86::GR8RegisterClass;
13237 } else if (VT == MVT::i32) {
13238 unsigned DestReg = 0;
13239 switch (Res.first) {
13241 case X86::AX: DestReg = X86::EAX; break;
13242 case X86::DX: DestReg = X86::EDX; break;
13243 case X86::CX: DestReg = X86::ECX; break;
13244 case X86::BX: DestReg = X86::EBX; break;
13245 case X86::SI: DestReg = X86::ESI; break;
13246 case X86::DI: DestReg = X86::EDI; break;
13247 case X86::BP: DestReg = X86::EBP; break;
13248 case X86::SP: DestReg = X86::ESP; break;
13251 Res.first = DestReg;
13252 Res.second = X86::GR32RegisterClass;
13254 } else if (VT == MVT::i64) {
13255 unsigned DestReg = 0;
13256 switch (Res.first) {
13258 case X86::AX: DestReg = X86::RAX; break;
13259 case X86::DX: DestReg = X86::RDX; break;
13260 case X86::CX: DestReg = X86::RCX; break;
13261 case X86::BX: DestReg = X86::RBX; break;
13262 case X86::SI: DestReg = X86::RSI; break;
13263 case X86::DI: DestReg = X86::RDI; break;
13264 case X86::BP: DestReg = X86::RBP; break;
13265 case X86::SP: DestReg = X86::RSP; break;
13268 Res.first = DestReg;
13269 Res.second = X86::GR64RegisterClass;
13272 } else if (Res.second == X86::FR32RegisterClass ||
13273 Res.second == X86::FR64RegisterClass ||
13274 Res.second == X86::VR128RegisterClass) {
13275 // Handle references to XMM physical registers that got mapped into the
13276 // wrong class. This can happen with constraints like {xmm0} where the
13277 // target independent register mapper will just pick the first match it can
13278 // find, ignoring the required type.
13279 if (VT == MVT::f32)
13280 Res.second = X86::FR32RegisterClass;
13281 else if (VT == MVT::f64)
13282 Res.second = X86::FR64RegisterClass;
13283 else if (X86::VR128RegisterClass->hasType(VT))
13284 Res.second = X86::VR128RegisterClass;