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 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
75 /// sets things up to match to an AVX VEXTRACTF128 instruction or a
76 /// simple subregister reference. Idx is an index in the 128 bits we
77 /// want. It need not be aligned to a 128-bit bounday. That makes
78 /// lowering EXTRACT_VECTOR_ELT operations easier.
79 static SDValue Extract128BitVector(SDValue Vec,
83 EVT VT = Vec.getValueType();
84 assert(VT.getSizeInBits() == 256 && "Unexpected vector size!");
85 EVT ElVT = VT.getVectorElementType();
86 int Factor = VT.getSizeInBits()/128;
87 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
88 VT.getVectorNumElements()/Factor);
90 // Extract from UNDEF is UNDEF.
91 if (Vec.getOpcode() == ISD::UNDEF)
92 return DAG.getNode(ISD::UNDEF, dl, ResultVT);
94 if (isa<ConstantSDNode>(Idx)) {
95 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
97 // Extract the relevant 128 bits. Generate an EXTRACT_SUBVECTOR
98 // we can match to VEXTRACTF128.
99 unsigned ElemsPerChunk = 128 / ElVT.getSizeInBits();
101 // This is the index of the first element of the 128-bit chunk
103 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / 128)
106 SDValue VecIdx = DAG.getConstant(NormalizedIdxVal, MVT::i32);
107 SDValue Result = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec,
116 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
117 /// sets things up to match to an AVX VINSERTF128 instruction or a
118 /// simple superregister reference. Idx is an index in the 128 bits
119 /// we want. It need not be aligned to a 128-bit bounday. That makes
120 /// lowering INSERT_VECTOR_ELT operations easier.
121 static SDValue Insert128BitVector(SDValue Result,
126 if (isa<ConstantSDNode>(Idx)) {
127 EVT VT = Vec.getValueType();
128 assert(VT.getSizeInBits() == 128 && "Unexpected vector size!");
130 EVT ElVT = VT.getVectorElementType();
131 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
132 EVT ResultVT = Result.getValueType();
134 // Insert the relevant 128 bits.
135 unsigned ElemsPerChunk = 128/ElVT.getSizeInBits();
137 // This is the index of the first element of the 128-bit chunk
139 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/128)
142 SDValue VecIdx = DAG.getConstant(NormalizedIdxVal, MVT::i32);
143 Result = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec,
151 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
152 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
153 bool is64Bit = Subtarget->is64Bit();
155 if (Subtarget->isTargetEnvMacho()) {
157 return new X8664_MachoTargetObjectFile();
158 return new TargetLoweringObjectFileMachO();
161 if (Subtarget->isTargetELF())
162 return new TargetLoweringObjectFileELF();
163 if (Subtarget->isTargetCOFF() && !Subtarget->isTargetEnvMacho())
164 return new TargetLoweringObjectFileCOFF();
165 llvm_unreachable("unknown subtarget type");
168 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
169 : TargetLowering(TM, createTLOF(TM)) {
170 Subtarget = &TM.getSubtarget<X86Subtarget>();
171 X86ScalarSSEf64 = Subtarget->hasXMMInt() || Subtarget->hasAVX();
172 X86ScalarSSEf32 = Subtarget->hasXMM() || Subtarget->hasAVX();
173 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
175 RegInfo = TM.getRegisterInfo();
176 TD = getTargetData();
178 // Set up the TargetLowering object.
179 static MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
181 // X86 is weird, it always uses i8 for shift amounts and setcc results.
182 setBooleanContents(ZeroOrOneBooleanContent);
184 // For 64-bit since we have so many registers use the ILP scheduler, for
185 // 32-bit code use the register pressure specific scheduling.
186 if (Subtarget->is64Bit())
187 setSchedulingPreference(Sched::ILP);
189 setSchedulingPreference(Sched::RegPressure);
190 setStackPointerRegisterToSaveRestore(X86StackPtr);
192 if (Subtarget->isTargetWindows() && !Subtarget->isTargetCygMing()) {
193 // Setup Windows compiler runtime calls.
194 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
195 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
196 setLibcallName(RTLIB::SREM_I64, "_allrem");
197 setLibcallName(RTLIB::UREM_I64, "_aullrem");
198 setLibcallName(RTLIB::MUL_I64, "_allmul");
199 setLibcallName(RTLIB::FPTOUINT_F64_I64, "_ftol2");
200 setLibcallName(RTLIB::FPTOUINT_F32_I64, "_ftol2");
201 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
202 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
203 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
204 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
205 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
206 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::C);
207 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::C);
210 if (Subtarget->isTargetDarwin()) {
211 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
212 setUseUnderscoreSetJmp(false);
213 setUseUnderscoreLongJmp(false);
214 } else if (Subtarget->isTargetMingw()) {
215 // MS runtime is weird: it exports _setjmp, but longjmp!
216 setUseUnderscoreSetJmp(true);
217 setUseUnderscoreLongJmp(false);
219 setUseUnderscoreSetJmp(true);
220 setUseUnderscoreLongJmp(true);
223 // Set up the register classes.
224 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
225 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
226 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
227 if (Subtarget->is64Bit())
228 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
230 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
232 // We don't accept any truncstore of integer registers.
233 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
234 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
235 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
236 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
237 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
238 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
240 // SETOEQ and SETUNE require checking two conditions.
241 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
242 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
243 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
244 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
245 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
246 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
248 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
250 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
251 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
252 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
254 if (Subtarget->is64Bit()) {
255 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
256 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
257 } else if (!UseSoftFloat) {
258 // We have an algorithm for SSE2->double, and we turn this into a
259 // 64-bit FILD followed by conditional FADD for other targets.
260 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
261 // We have an algorithm for SSE2, and we turn this into a 64-bit
262 // FILD for other targets.
263 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
266 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
268 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
269 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
272 // SSE has no i16 to fp conversion, only i32
273 if (X86ScalarSSEf32) {
274 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
275 // f32 and f64 cases are Legal, f80 case is not
276 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
278 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
279 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
282 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
283 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
286 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
287 // are Legal, f80 is custom lowered.
288 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
289 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
291 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
293 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
294 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
296 if (X86ScalarSSEf32) {
297 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
298 // f32 and f64 cases are Legal, f80 case is not
299 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
301 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
302 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
305 // Handle FP_TO_UINT by promoting the destination to a larger signed
307 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
308 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
309 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
311 if (Subtarget->is64Bit()) {
312 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
313 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
314 } else if (!UseSoftFloat) {
315 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
316 // Expand FP_TO_UINT into a select.
317 // FIXME: We would like to use a Custom expander here eventually to do
318 // the optimal thing for SSE vs. the default expansion in the legalizer.
319 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
321 // With SSE3 we can use fisttpll to convert to a signed i64; without
322 // SSE, we're stuck with a fistpll.
323 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
326 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
327 if (!X86ScalarSSEf64) {
328 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
329 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
330 if (Subtarget->is64Bit()) {
331 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
332 // Without SSE, i64->f64 goes through memory.
333 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
337 // Scalar integer divide and remainder are lowered to use operations that
338 // produce two results, to match the available instructions. This exposes
339 // the two-result form to trivial CSE, which is able to combine x/y and x%y
340 // into a single instruction.
342 // Scalar integer multiply-high is also lowered to use two-result
343 // operations, to match the available instructions. However, plain multiply
344 // (low) operations are left as Legal, as there are single-result
345 // instructions for this in x86. Using the two-result multiply instructions
346 // when both high and low results are needed must be arranged by dagcombine.
347 for (unsigned i = 0, e = 4; i != e; ++i) {
349 setOperationAction(ISD::MULHS, VT, Expand);
350 setOperationAction(ISD::MULHU, VT, Expand);
351 setOperationAction(ISD::SDIV, VT, Expand);
352 setOperationAction(ISD::UDIV, VT, Expand);
353 setOperationAction(ISD::SREM, VT, Expand);
354 setOperationAction(ISD::UREM, VT, Expand);
356 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
357 setOperationAction(ISD::ADDC, VT, Custom);
358 setOperationAction(ISD::ADDE, VT, Custom);
359 setOperationAction(ISD::SUBC, VT, Custom);
360 setOperationAction(ISD::SUBE, VT, Custom);
363 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
364 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
365 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
366 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
367 if (Subtarget->is64Bit())
368 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
369 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
370 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
371 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
372 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
373 setOperationAction(ISD::FREM , MVT::f32 , Expand);
374 setOperationAction(ISD::FREM , MVT::f64 , Expand);
375 setOperationAction(ISD::FREM , MVT::f80 , Expand);
376 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
378 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
379 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
380 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
381 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
382 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
383 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
384 if (Subtarget->is64Bit()) {
385 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
386 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
389 if (Subtarget->hasPOPCNT()) {
390 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
392 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
393 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
394 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
395 if (Subtarget->is64Bit())
396 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
399 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
400 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
402 // These should be promoted to a larger select which is supported.
403 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
404 // X86 wants to expand cmov itself.
405 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
406 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
407 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
408 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
409 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
410 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
411 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
412 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
413 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
414 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
415 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
416 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
417 if (Subtarget->is64Bit()) {
418 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
419 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
421 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
424 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
425 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
426 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
427 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
428 if (Subtarget->is64Bit())
429 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
430 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
431 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
432 if (Subtarget->is64Bit()) {
433 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
434 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
435 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
436 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
437 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
439 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
440 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
441 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
442 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
443 if (Subtarget->is64Bit()) {
444 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
445 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
446 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
449 if (Subtarget->hasXMM())
450 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
452 setOperationAction(ISD::MEMBARRIER , MVT::Other, Custom);
453 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
455 // On X86 and X86-64, atomic operations are lowered to locked instructions.
456 // Locked instructions, in turn, have implicit fence semantics (all memory
457 // operations are flushed before issuing the locked instruction, and they
458 // are not buffered), so we can fold away the common pattern of
459 // fence-atomic-fence.
460 setShouldFoldAtomicFences(true);
462 // Expand certain atomics
463 for (unsigned i = 0, e = 4; i != e; ++i) {
465 setOperationAction(ISD::ATOMIC_CMP_SWAP, VT, Custom);
466 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
469 if (!Subtarget->is64Bit()) {
470 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
471 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
472 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
473 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
474 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
475 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
476 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
479 // FIXME - use subtarget debug flags
480 if (!Subtarget->isTargetDarwin() &&
481 !Subtarget->isTargetELF() &&
482 !Subtarget->isTargetCygMing()) {
483 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
486 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
487 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
488 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
489 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
490 if (Subtarget->is64Bit()) {
491 setExceptionPointerRegister(X86::RAX);
492 setExceptionSelectorRegister(X86::RDX);
494 setExceptionPointerRegister(X86::EAX);
495 setExceptionSelectorRegister(X86::EDX);
497 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
498 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
500 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
502 setOperationAction(ISD::TRAP, MVT::Other, Legal);
504 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
505 setOperationAction(ISD::VASTART , MVT::Other, Custom);
506 setOperationAction(ISD::VAEND , MVT::Other, Expand);
507 if (Subtarget->is64Bit()) {
508 setOperationAction(ISD::VAARG , MVT::Other, Custom);
509 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
511 setOperationAction(ISD::VAARG , MVT::Other, Expand);
512 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
515 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
516 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
517 setOperationAction(ISD::DYNAMIC_STACKALLOC,
518 (Subtarget->is64Bit() ? MVT::i64 : MVT::i32),
519 (Subtarget->isTargetCOFF()
520 && !Subtarget->isTargetEnvMacho()
523 if (!UseSoftFloat && X86ScalarSSEf64) {
524 // f32 and f64 use SSE.
525 // Set up the FP register classes.
526 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
527 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
529 // Use ANDPD to simulate FABS.
530 setOperationAction(ISD::FABS , MVT::f64, Custom);
531 setOperationAction(ISD::FABS , MVT::f32, Custom);
533 // Use XORP to simulate FNEG.
534 setOperationAction(ISD::FNEG , MVT::f64, Custom);
535 setOperationAction(ISD::FNEG , MVT::f32, Custom);
537 // Use ANDPD and ORPD to simulate FCOPYSIGN.
538 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
539 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
541 // Lower this to FGETSIGNx86 plus an AND.
542 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
543 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
545 // We don't support sin/cos/fmod
546 setOperationAction(ISD::FSIN , MVT::f64, Expand);
547 setOperationAction(ISD::FCOS , MVT::f64, Expand);
548 setOperationAction(ISD::FSIN , MVT::f32, Expand);
549 setOperationAction(ISD::FCOS , MVT::f32, Expand);
551 // Expand FP immediates into loads from the stack, except for the special
553 addLegalFPImmediate(APFloat(+0.0)); // xorpd
554 addLegalFPImmediate(APFloat(+0.0f)); // xorps
555 } else if (!UseSoftFloat && X86ScalarSSEf32) {
556 // Use SSE for f32, x87 for f64.
557 // Set up the FP register classes.
558 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
559 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
561 // Use ANDPS to simulate FABS.
562 setOperationAction(ISD::FABS , MVT::f32, Custom);
564 // Use XORP to simulate FNEG.
565 setOperationAction(ISD::FNEG , MVT::f32, Custom);
567 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
569 // Use ANDPS and ORPS to simulate FCOPYSIGN.
570 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
571 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
573 // We don't support sin/cos/fmod
574 setOperationAction(ISD::FSIN , MVT::f32, Expand);
575 setOperationAction(ISD::FCOS , MVT::f32, Expand);
577 // Special cases we handle for FP constants.
578 addLegalFPImmediate(APFloat(+0.0f)); // xorps
579 addLegalFPImmediate(APFloat(+0.0)); // FLD0
580 addLegalFPImmediate(APFloat(+1.0)); // FLD1
581 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
582 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
585 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
586 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
588 } else if (!UseSoftFloat) {
589 // f32 and f64 in x87.
590 // Set up the FP register classes.
591 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
592 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
594 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
595 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
596 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
597 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
600 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
601 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
603 addLegalFPImmediate(APFloat(+0.0)); // FLD0
604 addLegalFPImmediate(APFloat(+1.0)); // FLD1
605 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
606 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
607 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
608 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
609 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
610 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
613 // We don't support FMA.
614 setOperationAction(ISD::FMA, MVT::f64, Expand);
615 setOperationAction(ISD::FMA, MVT::f32, Expand);
617 // Long double always uses X87.
619 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
620 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
621 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
623 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
624 addLegalFPImmediate(TmpFlt); // FLD0
626 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
629 APFloat TmpFlt2(+1.0);
630 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
632 addLegalFPImmediate(TmpFlt2); // FLD1
633 TmpFlt2.changeSign();
634 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
638 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
639 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
642 setOperationAction(ISD::FMA, MVT::f80, Expand);
645 // Always use a library call for pow.
646 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
647 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
648 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
650 setOperationAction(ISD::FLOG, MVT::f80, Expand);
651 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
652 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
653 setOperationAction(ISD::FEXP, MVT::f80, Expand);
654 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
656 // First set operation action for all vector types to either promote
657 // (for widening) or expand (for scalarization). Then we will selectively
658 // turn on ones that can be effectively codegen'd.
659 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
660 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
661 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
662 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
663 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
664 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
665 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
666 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
667 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
668 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
669 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
670 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
671 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
672 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
673 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
674 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
675 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
676 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
677 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
678 setOperationAction(ISD::INSERT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
679 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
680 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
681 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
682 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
683 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
684 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
685 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
686 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
687 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
688 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
689 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
690 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
691 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
692 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
693 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
694 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
695 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
696 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
697 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
698 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
699 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
700 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
701 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
702 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
703 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
704 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
705 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
706 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
707 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
708 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
709 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
710 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
711 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
712 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
713 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
714 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
715 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
716 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
717 setTruncStoreAction((MVT::SimpleValueType)VT,
718 (MVT::SimpleValueType)InnerVT, Expand);
719 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
720 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
721 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
724 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
725 // with -msoft-float, disable use of MMX as well.
726 if (!UseSoftFloat && Subtarget->hasMMX()) {
727 addRegisterClass(MVT::x86mmx, X86::VR64RegisterClass);
728 // No operations on x86mmx supported, everything uses intrinsics.
731 // MMX-sized vectors (other than x86mmx) are expected to be expanded
732 // into smaller operations.
733 setOperationAction(ISD::MULHS, MVT::v8i8, Expand);
734 setOperationAction(ISD::MULHS, MVT::v4i16, Expand);
735 setOperationAction(ISD::MULHS, MVT::v2i32, Expand);
736 setOperationAction(ISD::MULHS, MVT::v1i64, Expand);
737 setOperationAction(ISD::AND, MVT::v8i8, Expand);
738 setOperationAction(ISD::AND, MVT::v4i16, Expand);
739 setOperationAction(ISD::AND, MVT::v2i32, Expand);
740 setOperationAction(ISD::AND, MVT::v1i64, Expand);
741 setOperationAction(ISD::OR, MVT::v8i8, Expand);
742 setOperationAction(ISD::OR, MVT::v4i16, Expand);
743 setOperationAction(ISD::OR, MVT::v2i32, Expand);
744 setOperationAction(ISD::OR, MVT::v1i64, Expand);
745 setOperationAction(ISD::XOR, MVT::v8i8, Expand);
746 setOperationAction(ISD::XOR, MVT::v4i16, Expand);
747 setOperationAction(ISD::XOR, MVT::v2i32, Expand);
748 setOperationAction(ISD::XOR, MVT::v1i64, Expand);
749 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Expand);
750 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Expand);
751 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Expand);
752 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Expand);
753 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
754 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
755 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
756 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
757 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
758 setOperationAction(ISD::BITCAST, MVT::v8i8, Expand);
759 setOperationAction(ISD::BITCAST, MVT::v4i16, Expand);
760 setOperationAction(ISD::BITCAST, MVT::v2i32, Expand);
761 setOperationAction(ISD::BITCAST, MVT::v1i64, Expand);
763 if (!UseSoftFloat && Subtarget->hasXMM()) {
764 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
766 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
767 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
768 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
769 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
770 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
771 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
772 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
773 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
774 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
775 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
776 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
777 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
780 if (!UseSoftFloat && Subtarget->hasXMMInt()) {
781 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
783 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
784 // registers cannot be used even for integer operations.
785 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
786 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
787 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
788 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
790 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
791 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
792 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
793 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
794 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
795 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
796 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
797 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
798 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
799 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
800 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
801 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
802 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
803 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
804 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
805 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
807 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
808 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
809 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
810 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
812 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
813 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
814 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
815 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
816 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
818 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
819 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
820 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
821 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
822 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
824 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
825 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
826 EVT VT = (MVT::SimpleValueType)i;
827 // Do not attempt to custom lower non-power-of-2 vectors
828 if (!isPowerOf2_32(VT.getVectorNumElements()))
830 // Do not attempt to custom lower non-128-bit vectors
831 if (!VT.is128BitVector())
833 setOperationAction(ISD::BUILD_VECTOR,
834 VT.getSimpleVT().SimpleTy, Custom);
835 setOperationAction(ISD::VECTOR_SHUFFLE,
836 VT.getSimpleVT().SimpleTy, Custom);
837 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
838 VT.getSimpleVT().SimpleTy, Custom);
841 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
842 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
843 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
844 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
845 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
846 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
848 if (Subtarget->is64Bit()) {
849 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
850 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
853 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
854 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
855 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
858 // Do not attempt to promote non-128-bit vectors
859 if (!VT.is128BitVector())
862 setOperationAction(ISD::AND, SVT, Promote);
863 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
864 setOperationAction(ISD::OR, SVT, Promote);
865 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
866 setOperationAction(ISD::XOR, SVT, Promote);
867 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
868 setOperationAction(ISD::LOAD, SVT, Promote);
869 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
870 setOperationAction(ISD::SELECT, SVT, Promote);
871 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
874 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
876 // Custom lower v2i64 and v2f64 selects.
877 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
878 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
879 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
880 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
882 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
883 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
886 if (Subtarget->hasSSE41() || Subtarget->hasAVX()) {
887 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
888 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
889 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
890 setOperationAction(ISD::FRINT, MVT::f32, Legal);
891 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
892 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
893 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
894 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
895 setOperationAction(ISD::FRINT, MVT::f64, Legal);
896 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
898 // FIXME: Do we need to handle scalar-to-vector here?
899 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
901 // Can turn SHL into an integer multiply.
902 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
903 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
905 // i8 and i16 vectors are custom , because the source register and source
906 // source memory operand types are not the same width. f32 vectors are
907 // custom since the immediate controlling the insert encodes additional
909 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
910 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
911 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
912 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
914 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
915 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
916 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
917 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
919 if (Subtarget->is64Bit()) {
920 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
921 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
925 if (Subtarget->hasSSE2() || Subtarget->hasAVX()) {
926 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
927 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
928 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
929 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
931 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
932 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
933 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
935 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
936 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
939 if (Subtarget->hasSSE42() || Subtarget->hasAVX())
940 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
942 if (!UseSoftFloat && Subtarget->hasAVX()) {
943 addRegisterClass(MVT::v32i8, X86::VR256RegisterClass);
944 addRegisterClass(MVT::v16i16, X86::VR256RegisterClass);
945 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
946 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
947 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
948 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
950 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
951 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
952 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
954 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
955 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
956 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
957 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
958 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
959 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
961 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
962 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
963 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
964 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
965 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
966 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
968 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
969 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
970 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
972 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f64, Custom);
973 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i64, Custom);
974 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
975 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
976 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i8, Custom);
977 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i16, Custom);
979 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
980 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
981 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
982 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
984 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
985 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
986 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
987 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
989 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
990 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
992 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
993 setOperationAction(ISD::VSETCC, MVT::v4i64, Custom);
995 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
996 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
997 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
999 // Custom lower several nodes for 256-bit types.
1000 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1001 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
1002 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
1005 // Extract subvector is special because the value type
1006 // (result) is 128-bit but the source is 256-bit wide.
1007 if (VT.is128BitVector())
1008 setOperationAction(ISD::EXTRACT_SUBVECTOR, SVT, Custom);
1010 // Do not attempt to custom lower other non-256-bit vectors
1011 if (!VT.is256BitVector())
1014 setOperationAction(ISD::BUILD_VECTOR, SVT, Custom);
1015 setOperationAction(ISD::VECTOR_SHUFFLE, SVT, Custom);
1016 setOperationAction(ISD::INSERT_VECTOR_ELT, SVT, Custom);
1017 setOperationAction(ISD::EXTRACT_VECTOR_ELT, SVT, Custom);
1018 setOperationAction(ISD::SCALAR_TO_VECTOR, SVT, Custom);
1019 setOperationAction(ISD::INSERT_SUBVECTOR, SVT, Custom);
1022 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1023 for (unsigned i = (unsigned)MVT::v32i8; i != (unsigned)MVT::v4i64; ++i) {
1024 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
1027 // Do not attempt to promote non-256-bit vectors
1028 if (!VT.is256BitVector())
1031 setOperationAction(ISD::AND, SVT, Promote);
1032 AddPromotedToType (ISD::AND, SVT, MVT::v4i64);
1033 setOperationAction(ISD::OR, SVT, Promote);
1034 AddPromotedToType (ISD::OR, SVT, MVT::v4i64);
1035 setOperationAction(ISD::XOR, SVT, Promote);
1036 AddPromotedToType (ISD::XOR, SVT, MVT::v4i64);
1037 setOperationAction(ISD::LOAD, SVT, Promote);
1038 AddPromotedToType (ISD::LOAD, SVT, MVT::v4i64);
1039 setOperationAction(ISD::SELECT, SVT, Promote);
1040 AddPromotedToType (ISD::SELECT, SVT, MVT::v4i64);
1044 // SIGN_EXTEND_INREGs are evaluated by the extend type. Handle the expansion
1045 // of this type with custom code.
1046 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1047 VT != (unsigned)MVT::LAST_VECTOR_VALUETYPE; VT++) {
1048 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT, Custom);
1051 // We want to custom lower some of our intrinsics.
1052 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1055 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1056 // handle type legalization for these operations here.
1058 // FIXME: We really should do custom legalization for addition and
1059 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1060 // than generic legalization for 64-bit multiplication-with-overflow, though.
1061 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1062 // Add/Sub/Mul with overflow operations are custom lowered.
1064 setOperationAction(ISD::SADDO, VT, Custom);
1065 setOperationAction(ISD::UADDO, VT, Custom);
1066 setOperationAction(ISD::SSUBO, VT, Custom);
1067 setOperationAction(ISD::USUBO, VT, Custom);
1068 setOperationAction(ISD::SMULO, VT, Custom);
1069 setOperationAction(ISD::UMULO, VT, Custom);
1072 // There are no 8-bit 3-address imul/mul instructions
1073 setOperationAction(ISD::SMULO, MVT::i8, Expand);
1074 setOperationAction(ISD::UMULO, MVT::i8, Expand);
1076 if (!Subtarget->is64Bit()) {
1077 // These libcalls are not available in 32-bit.
1078 setLibcallName(RTLIB::SHL_I128, 0);
1079 setLibcallName(RTLIB::SRL_I128, 0);
1080 setLibcallName(RTLIB::SRA_I128, 0);
1083 // We have target-specific dag combine patterns for the following nodes:
1084 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1085 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1086 setTargetDAGCombine(ISD::BUILD_VECTOR);
1087 setTargetDAGCombine(ISD::SELECT);
1088 setTargetDAGCombine(ISD::SHL);
1089 setTargetDAGCombine(ISD::SRA);
1090 setTargetDAGCombine(ISD::SRL);
1091 setTargetDAGCombine(ISD::OR);
1092 setTargetDAGCombine(ISD::AND);
1093 setTargetDAGCombine(ISD::ADD);
1094 setTargetDAGCombine(ISD::SUB);
1095 setTargetDAGCombine(ISD::STORE);
1096 setTargetDAGCombine(ISD::ZERO_EXTEND);
1097 setTargetDAGCombine(ISD::SINT_TO_FP);
1098 if (Subtarget->is64Bit())
1099 setTargetDAGCombine(ISD::MUL);
1101 computeRegisterProperties();
1103 // On Darwin, -Os means optimize for size without hurting performance,
1104 // do not reduce the limit.
1105 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1106 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1107 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1108 maxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1109 maxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1110 maxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1111 setPrefLoopAlignment(16);
1112 benefitFromCodePlacementOpt = true;
1114 setPrefFunctionAlignment(4);
1118 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1123 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1124 /// the desired ByVal argument alignment.
1125 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1128 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1129 if (VTy->getBitWidth() == 128)
1131 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1132 unsigned EltAlign = 0;
1133 getMaxByValAlign(ATy->getElementType(), EltAlign);
1134 if (EltAlign > MaxAlign)
1135 MaxAlign = EltAlign;
1136 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1137 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1138 unsigned EltAlign = 0;
1139 getMaxByValAlign(STy->getElementType(i), EltAlign);
1140 if (EltAlign > MaxAlign)
1141 MaxAlign = EltAlign;
1149 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1150 /// function arguments in the caller parameter area. For X86, aggregates
1151 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1152 /// are at 4-byte boundaries.
1153 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1154 if (Subtarget->is64Bit()) {
1155 // Max of 8 and alignment of type.
1156 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1163 if (Subtarget->hasXMM())
1164 getMaxByValAlign(Ty, Align);
1168 /// getOptimalMemOpType - Returns the target specific optimal type for load
1169 /// and store operations as a result of memset, memcpy, and memmove
1170 /// lowering. If DstAlign is zero that means it's safe to destination
1171 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1172 /// means there isn't a need to check it against alignment requirement,
1173 /// probably because the source does not need to be loaded. If
1174 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1175 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1176 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1177 /// constant so it does not need to be loaded.
1178 /// It returns EVT::Other if the type should be determined using generic
1179 /// target-independent logic.
1181 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1182 unsigned DstAlign, unsigned SrcAlign,
1183 bool NonScalarIntSafe,
1185 MachineFunction &MF) const {
1186 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1187 // linux. This is because the stack realignment code can't handle certain
1188 // cases like PR2962. This should be removed when PR2962 is fixed.
1189 const Function *F = MF.getFunction();
1190 if (NonScalarIntSafe &&
1191 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1193 (Subtarget->isUnalignedMemAccessFast() ||
1194 ((DstAlign == 0 || DstAlign >= 16) &&
1195 (SrcAlign == 0 || SrcAlign >= 16))) &&
1196 Subtarget->getStackAlignment() >= 16) {
1197 if (Subtarget->hasSSE2())
1199 if (Subtarget->hasSSE1())
1201 } else if (!MemcpyStrSrc && Size >= 8 &&
1202 !Subtarget->is64Bit() &&
1203 Subtarget->getStackAlignment() >= 8 &&
1204 Subtarget->hasXMMInt()) {
1205 // Do not use f64 to lower memcpy if source is string constant. It's
1206 // better to use i32 to avoid the loads.
1210 if (Subtarget->is64Bit() && Size >= 8)
1215 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1216 /// current function. The returned value is a member of the
1217 /// MachineJumpTableInfo::JTEntryKind enum.
1218 unsigned X86TargetLowering::getJumpTableEncoding() const {
1219 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1221 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1222 Subtarget->isPICStyleGOT())
1223 return MachineJumpTableInfo::EK_Custom32;
1225 // Otherwise, use the normal jump table encoding heuristics.
1226 return TargetLowering::getJumpTableEncoding();
1230 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1231 const MachineBasicBlock *MBB,
1232 unsigned uid,MCContext &Ctx) const{
1233 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1234 Subtarget->isPICStyleGOT());
1235 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1237 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1238 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1241 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1243 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1244 SelectionDAG &DAG) const {
1245 if (!Subtarget->is64Bit())
1246 // This doesn't have DebugLoc associated with it, but is not really the
1247 // same as a Register.
1248 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1252 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1253 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1255 const MCExpr *X86TargetLowering::
1256 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1257 MCContext &Ctx) const {
1258 // X86-64 uses RIP relative addressing based on the jump table label.
1259 if (Subtarget->isPICStyleRIPRel())
1260 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1262 // Otherwise, the reference is relative to the PIC base.
1263 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1266 // FIXME: Why this routine is here? Move to RegInfo!
1267 std::pair<const TargetRegisterClass*, uint8_t>
1268 X86TargetLowering::findRepresentativeClass(EVT VT) const{
1269 const TargetRegisterClass *RRC = 0;
1271 switch (VT.getSimpleVT().SimpleTy) {
1273 return TargetLowering::findRepresentativeClass(VT);
1274 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1275 RRC = (Subtarget->is64Bit()
1276 ? X86::GR64RegisterClass : X86::GR32RegisterClass);
1279 RRC = X86::VR64RegisterClass;
1281 case MVT::f32: case MVT::f64:
1282 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1283 case MVT::v4f32: case MVT::v2f64:
1284 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1286 RRC = X86::VR128RegisterClass;
1289 return std::make_pair(RRC, Cost);
1292 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1293 unsigned &Offset) const {
1294 if (!Subtarget->isTargetLinux())
1297 if (Subtarget->is64Bit()) {
1298 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1300 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1313 //===----------------------------------------------------------------------===//
1314 // Return Value Calling Convention Implementation
1315 //===----------------------------------------------------------------------===//
1317 #include "X86GenCallingConv.inc"
1320 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1321 MachineFunction &MF, bool isVarArg,
1322 const SmallVectorImpl<ISD::OutputArg> &Outs,
1323 LLVMContext &Context) const {
1324 SmallVector<CCValAssign, 16> RVLocs;
1325 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1327 return CCInfo.CheckReturn(Outs, RetCC_X86);
1331 X86TargetLowering::LowerReturn(SDValue Chain,
1332 CallingConv::ID CallConv, bool isVarArg,
1333 const SmallVectorImpl<ISD::OutputArg> &Outs,
1334 const SmallVectorImpl<SDValue> &OutVals,
1335 DebugLoc dl, SelectionDAG &DAG) const {
1336 MachineFunction &MF = DAG.getMachineFunction();
1337 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1339 SmallVector<CCValAssign, 16> RVLocs;
1340 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1341 RVLocs, *DAG.getContext());
1342 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1344 // Add the regs to the liveout set for the function.
1345 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1346 for (unsigned i = 0; i != RVLocs.size(); ++i)
1347 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1348 MRI.addLiveOut(RVLocs[i].getLocReg());
1352 SmallVector<SDValue, 6> RetOps;
1353 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1354 // Operand #1 = Bytes To Pop
1355 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1358 // Copy the result values into the output registers.
1359 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1360 CCValAssign &VA = RVLocs[i];
1361 assert(VA.isRegLoc() && "Can only return in registers!");
1362 SDValue ValToCopy = OutVals[i];
1363 EVT ValVT = ValToCopy.getValueType();
1365 // If this is x86-64, and we disabled SSE, we can't return FP values,
1366 // or SSE or MMX vectors.
1367 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1368 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1369 (Subtarget->is64Bit() && !Subtarget->hasXMM())) {
1370 report_fatal_error("SSE register return with SSE disabled");
1372 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1373 // llvm-gcc has never done it right and no one has noticed, so this
1374 // should be OK for now.
1375 if (ValVT == MVT::f64 &&
1376 (Subtarget->is64Bit() && !Subtarget->hasXMMInt()))
1377 report_fatal_error("SSE2 register return with SSE2 disabled");
1379 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1380 // the RET instruction and handled by the FP Stackifier.
1381 if (VA.getLocReg() == X86::ST0 ||
1382 VA.getLocReg() == X86::ST1) {
1383 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1384 // change the value to the FP stack register class.
1385 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1386 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1387 RetOps.push_back(ValToCopy);
1388 // Don't emit a copytoreg.
1392 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1393 // which is returned in RAX / RDX.
1394 if (Subtarget->is64Bit()) {
1395 if (ValVT == MVT::x86mmx) {
1396 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1397 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
1398 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1400 // If we don't have SSE2 available, convert to v4f32 so the generated
1401 // register is legal.
1402 if (!Subtarget->hasSSE2())
1403 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
1408 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1409 Flag = Chain.getValue(1);
1412 // The x86-64 ABI for returning structs by value requires that we copy
1413 // the sret argument into %rax for the return. We saved the argument into
1414 // a virtual register in the entry block, so now we copy the value out
1416 if (Subtarget->is64Bit() &&
1417 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1418 MachineFunction &MF = DAG.getMachineFunction();
1419 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1420 unsigned Reg = FuncInfo->getSRetReturnReg();
1422 "SRetReturnReg should have been set in LowerFormalArguments().");
1423 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1425 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1426 Flag = Chain.getValue(1);
1428 // RAX now acts like a return value.
1429 MRI.addLiveOut(X86::RAX);
1432 RetOps[0] = Chain; // Update chain.
1434 // Add the flag if we have it.
1436 RetOps.push_back(Flag);
1438 return DAG.getNode(X86ISD::RET_FLAG, dl,
1439 MVT::Other, &RetOps[0], RetOps.size());
1442 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N) const {
1443 if (N->getNumValues() != 1)
1445 if (!N->hasNUsesOfValue(1, 0))
1448 SDNode *Copy = *N->use_begin();
1449 if (Copy->getOpcode() != ISD::CopyToReg &&
1450 Copy->getOpcode() != ISD::FP_EXTEND)
1453 bool HasRet = false;
1454 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1456 if (UI->getOpcode() != X86ISD::RET_FLAG)
1465 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
1466 ISD::NodeType ExtendKind) const {
1468 // TODO: Is this also valid on 32-bit?
1469 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
1470 ReturnMVT = MVT::i8;
1472 ReturnMVT = MVT::i32;
1474 EVT MinVT = getRegisterType(Context, ReturnMVT);
1475 return VT.bitsLT(MinVT) ? MinVT : VT;
1478 /// LowerCallResult - Lower the result values of a call into the
1479 /// appropriate copies out of appropriate physical registers.
1482 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1483 CallingConv::ID CallConv, bool isVarArg,
1484 const SmallVectorImpl<ISD::InputArg> &Ins,
1485 DebugLoc dl, SelectionDAG &DAG,
1486 SmallVectorImpl<SDValue> &InVals) const {
1488 // Assign locations to each value returned by this call.
1489 SmallVector<CCValAssign, 16> RVLocs;
1490 bool Is64Bit = Subtarget->is64Bit();
1491 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1492 getTargetMachine(), RVLocs, *DAG.getContext());
1493 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1495 // Copy all of the result registers out of their specified physreg.
1496 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1497 CCValAssign &VA = RVLocs[i];
1498 EVT CopyVT = VA.getValVT();
1500 // If this is x86-64, and we disabled SSE, we can't return FP values
1501 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1502 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasXMM())) {
1503 report_fatal_error("SSE register return with SSE disabled");
1508 // If this is a call to a function that returns an fp value on the floating
1509 // point stack, we must guarantee the the value is popped from the stack, so
1510 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1511 // if the return value is not used. We use the FpPOP_RETVAL instruction
1513 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1514 // If we prefer to use the value in xmm registers, copy it out as f80 and
1515 // use a truncate to move it from fp stack reg to xmm reg.
1516 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1517 SDValue Ops[] = { Chain, InFlag };
1518 Chain = SDValue(DAG.getMachineNode(X86::FpPOP_RETVAL, dl, CopyVT,
1519 MVT::Other, MVT::Glue, Ops, 2), 1);
1520 Val = Chain.getValue(0);
1522 // Round the f80 to the right size, which also moves it to the appropriate
1524 if (CopyVT != VA.getValVT())
1525 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1526 // This truncation won't change the value.
1527 DAG.getIntPtrConstant(1));
1529 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1530 CopyVT, InFlag).getValue(1);
1531 Val = Chain.getValue(0);
1533 InFlag = Chain.getValue(2);
1534 InVals.push_back(Val);
1541 //===----------------------------------------------------------------------===//
1542 // C & StdCall & Fast Calling Convention implementation
1543 //===----------------------------------------------------------------------===//
1544 // StdCall calling convention seems to be standard for many Windows' API
1545 // routines and around. It differs from C calling convention just a little:
1546 // callee should clean up the stack, not caller. Symbols should be also
1547 // decorated in some fancy way :) It doesn't support any vector arguments.
1548 // For info on fast calling convention see Fast Calling Convention (tail call)
1549 // implementation LowerX86_32FastCCCallTo.
1551 /// CallIsStructReturn - Determines whether a call uses struct return
1553 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1557 return Outs[0].Flags.isSRet();
1560 /// ArgsAreStructReturn - Determines whether a function uses struct
1561 /// return semantics.
1563 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1567 return Ins[0].Flags.isSRet();
1570 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1571 /// by "Src" to address "Dst" with size and alignment information specified by
1572 /// the specific parameter attribute. The copy will be passed as a byval
1573 /// function parameter.
1575 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1576 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1578 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1580 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1581 /*isVolatile*/false, /*AlwaysInline=*/true,
1582 MachinePointerInfo(), MachinePointerInfo());
1585 /// IsTailCallConvention - Return true if the calling convention is one that
1586 /// supports tail call optimization.
1587 static bool IsTailCallConvention(CallingConv::ID CC) {
1588 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1591 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1592 if (!CI->isTailCall())
1596 CallingConv::ID CalleeCC = CS.getCallingConv();
1597 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
1603 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1604 /// a tailcall target by changing its ABI.
1605 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1606 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1610 X86TargetLowering::LowerMemArgument(SDValue Chain,
1611 CallingConv::ID CallConv,
1612 const SmallVectorImpl<ISD::InputArg> &Ins,
1613 DebugLoc dl, SelectionDAG &DAG,
1614 const CCValAssign &VA,
1615 MachineFrameInfo *MFI,
1617 // Create the nodes corresponding to a load from this parameter slot.
1618 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1619 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1620 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1623 // If value is passed by pointer we have address passed instead of the value
1625 if (VA.getLocInfo() == CCValAssign::Indirect)
1626 ValVT = VA.getLocVT();
1628 ValVT = VA.getValVT();
1630 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1631 // changed with more analysis.
1632 // In case of tail call optimization mark all arguments mutable. Since they
1633 // could be overwritten by lowering of arguments in case of a tail call.
1634 if (Flags.isByVal()) {
1635 unsigned Bytes = Flags.getByValSize();
1636 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
1637 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
1638 return DAG.getFrameIndex(FI, getPointerTy());
1640 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1641 VA.getLocMemOffset(), isImmutable);
1642 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1643 return DAG.getLoad(ValVT, dl, Chain, FIN,
1644 MachinePointerInfo::getFixedStack(FI),
1650 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1651 CallingConv::ID CallConv,
1653 const SmallVectorImpl<ISD::InputArg> &Ins,
1656 SmallVectorImpl<SDValue> &InVals)
1658 MachineFunction &MF = DAG.getMachineFunction();
1659 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1661 const Function* Fn = MF.getFunction();
1662 if (Fn->hasExternalLinkage() &&
1663 Subtarget->isTargetCygMing() &&
1664 Fn->getName() == "main")
1665 FuncInfo->setForceFramePointer(true);
1667 MachineFrameInfo *MFI = MF.getFrameInfo();
1668 bool Is64Bit = Subtarget->is64Bit();
1669 bool IsWin64 = Subtarget->isTargetWin64();
1671 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1672 "Var args not supported with calling convention fastcc or ghc");
1674 // Assign locations to all of the incoming arguments.
1675 SmallVector<CCValAssign, 16> ArgLocs;
1676 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1677 ArgLocs, *DAG.getContext());
1679 // Allocate shadow area for Win64
1681 CCInfo.AllocateStack(32, 8);
1684 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
1686 unsigned LastVal = ~0U;
1688 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1689 CCValAssign &VA = ArgLocs[i];
1690 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1692 assert(VA.getValNo() != LastVal &&
1693 "Don't support value assigned to multiple locs yet");
1694 LastVal = VA.getValNo();
1696 if (VA.isRegLoc()) {
1697 EVT RegVT = VA.getLocVT();
1698 TargetRegisterClass *RC = NULL;
1699 if (RegVT == MVT::i32)
1700 RC = X86::GR32RegisterClass;
1701 else if (Is64Bit && RegVT == MVT::i64)
1702 RC = X86::GR64RegisterClass;
1703 else if (RegVT == MVT::f32)
1704 RC = X86::FR32RegisterClass;
1705 else if (RegVT == MVT::f64)
1706 RC = X86::FR64RegisterClass;
1707 else if (RegVT.isVector() && RegVT.getSizeInBits() == 256)
1708 RC = X86::VR256RegisterClass;
1709 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1710 RC = X86::VR128RegisterClass;
1711 else if (RegVT == MVT::x86mmx)
1712 RC = X86::VR64RegisterClass;
1714 llvm_unreachable("Unknown argument type!");
1716 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1717 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1719 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1720 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1722 if (VA.getLocInfo() == CCValAssign::SExt)
1723 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1724 DAG.getValueType(VA.getValVT()));
1725 else if (VA.getLocInfo() == CCValAssign::ZExt)
1726 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1727 DAG.getValueType(VA.getValVT()));
1728 else if (VA.getLocInfo() == CCValAssign::BCvt)
1729 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
1731 if (VA.isExtInLoc()) {
1732 // Handle MMX values passed in XMM regs.
1733 if (RegVT.isVector()) {
1734 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(),
1737 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1740 assert(VA.isMemLoc());
1741 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1744 // If value is passed via pointer - do a load.
1745 if (VA.getLocInfo() == CCValAssign::Indirect)
1746 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
1747 MachinePointerInfo(), false, false, 0);
1749 InVals.push_back(ArgValue);
1752 // The x86-64 ABI for returning structs by value requires that we copy
1753 // the sret argument into %rax for the return. Save the argument into
1754 // a virtual register so that we can access it from the return points.
1755 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1756 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1757 unsigned Reg = FuncInfo->getSRetReturnReg();
1759 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1760 FuncInfo->setSRetReturnReg(Reg);
1762 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1763 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1766 unsigned StackSize = CCInfo.getNextStackOffset();
1767 // Align stack specially for tail calls.
1768 if (FuncIsMadeTailCallSafe(CallConv))
1769 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1771 // If the function takes variable number of arguments, make a frame index for
1772 // the start of the first vararg value... for expansion of llvm.va_start.
1774 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1775 CallConv != CallingConv::X86_ThisCall)) {
1776 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1779 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1781 // FIXME: We should really autogenerate these arrays
1782 static const unsigned GPR64ArgRegsWin64[] = {
1783 X86::RCX, X86::RDX, X86::R8, X86::R9
1785 static const unsigned GPR64ArgRegs64Bit[] = {
1786 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1788 static const unsigned XMMArgRegs64Bit[] = {
1789 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1790 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1792 const unsigned *GPR64ArgRegs;
1793 unsigned NumXMMRegs = 0;
1796 // The XMM registers which might contain var arg parameters are shadowed
1797 // in their paired GPR. So we only need to save the GPR to their home
1799 TotalNumIntRegs = 4;
1800 GPR64ArgRegs = GPR64ArgRegsWin64;
1802 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1803 GPR64ArgRegs = GPR64ArgRegs64Bit;
1805 NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs64Bit, TotalNumXMMRegs);
1807 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1810 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1811 assert(!(NumXMMRegs && !Subtarget->hasXMM()) &&
1812 "SSE register cannot be used when SSE is disabled!");
1813 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1814 "SSE register cannot be used when SSE is disabled!");
1815 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasXMM())
1816 // Kernel mode asks for SSE to be disabled, so don't push them
1818 TotalNumXMMRegs = 0;
1821 const TargetFrameLowering &TFI = *getTargetMachine().getFrameLowering();
1822 // Get to the caller-allocated home save location. Add 8 to account
1823 // for the return address.
1824 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
1825 FuncInfo->setRegSaveFrameIndex(
1826 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
1827 // Fixup to set vararg frame on shadow area (4 x i64).
1829 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
1831 // For X86-64, if there are vararg parameters that are passed via
1832 // registers, then we must store them to their spots on the stack so they
1833 // may be loaded by deferencing the result of va_next.
1834 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1835 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1836 FuncInfo->setRegSaveFrameIndex(
1837 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1841 // Store the integer parameter registers.
1842 SmallVector<SDValue, 8> MemOps;
1843 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1845 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1846 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1847 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1848 DAG.getIntPtrConstant(Offset));
1849 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1850 X86::GR64RegisterClass);
1851 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1853 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1854 MachinePointerInfo::getFixedStack(
1855 FuncInfo->getRegSaveFrameIndex(), Offset),
1857 MemOps.push_back(Store);
1861 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1862 // Now store the XMM (fp + vector) parameter registers.
1863 SmallVector<SDValue, 11> SaveXMMOps;
1864 SaveXMMOps.push_back(Chain);
1866 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1867 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1868 SaveXMMOps.push_back(ALVal);
1870 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1871 FuncInfo->getRegSaveFrameIndex()));
1872 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1873 FuncInfo->getVarArgsFPOffset()));
1875 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1876 unsigned VReg = MF.addLiveIn(XMMArgRegs64Bit[NumXMMRegs],
1877 X86::VR128RegisterClass);
1878 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1879 SaveXMMOps.push_back(Val);
1881 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1883 &SaveXMMOps[0], SaveXMMOps.size()));
1886 if (!MemOps.empty())
1887 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1888 &MemOps[0], MemOps.size());
1892 // Some CCs need callee pop.
1893 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, GuaranteedTailCallOpt)) {
1894 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1896 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1897 // If this is an sret function, the return should pop the hidden pointer.
1898 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1899 FuncInfo->setBytesToPopOnReturn(4);
1903 // RegSaveFrameIndex is X86-64 only.
1904 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1905 if (CallConv == CallingConv::X86_FastCall ||
1906 CallConv == CallingConv::X86_ThisCall)
1907 // fastcc functions can't have varargs.
1908 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1915 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1916 SDValue StackPtr, SDValue Arg,
1917 DebugLoc dl, SelectionDAG &DAG,
1918 const CCValAssign &VA,
1919 ISD::ArgFlagsTy Flags) const {
1920 unsigned LocMemOffset = VA.getLocMemOffset();
1921 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1922 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1923 if (Flags.isByVal())
1924 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1926 return DAG.getStore(Chain, dl, Arg, PtrOff,
1927 MachinePointerInfo::getStack(LocMemOffset),
1931 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1932 /// optimization is performed and it is required.
1934 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1935 SDValue &OutRetAddr, SDValue Chain,
1936 bool IsTailCall, bool Is64Bit,
1937 int FPDiff, DebugLoc dl) const {
1938 // Adjust the Return address stack slot.
1939 EVT VT = getPointerTy();
1940 OutRetAddr = getReturnAddressFrameIndex(DAG);
1942 // Load the "old" Return address.
1943 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
1945 return SDValue(OutRetAddr.getNode(), 1);
1948 /// EmitTailCallStoreRetAddr - Emit a store of the return address if tail call
1949 /// optimization is performed and it is required (FPDiff!=0).
1951 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1952 SDValue Chain, SDValue RetAddrFrIdx,
1953 bool Is64Bit, int FPDiff, DebugLoc dl) {
1954 // Store the return address to the appropriate stack slot.
1955 if (!FPDiff) return Chain;
1956 // Calculate the new stack slot for the return address.
1957 int SlotSize = Is64Bit ? 8 : 4;
1958 int NewReturnAddrFI =
1959 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1960 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1961 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1962 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1963 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
1969 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1970 CallingConv::ID CallConv, bool isVarArg,
1972 const SmallVectorImpl<ISD::OutputArg> &Outs,
1973 const SmallVectorImpl<SDValue> &OutVals,
1974 const SmallVectorImpl<ISD::InputArg> &Ins,
1975 DebugLoc dl, SelectionDAG &DAG,
1976 SmallVectorImpl<SDValue> &InVals) const {
1977 MachineFunction &MF = DAG.getMachineFunction();
1978 bool Is64Bit = Subtarget->is64Bit();
1979 bool IsWin64 = Subtarget->isTargetWin64();
1980 bool IsStructRet = CallIsStructReturn(Outs);
1981 bool IsSibcall = false;
1984 // Check if it's really possible to do a tail call.
1985 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1986 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1987 Outs, OutVals, Ins, DAG);
1989 // Sibcalls are automatically detected tailcalls which do not require
1991 if (!GuaranteedTailCallOpt && isTailCall)
1998 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1999 "Var args not supported with calling convention fastcc or ghc");
2001 // Analyze operands of the call, assigning locations to each operand.
2002 SmallVector<CCValAssign, 16> ArgLocs;
2003 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
2004 ArgLocs, *DAG.getContext());
2006 // Allocate shadow area for Win64
2008 CCInfo.AllocateStack(32, 8);
2011 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2013 // Get a count of how many bytes are to be pushed on the stack.
2014 unsigned NumBytes = CCInfo.getNextStackOffset();
2016 // This is a sibcall. The memory operands are available in caller's
2017 // own caller's stack.
2019 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
2020 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2023 if (isTailCall && !IsSibcall) {
2024 // Lower arguments at fp - stackoffset + fpdiff.
2025 unsigned NumBytesCallerPushed =
2026 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
2027 FPDiff = NumBytesCallerPushed - NumBytes;
2029 // Set the delta of movement of the returnaddr stackslot.
2030 // But only set if delta is greater than previous delta.
2031 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
2032 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
2036 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2038 SDValue RetAddrFrIdx;
2039 // Load return address for tail calls.
2040 if (isTailCall && FPDiff)
2041 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2042 Is64Bit, FPDiff, dl);
2044 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2045 SmallVector<SDValue, 8> MemOpChains;
2048 // Walk the register/memloc assignments, inserting copies/loads. In the case
2049 // of tail call optimization arguments are handle later.
2050 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2051 CCValAssign &VA = ArgLocs[i];
2052 EVT RegVT = VA.getLocVT();
2053 SDValue Arg = OutVals[i];
2054 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2055 bool isByVal = Flags.isByVal();
2057 // Promote the value if needed.
2058 switch (VA.getLocInfo()) {
2059 default: llvm_unreachable("Unknown loc info!");
2060 case CCValAssign::Full: break;
2061 case CCValAssign::SExt:
2062 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2064 case CCValAssign::ZExt:
2065 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2067 case CCValAssign::AExt:
2068 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
2069 // Special case: passing MMX values in XMM registers.
2070 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2071 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2072 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2074 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2076 case CCValAssign::BCvt:
2077 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2079 case CCValAssign::Indirect: {
2080 // Store the argument.
2081 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2082 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2083 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2084 MachinePointerInfo::getFixedStack(FI),
2091 if (VA.isRegLoc()) {
2092 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2093 if (isVarArg && IsWin64) {
2094 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2095 // shadow reg if callee is a varargs function.
2096 unsigned ShadowReg = 0;
2097 switch (VA.getLocReg()) {
2098 case X86::XMM0: ShadowReg = X86::RCX; break;
2099 case X86::XMM1: ShadowReg = X86::RDX; break;
2100 case X86::XMM2: ShadowReg = X86::R8; break;
2101 case X86::XMM3: ShadowReg = X86::R9; break;
2104 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2106 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2107 assert(VA.isMemLoc());
2108 if (StackPtr.getNode() == 0)
2109 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
2110 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2111 dl, DAG, VA, Flags));
2115 if (!MemOpChains.empty())
2116 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2117 &MemOpChains[0], MemOpChains.size());
2119 // Build a sequence of copy-to-reg nodes chained together with token chain
2120 // and flag operands which copy the outgoing args into registers.
2122 // Tail call byval lowering might overwrite argument registers so in case of
2123 // tail call optimization the copies to registers are lowered later.
2125 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2126 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2127 RegsToPass[i].second, InFlag);
2128 InFlag = Chain.getValue(1);
2131 if (Subtarget->isPICStyleGOT()) {
2132 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2135 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
2136 DAG.getNode(X86ISD::GlobalBaseReg,
2137 DebugLoc(), getPointerTy()),
2139 InFlag = Chain.getValue(1);
2141 // If we are tail calling and generating PIC/GOT style code load the
2142 // address of the callee into ECX. The value in ecx is used as target of
2143 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2144 // for tail calls on PIC/GOT architectures. Normally we would just put the
2145 // address of GOT into ebx and then call target@PLT. But for tail calls
2146 // ebx would be restored (since ebx is callee saved) before jumping to the
2149 // Note: The actual moving to ECX is done further down.
2150 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2151 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2152 !G->getGlobal()->hasProtectedVisibility())
2153 Callee = LowerGlobalAddress(Callee, DAG);
2154 else if (isa<ExternalSymbolSDNode>(Callee))
2155 Callee = LowerExternalSymbol(Callee, DAG);
2159 if (Is64Bit && isVarArg && !IsWin64) {
2160 // From AMD64 ABI document:
2161 // For calls that may call functions that use varargs or stdargs
2162 // (prototype-less calls or calls to functions containing ellipsis (...) in
2163 // the declaration) %al is used as hidden argument to specify the number
2164 // of SSE registers used. The contents of %al do not need to match exactly
2165 // the number of registers, but must be an ubound on the number of SSE
2166 // registers used and is in the range 0 - 8 inclusive.
2168 // Count the number of XMM registers allocated.
2169 static const unsigned XMMArgRegs[] = {
2170 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2171 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2173 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2174 assert((Subtarget->hasXMM() || !NumXMMRegs)
2175 && "SSE registers cannot be used when SSE is disabled");
2177 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2178 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2179 InFlag = Chain.getValue(1);
2183 // For tail calls lower the arguments to the 'real' stack slot.
2185 // Force all the incoming stack arguments to be loaded from the stack
2186 // before any new outgoing arguments are stored to the stack, because the
2187 // outgoing stack slots may alias the incoming argument stack slots, and
2188 // the alias isn't otherwise explicit. This is slightly more conservative
2189 // than necessary, because it means that each store effectively depends
2190 // on every argument instead of just those arguments it would clobber.
2191 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2193 SmallVector<SDValue, 8> MemOpChains2;
2196 // Do not flag preceding copytoreg stuff together with the following stuff.
2198 if (GuaranteedTailCallOpt) {
2199 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2200 CCValAssign &VA = ArgLocs[i];
2203 assert(VA.isMemLoc());
2204 SDValue Arg = OutVals[i];
2205 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2206 // Create frame index.
2207 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2208 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2209 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2210 FIN = DAG.getFrameIndex(FI, getPointerTy());
2212 if (Flags.isByVal()) {
2213 // Copy relative to framepointer.
2214 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2215 if (StackPtr.getNode() == 0)
2216 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2218 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2220 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2224 // Store relative to framepointer.
2225 MemOpChains2.push_back(
2226 DAG.getStore(ArgChain, dl, Arg, FIN,
2227 MachinePointerInfo::getFixedStack(FI),
2233 if (!MemOpChains2.empty())
2234 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2235 &MemOpChains2[0], MemOpChains2.size());
2237 // Copy arguments to their registers.
2238 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2239 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2240 RegsToPass[i].second, InFlag);
2241 InFlag = Chain.getValue(1);
2245 // Store the return address to the appropriate stack slot.
2246 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2250 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2251 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2252 // In the 64-bit large code model, we have to make all calls
2253 // through a register, since the call instruction's 32-bit
2254 // pc-relative offset may not be large enough to hold the whole
2256 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2257 // If the callee is a GlobalAddress node (quite common, every direct call
2258 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2261 // We should use extra load for direct calls to dllimported functions in
2263 const GlobalValue *GV = G->getGlobal();
2264 if (!GV->hasDLLImportLinkage()) {
2265 unsigned char OpFlags = 0;
2266 bool ExtraLoad = false;
2267 unsigned WrapperKind = ISD::DELETED_NODE;
2269 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2270 // external symbols most go through the PLT in PIC mode. If the symbol
2271 // has hidden or protected visibility, or if it is static or local, then
2272 // we don't need to use the PLT - we can directly call it.
2273 if (Subtarget->isTargetELF() &&
2274 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2275 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2276 OpFlags = X86II::MO_PLT;
2277 } else if (Subtarget->isPICStyleStubAny() &&
2278 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2279 (!Subtarget->getTargetTriple().isMacOSX() ||
2280 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2281 // PC-relative references to external symbols should go through $stub,
2282 // unless we're building with the leopard linker or later, which
2283 // automatically synthesizes these stubs.
2284 OpFlags = X86II::MO_DARWIN_STUB;
2285 } else if (Subtarget->isPICStyleRIPRel() &&
2286 isa<Function>(GV) &&
2287 cast<Function>(GV)->hasFnAttr(Attribute::NonLazyBind)) {
2288 // If the function is marked as non-lazy, generate an indirect call
2289 // which loads from the GOT directly. This avoids runtime overhead
2290 // at the cost of eager binding (and one extra byte of encoding).
2291 OpFlags = X86II::MO_GOTPCREL;
2292 WrapperKind = X86ISD::WrapperRIP;
2296 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2297 G->getOffset(), OpFlags);
2299 // Add a wrapper if needed.
2300 if (WrapperKind != ISD::DELETED_NODE)
2301 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
2302 // Add extra indirection if needed.
2304 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
2305 MachinePointerInfo::getGOT(),
2308 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2309 unsigned char OpFlags = 0;
2311 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
2312 // external symbols should go through the PLT.
2313 if (Subtarget->isTargetELF() &&
2314 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2315 OpFlags = X86II::MO_PLT;
2316 } else if (Subtarget->isPICStyleStubAny() &&
2317 (!Subtarget->getTargetTriple().isMacOSX() ||
2318 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2319 // PC-relative references to external symbols should go through $stub,
2320 // unless we're building with the leopard linker or later, which
2321 // automatically synthesizes these stubs.
2322 OpFlags = X86II::MO_DARWIN_STUB;
2325 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2329 // Returns a chain & a flag for retval copy to use.
2330 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2331 SmallVector<SDValue, 8> Ops;
2333 if (!IsSibcall && isTailCall) {
2334 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2335 DAG.getIntPtrConstant(0, true), InFlag);
2336 InFlag = Chain.getValue(1);
2339 Ops.push_back(Chain);
2340 Ops.push_back(Callee);
2343 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2345 // Add argument registers to the end of the list so that they are known live
2347 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2348 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2349 RegsToPass[i].second.getValueType()));
2351 // Add an implicit use GOT pointer in EBX.
2352 if (!isTailCall && Subtarget->isPICStyleGOT())
2353 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2355 // Add an implicit use of AL for non-Windows x86 64-bit vararg functions.
2356 if (Is64Bit && isVarArg && !IsWin64)
2357 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2359 if (InFlag.getNode())
2360 Ops.push_back(InFlag);
2364 //// If this is the first return lowered for this function, add the regs
2365 //// to the liveout set for the function.
2366 // This isn't right, although it's probably harmless on x86; liveouts
2367 // should be computed from returns not tail calls. Consider a void
2368 // function making a tail call to a function returning int.
2369 return DAG.getNode(X86ISD::TC_RETURN, dl,
2370 NodeTys, &Ops[0], Ops.size());
2373 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2374 InFlag = Chain.getValue(1);
2376 // Create the CALLSEQ_END node.
2377 unsigned NumBytesForCalleeToPush;
2378 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, GuaranteedTailCallOpt))
2379 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2380 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2381 // If this is a call to a struct-return function, the callee
2382 // pops the hidden struct pointer, so we have to push it back.
2383 // This is common for Darwin/X86, Linux & Mingw32 targets.
2384 NumBytesForCalleeToPush = 4;
2386 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2388 // Returns a flag for retval copy to use.
2390 Chain = DAG.getCALLSEQ_END(Chain,
2391 DAG.getIntPtrConstant(NumBytes, true),
2392 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2395 InFlag = Chain.getValue(1);
2398 // Handle result values, copying them out of physregs into vregs that we
2400 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2401 Ins, dl, DAG, InVals);
2405 //===----------------------------------------------------------------------===//
2406 // Fast Calling Convention (tail call) implementation
2407 //===----------------------------------------------------------------------===//
2409 // Like std call, callee cleans arguments, convention except that ECX is
2410 // reserved for storing the tail called function address. Only 2 registers are
2411 // free for argument passing (inreg). Tail call optimization is performed
2413 // * tailcallopt is enabled
2414 // * caller/callee are fastcc
2415 // On X86_64 architecture with GOT-style position independent code only local
2416 // (within module) calls are supported at the moment.
2417 // To keep the stack aligned according to platform abi the function
2418 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2419 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2420 // If a tail called function callee has more arguments than the caller the
2421 // caller needs to make sure that there is room to move the RETADDR to. This is
2422 // achieved by reserving an area the size of the argument delta right after the
2423 // original REtADDR, but before the saved framepointer or the spilled registers
2424 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2436 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2437 /// for a 16 byte align requirement.
2439 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2440 SelectionDAG& DAG) const {
2441 MachineFunction &MF = DAG.getMachineFunction();
2442 const TargetMachine &TM = MF.getTarget();
2443 const TargetFrameLowering &TFI = *TM.getFrameLowering();
2444 unsigned StackAlignment = TFI.getStackAlignment();
2445 uint64_t AlignMask = StackAlignment - 1;
2446 int64_t Offset = StackSize;
2447 uint64_t SlotSize = TD->getPointerSize();
2448 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2449 // Number smaller than 12 so just add the difference.
2450 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2452 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2453 Offset = ((~AlignMask) & Offset) + StackAlignment +
2454 (StackAlignment-SlotSize);
2459 /// MatchingStackOffset - Return true if the given stack call argument is
2460 /// already available in the same position (relatively) of the caller's
2461 /// incoming argument stack.
2463 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2464 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2465 const X86InstrInfo *TII) {
2466 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2468 if (Arg.getOpcode() == ISD::CopyFromReg) {
2469 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2470 if (!TargetRegisterInfo::isVirtualRegister(VR))
2472 MachineInstr *Def = MRI->getVRegDef(VR);
2475 if (!Flags.isByVal()) {
2476 if (!TII->isLoadFromStackSlot(Def, FI))
2479 unsigned Opcode = Def->getOpcode();
2480 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2481 Def->getOperand(1).isFI()) {
2482 FI = Def->getOperand(1).getIndex();
2483 Bytes = Flags.getByValSize();
2487 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2488 if (Flags.isByVal())
2489 // ByVal argument is passed in as a pointer but it's now being
2490 // dereferenced. e.g.
2491 // define @foo(%struct.X* %A) {
2492 // tail call @bar(%struct.X* byval %A)
2495 SDValue Ptr = Ld->getBasePtr();
2496 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2499 FI = FINode->getIndex();
2500 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
2501 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
2502 FI = FINode->getIndex();
2503 Bytes = Flags.getByValSize();
2507 assert(FI != INT_MAX);
2508 if (!MFI->isFixedObjectIndex(FI))
2510 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2513 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2514 /// for tail call optimization. Targets which want to do tail call
2515 /// optimization should implement this function.
2517 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2518 CallingConv::ID CalleeCC,
2520 bool isCalleeStructRet,
2521 bool isCallerStructRet,
2522 const SmallVectorImpl<ISD::OutputArg> &Outs,
2523 const SmallVectorImpl<SDValue> &OutVals,
2524 const SmallVectorImpl<ISD::InputArg> &Ins,
2525 SelectionDAG& DAG) const {
2526 if (!IsTailCallConvention(CalleeCC) &&
2527 CalleeCC != CallingConv::C)
2530 // If -tailcallopt is specified, make fastcc functions tail-callable.
2531 const MachineFunction &MF = DAG.getMachineFunction();
2532 const Function *CallerF = DAG.getMachineFunction().getFunction();
2533 CallingConv::ID CallerCC = CallerF->getCallingConv();
2534 bool CCMatch = CallerCC == CalleeCC;
2536 if (GuaranteedTailCallOpt) {
2537 if (IsTailCallConvention(CalleeCC) && CCMatch)
2542 // Look for obvious safe cases to perform tail call optimization that do not
2543 // require ABI changes. This is what gcc calls sibcall.
2545 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2546 // emit a special epilogue.
2547 if (RegInfo->needsStackRealignment(MF))
2550 // Also avoid sibcall optimization if either caller or callee uses struct
2551 // return semantics.
2552 if (isCalleeStructRet || isCallerStructRet)
2555 // An stdcall caller is expected to clean up its arguments; the callee
2556 // isn't going to do that.
2557 if (!CCMatch && CallerCC==CallingConv::X86_StdCall)
2560 // Do not sibcall optimize vararg calls unless all arguments are passed via
2562 if (isVarArg && !Outs.empty()) {
2564 // Optimizing for varargs on Win64 is unlikely to be safe without
2565 // additional testing.
2566 if (Subtarget->isTargetWin64())
2569 SmallVector<CCValAssign, 16> ArgLocs;
2570 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2571 getTargetMachine(), ArgLocs, *DAG.getContext());
2573 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2574 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
2575 if (!ArgLocs[i].isRegLoc())
2579 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2580 // Therefore if it's not used by the call it is not safe to optimize this into
2582 bool Unused = false;
2583 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2590 SmallVector<CCValAssign, 16> RVLocs;
2591 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(),
2592 getTargetMachine(), RVLocs, *DAG.getContext());
2593 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2594 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2595 CCValAssign &VA = RVLocs[i];
2596 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2601 // If the calling conventions do not match, then we'd better make sure the
2602 // results are returned in the same way as what the caller expects.
2604 SmallVector<CCValAssign, 16> RVLocs1;
2605 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
2606 getTargetMachine(), RVLocs1, *DAG.getContext());
2607 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2609 SmallVector<CCValAssign, 16> RVLocs2;
2610 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
2611 getTargetMachine(), RVLocs2, *DAG.getContext());
2612 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2614 if (RVLocs1.size() != RVLocs2.size())
2616 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2617 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2619 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2621 if (RVLocs1[i].isRegLoc()) {
2622 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2625 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2631 // If the callee takes no arguments then go on to check the results of the
2633 if (!Outs.empty()) {
2634 // Check if stack adjustment is needed. For now, do not do this if any
2635 // argument is passed on the stack.
2636 SmallVector<CCValAssign, 16> ArgLocs;
2637 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2638 getTargetMachine(), ArgLocs, *DAG.getContext());
2640 // Allocate shadow area for Win64
2641 if (Subtarget->isTargetWin64()) {
2642 CCInfo.AllocateStack(32, 8);
2645 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2646 if (CCInfo.getNextStackOffset()) {
2647 MachineFunction &MF = DAG.getMachineFunction();
2648 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2651 // Check if the arguments are already laid out in the right way as
2652 // the caller's fixed stack objects.
2653 MachineFrameInfo *MFI = MF.getFrameInfo();
2654 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2655 const X86InstrInfo *TII =
2656 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2657 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2658 CCValAssign &VA = ArgLocs[i];
2659 SDValue Arg = OutVals[i];
2660 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2661 if (VA.getLocInfo() == CCValAssign::Indirect)
2663 if (!VA.isRegLoc()) {
2664 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2671 // If the tailcall address may be in a register, then make sure it's
2672 // possible to register allocate for it. In 32-bit, the call address can
2673 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2674 // callee-saved registers are restored. These happen to be the same
2675 // registers used to pass 'inreg' arguments so watch out for those.
2676 if (!Subtarget->is64Bit() &&
2677 !isa<GlobalAddressSDNode>(Callee) &&
2678 !isa<ExternalSymbolSDNode>(Callee)) {
2679 unsigned NumInRegs = 0;
2680 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2681 CCValAssign &VA = ArgLocs[i];
2684 unsigned Reg = VA.getLocReg();
2687 case X86::EAX: case X86::EDX: case X86::ECX:
2688 if (++NumInRegs == 3)
2700 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2701 return X86::createFastISel(funcInfo);
2705 //===----------------------------------------------------------------------===//
2706 // Other Lowering Hooks
2707 //===----------------------------------------------------------------------===//
2709 static bool MayFoldLoad(SDValue Op) {
2710 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
2713 static bool MayFoldIntoStore(SDValue Op) {
2714 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
2717 static bool isTargetShuffle(unsigned Opcode) {
2719 default: return false;
2720 case X86ISD::PSHUFD:
2721 case X86ISD::PSHUFHW:
2722 case X86ISD::PSHUFLW:
2723 case X86ISD::SHUFPD:
2724 case X86ISD::PALIGN:
2725 case X86ISD::SHUFPS:
2726 case X86ISD::MOVLHPS:
2727 case X86ISD::MOVLHPD:
2728 case X86ISD::MOVHLPS:
2729 case X86ISD::MOVLPS:
2730 case X86ISD::MOVLPD:
2731 case X86ISD::MOVSHDUP:
2732 case X86ISD::MOVSLDUP:
2733 case X86ISD::MOVDDUP:
2736 case X86ISD::UNPCKLPS:
2737 case X86ISD::UNPCKLPD:
2738 case X86ISD::VUNPCKLPSY:
2739 case X86ISD::VUNPCKLPDY:
2740 case X86ISD::PUNPCKLWD:
2741 case X86ISD::PUNPCKLBW:
2742 case X86ISD::PUNPCKLDQ:
2743 case X86ISD::PUNPCKLQDQ:
2744 case X86ISD::UNPCKHPS:
2745 case X86ISD::UNPCKHPD:
2746 case X86ISD::VUNPCKHPSY:
2747 case X86ISD::VUNPCKHPDY:
2748 case X86ISD::PUNPCKHWD:
2749 case X86ISD::PUNPCKHBW:
2750 case X86ISD::PUNPCKHDQ:
2751 case X86ISD::PUNPCKHQDQ:
2752 case X86ISD::VPERMILPS:
2753 case X86ISD::VPERMILPSY:
2754 case X86ISD::VPERMILPD:
2755 case X86ISD::VPERMILPDY:
2761 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2762 SDValue V1, SelectionDAG &DAG) {
2764 default: llvm_unreachable("Unknown x86 shuffle node");
2765 case X86ISD::MOVSHDUP:
2766 case X86ISD::MOVSLDUP:
2767 case X86ISD::MOVDDUP:
2768 return DAG.getNode(Opc, dl, VT, V1);
2774 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2775 SDValue V1, unsigned TargetMask, SelectionDAG &DAG) {
2777 default: llvm_unreachable("Unknown x86 shuffle node");
2778 case X86ISD::PSHUFD:
2779 case X86ISD::PSHUFHW:
2780 case X86ISD::PSHUFLW:
2781 case X86ISD::VPERMILPS:
2782 case X86ISD::VPERMILPSY:
2783 case X86ISD::VPERMILPD:
2784 case X86ISD::VPERMILPDY:
2785 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
2791 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2792 SDValue V1, SDValue V2, unsigned TargetMask, SelectionDAG &DAG) {
2794 default: llvm_unreachable("Unknown x86 shuffle node");
2795 case X86ISD::PALIGN:
2796 case X86ISD::SHUFPD:
2797 case X86ISD::SHUFPS:
2798 return DAG.getNode(Opc, dl, VT, V1, V2,
2799 DAG.getConstant(TargetMask, MVT::i8));
2804 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2805 SDValue V1, SDValue V2, SelectionDAG &DAG) {
2807 default: llvm_unreachable("Unknown x86 shuffle node");
2808 case X86ISD::MOVLHPS:
2809 case X86ISD::MOVLHPD:
2810 case X86ISD::MOVHLPS:
2811 case X86ISD::MOVLPS:
2812 case X86ISD::MOVLPD:
2815 case X86ISD::UNPCKLPS:
2816 case X86ISD::UNPCKLPD:
2817 case X86ISD::VUNPCKLPSY:
2818 case X86ISD::VUNPCKLPDY:
2819 case X86ISD::PUNPCKLWD:
2820 case X86ISD::PUNPCKLBW:
2821 case X86ISD::PUNPCKLDQ:
2822 case X86ISD::PUNPCKLQDQ:
2823 case X86ISD::UNPCKHPS:
2824 case X86ISD::UNPCKHPD:
2825 case X86ISD::VUNPCKHPSY:
2826 case X86ISD::VUNPCKHPDY:
2827 case X86ISD::PUNPCKHWD:
2828 case X86ISD::PUNPCKHBW:
2829 case X86ISD::PUNPCKHDQ:
2830 case X86ISD::PUNPCKHQDQ:
2831 return DAG.getNode(Opc, dl, VT, V1, V2);
2836 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2837 MachineFunction &MF = DAG.getMachineFunction();
2838 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2839 int ReturnAddrIndex = FuncInfo->getRAIndex();
2841 if (ReturnAddrIndex == 0) {
2842 // Set up a frame object for the return address.
2843 uint64_t SlotSize = TD->getPointerSize();
2844 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2846 FuncInfo->setRAIndex(ReturnAddrIndex);
2849 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2853 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2854 bool hasSymbolicDisplacement) {
2855 // Offset should fit into 32 bit immediate field.
2856 if (!isInt<32>(Offset))
2859 // If we don't have a symbolic displacement - we don't have any extra
2861 if (!hasSymbolicDisplacement)
2864 // FIXME: Some tweaks might be needed for medium code model.
2865 if (M != CodeModel::Small && M != CodeModel::Kernel)
2868 // For small code model we assume that latest object is 16MB before end of 31
2869 // bits boundary. We may also accept pretty large negative constants knowing
2870 // that all objects are in the positive half of address space.
2871 if (M == CodeModel::Small && Offset < 16*1024*1024)
2874 // For kernel code model we know that all object resist in the negative half
2875 // of 32bits address space. We may not accept negative offsets, since they may
2876 // be just off and we may accept pretty large positive ones.
2877 if (M == CodeModel::Kernel && Offset > 0)
2883 /// isCalleePop - Determines whether the callee is required to pop its
2884 /// own arguments. Callee pop is necessary to support tail calls.
2885 bool X86::isCalleePop(CallingConv::ID CallingConv,
2886 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
2890 switch (CallingConv) {
2893 case CallingConv::X86_StdCall:
2895 case CallingConv::X86_FastCall:
2897 case CallingConv::X86_ThisCall:
2899 case CallingConv::Fast:
2901 case CallingConv::GHC:
2906 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2907 /// specific condition code, returning the condition code and the LHS/RHS of the
2908 /// comparison to make.
2909 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2910 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2912 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2913 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2914 // X > -1 -> X == 0, jump !sign.
2915 RHS = DAG.getConstant(0, RHS.getValueType());
2916 return X86::COND_NS;
2917 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2918 // X < 0 -> X == 0, jump on sign.
2920 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2922 RHS = DAG.getConstant(0, RHS.getValueType());
2923 return X86::COND_LE;
2927 switch (SetCCOpcode) {
2928 default: llvm_unreachable("Invalid integer condition!");
2929 case ISD::SETEQ: return X86::COND_E;
2930 case ISD::SETGT: return X86::COND_G;
2931 case ISD::SETGE: return X86::COND_GE;
2932 case ISD::SETLT: return X86::COND_L;
2933 case ISD::SETLE: return X86::COND_LE;
2934 case ISD::SETNE: return X86::COND_NE;
2935 case ISD::SETULT: return X86::COND_B;
2936 case ISD::SETUGT: return X86::COND_A;
2937 case ISD::SETULE: return X86::COND_BE;
2938 case ISD::SETUGE: return X86::COND_AE;
2942 // First determine if it is required or is profitable to flip the operands.
2944 // If LHS is a foldable load, but RHS is not, flip the condition.
2945 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
2946 !ISD::isNON_EXTLoad(RHS.getNode())) {
2947 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2948 std::swap(LHS, RHS);
2951 switch (SetCCOpcode) {
2957 std::swap(LHS, RHS);
2961 // On a floating point condition, the flags are set as follows:
2963 // 0 | 0 | 0 | X > Y
2964 // 0 | 0 | 1 | X < Y
2965 // 1 | 0 | 0 | X == Y
2966 // 1 | 1 | 1 | unordered
2967 switch (SetCCOpcode) {
2968 default: llvm_unreachable("Condcode should be pre-legalized away");
2970 case ISD::SETEQ: return X86::COND_E;
2971 case ISD::SETOLT: // flipped
2973 case ISD::SETGT: return X86::COND_A;
2974 case ISD::SETOLE: // flipped
2976 case ISD::SETGE: return X86::COND_AE;
2977 case ISD::SETUGT: // flipped
2979 case ISD::SETLT: return X86::COND_B;
2980 case ISD::SETUGE: // flipped
2982 case ISD::SETLE: return X86::COND_BE;
2984 case ISD::SETNE: return X86::COND_NE;
2985 case ISD::SETUO: return X86::COND_P;
2986 case ISD::SETO: return X86::COND_NP;
2988 case ISD::SETUNE: return X86::COND_INVALID;
2992 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2993 /// code. Current x86 isa includes the following FP cmov instructions:
2994 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2995 static bool hasFPCMov(unsigned X86CC) {
3011 /// isFPImmLegal - Returns true if the target can instruction select the
3012 /// specified FP immediate natively. If false, the legalizer will
3013 /// materialize the FP immediate as a load from a constant pool.
3014 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3015 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3016 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3022 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3023 /// the specified range (L, H].
3024 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3025 return (Val < 0) || (Val >= Low && Val < Hi);
3028 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3029 /// specified value.
3030 static bool isUndefOrEqual(int Val, int CmpVal) {
3031 if (Val < 0 || Val == CmpVal)
3036 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
3037 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
3038 /// the second operand.
3039 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3040 if (VT == MVT::v4f32 || VT == MVT::v4i32 )
3041 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
3042 if (VT == MVT::v2f64 || VT == MVT::v2i64)
3043 return (Mask[0] < 2 && Mask[1] < 2);
3047 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
3048 SmallVector<int, 8> M;
3050 return ::isPSHUFDMask(M, N->getValueType(0));
3053 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
3054 /// is suitable for input to PSHUFHW.
3055 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3056 if (VT != MVT::v8i16)
3059 // Lower quadword copied in order or undef.
3060 for (int i = 0; i != 4; ++i)
3061 if (Mask[i] >= 0 && Mask[i] != i)
3064 // Upper quadword shuffled.
3065 for (int i = 4; i != 8; ++i)
3066 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
3072 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
3073 SmallVector<int, 8> M;
3075 return ::isPSHUFHWMask(M, N->getValueType(0));
3078 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
3079 /// is suitable for input to PSHUFLW.
3080 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3081 if (VT != MVT::v8i16)
3084 // Upper quadword copied in order.
3085 for (int i = 4; i != 8; ++i)
3086 if (Mask[i] >= 0 && Mask[i] != i)
3089 // Lower quadword shuffled.
3090 for (int i = 0; i != 4; ++i)
3097 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
3098 SmallVector<int, 8> M;
3100 return ::isPSHUFLWMask(M, N->getValueType(0));
3103 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
3104 /// is suitable for input to PALIGNR.
3105 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
3107 int i, e = VT.getVectorNumElements();
3108 if (VT.getSizeInBits() != 128 && VT.getSizeInBits() != 64)
3111 // Do not handle v2i64 / v2f64 shuffles with palignr.
3112 if (e < 4 || !hasSSSE3)
3115 for (i = 0; i != e; ++i)
3119 // All undef, not a palignr.
3123 // Make sure we're shifting in the right direction.
3127 int s = Mask[i] - i;
3129 // Check the rest of the elements to see if they are consecutive.
3130 for (++i; i != e; ++i) {
3132 if (m >= 0 && m != s+i)
3138 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
3139 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
3140 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3141 int NumElems = VT.getVectorNumElements();
3142 if (NumElems != 2 && NumElems != 4)
3145 int Half = NumElems / 2;
3146 for (int i = 0; i < Half; ++i)
3147 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3149 for (int i = Half; i < NumElems; ++i)
3150 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3156 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
3157 SmallVector<int, 8> M;
3159 return ::isSHUFPMask(M, N->getValueType(0));
3162 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
3163 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
3164 /// half elements to come from vector 1 (which would equal the dest.) and
3165 /// the upper half to come from vector 2.
3166 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3167 int NumElems = VT.getVectorNumElements();
3169 if (NumElems != 2 && NumElems != 4)
3172 int Half = NumElems / 2;
3173 for (int i = 0; i < Half; ++i)
3174 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3176 for (int i = Half; i < NumElems; ++i)
3177 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3182 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
3183 SmallVector<int, 8> M;
3185 return isCommutedSHUFPMask(M, N->getValueType(0));
3188 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3189 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3190 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
3191 EVT VT = N->getValueType(0);
3192 unsigned NumElems = VT.getVectorNumElements();
3194 if (VT.getSizeInBits() != 128)
3200 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3201 return isUndefOrEqual(N->getMaskElt(0), 6) &&
3202 isUndefOrEqual(N->getMaskElt(1), 7) &&
3203 isUndefOrEqual(N->getMaskElt(2), 2) &&
3204 isUndefOrEqual(N->getMaskElt(3), 3);
3207 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3208 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3210 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
3211 EVT VT = N->getValueType(0);
3212 unsigned NumElems = VT.getVectorNumElements();
3214 if (VT.getSizeInBits() != 128)
3220 return isUndefOrEqual(N->getMaskElt(0), 2) &&
3221 isUndefOrEqual(N->getMaskElt(1), 3) &&
3222 isUndefOrEqual(N->getMaskElt(2), 2) &&
3223 isUndefOrEqual(N->getMaskElt(3), 3);
3226 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3227 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3228 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
3229 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3231 if (NumElems != 2 && NumElems != 4)
3234 for (unsigned i = 0; i < NumElems/2; ++i)
3235 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
3238 for (unsigned i = NumElems/2; i < NumElems; ++i)
3239 if (!isUndefOrEqual(N->getMaskElt(i), i))
3245 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3246 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3247 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
3248 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3250 if ((NumElems != 2 && NumElems != 4)
3251 || N->getValueType(0).getSizeInBits() > 128)
3254 for (unsigned i = 0; i < NumElems/2; ++i)
3255 if (!isUndefOrEqual(N->getMaskElt(i), i))
3258 for (unsigned i = 0; i < NumElems/2; ++i)
3259 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
3265 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3266 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3267 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3268 bool V2IsSplat = false) {
3269 int NumElts = VT.getVectorNumElements();
3271 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3272 "Unsupported vector type for unpckh");
3274 if (VT.getSizeInBits() == 256 && NumElts != 4 && NumElts != 8)
3277 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3278 // independently on 128-bit lanes.
3279 unsigned NumLanes = VT.getSizeInBits()/128;
3280 unsigned NumLaneElts = NumElts/NumLanes;
3283 unsigned End = NumLaneElts;
3284 for (unsigned s = 0; s < NumLanes; ++s) {
3285 for (unsigned i = Start, j = s * NumLaneElts;
3289 int BitI1 = Mask[i+1];
3290 if (!isUndefOrEqual(BitI, j))
3293 if (!isUndefOrEqual(BitI1, NumElts))
3296 if (!isUndefOrEqual(BitI1, j + NumElts))
3300 // Process the next 128 bits.
3301 Start += NumLaneElts;
3308 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3309 SmallVector<int, 8> M;
3311 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
3314 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3315 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3316 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3317 bool V2IsSplat = false) {
3318 int NumElts = VT.getVectorNumElements();
3320 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3321 "Unsupported vector type for unpckh");
3323 if (VT.getSizeInBits() == 256 && NumElts != 4 && NumElts != 8)
3326 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3327 // independently on 128-bit lanes.
3328 unsigned NumLanes = VT.getSizeInBits()/128;
3329 unsigned NumLaneElts = NumElts/NumLanes;
3332 unsigned End = NumLaneElts;
3333 for (unsigned l = 0; l != NumLanes; ++l) {
3334 for (unsigned i = Start, j = (l*NumLaneElts)+NumLaneElts/2;
3335 i != End; i += 2, ++j) {
3337 int BitI1 = Mask[i+1];
3338 if (!isUndefOrEqual(BitI, j))
3341 if (isUndefOrEqual(BitI1, NumElts))
3344 if (!isUndefOrEqual(BitI1, j+NumElts))
3348 // Process the next 128 bits.
3349 Start += NumLaneElts;
3355 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3356 SmallVector<int, 8> M;
3358 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3361 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3362 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3364 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3365 int NumElems = VT.getVectorNumElements();
3366 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3369 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3370 // independently on 128-bit lanes.
3371 unsigned NumLanes = VT.getSizeInBits() / 128;
3372 unsigned NumLaneElts = NumElems / NumLanes;
3374 for (unsigned s = 0; s < NumLanes; ++s) {
3375 for (unsigned i = s * NumLaneElts, j = s * NumLaneElts;
3376 i != NumLaneElts * (s + 1);
3379 int BitI1 = Mask[i+1];
3381 if (!isUndefOrEqual(BitI, j))
3383 if (!isUndefOrEqual(BitI1, j))
3391 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3392 SmallVector<int, 8> M;
3394 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3397 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3398 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3400 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3401 int NumElems = VT.getVectorNumElements();
3402 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3405 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3407 int BitI1 = Mask[i+1];
3408 if (!isUndefOrEqual(BitI, j))
3410 if (!isUndefOrEqual(BitI1, j))
3416 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3417 SmallVector<int, 8> M;
3419 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3422 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3423 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3424 /// MOVSD, and MOVD, i.e. setting the lowest element.
3425 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3426 if (VT.getVectorElementType().getSizeInBits() < 32)
3429 int NumElts = VT.getVectorNumElements();
3431 if (!isUndefOrEqual(Mask[0], NumElts))
3434 for (int i = 1; i < NumElts; ++i)
3435 if (!isUndefOrEqual(Mask[i], i))
3441 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3442 SmallVector<int, 8> M;
3444 return ::isMOVLMask(M, N->getValueType(0));
3447 /// isVPERMILPDMask - Return true if the specified VECTOR_SHUFFLE operand
3448 /// specifies a shuffle of elements that is suitable for input to VPERMILPD*.
3449 /// Note that VPERMIL mask matching is different depending whether theunderlying
3450 /// type is 32 or 64. In the VPERMILPS the high half of the mask should point
3451 /// to the same elements of the low, but to the higher half of the source.
3452 /// In VPERMILPD the two lanes could be shuffled independently of each other
3453 /// with the same restriction that lanes can't be crossed.
3454 static bool isVPERMILPDMask(const SmallVectorImpl<int> &Mask, EVT VT,
3455 const X86Subtarget *Subtarget) {
3456 int NumElts = VT.getVectorNumElements();
3457 int NumLanes = VT.getSizeInBits()/128;
3459 if (!Subtarget->hasAVX())
3462 // Match any permutation of 128-bit vector with 64-bit types
3463 if (NumLanes == 1 && NumElts != 2)
3466 // Only match 256-bit with 32 types
3467 if (VT.getSizeInBits() == 256 && NumElts != 4)
3470 // The mask on the high lane is independent of the low. Both can match
3471 // any element in inside its own lane, but can't cross.
3472 int LaneSize = NumElts/NumLanes;
3473 for (int l = 0; l < NumLanes; ++l)
3474 for (int i = l*LaneSize; i < LaneSize*(l+1); ++i) {
3475 int LaneStart = l*LaneSize;
3476 if (!isUndefOrInRange(Mask[i], LaneStart, LaneStart+LaneSize))
3483 /// isVPERMILPSMask - Return true if the specified VECTOR_SHUFFLE operand
3484 /// specifies a shuffle of elements that is suitable for input to VPERMILPS*.
3485 /// Note that VPERMIL mask matching is different depending whether theunderlying
3486 /// type is 32 or 64. In the VPERMILPS the high half of the mask should point
3487 /// to the same elements of the low, but to the higher half of the source.
3488 /// In VPERMILPD the two lanes could be shuffled independently of each other
3489 /// with the same restriction that lanes can't be crossed.
3490 static bool isVPERMILPSMask(const SmallVectorImpl<int> &Mask, EVT VT,
3491 const X86Subtarget *Subtarget) {
3492 unsigned NumElts = VT.getVectorNumElements();
3493 unsigned NumLanes = VT.getSizeInBits()/128;
3495 if (!Subtarget->hasAVX())
3498 // Match any permutation of 128-bit vector with 32-bit types
3499 if (NumLanes == 1 && NumElts != 4)
3502 // Only match 256-bit with 32 types
3503 if (VT.getSizeInBits() == 256 && NumElts != 8)
3506 // The mask on the high lane should be the same as the low. Actually,
3507 // they can differ if any of the corresponding index in a lane is undef
3508 // and the other stays in range.
3509 int LaneSize = NumElts/NumLanes;
3510 for (int i = 0; i < LaneSize; ++i) {
3511 int HighElt = i+LaneSize;
3512 if (Mask[i] < 0 && (isUndefOrInRange(Mask[HighElt], LaneSize, NumElts)))
3514 if (Mask[HighElt] < 0 && (isUndefOrInRange(Mask[i], 0, LaneSize)))
3516 if (Mask[HighElt]-Mask[i] != LaneSize)
3523 /// getShuffleVPERMILPSImmediate - Return the appropriate immediate to shuffle
3524 /// the specified VECTOR_MASK mask with VPERMILPS* instructions.
3525 static unsigned getShuffleVPERMILPSImmediate(SDNode *N) {
3526 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3527 EVT VT = SVOp->getValueType(0);
3529 int NumElts = VT.getVectorNumElements();
3530 int NumLanes = VT.getSizeInBits()/128;
3531 int LaneSize = NumElts/NumLanes;
3533 // Although the mask is equal for both lanes do it twice to get the cases
3534 // where a mask will match because the same mask element is undef on the
3535 // first half but valid on the second. This would get pathological cases
3536 // such as: shuffle <u, 0, 1, 2, 4, 4, 5, 6>, which is completely valid.
3538 for (int l = 0; l < NumLanes; ++l) {
3539 for (int i = 0; i < LaneSize; ++i) {
3540 int MaskElt = SVOp->getMaskElt(i+(l*LaneSize));
3543 if (MaskElt >= LaneSize)
3544 MaskElt -= LaneSize;
3545 Mask |= MaskElt << (i*2);
3552 /// getShuffleVPERMILPDImmediate - Return the appropriate immediate to shuffle
3553 /// the specified VECTOR_MASK mask with VPERMILPD* instructions.
3554 static unsigned getShuffleVPERMILPDImmediate(SDNode *N) {
3555 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3556 EVT VT = SVOp->getValueType(0);
3558 int NumElts = VT.getVectorNumElements();
3559 int NumLanes = VT.getSizeInBits()/128;
3562 int LaneSize = NumElts/NumLanes;
3563 for (int l = 0; l < NumLanes; ++l)
3564 for (int i = l*LaneSize; i < LaneSize*(l+1); ++i) {
3565 int MaskElt = SVOp->getMaskElt(i);
3568 Mask |= (MaskElt-l*LaneSize) << i;
3574 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3575 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3576 /// element of vector 2 and the other elements to come from vector 1 in order.
3577 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3578 bool V2IsSplat = false, bool V2IsUndef = false) {
3579 int NumOps = VT.getVectorNumElements();
3580 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3583 if (!isUndefOrEqual(Mask[0], 0))
3586 for (int i = 1; i < NumOps; ++i)
3587 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3588 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3589 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3595 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3596 bool V2IsUndef = false) {
3597 SmallVector<int, 8> M;
3599 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3602 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3603 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3604 /// Masks to match: <1, 1, 3, 3> or <1, 1, 3, 3, 5, 5, 7, 7>
3605 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N,
3606 const X86Subtarget *Subtarget) {
3607 if (!Subtarget->hasSSE3() && !Subtarget->hasAVX())
3610 // The second vector must be undef
3611 if (N->getOperand(1).getOpcode() != ISD::UNDEF)
3614 EVT VT = N->getValueType(0);
3615 unsigned NumElems = VT.getVectorNumElements();
3617 if ((VT.getSizeInBits() == 128 && NumElems != 4) ||
3618 (VT.getSizeInBits() == 256 && NumElems != 8))
3621 // "i+1" is the value the indexed mask element must have
3622 for (unsigned i = 0; i < NumElems; i += 2)
3623 if (!isUndefOrEqual(N->getMaskElt(i), i+1) ||
3624 !isUndefOrEqual(N->getMaskElt(i+1), i+1))
3630 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3631 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3632 /// Masks to match: <0, 0, 2, 2> or <0, 0, 2, 2, 4, 4, 6, 6>
3633 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N,
3634 const X86Subtarget *Subtarget) {
3635 if (!Subtarget->hasSSE3() && !Subtarget->hasAVX())
3638 // The second vector must be undef
3639 if (N->getOperand(1).getOpcode() != ISD::UNDEF)
3642 EVT VT = N->getValueType(0);
3643 unsigned NumElems = VT.getVectorNumElements();
3645 if ((VT.getSizeInBits() == 128 && NumElems != 4) ||
3646 (VT.getSizeInBits() == 256 && NumElems != 8))
3649 // "i" is the value the indexed mask element must have
3650 for (unsigned i = 0; i < NumElems; i += 2)
3651 if (!isUndefOrEqual(N->getMaskElt(i), i) ||
3652 !isUndefOrEqual(N->getMaskElt(i+1), i))
3658 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3659 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3660 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3661 int e = N->getValueType(0).getVectorNumElements() / 2;
3663 for (int i = 0; i < e; ++i)
3664 if (!isUndefOrEqual(N->getMaskElt(i), i))
3666 for (int i = 0; i < e; ++i)
3667 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3672 /// isVEXTRACTF128Index - Return true if the specified
3673 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
3674 /// suitable for input to VEXTRACTF128.
3675 bool X86::isVEXTRACTF128Index(SDNode *N) {
3676 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3679 // The index should be aligned on a 128-bit boundary.
3681 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3683 unsigned VL = N->getValueType(0).getVectorNumElements();
3684 unsigned VBits = N->getValueType(0).getSizeInBits();
3685 unsigned ElSize = VBits / VL;
3686 bool Result = (Index * ElSize) % 128 == 0;
3691 /// isVINSERTF128Index - Return true if the specified INSERT_SUBVECTOR
3692 /// operand specifies a subvector insert that is suitable for input to
3694 bool X86::isVINSERTF128Index(SDNode *N) {
3695 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3698 // The index should be aligned on a 128-bit boundary.
3700 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3702 unsigned VL = N->getValueType(0).getVectorNumElements();
3703 unsigned VBits = N->getValueType(0).getSizeInBits();
3704 unsigned ElSize = VBits / VL;
3705 bool Result = (Index * ElSize) % 128 == 0;
3710 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3711 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3712 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3713 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3714 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3716 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3718 for (int i = 0; i < NumOperands; ++i) {
3719 int Val = SVOp->getMaskElt(NumOperands-i-1);
3720 if (Val < 0) Val = 0;
3721 if (Val >= NumOperands) Val -= NumOperands;
3723 if (i != NumOperands - 1)
3729 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3730 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3731 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3732 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3734 // 8 nodes, but we only care about the last 4.
3735 for (unsigned i = 7; i >= 4; --i) {
3736 int Val = SVOp->getMaskElt(i);
3745 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3746 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3747 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3748 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3750 // 8 nodes, but we only care about the first 4.
3751 for (int i = 3; i >= 0; --i) {
3752 int Val = SVOp->getMaskElt(i);
3761 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3762 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3763 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3764 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3765 EVT VVT = N->getValueType(0);
3766 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3770 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3771 Val = SVOp->getMaskElt(i);
3775 assert(Val - i > 0 && "PALIGNR imm should be positive");
3776 return (Val - i) * EltSize;
3779 /// getExtractVEXTRACTF128Immediate - Return the appropriate immediate
3780 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
3782 unsigned X86::getExtractVEXTRACTF128Immediate(SDNode *N) {
3783 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3784 llvm_unreachable("Illegal extract subvector for VEXTRACTF128");
3787 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3789 EVT VecVT = N->getOperand(0).getValueType();
3790 EVT ElVT = VecVT.getVectorElementType();
3792 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
3793 return Index / NumElemsPerChunk;
3796 /// getInsertVINSERTF128Immediate - Return the appropriate immediate
3797 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
3799 unsigned X86::getInsertVINSERTF128Immediate(SDNode *N) {
3800 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3801 llvm_unreachable("Illegal insert subvector for VINSERTF128");
3804 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3806 EVT VecVT = N->getValueType(0);
3807 EVT ElVT = VecVT.getVectorElementType();
3809 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
3810 return Index / NumElemsPerChunk;
3813 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3815 bool X86::isZeroNode(SDValue Elt) {
3816 return ((isa<ConstantSDNode>(Elt) &&
3817 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3818 (isa<ConstantFPSDNode>(Elt) &&
3819 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3822 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3823 /// their permute mask.
3824 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3825 SelectionDAG &DAG) {
3826 EVT VT = SVOp->getValueType(0);
3827 unsigned NumElems = VT.getVectorNumElements();
3828 SmallVector<int, 8> MaskVec;
3830 for (unsigned i = 0; i != NumElems; ++i) {
3831 int idx = SVOp->getMaskElt(i);
3833 MaskVec.push_back(idx);
3834 else if (idx < (int)NumElems)
3835 MaskVec.push_back(idx + NumElems);
3837 MaskVec.push_back(idx - NumElems);
3839 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3840 SVOp->getOperand(0), &MaskVec[0]);
3843 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3844 /// the two vector operands have swapped position.
3845 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3846 unsigned NumElems = VT.getVectorNumElements();
3847 for (unsigned i = 0; i != NumElems; ++i) {
3851 else if (idx < (int)NumElems)
3852 Mask[i] = idx + NumElems;
3854 Mask[i] = idx - NumElems;
3858 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3859 /// match movhlps. The lower half elements should come from upper half of
3860 /// V1 (and in order), and the upper half elements should come from the upper
3861 /// half of V2 (and in order).
3862 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3863 if (Op->getValueType(0).getVectorNumElements() != 4)
3865 for (unsigned i = 0, e = 2; i != e; ++i)
3866 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3868 for (unsigned i = 2; i != 4; ++i)
3869 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3874 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3875 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3877 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3878 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3880 N = N->getOperand(0).getNode();
3881 if (!ISD::isNON_EXTLoad(N))
3884 *LD = cast<LoadSDNode>(N);
3888 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3889 /// match movlp{s|d}. The lower half elements should come from lower half of
3890 /// V1 (and in order), and the upper half elements should come from the upper
3891 /// half of V2 (and in order). And since V1 will become the source of the
3892 /// MOVLP, it must be either a vector load or a scalar load to vector.
3893 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3894 ShuffleVectorSDNode *Op) {
3895 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3897 // Is V2 is a vector load, don't do this transformation. We will try to use
3898 // load folding shufps op.
3899 if (ISD::isNON_EXTLoad(V2))
3902 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3904 if (NumElems != 2 && NumElems != 4)
3906 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3907 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3909 for (unsigned i = NumElems/2; i != NumElems; ++i)
3910 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3915 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3917 static bool isSplatVector(SDNode *N) {
3918 if (N->getOpcode() != ISD::BUILD_VECTOR)
3921 SDValue SplatValue = N->getOperand(0);
3922 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3923 if (N->getOperand(i) != SplatValue)
3928 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3929 /// to an zero vector.
3930 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3931 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3932 SDValue V1 = N->getOperand(0);
3933 SDValue V2 = N->getOperand(1);
3934 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3935 for (unsigned i = 0; i != NumElems; ++i) {
3936 int Idx = N->getMaskElt(i);
3937 if (Idx >= (int)NumElems) {
3938 unsigned Opc = V2.getOpcode();
3939 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3941 if (Opc != ISD::BUILD_VECTOR ||
3942 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3944 } else if (Idx >= 0) {
3945 unsigned Opc = V1.getOpcode();
3946 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3948 if (Opc != ISD::BUILD_VECTOR ||
3949 !X86::isZeroNode(V1.getOperand(Idx)))
3956 /// getZeroVector - Returns a vector of specified type with all zero elements.
3958 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3960 assert(VT.isVector() && "Expected a vector type");
3962 // Always build SSE zero vectors as <4 x i32> bitcasted
3963 // to their dest type. This ensures they get CSE'd.
3965 if (VT.getSizeInBits() == 128) { // SSE
3966 if (HasSSE2) { // SSE2
3967 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3968 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3970 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3971 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3973 } else if (VT.getSizeInBits() == 256) { // AVX
3974 // 256-bit logic and arithmetic instructions in AVX are
3975 // all floating-point, no support for integer ops. Default
3976 // to emitting fp zeroed vectors then.
3977 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3978 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3979 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
3981 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
3984 /// getOnesVector - Returns a vector of specified type with all bits set.
3985 /// Always build ones vectors as <4 x i32>. For 256-bit types, use two
3986 /// <4 x i32> inserted in a <8 x i32> appropriately. Then bitcast to their
3987 /// original type, ensuring they get CSE'd.
3988 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3989 assert(VT.isVector() && "Expected a vector type");
3990 assert((VT.is128BitVector() || VT.is256BitVector())
3991 && "Expected a 128-bit or 256-bit vector type");
3993 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3994 SDValue Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
3995 Cst, Cst, Cst, Cst);
3997 if (VT.is256BitVector()) {
3998 SDValue InsV = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, MVT::v8i32),
3999 Vec, DAG.getConstant(0, MVT::i32), DAG, dl);
4000 Vec = Insert128BitVector(InsV, Vec,
4001 DAG.getConstant(4 /* NumElems/2 */, MVT::i32), DAG, dl);
4004 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4007 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
4008 /// that point to V2 points to its first element.
4009 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4010 EVT VT = SVOp->getValueType(0);
4011 unsigned NumElems = VT.getVectorNumElements();
4013 bool Changed = false;
4014 SmallVector<int, 8> MaskVec;
4015 SVOp->getMask(MaskVec);
4017 for (unsigned i = 0; i != NumElems; ++i) {
4018 if (MaskVec[i] > (int)NumElems) {
4019 MaskVec[i] = NumElems;
4024 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
4025 SVOp->getOperand(1), &MaskVec[0]);
4026 return SDValue(SVOp, 0);
4029 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
4030 /// operation of specified width.
4031 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
4033 unsigned NumElems = VT.getVectorNumElements();
4034 SmallVector<int, 8> Mask;
4035 Mask.push_back(NumElems);
4036 for (unsigned i = 1; i != NumElems; ++i)
4038 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4041 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
4042 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
4044 unsigned NumElems = VT.getVectorNumElements();
4045 SmallVector<int, 8> Mask;
4046 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4048 Mask.push_back(i + NumElems);
4050 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4053 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
4054 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
4056 unsigned NumElems = VT.getVectorNumElements();
4057 unsigned Half = NumElems/2;
4058 SmallVector<int, 8> Mask;
4059 for (unsigned i = 0; i != Half; ++i) {
4060 Mask.push_back(i + Half);
4061 Mask.push_back(i + NumElems + Half);
4063 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4066 // PromoteSplatv8v16 - All i16 and i8 vector types can't be used directly by
4067 // a generic shuffle instruction because the target has no such instructions.
4068 // Generate shuffles which repeat i16 and i8 several times until they can be
4069 // represented by v4f32 and then be manipulated by target suported shuffles.
4070 static SDValue PromoteSplatv8v16(SDValue V, SelectionDAG &DAG, int &EltNo) {
4071 EVT VT = V.getValueType();
4072 int NumElems = VT.getVectorNumElements();
4073 DebugLoc dl = V.getDebugLoc();
4075 while (NumElems > 4) {
4076 if (EltNo < NumElems/2) {
4077 V = getUnpackl(DAG, dl, VT, V, V);
4079 V = getUnpackh(DAG, dl, VT, V, V);
4080 EltNo -= NumElems/2;
4087 /// getLegalSplat - Generate a legal splat with supported x86 shuffles
4088 static SDValue getLegalSplat(SelectionDAG &DAG, SDValue V, int EltNo) {
4089 EVT VT = V.getValueType();
4090 DebugLoc dl = V.getDebugLoc();
4091 assert((VT.getSizeInBits() == 128 || VT.getSizeInBits() == 256)
4092 && "Vector size not supported");
4094 bool Is128 = VT.getSizeInBits() == 128;
4095 EVT NVT = Is128 ? MVT::v4f32 : MVT::v8f32;
4096 V = DAG.getNode(ISD::BITCAST, dl, NVT, V);
4099 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
4100 V = DAG.getVectorShuffle(NVT, dl, V, DAG.getUNDEF(NVT), &SplatMask[0]);
4102 // The second half of indicies refer to the higher part, which is a
4103 // duplication of the lower one. This makes this shuffle a perfect match
4104 // for the VPERM instruction.
4105 int SplatMask[8] = { EltNo, EltNo, EltNo, EltNo,
4106 EltNo+4, EltNo+4, EltNo+4, EltNo+4 };
4107 V = DAG.getVectorShuffle(NVT, dl, V, DAG.getUNDEF(NVT), &SplatMask[0]);
4110 return DAG.getNode(ISD::BITCAST, dl, VT, V);
4113 /// PromoteVectorToScalarSplat - Since there's no native support for
4114 /// scalar_to_vector for 256-bit AVX, a 128-bit scalar_to_vector +
4115 /// INSERT_SUBVECTOR is generated. Recognize this idiom and do the
4116 /// shuffle before the insertion, this yields less instructions in the end.
4117 static SDValue PromoteVectorToScalarSplat(ShuffleVectorSDNode *SV,
4118 SelectionDAG &DAG) {
4119 EVT SrcVT = SV->getValueType(0);
4120 SDValue V1 = SV->getOperand(0);
4121 DebugLoc dl = SV->getDebugLoc();
4122 int NumElems = SrcVT.getVectorNumElements();
4124 assert(SrcVT.is256BitVector() && "unknown howto handle vector type");
4126 SmallVector<int, 4> Mask;
4127 for (int i = 0; i < NumElems/2; ++i)
4128 Mask.push_back(SV->getMaskElt(i));
4130 EVT SVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
4132 SDValue SV1 = DAG.getVectorShuffle(SVT, dl, V1.getOperand(1),
4133 DAG.getUNDEF(SVT), &Mask[0]);
4134 SDValue InsV = Insert128BitVector(DAG.getUNDEF(SrcVT), SV1,
4135 DAG.getConstant(0, MVT::i32), DAG, dl);
4137 return Insert128BitVector(InsV, SV1,
4138 DAG.getConstant(NumElems/2, MVT::i32), DAG, dl);
4141 /// PromoteSplat - Promote a splat of v4i32, v8i16 or v16i8 to v4f32 and
4142 /// v8i32, v16i16 or v32i8 to v8f32.
4143 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
4144 EVT SrcVT = SV->getValueType(0);
4145 SDValue V1 = SV->getOperand(0);
4146 DebugLoc dl = SV->getDebugLoc();
4148 int EltNo = SV->getSplatIndex();
4149 int NumElems = SrcVT.getVectorNumElements();
4150 unsigned Size = SrcVT.getSizeInBits();
4152 // Extract the 128-bit part containing the splat element and update
4153 // the splat element index when it refers to the higher register.
4155 unsigned Idx = (EltNo > NumElems/2) ? NumElems/2 : 0;
4156 V1 = Extract128BitVector(V1, DAG.getConstant(Idx, MVT::i32), DAG, dl);
4158 EltNo -= NumElems/2;
4161 // Make this 128-bit vector duplicate i8 and i16 elements
4163 V1 = PromoteSplatv8v16(V1, DAG, EltNo);
4165 // Recreate the 256-bit vector and place the same 128-bit vector
4166 // into the low and high part. This is necessary because we want
4167 // to use VPERM to shuffle the v8f32 vector, and VPERM only shuffles
4168 // inside each separate v4f32 lane.
4170 SDValue InsV = Insert128BitVector(DAG.getUNDEF(SrcVT), V1,
4171 DAG.getConstant(0, MVT::i32), DAG, dl);
4172 V1 = Insert128BitVector(InsV, V1,
4173 DAG.getConstant(NumElems/2, MVT::i32), DAG, dl);
4176 return getLegalSplat(DAG, V1, EltNo);
4179 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
4180 /// vector of zero or undef vector. This produces a shuffle where the low
4181 /// element of V2 is swizzled into the zero/undef vector, landing at element
4182 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4183 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4184 bool isZero, bool HasSSE2,
4185 SelectionDAG &DAG) {
4186 EVT VT = V2.getValueType();
4188 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
4189 unsigned NumElems = VT.getVectorNumElements();
4190 SmallVector<int, 16> MaskVec;
4191 for (unsigned i = 0; i != NumElems; ++i)
4192 // If this is the insertion idx, put the low elt of V2 here.
4193 MaskVec.push_back(i == Idx ? NumElems : i);
4194 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
4197 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4198 /// element of the result of the vector shuffle.
4199 static SDValue getShuffleScalarElt(SDNode *N, int Index, SelectionDAG &DAG,
4202 return SDValue(); // Limit search depth.
4204 SDValue V = SDValue(N, 0);
4205 EVT VT = V.getValueType();
4206 unsigned Opcode = V.getOpcode();
4208 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4209 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4210 Index = SV->getMaskElt(Index);
4213 return DAG.getUNDEF(VT.getVectorElementType());
4215 int NumElems = VT.getVectorNumElements();
4216 SDValue NewV = (Index < NumElems) ? SV->getOperand(0) : SV->getOperand(1);
4217 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG, Depth+1);
4220 // Recurse into target specific vector shuffles to find scalars.
4221 if (isTargetShuffle(Opcode)) {
4222 int NumElems = VT.getVectorNumElements();
4223 SmallVector<unsigned, 16> ShuffleMask;
4227 case X86ISD::SHUFPS:
4228 case X86ISD::SHUFPD:
4229 ImmN = N->getOperand(N->getNumOperands()-1);
4230 DecodeSHUFPSMask(NumElems,
4231 cast<ConstantSDNode>(ImmN)->getZExtValue(),
4234 case X86ISD::PUNPCKHBW:
4235 case X86ISD::PUNPCKHWD:
4236 case X86ISD::PUNPCKHDQ:
4237 case X86ISD::PUNPCKHQDQ:
4238 DecodePUNPCKHMask(NumElems, ShuffleMask);
4240 case X86ISD::UNPCKHPS:
4241 case X86ISD::UNPCKHPD:
4242 case X86ISD::VUNPCKHPSY:
4243 case X86ISD::VUNPCKHPDY:
4244 DecodeUNPCKHPMask(NumElems, ShuffleMask);
4246 case X86ISD::PUNPCKLBW:
4247 case X86ISD::PUNPCKLWD:
4248 case X86ISD::PUNPCKLDQ:
4249 case X86ISD::PUNPCKLQDQ:
4250 DecodePUNPCKLMask(VT, ShuffleMask);
4252 case X86ISD::UNPCKLPS:
4253 case X86ISD::UNPCKLPD:
4254 case X86ISD::VUNPCKLPSY:
4255 case X86ISD::VUNPCKLPDY:
4256 DecodeUNPCKLPMask(VT, ShuffleMask);
4258 case X86ISD::MOVHLPS:
4259 DecodeMOVHLPSMask(NumElems, ShuffleMask);
4261 case X86ISD::MOVLHPS:
4262 DecodeMOVLHPSMask(NumElems, ShuffleMask);
4264 case X86ISD::PSHUFD:
4265 ImmN = N->getOperand(N->getNumOperands()-1);
4266 DecodePSHUFMask(NumElems,
4267 cast<ConstantSDNode>(ImmN)->getZExtValue(),
4270 case X86ISD::PSHUFHW:
4271 ImmN = N->getOperand(N->getNumOperands()-1);
4272 DecodePSHUFHWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
4275 case X86ISD::PSHUFLW:
4276 ImmN = N->getOperand(N->getNumOperands()-1);
4277 DecodePSHUFLWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
4281 case X86ISD::MOVSD: {
4282 // The index 0 always comes from the first element of the second source,
4283 // this is why MOVSS and MOVSD are used in the first place. The other
4284 // elements come from the other positions of the first source vector.
4285 unsigned OpNum = (Index == 0) ? 1 : 0;
4286 return getShuffleScalarElt(V.getOperand(OpNum).getNode(), Index, DAG,
4289 case X86ISD::VPERMILPS:
4290 ImmN = N->getOperand(N->getNumOperands()-1);
4291 DecodeVPERMILPSMask(4, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4294 case X86ISD::VPERMILPSY:
4295 ImmN = N->getOperand(N->getNumOperands()-1);
4296 DecodeVPERMILPSMask(8, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4299 case X86ISD::VPERMILPD:
4300 ImmN = N->getOperand(N->getNumOperands()-1);
4301 DecodeVPERMILPDMask(2, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4304 case X86ISD::VPERMILPDY:
4305 ImmN = N->getOperand(N->getNumOperands()-1);
4306 DecodeVPERMILPDMask(4, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4310 assert("not implemented for target shuffle node");
4314 Index = ShuffleMask[Index];
4316 return DAG.getUNDEF(VT.getVectorElementType());
4318 SDValue NewV = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
4319 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG,
4323 // Actual nodes that may contain scalar elements
4324 if (Opcode == ISD::BITCAST) {
4325 V = V.getOperand(0);
4326 EVT SrcVT = V.getValueType();
4327 unsigned NumElems = VT.getVectorNumElements();
4329 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4333 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4334 return (Index == 0) ? V.getOperand(0)
4335 : DAG.getUNDEF(VT.getVectorElementType());
4337 if (V.getOpcode() == ISD::BUILD_VECTOR)
4338 return V.getOperand(Index);
4343 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
4344 /// shuffle operation which come from a consecutively from a zero. The
4345 /// search can start in two different directions, from left or right.
4347 unsigned getNumOfConsecutiveZeros(SDNode *N, int NumElems,
4348 bool ZerosFromLeft, SelectionDAG &DAG) {
4351 while (i < NumElems) {
4352 unsigned Index = ZerosFromLeft ? i : NumElems-i-1;
4353 SDValue Elt = getShuffleScalarElt(N, Index, DAG, 0);
4354 if (!(Elt.getNode() &&
4355 (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt))))
4363 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies from MaskI to
4364 /// MaskE correspond consecutively to elements from one of the vector operands,
4365 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
4367 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp, int MaskI, int MaskE,
4368 int OpIdx, int NumElems, unsigned &OpNum) {
4369 bool SeenV1 = false;
4370 bool SeenV2 = false;
4372 for (int i = MaskI; i <= MaskE; ++i, ++OpIdx) {
4373 int Idx = SVOp->getMaskElt(i);
4374 // Ignore undef indicies
4383 // Only accept consecutive elements from the same vector
4384 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
4388 OpNum = SeenV1 ? 0 : 1;
4392 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
4393 /// logical left shift of a vector.
4394 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4395 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4396 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4397 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4398 false /* check zeros from right */, DAG);
4404 // Considering the elements in the mask that are not consecutive zeros,
4405 // check if they consecutively come from only one of the source vectors.
4407 // V1 = {X, A, B, C} 0
4409 // vector_shuffle V1, V2 <1, 2, 3, X>
4411 if (!isShuffleMaskConsecutive(SVOp,
4412 0, // Mask Start Index
4413 NumElems-NumZeros-1, // Mask End Index
4414 NumZeros, // Where to start looking in the src vector
4415 NumElems, // Number of elements in vector
4416 OpSrc)) // Which source operand ?
4421 ShVal = SVOp->getOperand(OpSrc);
4425 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
4426 /// logical left shift of a vector.
4427 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4428 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4429 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4430 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4431 true /* check zeros from left */, DAG);
4437 // Considering the elements in the mask that are not consecutive zeros,
4438 // check if they consecutively come from only one of the source vectors.
4440 // 0 { A, B, X, X } = V2
4442 // vector_shuffle V1, V2 <X, X, 4, 5>
4444 if (!isShuffleMaskConsecutive(SVOp,
4445 NumZeros, // Mask Start Index
4446 NumElems-1, // Mask End Index
4447 0, // Where to start looking in the src vector
4448 NumElems, // Number of elements in vector
4449 OpSrc)) // Which source operand ?
4454 ShVal = SVOp->getOperand(OpSrc);
4458 /// isVectorShift - Returns true if the shuffle can be implemented as a
4459 /// logical left or right shift of a vector.
4460 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4461 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4462 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
4463 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
4469 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
4471 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4472 unsigned NumNonZero, unsigned NumZero,
4474 const TargetLowering &TLI) {
4478 DebugLoc dl = Op.getDebugLoc();
4481 for (unsigned i = 0; i < 16; ++i) {
4482 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4483 if (ThisIsNonZero && First) {
4485 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4487 V = DAG.getUNDEF(MVT::v8i16);
4492 SDValue ThisElt(0, 0), LastElt(0, 0);
4493 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4494 if (LastIsNonZero) {
4495 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4496 MVT::i16, Op.getOperand(i-1));
4498 if (ThisIsNonZero) {
4499 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4500 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4501 ThisElt, DAG.getConstant(8, MVT::i8));
4503 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4507 if (ThisElt.getNode())
4508 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4509 DAG.getIntPtrConstant(i/2));
4513 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
4516 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
4518 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4519 unsigned NumNonZero, unsigned NumZero,
4521 const TargetLowering &TLI) {
4525 DebugLoc dl = Op.getDebugLoc();
4528 for (unsigned i = 0; i < 8; ++i) {
4529 bool isNonZero = (NonZeros & (1 << i)) != 0;
4533 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4535 V = DAG.getUNDEF(MVT::v8i16);
4538 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4539 MVT::v8i16, V, Op.getOperand(i),
4540 DAG.getIntPtrConstant(i));
4547 /// getVShift - Return a vector logical shift node.
4549 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4550 unsigned NumBits, SelectionDAG &DAG,
4551 const TargetLowering &TLI, DebugLoc dl) {
4552 EVT ShVT = MVT::v2i64;
4553 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
4554 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
4555 return DAG.getNode(ISD::BITCAST, dl, VT,
4556 DAG.getNode(Opc, dl, ShVT, SrcOp,
4557 DAG.getConstant(NumBits,
4558 TLI.getShiftAmountTy(SrcOp.getValueType()))));
4562 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
4563 SelectionDAG &DAG) const {
4565 // Check if the scalar load can be widened into a vector load. And if
4566 // the address is "base + cst" see if the cst can be "absorbed" into
4567 // the shuffle mask.
4568 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4569 SDValue Ptr = LD->getBasePtr();
4570 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4572 EVT PVT = LD->getValueType(0);
4573 if (PVT != MVT::i32 && PVT != MVT::f32)
4578 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4579 FI = FINode->getIndex();
4581 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4582 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4583 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4584 Offset = Ptr.getConstantOperandVal(1);
4585 Ptr = Ptr.getOperand(0);
4590 // FIXME: 256-bit vector instructions don't require a strict alignment,
4591 // improve this code to support it better.
4592 unsigned RequiredAlign = VT.getSizeInBits()/8;
4593 SDValue Chain = LD->getChain();
4594 // Make sure the stack object alignment is at least 16 or 32.
4595 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4596 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
4597 if (MFI->isFixedObjectIndex(FI)) {
4598 // Can't change the alignment. FIXME: It's possible to compute
4599 // the exact stack offset and reference FI + adjust offset instead.
4600 // If someone *really* cares about this. That's the way to implement it.
4603 MFI->setObjectAlignment(FI, RequiredAlign);
4607 // (Offset % 16 or 32) must be multiple of 4. Then address is then
4608 // Ptr + (Offset & ~15).
4611 if ((Offset % RequiredAlign) & 3)
4613 int64_t StartOffset = Offset & ~(RequiredAlign-1);
4615 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
4616 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4618 int EltNo = (Offset - StartOffset) >> 2;
4619 int NumElems = VT.getVectorNumElements();
4621 EVT CanonVT = VT.getSizeInBits() == 128 ? MVT::v4i32 : MVT::v8i32;
4622 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
4623 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
4624 LD->getPointerInfo().getWithOffset(StartOffset),
4627 // Canonicalize it to a v4i32 or v8i32 shuffle.
4628 SmallVector<int, 8> Mask;
4629 for (int i = 0; i < NumElems; ++i)
4630 Mask.push_back(EltNo);
4632 V1 = DAG.getNode(ISD::BITCAST, dl, CanonVT, V1);
4633 return DAG.getNode(ISD::BITCAST, dl, NVT,
4634 DAG.getVectorShuffle(CanonVT, dl, V1,
4635 DAG.getUNDEF(CanonVT),&Mask[0]));
4641 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
4642 /// vector of type 'VT', see if the elements can be replaced by a single large
4643 /// load which has the same value as a build_vector whose operands are 'elts'.
4645 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4647 /// FIXME: we'd also like to handle the case where the last elements are zero
4648 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4649 /// There's even a handy isZeroNode for that purpose.
4650 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
4651 DebugLoc &DL, SelectionDAG &DAG) {
4652 EVT EltVT = VT.getVectorElementType();
4653 unsigned NumElems = Elts.size();
4655 LoadSDNode *LDBase = NULL;
4656 unsigned LastLoadedElt = -1U;
4658 // For each element in the initializer, see if we've found a load or an undef.
4659 // If we don't find an initial load element, or later load elements are
4660 // non-consecutive, bail out.
4661 for (unsigned i = 0; i < NumElems; ++i) {
4662 SDValue Elt = Elts[i];
4664 if (!Elt.getNode() ||
4665 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4668 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4670 LDBase = cast<LoadSDNode>(Elt.getNode());
4674 if (Elt.getOpcode() == ISD::UNDEF)
4677 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4678 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
4683 // If we have found an entire vector of loads and undefs, then return a large
4684 // load of the entire vector width starting at the base pointer. If we found
4685 // consecutive loads for the low half, generate a vzext_load node.
4686 if (LastLoadedElt == NumElems - 1) {
4687 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
4688 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4689 LDBase->getPointerInfo(),
4690 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
4691 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4692 LDBase->getPointerInfo(),
4693 LDBase->isVolatile(), LDBase->isNonTemporal(),
4694 LDBase->getAlignment());
4695 } else if (NumElems == 4 && LastLoadedElt == 1 &&
4696 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
4697 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4698 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4699 SDValue ResNode = DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys,
4701 LDBase->getMemOperand());
4702 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
4708 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
4709 DebugLoc dl = Op.getDebugLoc();
4711 EVT VT = Op.getValueType();
4712 EVT ExtVT = VT.getVectorElementType();
4713 unsigned NumElems = Op.getNumOperands();
4715 // Vectors containing all zeros can be matched by pxor and xorps later
4716 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
4717 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
4718 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
4719 if (Op.getValueType() == MVT::v4i32 ||
4720 Op.getValueType() == MVT::v8i32)
4723 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
4726 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
4727 // vectors or broken into v4i32 operations on 256-bit vectors.
4728 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
4729 if (Op.getValueType() == MVT::v4i32)
4732 return getOnesVector(Op.getValueType(), DAG, dl);
4735 unsigned EVTBits = ExtVT.getSizeInBits();
4737 unsigned NumZero = 0;
4738 unsigned NumNonZero = 0;
4739 unsigned NonZeros = 0;
4740 bool IsAllConstants = true;
4741 SmallSet<SDValue, 8> Values;
4742 for (unsigned i = 0; i < NumElems; ++i) {
4743 SDValue Elt = Op.getOperand(i);
4744 if (Elt.getOpcode() == ISD::UNDEF)
4747 if (Elt.getOpcode() != ISD::Constant &&
4748 Elt.getOpcode() != ISD::ConstantFP)
4749 IsAllConstants = false;
4750 if (X86::isZeroNode(Elt))
4753 NonZeros |= (1 << i);
4758 // All undef vector. Return an UNDEF. All zero vectors were handled above.
4759 if (NumNonZero == 0)
4760 return DAG.getUNDEF(VT);
4762 // Special case for single non-zero, non-undef, element.
4763 if (NumNonZero == 1) {
4764 unsigned Idx = CountTrailingZeros_32(NonZeros);
4765 SDValue Item = Op.getOperand(Idx);
4767 // If this is an insertion of an i64 value on x86-32, and if the top bits of
4768 // the value are obviously zero, truncate the value to i32 and do the
4769 // insertion that way. Only do this if the value is non-constant or if the
4770 // value is a constant being inserted into element 0. It is cheaper to do
4771 // a constant pool load than it is to do a movd + shuffle.
4772 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
4773 (!IsAllConstants || Idx == 0)) {
4774 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
4776 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
4777 EVT VecVT = MVT::v4i32;
4778 unsigned VecElts = 4;
4780 // Truncate the value (which may itself be a constant) to i32, and
4781 // convert it to a vector with movd (S2V+shuffle to zero extend).
4782 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
4783 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
4784 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4785 Subtarget->hasSSE2(), DAG);
4787 // Now we have our 32-bit value zero extended in the low element of
4788 // a vector. If Idx != 0, swizzle it into place.
4790 SmallVector<int, 4> Mask;
4791 Mask.push_back(Idx);
4792 for (unsigned i = 1; i != VecElts; ++i)
4794 Item = DAG.getVectorShuffle(VecVT, dl, Item,
4795 DAG.getUNDEF(Item.getValueType()),
4798 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Item);
4802 // If we have a constant or non-constant insertion into the low element of
4803 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
4804 // the rest of the elements. This will be matched as movd/movq/movss/movsd
4805 // depending on what the source datatype is.
4808 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4809 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
4810 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
4811 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4812 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
4813 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
4815 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
4816 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
4817 assert(VT.getSizeInBits() == 128 && "Expected an SSE value type!");
4818 EVT MiddleVT = MVT::v4i32;
4819 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
4820 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4821 Subtarget->hasSSE2(), DAG);
4822 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
4826 // Is it a vector logical left shift?
4827 if (NumElems == 2 && Idx == 1 &&
4828 X86::isZeroNode(Op.getOperand(0)) &&
4829 !X86::isZeroNode(Op.getOperand(1))) {
4830 unsigned NumBits = VT.getSizeInBits();
4831 return getVShift(true, VT,
4832 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4833 VT, Op.getOperand(1)),
4834 NumBits/2, DAG, *this, dl);
4837 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
4840 // Otherwise, if this is a vector with i32 or f32 elements, and the element
4841 // is a non-constant being inserted into an element other than the low one,
4842 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
4843 // movd/movss) to move this into the low element, then shuffle it into
4845 if (EVTBits == 32) {
4846 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4848 // Turn it into a shuffle of zero and zero-extended scalar to vector.
4849 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
4850 Subtarget->hasSSE2(), DAG);
4851 SmallVector<int, 8> MaskVec;
4852 for (unsigned i = 0; i < NumElems; i++)
4853 MaskVec.push_back(i == Idx ? 0 : 1);
4854 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
4858 // Splat is obviously ok. Let legalizer expand it to a shuffle.
4859 if (Values.size() == 1) {
4860 if (EVTBits == 32) {
4861 // Instead of a shuffle like this:
4862 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
4863 // Check if it's possible to issue this instead.
4864 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
4865 unsigned Idx = CountTrailingZeros_32(NonZeros);
4866 SDValue Item = Op.getOperand(Idx);
4867 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
4868 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
4873 // A vector full of immediates; various special cases are already
4874 // handled, so this is best done with a single constant-pool load.
4878 // For AVX-length vectors, build the individual 128-bit pieces and use
4879 // shuffles to put them in place.
4880 if (VT.getSizeInBits() == 256 && !ISD::isBuildVectorAllZeros(Op.getNode())) {
4881 SmallVector<SDValue, 32> V;
4882 for (unsigned i = 0; i < NumElems; ++i)
4883 V.push_back(Op.getOperand(i));
4885 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
4887 // Build both the lower and upper subvector.
4888 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[0], NumElems/2);
4889 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[NumElems / 2],
4892 // Recreate the wider vector with the lower and upper part.
4893 SDValue Vec = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT), Lower,
4894 DAG.getConstant(0, MVT::i32), DAG, dl);
4895 return Insert128BitVector(Vec, Upper, DAG.getConstant(NumElems/2, MVT::i32),
4899 // Let legalizer expand 2-wide build_vectors.
4900 if (EVTBits == 64) {
4901 if (NumNonZero == 1) {
4902 // One half is zero or undef.
4903 unsigned Idx = CountTrailingZeros_32(NonZeros);
4904 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
4905 Op.getOperand(Idx));
4906 return getShuffleVectorZeroOrUndef(V2, Idx, true,
4907 Subtarget->hasSSE2(), DAG);
4912 // If element VT is < 32 bits, convert it to inserts into a zero vector.
4913 if (EVTBits == 8 && NumElems == 16) {
4914 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
4916 if (V.getNode()) return V;
4919 if (EVTBits == 16 && NumElems == 8) {
4920 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
4922 if (V.getNode()) return V;
4925 // If element VT is == 32 bits, turn it into a number of shuffles.
4926 SmallVector<SDValue, 8> V;
4928 if (NumElems == 4 && NumZero > 0) {
4929 for (unsigned i = 0; i < 4; ++i) {
4930 bool isZero = !(NonZeros & (1 << i));
4932 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4934 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4937 for (unsigned i = 0; i < 2; ++i) {
4938 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
4941 V[i] = V[i*2]; // Must be a zero vector.
4944 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
4947 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
4950 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
4955 SmallVector<int, 8> MaskVec;
4956 bool Reverse = (NonZeros & 0x3) == 2;
4957 for (unsigned i = 0; i < 2; ++i)
4958 MaskVec.push_back(Reverse ? 1-i : i);
4959 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4960 for (unsigned i = 0; i < 2; ++i)
4961 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4962 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4965 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4966 // Check for a build vector of consecutive loads.
4967 for (unsigned i = 0; i < NumElems; ++i)
4968 V[i] = Op.getOperand(i);
4970 // Check for elements which are consecutive loads.
4971 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4975 // For SSE 4.1, use insertps to put the high elements into the low element.
4976 if (getSubtarget()->hasSSE41()) {
4978 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
4979 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
4981 Result = DAG.getUNDEF(VT);
4983 for (unsigned i = 1; i < NumElems; ++i) {
4984 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
4985 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
4986 Op.getOperand(i), DAG.getIntPtrConstant(i));
4991 // Otherwise, expand into a number of unpckl*, start by extending each of
4992 // our (non-undef) elements to the full vector width with the element in the
4993 // bottom slot of the vector (which generates no code for SSE).
4994 for (unsigned i = 0; i < NumElems; ++i) {
4995 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4996 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4998 V[i] = DAG.getUNDEF(VT);
5001 // Next, we iteratively mix elements, e.g. for v4f32:
5002 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
5003 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
5004 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
5005 unsigned EltStride = NumElems >> 1;
5006 while (EltStride != 0) {
5007 for (unsigned i = 0; i < EltStride; ++i) {
5008 // If V[i+EltStride] is undef and this is the first round of mixing,
5009 // then it is safe to just drop this shuffle: V[i] is already in the
5010 // right place, the one element (since it's the first round) being
5011 // inserted as undef can be dropped. This isn't safe for successive
5012 // rounds because they will permute elements within both vectors.
5013 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
5014 EltStride == NumElems/2)
5017 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
5026 // LowerMMXCONCAT_VECTORS - We support concatenate two MMX registers and place
5027 // them in a MMX register. This is better than doing a stack convert.
5028 static SDValue LowerMMXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5029 DebugLoc dl = Op.getDebugLoc();
5030 EVT ResVT = Op.getValueType();
5032 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
5033 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
5035 SDValue InVec = DAG.getNode(ISD::BITCAST,dl, MVT::v1i64, Op.getOperand(0));
5036 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
5037 InVec = Op.getOperand(1);
5038 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5039 unsigned NumElts = ResVT.getVectorNumElements();
5040 VecOp = DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
5041 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
5042 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
5044 InVec = DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, InVec);
5045 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
5046 Mask[0] = 0; Mask[1] = 2;
5047 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
5049 return DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
5052 // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
5053 // to create 256-bit vectors from two other 128-bit ones.
5054 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5055 DebugLoc dl = Op.getDebugLoc();
5056 EVT ResVT = Op.getValueType();
5058 assert(ResVT.getSizeInBits() == 256 && "Value type must be 256-bit wide");
5060 SDValue V1 = Op.getOperand(0);
5061 SDValue V2 = Op.getOperand(1);
5062 unsigned NumElems = ResVT.getVectorNumElements();
5064 SDValue V = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, ResVT), V1,
5065 DAG.getConstant(0, MVT::i32), DAG, dl);
5066 return Insert128BitVector(V, V2, DAG.getConstant(NumElems/2, MVT::i32),
5071 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
5072 EVT ResVT = Op.getValueType();
5074 assert(Op.getNumOperands() == 2);
5075 assert((ResVT.getSizeInBits() == 128 || ResVT.getSizeInBits() == 256) &&
5076 "Unsupported CONCAT_VECTORS for value type");
5078 // We support concatenate two MMX registers and place them in a MMX register.
5079 // This is better than doing a stack convert.
5080 if (ResVT.is128BitVector())
5081 return LowerMMXCONCAT_VECTORS(Op, DAG);
5083 // 256-bit AVX can use the vinsertf128 instruction to create 256-bit vectors
5084 // from two other 128-bit ones.
5085 return LowerAVXCONCAT_VECTORS(Op, DAG);
5088 // v8i16 shuffles - Prefer shuffles in the following order:
5089 // 1. [all] pshuflw, pshufhw, optional move
5090 // 2. [ssse3] 1 x pshufb
5091 // 3. [ssse3] 2 x pshufb + 1 x por
5092 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
5094 X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
5095 SelectionDAG &DAG) const {
5096 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5097 SDValue V1 = SVOp->getOperand(0);
5098 SDValue V2 = SVOp->getOperand(1);
5099 DebugLoc dl = SVOp->getDebugLoc();
5100 SmallVector<int, 8> MaskVals;
5102 // Determine if more than 1 of the words in each of the low and high quadwords
5103 // of the result come from the same quadword of one of the two inputs. Undef
5104 // mask values count as coming from any quadword, for better codegen.
5105 SmallVector<unsigned, 4> LoQuad(4);
5106 SmallVector<unsigned, 4> HiQuad(4);
5107 BitVector InputQuads(4);
5108 for (unsigned i = 0; i < 8; ++i) {
5109 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
5110 int EltIdx = SVOp->getMaskElt(i);
5111 MaskVals.push_back(EltIdx);
5120 InputQuads.set(EltIdx / 4);
5123 int BestLoQuad = -1;
5124 unsigned MaxQuad = 1;
5125 for (unsigned i = 0; i < 4; ++i) {
5126 if (LoQuad[i] > MaxQuad) {
5128 MaxQuad = LoQuad[i];
5132 int BestHiQuad = -1;
5134 for (unsigned i = 0; i < 4; ++i) {
5135 if (HiQuad[i] > MaxQuad) {
5137 MaxQuad = HiQuad[i];
5141 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
5142 // of the two input vectors, shuffle them into one input vector so only a
5143 // single pshufb instruction is necessary. If There are more than 2 input
5144 // quads, disable the next transformation since it does not help SSSE3.
5145 bool V1Used = InputQuads[0] || InputQuads[1];
5146 bool V2Used = InputQuads[2] || InputQuads[3];
5147 if (Subtarget->hasSSSE3()) {
5148 if (InputQuads.count() == 2 && V1Used && V2Used) {
5149 BestLoQuad = InputQuads.find_first();
5150 BestHiQuad = InputQuads.find_next(BestLoQuad);
5152 if (InputQuads.count() > 2) {
5158 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
5159 // the shuffle mask. If a quad is scored as -1, that means that it contains
5160 // words from all 4 input quadwords.
5162 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
5163 SmallVector<int, 8> MaskV;
5164 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
5165 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
5166 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
5167 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
5168 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
5169 NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
5171 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
5172 // source words for the shuffle, to aid later transformations.
5173 bool AllWordsInNewV = true;
5174 bool InOrder[2] = { true, true };
5175 for (unsigned i = 0; i != 8; ++i) {
5176 int idx = MaskVals[i];
5178 InOrder[i/4] = false;
5179 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
5181 AllWordsInNewV = false;
5185 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
5186 if (AllWordsInNewV) {
5187 for (int i = 0; i != 8; ++i) {
5188 int idx = MaskVals[i];
5191 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
5192 if ((idx != i) && idx < 4)
5194 if ((idx != i) && idx > 3)
5203 // If we've eliminated the use of V2, and the new mask is a pshuflw or
5204 // pshufhw, that's as cheap as it gets. Return the new shuffle.
5205 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
5206 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
5207 unsigned TargetMask = 0;
5208 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
5209 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
5210 TargetMask = pshufhw ? X86::getShufflePSHUFHWImmediate(NewV.getNode()):
5211 X86::getShufflePSHUFLWImmediate(NewV.getNode());
5212 V1 = NewV.getOperand(0);
5213 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
5217 // If we have SSSE3, and all words of the result are from 1 input vector,
5218 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
5219 // is present, fall back to case 4.
5220 if (Subtarget->hasSSSE3()) {
5221 SmallVector<SDValue,16> pshufbMask;
5223 // If we have elements from both input vectors, set the high bit of the
5224 // shuffle mask element to zero out elements that come from V2 in the V1
5225 // mask, and elements that come from V1 in the V2 mask, so that the two
5226 // results can be OR'd together.
5227 bool TwoInputs = V1Used && V2Used;
5228 for (unsigned i = 0; i != 8; ++i) {
5229 int EltIdx = MaskVals[i] * 2;
5230 if (TwoInputs && (EltIdx >= 16)) {
5231 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5232 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5235 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
5236 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
5238 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V1);
5239 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
5240 DAG.getNode(ISD::BUILD_VECTOR, dl,
5241 MVT::v16i8, &pshufbMask[0], 16));
5243 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5245 // Calculate the shuffle mask for the second input, shuffle it, and
5246 // OR it with the first shuffled input.
5248 for (unsigned i = 0; i != 8; ++i) {
5249 int EltIdx = MaskVals[i] * 2;
5251 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5252 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5255 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
5256 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
5258 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V2);
5259 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
5260 DAG.getNode(ISD::BUILD_VECTOR, dl,
5261 MVT::v16i8, &pshufbMask[0], 16));
5262 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
5263 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5266 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
5267 // and update MaskVals with new element order.
5268 BitVector InOrder(8);
5269 if (BestLoQuad >= 0) {
5270 SmallVector<int, 8> MaskV;
5271 for (int i = 0; i != 4; ++i) {
5272 int idx = MaskVals[i];
5274 MaskV.push_back(-1);
5276 } else if ((idx / 4) == BestLoQuad) {
5277 MaskV.push_back(idx & 3);
5280 MaskV.push_back(-1);
5283 for (unsigned i = 4; i != 8; ++i)
5285 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
5288 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
5289 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
5291 X86::getShufflePSHUFLWImmediate(NewV.getNode()),
5295 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
5296 // and update MaskVals with the new element order.
5297 if (BestHiQuad >= 0) {
5298 SmallVector<int, 8> MaskV;
5299 for (unsigned i = 0; i != 4; ++i)
5301 for (unsigned i = 4; i != 8; ++i) {
5302 int idx = MaskVals[i];
5304 MaskV.push_back(-1);
5306 } else if ((idx / 4) == BestHiQuad) {
5307 MaskV.push_back((idx & 3) + 4);
5310 MaskV.push_back(-1);
5313 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
5316 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
5317 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
5319 X86::getShufflePSHUFHWImmediate(NewV.getNode()),
5323 // In case BestHi & BestLo were both -1, which means each quadword has a word
5324 // from each of the four input quadwords, calculate the InOrder bitvector now
5325 // before falling through to the insert/extract cleanup.
5326 if (BestLoQuad == -1 && BestHiQuad == -1) {
5328 for (int i = 0; i != 8; ++i)
5329 if (MaskVals[i] < 0 || MaskVals[i] == i)
5333 // The other elements are put in the right place using pextrw and pinsrw.
5334 for (unsigned i = 0; i != 8; ++i) {
5337 int EltIdx = MaskVals[i];
5340 SDValue ExtOp = (EltIdx < 8)
5341 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
5342 DAG.getIntPtrConstant(EltIdx))
5343 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
5344 DAG.getIntPtrConstant(EltIdx - 8));
5345 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
5346 DAG.getIntPtrConstant(i));
5351 // v16i8 shuffles - Prefer shuffles in the following order:
5352 // 1. [ssse3] 1 x pshufb
5353 // 2. [ssse3] 2 x pshufb + 1 x por
5354 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
5356 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
5358 const X86TargetLowering &TLI) {
5359 SDValue V1 = SVOp->getOperand(0);
5360 SDValue V2 = SVOp->getOperand(1);
5361 DebugLoc dl = SVOp->getDebugLoc();
5362 SmallVector<int, 16> MaskVals;
5363 SVOp->getMask(MaskVals);
5365 // If we have SSSE3, case 1 is generated when all result bytes come from
5366 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
5367 // present, fall back to case 3.
5368 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
5371 for (unsigned i = 0; i < 16; ++i) {
5372 int EltIdx = MaskVals[i];
5381 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
5382 if (TLI.getSubtarget()->hasSSSE3()) {
5383 SmallVector<SDValue,16> pshufbMask;
5385 // If all result elements are from one input vector, then only translate
5386 // undef mask values to 0x80 (zero out result) in the pshufb mask.
5388 // Otherwise, we have elements from both input vectors, and must zero out
5389 // elements that come from V2 in the first mask, and V1 in the second mask
5390 // so that we can OR them together.
5391 bool TwoInputs = !(V1Only || V2Only);
5392 for (unsigned i = 0; i != 16; ++i) {
5393 int EltIdx = MaskVals[i];
5394 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
5395 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5398 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
5400 // If all the elements are from V2, assign it to V1 and return after
5401 // building the first pshufb.
5404 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
5405 DAG.getNode(ISD::BUILD_VECTOR, dl,
5406 MVT::v16i8, &pshufbMask[0], 16));
5410 // Calculate the shuffle mask for the second input, shuffle it, and
5411 // OR it with the first shuffled input.
5413 for (unsigned i = 0; i != 16; ++i) {
5414 int EltIdx = MaskVals[i];
5416 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5419 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
5421 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
5422 DAG.getNode(ISD::BUILD_VECTOR, dl,
5423 MVT::v16i8, &pshufbMask[0], 16));
5424 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
5427 // No SSSE3 - Calculate in place words and then fix all out of place words
5428 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
5429 // the 16 different words that comprise the two doublequadword input vectors.
5430 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5431 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
5432 SDValue NewV = V2Only ? V2 : V1;
5433 for (int i = 0; i != 8; ++i) {
5434 int Elt0 = MaskVals[i*2];
5435 int Elt1 = MaskVals[i*2+1];
5437 // This word of the result is all undef, skip it.
5438 if (Elt0 < 0 && Elt1 < 0)
5441 // This word of the result is already in the correct place, skip it.
5442 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
5444 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
5447 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
5448 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
5451 // If Elt0 and Elt1 are defined, are consecutive, and can be load
5452 // using a single extract together, load it and store it.
5453 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
5454 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
5455 DAG.getIntPtrConstant(Elt1 / 2));
5456 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5457 DAG.getIntPtrConstant(i));
5461 // If Elt1 is defined, extract it from the appropriate source. If the
5462 // source byte is not also odd, shift the extracted word left 8 bits
5463 // otherwise clear the bottom 8 bits if we need to do an or.
5465 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
5466 DAG.getIntPtrConstant(Elt1 / 2));
5467 if ((Elt1 & 1) == 0)
5468 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
5470 TLI.getShiftAmountTy(InsElt.getValueType())));
5472 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
5473 DAG.getConstant(0xFF00, MVT::i16));
5475 // If Elt0 is defined, extract it from the appropriate source. If the
5476 // source byte is not also even, shift the extracted word right 8 bits. If
5477 // Elt1 was also defined, OR the extracted values together before
5478 // inserting them in the result.
5480 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
5481 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
5482 if ((Elt0 & 1) != 0)
5483 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
5485 TLI.getShiftAmountTy(InsElt0.getValueType())));
5487 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
5488 DAG.getConstant(0x00FF, MVT::i16));
5489 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
5492 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5493 DAG.getIntPtrConstant(i));
5495 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
5498 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
5499 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
5500 /// done when every pair / quad of shuffle mask elements point to elements in
5501 /// the right sequence. e.g.
5502 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
5504 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
5505 SelectionDAG &DAG, DebugLoc dl) {
5506 EVT VT = SVOp->getValueType(0);
5507 SDValue V1 = SVOp->getOperand(0);
5508 SDValue V2 = SVOp->getOperand(1);
5509 unsigned NumElems = VT.getVectorNumElements();
5510 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
5512 switch (VT.getSimpleVT().SimpleTy) {
5513 default: assert(false && "Unexpected!");
5514 case MVT::v4f32: NewVT = MVT::v2f64; break;
5515 case MVT::v4i32: NewVT = MVT::v2i64; break;
5516 case MVT::v8i16: NewVT = MVT::v4i32; break;
5517 case MVT::v16i8: NewVT = MVT::v4i32; break;
5520 int Scale = NumElems / NewWidth;
5521 SmallVector<int, 8> MaskVec;
5522 for (unsigned i = 0; i < NumElems; i += Scale) {
5524 for (int j = 0; j < Scale; ++j) {
5525 int EltIdx = SVOp->getMaskElt(i+j);
5529 StartIdx = EltIdx - (EltIdx % Scale);
5530 if (EltIdx != StartIdx + j)
5534 MaskVec.push_back(-1);
5536 MaskVec.push_back(StartIdx / Scale);
5539 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
5540 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
5541 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
5544 /// getVZextMovL - Return a zero-extending vector move low node.
5546 static SDValue getVZextMovL(EVT VT, EVT OpVT,
5547 SDValue SrcOp, SelectionDAG &DAG,
5548 const X86Subtarget *Subtarget, DebugLoc dl) {
5549 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
5550 LoadSDNode *LD = NULL;
5551 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
5552 LD = dyn_cast<LoadSDNode>(SrcOp);
5554 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
5556 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
5557 if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
5558 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
5559 SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
5560 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
5562 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
5563 return DAG.getNode(ISD::BITCAST, dl, VT,
5564 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5565 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5573 return DAG.getNode(ISD::BITCAST, dl, VT,
5574 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5575 DAG.getNode(ISD::BITCAST, dl,
5579 /// LowerVECTOR_SHUFFLE_256 - Handle all 256-bit wide vectors shuffles
5580 /// which could not be matched by any known target speficic shuffle
5582 LowerVECTOR_SHUFFLE_256(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
5586 /// LowerVECTOR_SHUFFLE_128v4 - Handle all 128-bit wide vectors with
5587 /// 4 elements, and match them with several different shuffle types.
5589 LowerVECTOR_SHUFFLE_128v4(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
5590 SDValue V1 = SVOp->getOperand(0);
5591 SDValue V2 = SVOp->getOperand(1);
5592 DebugLoc dl = SVOp->getDebugLoc();
5593 EVT VT = SVOp->getValueType(0);
5595 assert(VT.getSizeInBits() == 128 && "Unsupported vector size");
5597 SmallVector<std::pair<int, int>, 8> Locs;
5599 SmallVector<int, 8> Mask1(4U, -1);
5600 SmallVector<int, 8> PermMask;
5601 SVOp->getMask(PermMask);
5605 for (unsigned i = 0; i != 4; ++i) {
5606 int Idx = PermMask[i];
5608 Locs[i] = std::make_pair(-1, -1);
5610 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
5612 Locs[i] = std::make_pair(0, NumLo);
5616 Locs[i] = std::make_pair(1, NumHi);
5618 Mask1[2+NumHi] = Idx;
5624 if (NumLo <= 2 && NumHi <= 2) {
5625 // If no more than two elements come from either vector. This can be
5626 // implemented with two shuffles. First shuffle gather the elements.
5627 // The second shuffle, which takes the first shuffle as both of its
5628 // vector operands, put the elements into the right order.
5629 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5631 SmallVector<int, 8> Mask2(4U, -1);
5633 for (unsigned i = 0; i != 4; ++i) {
5634 if (Locs[i].first == -1)
5637 unsigned Idx = (i < 2) ? 0 : 4;
5638 Idx += Locs[i].first * 2 + Locs[i].second;
5643 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
5644 } else if (NumLo == 3 || NumHi == 3) {
5645 // Otherwise, we must have three elements from one vector, call it X, and
5646 // one element from the other, call it Y. First, use a shufps to build an
5647 // intermediate vector with the one element from Y and the element from X
5648 // that will be in the same half in the final destination (the indexes don't
5649 // matter). Then, use a shufps to build the final vector, taking the half
5650 // containing the element from Y from the intermediate, and the other half
5653 // Normalize it so the 3 elements come from V1.
5654 CommuteVectorShuffleMask(PermMask, VT);
5658 // Find the element from V2.
5660 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
5661 int Val = PermMask[HiIndex];
5668 Mask1[0] = PermMask[HiIndex];
5670 Mask1[2] = PermMask[HiIndex^1];
5672 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5675 Mask1[0] = PermMask[0];
5676 Mask1[1] = PermMask[1];
5677 Mask1[2] = HiIndex & 1 ? 6 : 4;
5678 Mask1[3] = HiIndex & 1 ? 4 : 6;
5679 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5681 Mask1[0] = HiIndex & 1 ? 2 : 0;
5682 Mask1[1] = HiIndex & 1 ? 0 : 2;
5683 Mask1[2] = PermMask[2];
5684 Mask1[3] = PermMask[3];
5689 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
5693 // Break it into (shuffle shuffle_hi, shuffle_lo).
5696 SmallVector<int,8> LoMask(4U, -1);
5697 SmallVector<int,8> HiMask(4U, -1);
5699 SmallVector<int,8> *MaskPtr = &LoMask;
5700 unsigned MaskIdx = 0;
5703 for (unsigned i = 0; i != 4; ++i) {
5710 int Idx = PermMask[i];
5712 Locs[i] = std::make_pair(-1, -1);
5713 } else if (Idx < 4) {
5714 Locs[i] = std::make_pair(MaskIdx, LoIdx);
5715 (*MaskPtr)[LoIdx] = Idx;
5718 Locs[i] = std::make_pair(MaskIdx, HiIdx);
5719 (*MaskPtr)[HiIdx] = Idx;
5724 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
5725 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
5726 SmallVector<int, 8> MaskOps;
5727 for (unsigned i = 0; i != 4; ++i) {
5728 if (Locs[i].first == -1) {
5729 MaskOps.push_back(-1);
5731 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
5732 MaskOps.push_back(Idx);
5735 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
5738 static bool MayFoldVectorLoad(SDValue V) {
5739 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
5740 V = V.getOperand(0);
5741 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5742 V = V.getOperand(0);
5748 // FIXME: the version above should always be used. Since there's
5749 // a bug where several vector shuffles can't be folded because the
5750 // DAG is not updated during lowering and a node claims to have two
5751 // uses while it only has one, use this version, and let isel match
5752 // another instruction if the load really happens to have more than
5753 // one use. Remove this version after this bug get fixed.
5754 // rdar://8434668, PR8156
5755 static bool RelaxedMayFoldVectorLoad(SDValue V) {
5756 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
5757 V = V.getOperand(0);
5758 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5759 V = V.getOperand(0);
5760 if (ISD::isNormalLoad(V.getNode()))
5765 /// CanFoldShuffleIntoVExtract - Check if the current shuffle is used by
5766 /// a vector extract, and if both can be later optimized into a single load.
5767 /// This is done in visitEXTRACT_VECTOR_ELT and the conditions are checked
5768 /// here because otherwise a target specific shuffle node is going to be
5769 /// emitted for this shuffle, and the optimization not done.
5770 /// FIXME: This is probably not the best approach, but fix the problem
5771 /// until the right path is decided.
5773 bool CanXFormVExtractWithShuffleIntoLoad(SDValue V, SelectionDAG &DAG,
5774 const TargetLowering &TLI) {
5775 EVT VT = V.getValueType();
5776 ShuffleVectorSDNode *SVOp = dyn_cast<ShuffleVectorSDNode>(V);
5778 // Be sure that the vector shuffle is present in a pattern like this:
5779 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), c) -> (f32 load $addr)
5783 SDNode *N = *V.getNode()->use_begin();
5784 if (N->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5787 SDValue EltNo = N->getOperand(1);
5788 if (!isa<ConstantSDNode>(EltNo))
5791 // If the bit convert changed the number of elements, it is unsafe
5792 // to examine the mask.
5793 bool HasShuffleIntoBitcast = false;
5794 if (V.getOpcode() == ISD::BITCAST) {
5795 EVT SrcVT = V.getOperand(0).getValueType();
5796 if (SrcVT.getVectorNumElements() != VT.getVectorNumElements())
5798 V = V.getOperand(0);
5799 HasShuffleIntoBitcast = true;
5802 // Select the input vector, guarding against out of range extract vector.
5803 unsigned NumElems = VT.getVectorNumElements();
5804 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5805 int Idx = (Elt > NumElems) ? -1 : SVOp->getMaskElt(Elt);
5806 V = (Idx < (int)NumElems) ? V.getOperand(0) : V.getOperand(1);
5808 // Skip one more bit_convert if necessary
5809 if (V.getOpcode() == ISD::BITCAST)
5810 V = V.getOperand(0);
5812 if (ISD::isNormalLoad(V.getNode())) {
5813 // Is the original load suitable?
5814 LoadSDNode *LN0 = cast<LoadSDNode>(V);
5816 // FIXME: avoid the multi-use bug that is preventing lots of
5817 // of foldings to be detected, this is still wrong of course, but
5818 // give the temporary desired behavior, and if it happens that
5819 // the load has real more uses, during isel it will not fold, and
5820 // will generate poor code.
5821 if (!LN0 || LN0->isVolatile()) // || !LN0->hasOneUse()
5824 if (!HasShuffleIntoBitcast)
5827 // If there's a bitcast before the shuffle, check if the load type and
5828 // alignment is valid.
5829 unsigned Align = LN0->getAlignment();
5831 TLI.getTargetData()->getABITypeAlignment(
5832 VT.getTypeForEVT(*DAG.getContext()));
5834 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
5842 SDValue getMOVDDup(SDValue &Op, DebugLoc &dl, SDValue V1, SelectionDAG &DAG) {
5843 EVT VT = Op.getValueType();
5845 // Canonizalize to v2f64.
5846 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
5847 return DAG.getNode(ISD::BITCAST, dl, VT,
5848 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
5853 SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
5855 SDValue V1 = Op.getOperand(0);
5856 SDValue V2 = Op.getOperand(1);
5857 EVT VT = Op.getValueType();
5859 assert(VT != MVT::v2i64 && "unsupported shuffle type");
5861 if (HasSSE2 && VT == MVT::v2f64)
5862 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
5865 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V2, DAG);
5869 SDValue getMOVHighToLow(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG) {
5870 SDValue V1 = Op.getOperand(0);
5871 SDValue V2 = Op.getOperand(1);
5872 EVT VT = Op.getValueType();
5874 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
5875 "unsupported shuffle type");
5877 if (V2.getOpcode() == ISD::UNDEF)
5881 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
5885 SDValue getMOVLP(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
5886 SDValue V1 = Op.getOperand(0);
5887 SDValue V2 = Op.getOperand(1);
5888 EVT VT = Op.getValueType();
5889 unsigned NumElems = VT.getVectorNumElements();
5891 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
5892 // operand of these instructions is only memory, so check if there's a
5893 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
5895 bool CanFoldLoad = false;
5897 // Trivial case, when V2 comes from a load.
5898 if (MayFoldVectorLoad(V2))
5901 // When V1 is a load, it can be folded later into a store in isel, example:
5902 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
5904 // (MOVLPSmr addr:$src1, VR128:$src2)
5905 // So, recognize this potential and also use MOVLPS or MOVLPD
5906 if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
5909 // Both of them can't be memory operations though.
5910 if (MayFoldVectorLoad(V1) && MayFoldVectorLoad(V2))
5911 CanFoldLoad = false;
5914 if (HasSSE2 && NumElems == 2)
5915 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
5918 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
5921 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5922 // movl and movlp will both match v2i64, but v2i64 is never matched by
5923 // movl earlier because we make it strict to avoid messing with the movlp load
5924 // folding logic (see the code above getMOVLP call). Match it here then,
5925 // this is horrible, but will stay like this until we move all shuffle
5926 // matching to x86 specific nodes. Note that for the 1st condition all
5927 // types are matched with movsd.
5928 if ((HasSSE2 && NumElems == 2) || !X86::isMOVLMask(SVOp))
5929 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5931 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5934 assert(VT != MVT::v4i32 && "unsupported shuffle type");
5936 // Invert the operand order and use SHUFPS to match it.
5937 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V2, V1,
5938 X86::getShuffleSHUFImmediate(SVOp), DAG);
5941 static inline unsigned getUNPCKLOpcode(EVT VT) {
5942 switch(VT.getSimpleVT().SimpleTy) {
5943 case MVT::v4i32: return X86ISD::PUNPCKLDQ;
5944 case MVT::v2i64: return X86ISD::PUNPCKLQDQ;
5945 case MVT::v4f32: return X86ISD::UNPCKLPS;
5946 case MVT::v2f64: return X86ISD::UNPCKLPD;
5947 case MVT::v8i32: // Use fp unit for int unpack.
5948 case MVT::v8f32: return X86ISD::VUNPCKLPSY;
5949 case MVT::v4i64: // Use fp unit for int unpack.
5950 case MVT::v4f64: return X86ISD::VUNPCKLPDY;
5951 case MVT::v16i8: return X86ISD::PUNPCKLBW;
5952 case MVT::v8i16: return X86ISD::PUNPCKLWD;
5954 llvm_unreachable("Unknown type for unpckl");
5959 static inline unsigned getUNPCKHOpcode(EVT VT) {
5960 switch(VT.getSimpleVT().SimpleTy) {
5961 case MVT::v4i32: return X86ISD::PUNPCKHDQ;
5962 case MVT::v2i64: return X86ISD::PUNPCKHQDQ;
5963 case MVT::v4f32: return X86ISD::UNPCKHPS;
5964 case MVT::v2f64: return X86ISD::UNPCKHPD;
5965 case MVT::v8i32: // Use fp unit for int unpack.
5966 case MVT::v8f32: return X86ISD::VUNPCKHPSY;
5967 case MVT::v4i64: // Use fp unit for int unpack.
5968 case MVT::v4f64: return X86ISD::VUNPCKHPDY;
5969 case MVT::v16i8: return X86ISD::PUNPCKHBW;
5970 case MVT::v8i16: return X86ISD::PUNPCKHWD;
5972 llvm_unreachable("Unknown type for unpckh");
5977 static inline unsigned getVPERMILOpcode(EVT VT) {
5978 switch(VT.getSimpleVT().SimpleTy) {
5980 case MVT::v4f32: return X86ISD::VPERMILPS;
5982 case MVT::v2f64: return X86ISD::VPERMILPD;
5984 case MVT::v8f32: return X86ISD::VPERMILPSY;
5986 case MVT::v4f64: return X86ISD::VPERMILPDY;
5988 llvm_unreachable("Unknown type for vpermil");
5994 SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG,
5995 const TargetLowering &TLI,
5996 const X86Subtarget *Subtarget) {
5997 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5998 EVT VT = Op.getValueType();
5999 DebugLoc dl = Op.getDebugLoc();
6000 SDValue V1 = Op.getOperand(0);
6001 SDValue V2 = Op.getOperand(1);
6003 if (isZeroShuffle(SVOp))
6004 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
6006 // Handle splat operations
6007 if (SVOp->isSplat()) {
6008 unsigned NumElem = VT.getVectorNumElements();
6009 // Special case, this is the only place now where it's allowed to return
6010 // a vector_shuffle operation without using a target specific node, because
6011 // *hopefully* it will be optimized away by the dag combiner. FIXME: should
6012 // this be moved to DAGCombine instead?
6013 if (NumElem <= 4 && CanXFormVExtractWithShuffleIntoLoad(Op, DAG, TLI))
6016 // Since there's no native support for scalar_to_vector for 256-bit AVX, a
6017 // 128-bit scalar_to_vector + INSERT_SUBVECTOR is generated. Recognize this
6018 // idiom and do the shuffle before the insertion, this yields less
6019 // instructions in the end.
6020 if (VT.is256BitVector() &&
6021 V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
6022 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
6023 V1.getOperand(1).getOpcode() == ISD::SCALAR_TO_VECTOR)
6024 return PromoteVectorToScalarSplat(SVOp, DAG);
6026 // Handle splats by matching through known shuffle masks
6027 if ((VT.is128BitVector() && NumElem <= 4) ||
6028 (VT.is256BitVector() && NumElem <= 8))
6031 // All i16 and i8 vector types can't be used directly by a generic shuffle
6032 // instruction because the target has no such instruction. Generate shuffles
6033 // which repeat i16 and i8 several times until they fit in i32, and then can
6034 // be manipulated by target suported shuffles. After the insertion of the
6035 // necessary shuffles, the result is bitcasted back to v4f32 or v8f32.
6036 return PromoteSplat(SVOp, DAG);
6039 // If the shuffle can be profitably rewritten as a narrower shuffle, then
6041 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
6042 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
6043 if (NewOp.getNode())
6044 return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
6045 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
6046 // FIXME: Figure out a cleaner way to do this.
6047 // Try to make use of movq to zero out the top part.
6048 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
6049 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
6050 if (NewOp.getNode()) {
6051 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
6052 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
6053 DAG, Subtarget, dl);
6055 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
6056 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
6057 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
6058 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
6059 DAG, Subtarget, dl);
6066 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
6067 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6068 SDValue V1 = Op.getOperand(0);
6069 SDValue V2 = Op.getOperand(1);
6070 EVT VT = Op.getValueType();
6071 DebugLoc dl = Op.getDebugLoc();
6072 unsigned NumElems = VT.getVectorNumElements();
6073 bool isMMX = VT.getSizeInBits() == 64;
6074 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
6075 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
6076 bool V1IsSplat = false;
6077 bool V2IsSplat = false;
6078 bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
6079 bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
6080 bool HasSSSE3 = Subtarget->hasSSSE3() || Subtarget->hasAVX();
6081 MachineFunction &MF = DAG.getMachineFunction();
6082 bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
6084 // Shuffle operations on MMX not supported.
6088 // Vector shuffle lowering takes 3 steps:
6090 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
6091 // narrowing and commutation of operands should be handled.
6092 // 2) Matching of shuffles with known shuffle masks to x86 target specific
6094 // 3) Rewriting of unmatched masks into new generic shuffle operations,
6095 // so the shuffle can be broken into other shuffles and the legalizer can
6096 // try the lowering again.
6098 // The general ideia is that no vector_shuffle operation should be left to
6099 // be matched during isel, all of them must be converted to a target specific
6102 // Normalize the input vectors. Here splats, zeroed vectors, profitable
6103 // narrowing and commutation of operands should be handled. The actual code
6104 // doesn't include all of those, work in progress...
6105 SDValue NewOp = NormalizeVectorShuffle(Op, DAG, *this, Subtarget);
6106 if (NewOp.getNode())
6109 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
6110 // unpckh_undef). Only use pshufd if speed is more important than size.
6111 if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp))
6112 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
6113 if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp))
6114 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
6116 if (X86::isMOVDDUPMask(SVOp) && HasSSE3 && V2IsUndef &&
6117 RelaxedMayFoldVectorLoad(V1))
6118 return getMOVDDup(Op, dl, V1, DAG);
6120 if (X86::isMOVHLPS_v_undef_Mask(SVOp))
6121 return getMOVHighToLow(Op, dl, DAG);
6123 // Use to match splats
6124 if (HasSSE2 && X86::isUNPCKHMask(SVOp) && V2IsUndef &&
6125 (VT == MVT::v2f64 || VT == MVT::v2i64))
6126 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
6128 if (X86::isPSHUFDMask(SVOp)) {
6129 // The actual implementation will match the mask in the if above and then
6130 // during isel it can match several different instructions, not only pshufd
6131 // as its name says, sad but true, emulate the behavior for now...
6132 if (X86::isMOVDDUPMask(SVOp) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
6133 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
6135 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
6137 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
6138 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
6140 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
6141 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V1,
6144 if (VT == MVT::v4f32)
6145 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V1,
6149 // Check if this can be converted into a logical shift.
6150 bool isLeft = false;
6153 bool isShift = getSubtarget()->hasSSE2() &&
6154 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
6155 if (isShift && ShVal.hasOneUse()) {
6156 // If the shifted value has multiple uses, it may be cheaper to use
6157 // v_set0 + movlhps or movhlps, etc.
6158 EVT EltVT = VT.getVectorElementType();
6159 ShAmt *= EltVT.getSizeInBits();
6160 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
6163 if (X86::isMOVLMask(SVOp)) {
6166 if (ISD::isBuildVectorAllZeros(V1.getNode()))
6167 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
6168 if (!X86::isMOVLPMask(SVOp)) {
6169 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
6170 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
6172 if (VT == MVT::v4i32 || VT == MVT::v4f32)
6173 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
6177 // FIXME: fold these into legal mask.
6178 if (X86::isMOVLHPSMask(SVOp) && !X86::isUNPCKLMask(SVOp))
6179 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
6181 if (X86::isMOVHLPSMask(SVOp))
6182 return getMOVHighToLow(Op, dl, DAG);
6184 if (X86::isMOVSHDUPMask(SVOp, Subtarget))
6185 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
6187 if (X86::isMOVSLDUPMask(SVOp, Subtarget))
6188 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
6190 if (X86::isMOVLPMask(SVOp))
6191 return getMOVLP(Op, dl, DAG, HasSSE2);
6193 if (ShouldXformToMOVHLPS(SVOp) ||
6194 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
6195 return CommuteVectorShuffle(SVOp, DAG);
6198 // No better options. Use a vshl / vsrl.
6199 EVT EltVT = VT.getVectorElementType();
6200 ShAmt *= EltVT.getSizeInBits();
6201 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
6204 bool Commuted = false;
6205 // FIXME: This should also accept a bitcast of a splat? Be careful, not
6206 // 1,1,1,1 -> v8i16 though.
6207 V1IsSplat = isSplatVector(V1.getNode());
6208 V2IsSplat = isSplatVector(V2.getNode());
6210 // Canonicalize the splat or undef, if present, to be on the RHS.
6211 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
6212 Op = CommuteVectorShuffle(SVOp, DAG);
6213 SVOp = cast<ShuffleVectorSDNode>(Op);
6214 V1 = SVOp->getOperand(0);
6215 V2 = SVOp->getOperand(1);
6216 std::swap(V1IsSplat, V2IsSplat);
6217 std::swap(V1IsUndef, V2IsUndef);
6221 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
6222 // Shuffling low element of v1 into undef, just return v1.
6225 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
6226 // the instruction selector will not match, so get a canonical MOVL with
6227 // swapped operands to undo the commute.
6228 return getMOVL(DAG, dl, VT, V2, V1);
6231 if (X86::isUNPCKLMask(SVOp))
6232 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V2, DAG);
6234 if (X86::isUNPCKHMask(SVOp))
6235 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V2, DAG);
6238 // Normalize mask so all entries that point to V2 points to its first
6239 // element then try to match unpck{h|l} again. If match, return a
6240 // new vector_shuffle with the corrected mask.
6241 SDValue NewMask = NormalizeMask(SVOp, DAG);
6242 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
6243 if (NSVOp != SVOp) {
6244 if (X86::isUNPCKLMask(NSVOp, true)) {
6246 } else if (X86::isUNPCKHMask(NSVOp, true)) {
6253 // Commute is back and try unpck* again.
6254 // FIXME: this seems wrong.
6255 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
6256 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
6258 if (X86::isUNPCKLMask(NewSVOp))
6259 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V2, V1, DAG);
6261 if (X86::isUNPCKHMask(NewSVOp))
6262 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V2, V1, DAG);
6265 // Normalize the node to match x86 shuffle ops if needed
6266 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
6267 return CommuteVectorShuffle(SVOp, DAG);
6269 // The checks below are all present in isShuffleMaskLegal, but they are
6270 // inlined here right now to enable us to directly emit target specific
6271 // nodes, and remove one by one until they don't return Op anymore.
6272 SmallVector<int, 16> M;
6275 if (isPALIGNRMask(M, VT, HasSSSE3))
6276 return getTargetShuffleNode(X86ISD::PALIGN, dl, VT, V1, V2,
6277 X86::getShufflePALIGNRImmediate(SVOp),
6280 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
6281 SVOp->getSplatIndex() == 0 && V2IsUndef) {
6282 if (VT == MVT::v2f64)
6283 return getTargetShuffleNode(X86ISD::UNPCKLPD, dl, VT, V1, V1, DAG);
6284 if (VT == MVT::v2i64)
6285 return getTargetShuffleNode(X86ISD::PUNPCKLQDQ, dl, VT, V1, V1, DAG);
6288 if (isPSHUFHWMask(M, VT))
6289 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
6290 X86::getShufflePSHUFHWImmediate(SVOp),
6293 if (isPSHUFLWMask(M, VT))
6294 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
6295 X86::getShufflePSHUFLWImmediate(SVOp),
6298 if (isSHUFPMask(M, VT)) {
6299 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
6300 if (VT == MVT::v4f32 || VT == MVT::v4i32)
6301 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V2,
6303 if (VT == MVT::v2f64 || VT == MVT::v2i64)
6304 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V2,
6308 if (X86::isUNPCKL_v_undef_Mask(SVOp))
6309 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
6310 if (X86::isUNPCKH_v_undef_Mask(SVOp))
6311 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
6313 //===--------------------------------------------------------------------===//
6314 // Generate target specific nodes for 128 or 256-bit shuffles only
6315 // supported in the AVX instruction set.
6318 // Handle VPERMILPS* permutations
6319 if (isVPERMILPSMask(M, VT, Subtarget))
6320 return getTargetShuffleNode(getVPERMILOpcode(VT), dl, VT, V1,
6321 getShuffleVPERMILPSImmediate(SVOp), DAG);
6323 // Handle VPERMILPD* permutations
6324 if (isVPERMILPDMask(M, VT, Subtarget))
6325 return getTargetShuffleNode(getVPERMILOpcode(VT), dl, VT, V1,
6326 getShuffleVPERMILPDImmediate(SVOp), DAG);
6328 //===--------------------------------------------------------------------===//
6329 // Since no target specific shuffle was selected for this generic one,
6330 // lower it into other known shuffles. FIXME: this isn't true yet, but
6331 // this is the plan.
6334 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
6335 if (VT == MVT::v8i16) {
6336 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
6337 if (NewOp.getNode())
6341 if (VT == MVT::v16i8) {
6342 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
6343 if (NewOp.getNode())
6347 // Handle all 128-bit wide vectors with 4 elements, and match them with
6348 // several different shuffle types.
6349 if (NumElems == 4 && VT.getSizeInBits() == 128)
6350 return LowerVECTOR_SHUFFLE_128v4(SVOp, DAG);
6352 // Handle general 256-bit shuffles
6353 if (VT.is256BitVector())
6354 return LowerVECTOR_SHUFFLE_256(SVOp, DAG);
6360 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
6361 SelectionDAG &DAG) const {
6362 EVT VT = Op.getValueType();
6363 DebugLoc dl = Op.getDebugLoc();
6365 if (Op.getOperand(0).getValueType().getSizeInBits() != 128)
6368 if (VT.getSizeInBits() == 8) {
6369 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
6370 Op.getOperand(0), Op.getOperand(1));
6371 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
6372 DAG.getValueType(VT));
6373 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6374 } else if (VT.getSizeInBits() == 16) {
6375 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6376 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
6378 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
6379 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6380 DAG.getNode(ISD::BITCAST, dl,
6384 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
6385 Op.getOperand(0), Op.getOperand(1));
6386 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
6387 DAG.getValueType(VT));
6388 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6389 } else if (VT == MVT::f32) {
6390 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
6391 // the result back to FR32 register. It's only worth matching if the
6392 // result has a single use which is a store or a bitcast to i32. And in
6393 // the case of a store, it's not worth it if the index is a constant 0,
6394 // because a MOVSSmr can be used instead, which is smaller and faster.
6395 if (!Op.hasOneUse())
6397 SDNode *User = *Op.getNode()->use_begin();
6398 if ((User->getOpcode() != ISD::STORE ||
6399 (isa<ConstantSDNode>(Op.getOperand(1)) &&
6400 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
6401 (User->getOpcode() != ISD::BITCAST ||
6402 User->getValueType(0) != MVT::i32))
6404 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6405 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
6408 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
6409 } else if (VT == MVT::i32) {
6410 // ExtractPS works with constant index.
6411 if (isa<ConstantSDNode>(Op.getOperand(1)))
6419 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
6420 SelectionDAG &DAG) const {
6421 if (!isa<ConstantSDNode>(Op.getOperand(1)))
6424 SDValue Vec = Op.getOperand(0);
6425 EVT VecVT = Vec.getValueType();
6427 // If this is a 256-bit vector result, first extract the 128-bit vector and
6428 // then extract the element from the 128-bit vector.
6429 if (VecVT.getSizeInBits() == 256) {
6430 DebugLoc dl = Op.getNode()->getDebugLoc();
6431 unsigned NumElems = VecVT.getVectorNumElements();
6432 SDValue Idx = Op.getOperand(1);
6433 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
6435 // Get the 128-bit vector.
6436 bool Upper = IdxVal >= NumElems/2;
6437 Vec = Extract128BitVector(Vec,
6438 DAG.getConstant(Upper ? NumElems/2 : 0, MVT::i32), DAG, dl);
6440 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
6441 Upper ? DAG.getConstant(IdxVal-NumElems/2, MVT::i32) : Idx);
6444 assert(Vec.getValueSizeInBits() <= 128 && "Unexpected vector length");
6446 if (Subtarget->hasSSE41() || Subtarget->hasAVX()) {
6447 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
6452 EVT VT = Op.getValueType();
6453 DebugLoc dl = Op.getDebugLoc();
6454 // TODO: handle v16i8.
6455 if (VT.getSizeInBits() == 16) {
6456 SDValue Vec = Op.getOperand(0);
6457 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6459 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
6460 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6461 DAG.getNode(ISD::BITCAST, dl,
6464 // Transform it so it match pextrw which produces a 32-bit result.
6465 EVT EltVT = MVT::i32;
6466 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
6467 Op.getOperand(0), Op.getOperand(1));
6468 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
6469 DAG.getValueType(VT));
6470 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6471 } else if (VT.getSizeInBits() == 32) {
6472 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6476 // SHUFPS the element to the lowest double word, then movss.
6477 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
6478 EVT VVT = Op.getOperand(0).getValueType();
6479 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
6480 DAG.getUNDEF(VVT), Mask);
6481 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
6482 DAG.getIntPtrConstant(0));
6483 } else if (VT.getSizeInBits() == 64) {
6484 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
6485 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
6486 // to match extract_elt for f64.
6487 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6491 // UNPCKHPD the element to the lowest double word, then movsd.
6492 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
6493 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
6494 int Mask[2] = { 1, -1 };
6495 EVT VVT = Op.getOperand(0).getValueType();
6496 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
6497 DAG.getUNDEF(VVT), Mask);
6498 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
6499 DAG.getIntPtrConstant(0));
6506 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
6507 SelectionDAG &DAG) const {
6508 EVT VT = Op.getValueType();
6509 EVT EltVT = VT.getVectorElementType();
6510 DebugLoc dl = Op.getDebugLoc();
6512 SDValue N0 = Op.getOperand(0);
6513 SDValue N1 = Op.getOperand(1);
6514 SDValue N2 = Op.getOperand(2);
6516 if (VT.getSizeInBits() == 256)
6519 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
6520 isa<ConstantSDNode>(N2)) {
6522 if (VT == MVT::v8i16)
6523 Opc = X86ISD::PINSRW;
6524 else if (VT == MVT::v16i8)
6525 Opc = X86ISD::PINSRB;
6527 Opc = X86ISD::PINSRB;
6529 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
6531 if (N1.getValueType() != MVT::i32)
6532 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
6533 if (N2.getValueType() != MVT::i32)
6534 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
6535 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
6536 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
6537 // Bits [7:6] of the constant are the source select. This will always be
6538 // zero here. The DAG Combiner may combine an extract_elt index into these
6539 // bits. For example (insert (extract, 3), 2) could be matched by putting
6540 // the '3' into bits [7:6] of X86ISD::INSERTPS.
6541 // Bits [5:4] of the constant are the destination select. This is the
6542 // value of the incoming immediate.
6543 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
6544 // combine either bitwise AND or insert of float 0.0 to set these bits.
6545 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
6546 // Create this as a scalar to vector..
6547 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
6548 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
6549 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
6550 // PINSR* works with constant index.
6557 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
6558 EVT VT = Op.getValueType();
6559 EVT EltVT = VT.getVectorElementType();
6561 DebugLoc dl = Op.getDebugLoc();
6562 SDValue N0 = Op.getOperand(0);
6563 SDValue N1 = Op.getOperand(1);
6564 SDValue N2 = Op.getOperand(2);
6566 // If this is a 256-bit vector result, first extract the 128-bit vector,
6567 // insert the element into the extracted half and then place it back.
6568 if (VT.getSizeInBits() == 256) {
6569 if (!isa<ConstantSDNode>(N2))
6572 // Get the desired 128-bit vector half.
6573 unsigned NumElems = VT.getVectorNumElements();
6574 unsigned IdxVal = cast<ConstantSDNode>(N2)->getZExtValue();
6575 bool Upper = IdxVal >= NumElems/2;
6576 SDValue Ins128Idx = DAG.getConstant(Upper ? NumElems/2 : 0, MVT::i32);
6577 SDValue V = Extract128BitVector(N0, Ins128Idx, DAG, dl);
6579 // Insert the element into the desired half.
6580 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V,
6581 N1, Upper ? DAG.getConstant(IdxVal-NumElems/2, MVT::i32) : N2);
6583 // Insert the changed part back to the 256-bit vector
6584 return Insert128BitVector(N0, V, Ins128Idx, DAG, dl);
6587 if (Subtarget->hasSSE41() || Subtarget->hasAVX())
6588 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
6590 if (EltVT == MVT::i8)
6593 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
6594 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
6595 // as its second argument.
6596 if (N1.getValueType() != MVT::i32)
6597 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
6598 if (N2.getValueType() != MVT::i32)
6599 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
6600 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
6606 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6607 LLVMContext *Context = DAG.getContext();
6608 DebugLoc dl = Op.getDebugLoc();
6609 EVT OpVT = Op.getValueType();
6611 // If this is a 256-bit vector result, first insert into a 128-bit
6612 // vector and then insert into the 256-bit vector.
6613 if (OpVT.getSizeInBits() > 128) {
6614 // Insert into a 128-bit vector.
6615 EVT VT128 = EVT::getVectorVT(*Context,
6616 OpVT.getVectorElementType(),
6617 OpVT.getVectorNumElements() / 2);
6619 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
6621 // Insert the 128-bit vector.
6622 return Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, OpVT), Op,
6623 DAG.getConstant(0, MVT::i32),
6627 if (Op.getValueType() == MVT::v1i64 &&
6628 Op.getOperand(0).getValueType() == MVT::i64)
6629 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
6631 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
6632 assert(Op.getValueType().getSimpleVT().getSizeInBits() == 128 &&
6633 "Expected an SSE type!");
6634 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(),
6635 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
6638 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
6639 // a simple subregister reference or explicit instructions to grab
6640 // upper bits of a vector.
6642 X86TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
6643 if (Subtarget->hasAVX()) {
6644 DebugLoc dl = Op.getNode()->getDebugLoc();
6645 SDValue Vec = Op.getNode()->getOperand(0);
6646 SDValue Idx = Op.getNode()->getOperand(1);
6648 if (Op.getNode()->getValueType(0).getSizeInBits() == 128
6649 && Vec.getNode()->getValueType(0).getSizeInBits() == 256) {
6650 return Extract128BitVector(Vec, Idx, DAG, dl);
6656 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
6657 // simple superregister reference or explicit instructions to insert
6658 // the upper bits of a vector.
6660 X86TargetLowering::LowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
6661 if (Subtarget->hasAVX()) {
6662 DebugLoc dl = Op.getNode()->getDebugLoc();
6663 SDValue Vec = Op.getNode()->getOperand(0);
6664 SDValue SubVec = Op.getNode()->getOperand(1);
6665 SDValue Idx = Op.getNode()->getOperand(2);
6667 if (Op.getNode()->getValueType(0).getSizeInBits() == 256
6668 && SubVec.getNode()->getValueType(0).getSizeInBits() == 128) {
6669 return Insert128BitVector(Vec, SubVec, Idx, DAG, dl);
6675 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
6676 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
6677 // one of the above mentioned nodes. It has to be wrapped because otherwise
6678 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
6679 // be used to form addressing mode. These wrapped nodes will be selected
6682 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
6683 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
6685 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6687 unsigned char OpFlag = 0;
6688 unsigned WrapperKind = X86ISD::Wrapper;
6689 CodeModel::Model M = getTargetMachine().getCodeModel();
6691 if (Subtarget->isPICStyleRIPRel() &&
6692 (M == CodeModel::Small || M == CodeModel::Kernel))
6693 WrapperKind = X86ISD::WrapperRIP;
6694 else if (Subtarget->isPICStyleGOT())
6695 OpFlag = X86II::MO_GOTOFF;
6696 else if (Subtarget->isPICStyleStubPIC())
6697 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6699 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
6701 CP->getOffset(), OpFlag);
6702 DebugLoc DL = CP->getDebugLoc();
6703 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6704 // With PIC, the address is actually $g + Offset.
6706 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6707 DAG.getNode(X86ISD::GlobalBaseReg,
6708 DebugLoc(), getPointerTy()),
6715 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
6716 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
6718 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6720 unsigned char OpFlag = 0;
6721 unsigned WrapperKind = X86ISD::Wrapper;
6722 CodeModel::Model M = getTargetMachine().getCodeModel();
6724 if (Subtarget->isPICStyleRIPRel() &&
6725 (M == CodeModel::Small || M == CodeModel::Kernel))
6726 WrapperKind = X86ISD::WrapperRIP;
6727 else if (Subtarget->isPICStyleGOT())
6728 OpFlag = X86II::MO_GOTOFF;
6729 else if (Subtarget->isPICStyleStubPIC())
6730 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6732 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
6734 DebugLoc DL = JT->getDebugLoc();
6735 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6737 // With PIC, the address is actually $g + Offset.
6739 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6740 DAG.getNode(X86ISD::GlobalBaseReg,
6741 DebugLoc(), getPointerTy()),
6748 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
6749 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
6751 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6753 unsigned char OpFlag = 0;
6754 unsigned WrapperKind = X86ISD::Wrapper;
6755 CodeModel::Model M = getTargetMachine().getCodeModel();
6757 if (Subtarget->isPICStyleRIPRel() &&
6758 (M == CodeModel::Small || M == CodeModel::Kernel))
6759 WrapperKind = X86ISD::WrapperRIP;
6760 else if (Subtarget->isPICStyleGOT())
6761 OpFlag = X86II::MO_GOTOFF;
6762 else if (Subtarget->isPICStyleStubPIC())
6763 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6765 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
6767 DebugLoc DL = Op.getDebugLoc();
6768 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6771 // With PIC, the address is actually $g + Offset.
6772 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
6773 !Subtarget->is64Bit()) {
6774 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6775 DAG.getNode(X86ISD::GlobalBaseReg,
6776 DebugLoc(), getPointerTy()),
6784 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
6785 // Create the TargetBlockAddressAddress node.
6786 unsigned char OpFlags =
6787 Subtarget->ClassifyBlockAddressReference();
6788 CodeModel::Model M = getTargetMachine().getCodeModel();
6789 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
6790 DebugLoc dl = Op.getDebugLoc();
6791 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
6792 /*isTarget=*/true, OpFlags);
6794 if (Subtarget->isPICStyleRIPRel() &&
6795 (M == CodeModel::Small || M == CodeModel::Kernel))
6796 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6798 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6800 // With PIC, the address is actually $g + Offset.
6801 if (isGlobalRelativeToPICBase(OpFlags)) {
6802 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6803 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6811 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
6813 SelectionDAG &DAG) const {
6814 // Create the TargetGlobalAddress node, folding in the constant
6815 // offset if it is legal.
6816 unsigned char OpFlags =
6817 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
6818 CodeModel::Model M = getTargetMachine().getCodeModel();
6820 if (OpFlags == X86II::MO_NO_FLAG &&
6821 X86::isOffsetSuitableForCodeModel(Offset, M)) {
6822 // A direct static reference to a global.
6823 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
6826 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
6829 if (Subtarget->isPICStyleRIPRel() &&
6830 (M == CodeModel::Small || M == CodeModel::Kernel))
6831 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6833 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6835 // With PIC, the address is actually $g + Offset.
6836 if (isGlobalRelativeToPICBase(OpFlags)) {
6837 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6838 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6842 // For globals that require a load from a stub to get the address, emit the
6844 if (isGlobalStubReference(OpFlags))
6845 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
6846 MachinePointerInfo::getGOT(), false, false, 0);
6848 // If there was a non-zero offset that we didn't fold, create an explicit
6851 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
6852 DAG.getConstant(Offset, getPointerTy()));
6858 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
6859 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
6860 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
6861 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
6865 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
6866 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
6867 unsigned char OperandFlags) {
6868 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6869 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6870 DebugLoc dl = GA->getDebugLoc();
6871 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6872 GA->getValueType(0),
6876 SDValue Ops[] = { Chain, TGA, *InFlag };
6877 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
6879 SDValue Ops[] = { Chain, TGA };
6880 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
6883 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
6884 MFI->setAdjustsStack(true);
6886 SDValue Flag = Chain.getValue(1);
6887 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
6890 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
6892 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6895 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
6896 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
6897 DAG.getNode(X86ISD::GlobalBaseReg,
6898 DebugLoc(), PtrVT), InFlag);
6899 InFlag = Chain.getValue(1);
6901 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
6904 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
6906 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6908 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
6909 X86::RAX, X86II::MO_TLSGD);
6912 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
6913 // "local exec" model.
6914 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6915 const EVT PtrVT, TLSModel::Model model,
6917 DebugLoc dl = GA->getDebugLoc();
6919 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
6920 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
6921 is64Bit ? 257 : 256));
6923 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
6924 DAG.getIntPtrConstant(0),
6925 MachinePointerInfo(Ptr), false, false, 0);
6927 unsigned char OperandFlags = 0;
6928 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
6930 unsigned WrapperKind = X86ISD::Wrapper;
6931 if (model == TLSModel::LocalExec) {
6932 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
6933 } else if (is64Bit) {
6934 assert(model == TLSModel::InitialExec);
6935 OperandFlags = X86II::MO_GOTTPOFF;
6936 WrapperKind = X86ISD::WrapperRIP;
6938 assert(model == TLSModel::InitialExec);
6939 OperandFlags = X86II::MO_INDNTPOFF;
6942 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
6944 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6945 GA->getValueType(0),
6946 GA->getOffset(), OperandFlags);
6947 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
6949 if (model == TLSModel::InitialExec)
6950 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
6951 MachinePointerInfo::getGOT(), false, false, 0);
6953 // The address of the thread local variable is the add of the thread
6954 // pointer with the offset of the variable.
6955 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
6959 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
6961 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
6962 const GlobalValue *GV = GA->getGlobal();
6964 if (Subtarget->isTargetELF()) {
6965 // TODO: implement the "local dynamic" model
6966 // TODO: implement the "initial exec"model for pic executables
6968 // If GV is an alias then use the aliasee for determining
6969 // thread-localness.
6970 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
6971 GV = GA->resolveAliasedGlobal(false);
6973 TLSModel::Model model
6974 = getTLSModel(GV, getTargetMachine().getRelocationModel());
6977 case TLSModel::GeneralDynamic:
6978 case TLSModel::LocalDynamic: // not implemented
6979 if (Subtarget->is64Bit())
6980 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
6981 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
6983 case TLSModel::InitialExec:
6984 case TLSModel::LocalExec:
6985 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
6986 Subtarget->is64Bit());
6988 } else if (Subtarget->isTargetDarwin()) {
6989 // Darwin only has one model of TLS. Lower to that.
6990 unsigned char OpFlag = 0;
6991 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
6992 X86ISD::WrapperRIP : X86ISD::Wrapper;
6994 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6996 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
6997 !Subtarget->is64Bit();
6999 OpFlag = X86II::MO_TLVP_PIC_BASE;
7001 OpFlag = X86II::MO_TLVP;
7002 DebugLoc DL = Op.getDebugLoc();
7003 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
7004 GA->getValueType(0),
7005 GA->getOffset(), OpFlag);
7006 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7008 // With PIC32, the address is actually $g + Offset.
7010 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7011 DAG.getNode(X86ISD::GlobalBaseReg,
7012 DebugLoc(), getPointerTy()),
7015 // Lowering the machine isd will make sure everything is in the right
7017 SDValue Chain = DAG.getEntryNode();
7018 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
7019 SDValue Args[] = { Chain, Offset };
7020 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args, 2);
7022 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
7023 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7024 MFI->setAdjustsStack(true);
7026 // And our return value (tls address) is in the standard call return value
7028 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
7029 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
7033 "TLS not implemented for this target.");
7035 llvm_unreachable("Unreachable");
7040 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values and
7041 /// take a 2 x i32 value to shift plus a shift amount.
7042 SDValue X86TargetLowering::LowerShiftParts(SDValue Op, SelectionDAG &DAG) const {
7043 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
7044 EVT VT = Op.getValueType();
7045 unsigned VTBits = VT.getSizeInBits();
7046 DebugLoc dl = Op.getDebugLoc();
7047 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
7048 SDValue ShOpLo = Op.getOperand(0);
7049 SDValue ShOpHi = Op.getOperand(1);
7050 SDValue ShAmt = Op.getOperand(2);
7051 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
7052 DAG.getConstant(VTBits - 1, MVT::i8))
7053 : DAG.getConstant(0, VT);
7056 if (Op.getOpcode() == ISD::SHL_PARTS) {
7057 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
7058 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
7060 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
7061 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
7064 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
7065 DAG.getConstant(VTBits, MVT::i8));
7066 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
7067 AndNode, DAG.getConstant(0, MVT::i8));
7070 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7071 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
7072 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
7074 if (Op.getOpcode() == ISD::SHL_PARTS) {
7075 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
7076 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
7078 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
7079 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
7082 SDValue Ops[2] = { Lo, Hi };
7083 return DAG.getMergeValues(Ops, 2, dl);
7086 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
7087 SelectionDAG &DAG) const {
7088 EVT SrcVT = Op.getOperand(0).getValueType();
7090 if (SrcVT.isVector())
7093 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
7094 "Unknown SINT_TO_FP to lower!");
7096 // These are really Legal; return the operand so the caller accepts it as
7098 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
7100 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
7101 Subtarget->is64Bit()) {
7105 DebugLoc dl = Op.getDebugLoc();
7106 unsigned Size = SrcVT.getSizeInBits()/8;
7107 MachineFunction &MF = DAG.getMachineFunction();
7108 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
7109 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7110 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7112 MachinePointerInfo::getFixedStack(SSFI),
7114 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
7117 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
7119 SelectionDAG &DAG) const {
7121 DebugLoc DL = Op.getDebugLoc();
7123 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
7125 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
7127 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
7129 unsigned ByteSize = SrcVT.getSizeInBits()/8;
7131 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
7132 MachineMemOperand *MMO;
7134 int SSFI = FI->getIndex();
7136 DAG.getMachineFunction()
7137 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7138 MachineMemOperand::MOLoad, ByteSize, ByteSize);
7140 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
7141 StackSlot = StackSlot.getOperand(1);
7143 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
7144 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
7146 Tys, Ops, array_lengthof(Ops),
7150 Chain = Result.getValue(1);
7151 SDValue InFlag = Result.getValue(2);
7153 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
7154 // shouldn't be necessary except that RFP cannot be live across
7155 // multiple blocks. When stackifier is fixed, they can be uncoupled.
7156 MachineFunction &MF = DAG.getMachineFunction();
7157 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
7158 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
7159 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7160 Tys = DAG.getVTList(MVT::Other);
7162 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
7164 MachineMemOperand *MMO =
7165 DAG.getMachineFunction()
7166 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7167 MachineMemOperand::MOStore, SSFISize, SSFISize);
7169 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
7170 Ops, array_lengthof(Ops),
7171 Op.getValueType(), MMO);
7172 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
7173 MachinePointerInfo::getFixedStack(SSFI),
7180 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
7181 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
7182 SelectionDAG &DAG) const {
7183 // This algorithm is not obvious. Here it is in C code, more or less:
7185 double uint64_to_double( uint32_t hi, uint32_t lo ) {
7186 static const __m128i exp = { 0x4330000045300000ULL, 0 };
7187 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
7189 // Copy ints to xmm registers.
7190 __m128i xh = _mm_cvtsi32_si128( hi );
7191 __m128i xl = _mm_cvtsi32_si128( lo );
7193 // Combine into low half of a single xmm register.
7194 __m128i x = _mm_unpacklo_epi32( xh, xl );
7198 // Merge in appropriate exponents to give the integer bits the right
7200 x = _mm_unpacklo_epi32( x, exp );
7202 // Subtract away the biases to deal with the IEEE-754 double precision
7204 d = _mm_sub_pd( (__m128d) x, bias );
7206 // All conversions up to here are exact. The correctly rounded result is
7207 // calculated using the current rounding mode using the following
7209 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
7210 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
7211 // store doesn't really need to be here (except
7212 // maybe to zero the other double)
7217 DebugLoc dl = Op.getDebugLoc();
7218 LLVMContext *Context = DAG.getContext();
7220 // Build some magic constants.
7221 std::vector<Constant*> CV0;
7222 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
7223 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
7224 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
7225 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
7226 Constant *C0 = ConstantVector::get(CV0);
7227 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
7229 std::vector<Constant*> CV1;
7231 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
7233 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
7234 Constant *C1 = ConstantVector::get(CV1);
7235 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
7237 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
7238 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7240 DAG.getIntPtrConstant(1)));
7241 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
7242 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7244 DAG.getIntPtrConstant(0)));
7245 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
7246 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
7247 MachinePointerInfo::getConstantPool(),
7249 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
7250 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck2);
7251 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
7252 MachinePointerInfo::getConstantPool(),
7254 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
7256 // Add the halves; easiest way is to swap them into another reg first.
7257 int ShufMask[2] = { 1, -1 };
7258 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
7259 DAG.getUNDEF(MVT::v2f64), ShufMask);
7260 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
7261 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
7262 DAG.getIntPtrConstant(0));
7265 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
7266 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
7267 SelectionDAG &DAG) const {
7268 DebugLoc dl = Op.getDebugLoc();
7269 // FP constant to bias correct the final result.
7270 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
7273 // Load the 32-bit value into an XMM register.
7274 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
7277 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7278 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
7279 DAG.getIntPtrConstant(0));
7281 // Or the load with the bias.
7282 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
7283 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7284 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
7286 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7287 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
7288 MVT::v2f64, Bias)));
7289 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7290 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
7291 DAG.getIntPtrConstant(0));
7293 // Subtract the bias.
7294 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
7296 // Handle final rounding.
7297 EVT DestVT = Op.getValueType();
7299 if (DestVT.bitsLT(MVT::f64)) {
7300 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
7301 DAG.getIntPtrConstant(0));
7302 } else if (DestVT.bitsGT(MVT::f64)) {
7303 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
7306 // Handle final rounding.
7310 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
7311 SelectionDAG &DAG) const {
7312 SDValue N0 = Op.getOperand(0);
7313 DebugLoc dl = Op.getDebugLoc();
7315 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
7316 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
7317 // the optimization here.
7318 if (DAG.SignBitIsZero(N0))
7319 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
7321 EVT SrcVT = N0.getValueType();
7322 EVT DstVT = Op.getValueType();
7323 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
7324 return LowerUINT_TO_FP_i64(Op, DAG);
7325 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
7326 return LowerUINT_TO_FP_i32(Op, DAG);
7328 // Make a 64-bit buffer, and use it to build an FILD.
7329 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
7330 if (SrcVT == MVT::i32) {
7331 SDValue WordOff = DAG.getConstant(4, getPointerTy());
7332 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
7333 getPointerTy(), StackSlot, WordOff);
7334 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7335 StackSlot, MachinePointerInfo(),
7337 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
7338 OffsetSlot, MachinePointerInfo(),
7340 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
7344 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
7345 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7346 StackSlot, MachinePointerInfo(),
7348 // For i64 source, we need to add the appropriate power of 2 if the input
7349 // was negative. This is the same as the optimization in
7350 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
7351 // we must be careful to do the computation in x87 extended precision, not
7352 // in SSE. (The generic code can't know it's OK to do this, or how to.)
7353 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
7354 MachineMemOperand *MMO =
7355 DAG.getMachineFunction()
7356 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7357 MachineMemOperand::MOLoad, 8, 8);
7359 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
7360 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
7361 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops, 3,
7364 APInt FF(32, 0x5F800000ULL);
7366 // Check whether the sign bit is set.
7367 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
7368 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
7371 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
7372 SDValue FudgePtr = DAG.getConstantPool(
7373 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
7376 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
7377 SDValue Zero = DAG.getIntPtrConstant(0);
7378 SDValue Four = DAG.getIntPtrConstant(4);
7379 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
7381 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
7383 // Load the value out, extending it from f32 to f80.
7384 // FIXME: Avoid the extend by constructing the right constant pool?
7385 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
7386 FudgePtr, MachinePointerInfo::getConstantPool(),
7387 MVT::f32, false, false, 4);
7388 // Extend everything to 80 bits to force it to be done on x87.
7389 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
7390 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
7393 std::pair<SDValue,SDValue> X86TargetLowering::
7394 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
7395 DebugLoc DL = Op.getDebugLoc();
7397 EVT DstTy = Op.getValueType();
7400 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
7404 assert(DstTy.getSimpleVT() <= MVT::i64 &&
7405 DstTy.getSimpleVT() >= MVT::i16 &&
7406 "Unknown FP_TO_SINT to lower!");
7408 // These are really Legal.
7409 if (DstTy == MVT::i32 &&
7410 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
7411 return std::make_pair(SDValue(), SDValue());
7412 if (Subtarget->is64Bit() &&
7413 DstTy == MVT::i64 &&
7414 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
7415 return std::make_pair(SDValue(), SDValue());
7417 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
7419 MachineFunction &MF = DAG.getMachineFunction();
7420 unsigned MemSize = DstTy.getSizeInBits()/8;
7421 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
7422 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7427 switch (DstTy.getSimpleVT().SimpleTy) {
7428 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
7429 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
7430 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
7431 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
7434 SDValue Chain = DAG.getEntryNode();
7435 SDValue Value = Op.getOperand(0);
7436 EVT TheVT = Op.getOperand(0).getValueType();
7437 if (isScalarFPTypeInSSEReg(TheVT)) {
7438 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
7439 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
7440 MachinePointerInfo::getFixedStack(SSFI),
7442 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
7444 Chain, StackSlot, DAG.getValueType(TheVT)
7447 MachineMemOperand *MMO =
7448 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7449 MachineMemOperand::MOLoad, MemSize, MemSize);
7450 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, 3,
7452 Chain = Value.getValue(1);
7453 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
7454 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7457 MachineMemOperand *MMO =
7458 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7459 MachineMemOperand::MOStore, MemSize, MemSize);
7461 // Build the FP_TO_INT*_IN_MEM
7462 SDValue Ops[] = { Chain, Value, StackSlot };
7463 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
7464 Ops, 3, DstTy, MMO);
7466 return std::make_pair(FIST, StackSlot);
7469 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
7470 SelectionDAG &DAG) const {
7471 if (Op.getValueType().isVector())
7474 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
7475 SDValue FIST = Vals.first, StackSlot = Vals.second;
7476 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
7477 if (FIST.getNode() == 0) return Op;
7480 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
7481 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
7484 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
7485 SelectionDAG &DAG) const {
7486 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
7487 SDValue FIST = Vals.first, StackSlot = Vals.second;
7488 assert(FIST.getNode() && "Unexpected failure");
7491 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
7492 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
7495 SDValue X86TargetLowering::LowerFABS(SDValue Op,
7496 SelectionDAG &DAG) const {
7497 LLVMContext *Context = DAG.getContext();
7498 DebugLoc dl = Op.getDebugLoc();
7499 EVT VT = Op.getValueType();
7502 EltVT = VT.getVectorElementType();
7503 std::vector<Constant*> CV;
7504 if (EltVT == MVT::f64) {
7505 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
7509 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
7515 Constant *C = ConstantVector::get(CV);
7516 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7517 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7518 MachinePointerInfo::getConstantPool(),
7520 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
7523 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
7524 LLVMContext *Context = DAG.getContext();
7525 DebugLoc dl = Op.getDebugLoc();
7526 EVT VT = Op.getValueType();
7529 EltVT = VT.getVectorElementType();
7530 std::vector<Constant*> CV;
7531 if (EltVT == MVT::f64) {
7532 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
7536 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
7542 Constant *C = ConstantVector::get(CV);
7543 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7544 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7545 MachinePointerInfo::getConstantPool(),
7547 if (VT.isVector()) {
7548 return DAG.getNode(ISD::BITCAST, dl, VT,
7549 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
7550 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7552 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, Mask)));
7554 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
7558 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
7559 LLVMContext *Context = DAG.getContext();
7560 SDValue Op0 = Op.getOperand(0);
7561 SDValue Op1 = Op.getOperand(1);
7562 DebugLoc dl = Op.getDebugLoc();
7563 EVT VT = Op.getValueType();
7564 EVT SrcVT = Op1.getValueType();
7566 // If second operand is smaller, extend it first.
7567 if (SrcVT.bitsLT(VT)) {
7568 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
7571 // And if it is bigger, shrink it first.
7572 if (SrcVT.bitsGT(VT)) {
7573 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
7577 // At this point the operands and the result should have the same
7578 // type, and that won't be f80 since that is not custom lowered.
7580 // First get the sign bit of second operand.
7581 std::vector<Constant*> CV;
7582 if (SrcVT == MVT::f64) {
7583 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
7584 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
7586 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
7587 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7588 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7589 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7591 Constant *C = ConstantVector::get(CV);
7592 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7593 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
7594 MachinePointerInfo::getConstantPool(),
7596 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
7598 // Shift sign bit right or left if the two operands have different types.
7599 if (SrcVT.bitsGT(VT)) {
7600 // Op0 is MVT::f32, Op1 is MVT::f64.
7601 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
7602 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
7603 DAG.getConstant(32, MVT::i32));
7604 SignBit = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, SignBit);
7605 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
7606 DAG.getIntPtrConstant(0));
7609 // Clear first operand sign bit.
7611 if (VT == MVT::f64) {
7612 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
7613 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
7615 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
7616 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7617 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7618 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7620 C = ConstantVector::get(CV);
7621 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7622 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7623 MachinePointerInfo::getConstantPool(),
7625 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
7627 // Or the value with the sign bit.
7628 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
7631 SDValue X86TargetLowering::LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) const {
7632 SDValue N0 = Op.getOperand(0);
7633 DebugLoc dl = Op.getDebugLoc();
7634 EVT VT = Op.getValueType();
7636 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
7637 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
7638 DAG.getConstant(1, VT));
7639 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT));
7642 /// Emit nodes that will be selected as "test Op0,Op0", or something
7644 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
7645 SelectionDAG &DAG) const {
7646 DebugLoc dl = Op.getDebugLoc();
7648 // CF and OF aren't always set the way we want. Determine which
7649 // of these we need.
7650 bool NeedCF = false;
7651 bool NeedOF = false;
7654 case X86::COND_A: case X86::COND_AE:
7655 case X86::COND_B: case X86::COND_BE:
7658 case X86::COND_G: case X86::COND_GE:
7659 case X86::COND_L: case X86::COND_LE:
7660 case X86::COND_O: case X86::COND_NO:
7665 // See if we can use the EFLAGS value from the operand instead of
7666 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
7667 // we prove that the arithmetic won't overflow, we can't use OF or CF.
7668 if (Op.getResNo() != 0 || NeedOF || NeedCF)
7669 // Emit a CMP with 0, which is the TEST pattern.
7670 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
7671 DAG.getConstant(0, Op.getValueType()));
7673 unsigned Opcode = 0;
7674 unsigned NumOperands = 0;
7675 switch (Op.getNode()->getOpcode()) {
7677 // Due to an isel shortcoming, be conservative if this add is likely to be
7678 // selected as part of a load-modify-store instruction. When the root node
7679 // in a match is a store, isel doesn't know how to remap non-chain non-flag
7680 // uses of other nodes in the match, such as the ADD in this case. This
7681 // leads to the ADD being left around and reselected, with the result being
7682 // two adds in the output. Alas, even if none our users are stores, that
7683 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
7684 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
7685 // climbing the DAG back to the root, and it doesn't seem to be worth the
7687 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7688 UE = Op.getNode()->use_end(); UI != UE; ++UI)
7689 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
7692 if (ConstantSDNode *C =
7693 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
7694 // An add of one will be selected as an INC.
7695 if (C->getAPIntValue() == 1) {
7696 Opcode = X86ISD::INC;
7701 // An add of negative one (subtract of one) will be selected as a DEC.
7702 if (C->getAPIntValue().isAllOnesValue()) {
7703 Opcode = X86ISD::DEC;
7709 // Otherwise use a regular EFLAGS-setting add.
7710 Opcode = X86ISD::ADD;
7714 // If the primary and result isn't used, don't bother using X86ISD::AND,
7715 // because a TEST instruction will be better.
7716 bool NonFlagUse = false;
7717 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7718 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
7720 unsigned UOpNo = UI.getOperandNo();
7721 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
7722 // Look pass truncate.
7723 UOpNo = User->use_begin().getOperandNo();
7724 User = *User->use_begin();
7727 if (User->getOpcode() != ISD::BRCOND &&
7728 User->getOpcode() != ISD::SETCC &&
7729 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
7742 // Due to the ISEL shortcoming noted above, be conservative if this op is
7743 // likely to be selected as part of a load-modify-store instruction.
7744 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7745 UE = Op.getNode()->use_end(); UI != UE; ++UI)
7746 if (UI->getOpcode() == ISD::STORE)
7749 // Otherwise use a regular EFLAGS-setting instruction.
7750 switch (Op.getNode()->getOpcode()) {
7751 default: llvm_unreachable("unexpected operator!");
7752 case ISD::SUB: Opcode = X86ISD::SUB; break;
7753 case ISD::OR: Opcode = X86ISD::OR; break;
7754 case ISD::XOR: Opcode = X86ISD::XOR; break;
7755 case ISD::AND: Opcode = X86ISD::AND; break;
7767 return SDValue(Op.getNode(), 1);
7774 // Emit a CMP with 0, which is the TEST pattern.
7775 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
7776 DAG.getConstant(0, Op.getValueType()));
7778 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
7779 SmallVector<SDValue, 4> Ops;
7780 for (unsigned i = 0; i != NumOperands; ++i)
7781 Ops.push_back(Op.getOperand(i));
7783 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
7784 DAG.ReplaceAllUsesWith(Op, New);
7785 return SDValue(New.getNode(), 1);
7788 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
7790 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
7791 SelectionDAG &DAG) const {
7792 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
7793 if (C->getAPIntValue() == 0)
7794 return EmitTest(Op0, X86CC, DAG);
7796 DebugLoc dl = Op0.getDebugLoc();
7797 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
7800 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
7801 /// if it's possible.
7802 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
7803 DebugLoc dl, SelectionDAG &DAG) const {
7804 SDValue Op0 = And.getOperand(0);
7805 SDValue Op1 = And.getOperand(1);
7806 if (Op0.getOpcode() == ISD::TRUNCATE)
7807 Op0 = Op0.getOperand(0);
7808 if (Op1.getOpcode() == ISD::TRUNCATE)
7809 Op1 = Op1.getOperand(0);
7812 if (Op1.getOpcode() == ISD::SHL)
7813 std::swap(Op0, Op1);
7814 if (Op0.getOpcode() == ISD::SHL) {
7815 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
7816 if (And00C->getZExtValue() == 1) {
7817 // If we looked past a truncate, check that it's only truncating away
7819 unsigned BitWidth = Op0.getValueSizeInBits();
7820 unsigned AndBitWidth = And.getValueSizeInBits();
7821 if (BitWidth > AndBitWidth) {
7822 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
7823 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
7824 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
7828 RHS = Op0.getOperand(1);
7830 } else if (Op1.getOpcode() == ISD::Constant) {
7831 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
7832 SDValue AndLHS = Op0;
7833 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
7834 LHS = AndLHS.getOperand(0);
7835 RHS = AndLHS.getOperand(1);
7839 if (LHS.getNode()) {
7840 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
7841 // instruction. Since the shift amount is in-range-or-undefined, we know
7842 // that doing a bittest on the i32 value is ok. We extend to i32 because
7843 // the encoding for the i16 version is larger than the i32 version.
7844 // Also promote i16 to i32 for performance / code size reason.
7845 if (LHS.getValueType() == MVT::i8 ||
7846 LHS.getValueType() == MVT::i16)
7847 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
7849 // If the operand types disagree, extend the shift amount to match. Since
7850 // BT ignores high bits (like shifts) we can use anyextend.
7851 if (LHS.getValueType() != RHS.getValueType())
7852 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
7854 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
7855 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
7856 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7857 DAG.getConstant(Cond, MVT::i8), BT);
7863 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
7864 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
7865 SDValue Op0 = Op.getOperand(0);
7866 SDValue Op1 = Op.getOperand(1);
7867 DebugLoc dl = Op.getDebugLoc();
7868 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
7870 // Optimize to BT if possible.
7871 // Lower (X & (1 << N)) == 0 to BT(X, N).
7872 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
7873 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
7874 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
7875 Op1.getOpcode() == ISD::Constant &&
7876 cast<ConstantSDNode>(Op1)->isNullValue() &&
7877 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7878 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
7879 if (NewSetCC.getNode())
7883 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
7885 if (Op1.getOpcode() == ISD::Constant &&
7886 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
7887 cast<ConstantSDNode>(Op1)->isNullValue()) &&
7888 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7890 // If the input is a setcc, then reuse the input setcc or use a new one with
7891 // the inverted condition.
7892 if (Op0.getOpcode() == X86ISD::SETCC) {
7893 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
7894 bool Invert = (CC == ISD::SETNE) ^
7895 cast<ConstantSDNode>(Op1)->isNullValue();
7896 if (!Invert) return Op0;
7898 CCode = X86::GetOppositeBranchCondition(CCode);
7899 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7900 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
7904 bool isFP = Op1.getValueType().isFloatingPoint();
7905 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
7906 if (X86CC == X86::COND_INVALID)
7909 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
7910 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7911 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
7914 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
7916 SDValue Op0 = Op.getOperand(0);
7917 SDValue Op1 = Op.getOperand(1);
7918 SDValue CC = Op.getOperand(2);
7919 EVT VT = Op.getValueType();
7920 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
7921 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
7922 DebugLoc dl = Op.getDebugLoc();
7926 EVT EltVT = Op0.getValueType().getVectorElementType();
7927 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
7929 unsigned Opc = EltVT == MVT::f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
7932 switch (SetCCOpcode) {
7935 case ISD::SETEQ: SSECC = 0; break;
7937 case ISD::SETGT: Swap = true; // Fallthrough
7939 case ISD::SETOLT: SSECC = 1; break;
7941 case ISD::SETGE: Swap = true; // Fallthrough
7943 case ISD::SETOLE: SSECC = 2; break;
7944 case ISD::SETUO: SSECC = 3; break;
7946 case ISD::SETNE: SSECC = 4; break;
7947 case ISD::SETULE: Swap = true;
7948 case ISD::SETUGE: SSECC = 5; break;
7949 case ISD::SETULT: Swap = true;
7950 case ISD::SETUGT: SSECC = 6; break;
7951 case ISD::SETO: SSECC = 7; break;
7954 std::swap(Op0, Op1);
7956 // In the two special cases we can't handle, emit two comparisons.
7958 if (SetCCOpcode == ISD::SETUEQ) {
7960 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
7961 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
7962 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
7964 else if (SetCCOpcode == ISD::SETONE) {
7966 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
7967 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
7968 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
7970 llvm_unreachable("Illegal FP comparison");
7972 // Handle all other FP comparisons here.
7973 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
7976 if (!isFP && VT.getSizeInBits() == 256)
7979 // We are handling one of the integer comparisons here. Since SSE only has
7980 // GT and EQ comparisons for integer, swapping operands and multiple
7981 // operations may be required for some comparisons.
7982 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
7983 bool Swap = false, Invert = false, FlipSigns = false;
7985 switch (VT.getSimpleVT().SimpleTy) {
7987 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
7988 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
7989 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
7990 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
7993 switch (SetCCOpcode) {
7995 case ISD::SETNE: Invert = true;
7996 case ISD::SETEQ: Opc = EQOpc; break;
7997 case ISD::SETLT: Swap = true;
7998 case ISD::SETGT: Opc = GTOpc; break;
7999 case ISD::SETGE: Swap = true;
8000 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
8001 case ISD::SETULT: Swap = true;
8002 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
8003 case ISD::SETUGE: Swap = true;
8004 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
8007 std::swap(Op0, Op1);
8009 // Since SSE has no unsigned integer comparisons, we need to flip the sign
8010 // bits of the inputs before performing those operations.
8012 EVT EltVT = VT.getVectorElementType();
8013 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
8015 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
8016 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
8018 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
8019 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
8022 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
8024 // If the logical-not of the result is required, perform that now.
8026 Result = DAG.getNOT(dl, Result, VT);
8031 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
8032 static bool isX86LogicalCmp(SDValue Op) {
8033 unsigned Opc = Op.getNode()->getOpcode();
8034 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
8036 if (Op.getResNo() == 1 &&
8037 (Opc == X86ISD::ADD ||
8038 Opc == X86ISD::SUB ||
8039 Opc == X86ISD::ADC ||
8040 Opc == X86ISD::SBB ||
8041 Opc == X86ISD::SMUL ||
8042 Opc == X86ISD::UMUL ||
8043 Opc == X86ISD::INC ||
8044 Opc == X86ISD::DEC ||
8045 Opc == X86ISD::OR ||
8046 Opc == X86ISD::XOR ||
8047 Opc == X86ISD::AND))
8050 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
8056 static bool isZero(SDValue V) {
8057 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
8058 return C && C->isNullValue();
8061 static bool isAllOnes(SDValue V) {
8062 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
8063 return C && C->isAllOnesValue();
8066 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
8067 bool addTest = true;
8068 SDValue Cond = Op.getOperand(0);
8069 SDValue Op1 = Op.getOperand(1);
8070 SDValue Op2 = Op.getOperand(2);
8071 DebugLoc DL = Op.getDebugLoc();
8074 if (Cond.getOpcode() == ISD::SETCC) {
8075 SDValue NewCond = LowerSETCC(Cond, DAG);
8076 if (NewCond.getNode())
8080 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
8081 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
8082 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
8083 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
8084 if (Cond.getOpcode() == X86ISD::SETCC &&
8085 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
8086 isZero(Cond.getOperand(1).getOperand(1))) {
8087 SDValue Cmp = Cond.getOperand(1);
8089 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
8091 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
8092 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
8093 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
8095 SDValue CmpOp0 = Cmp.getOperand(0);
8096 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
8097 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
8099 SDValue Res = // Res = 0 or -1.
8100 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
8101 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
8103 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
8104 Res = DAG.getNOT(DL, Res, Res.getValueType());
8106 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
8107 if (N2C == 0 || !N2C->isNullValue())
8108 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
8113 // Look past (and (setcc_carry (cmp ...)), 1).
8114 if (Cond.getOpcode() == ISD::AND &&
8115 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
8116 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
8117 if (C && C->getAPIntValue() == 1)
8118 Cond = Cond.getOperand(0);
8121 // If condition flag is set by a X86ISD::CMP, then use it as the condition
8122 // setting operand in place of the X86ISD::SETCC.
8123 if (Cond.getOpcode() == X86ISD::SETCC ||
8124 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
8125 CC = Cond.getOperand(0);
8127 SDValue Cmp = Cond.getOperand(1);
8128 unsigned Opc = Cmp.getOpcode();
8129 EVT VT = Op.getValueType();
8131 bool IllegalFPCMov = false;
8132 if (VT.isFloatingPoint() && !VT.isVector() &&
8133 !isScalarFPTypeInSSEReg(VT)) // FPStack?
8134 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
8136 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
8137 Opc == X86ISD::BT) { // FIXME
8144 // Look pass the truncate.
8145 if (Cond.getOpcode() == ISD::TRUNCATE)
8146 Cond = Cond.getOperand(0);
8148 // We know the result of AND is compared against zero. Try to match
8150 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
8151 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
8152 if (NewSetCC.getNode()) {
8153 CC = NewSetCC.getOperand(0);
8154 Cond = NewSetCC.getOperand(1);
8161 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
8162 Cond = EmitTest(Cond, X86::COND_NE, DAG);
8165 // a < b ? -1 : 0 -> RES = ~setcc_carry
8166 // a < b ? 0 : -1 -> RES = setcc_carry
8167 // a >= b ? -1 : 0 -> RES = setcc_carry
8168 // a >= b ? 0 : -1 -> RES = ~setcc_carry
8169 if (Cond.getOpcode() == X86ISD::CMP) {
8170 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
8172 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
8173 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
8174 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
8175 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
8176 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
8177 return DAG.getNOT(DL, Res, Res.getValueType());
8182 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
8183 // condition is true.
8184 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
8185 SDValue Ops[] = { Op2, Op1, CC, Cond };
8186 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops, array_lengthof(Ops));
8189 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
8190 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
8191 // from the AND / OR.
8192 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
8193 Opc = Op.getOpcode();
8194 if (Opc != ISD::OR && Opc != ISD::AND)
8196 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
8197 Op.getOperand(0).hasOneUse() &&
8198 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
8199 Op.getOperand(1).hasOneUse());
8202 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
8203 // 1 and that the SETCC node has a single use.
8204 static bool isXor1OfSetCC(SDValue Op) {
8205 if (Op.getOpcode() != ISD::XOR)
8207 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
8208 if (N1C && N1C->getAPIntValue() == 1) {
8209 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
8210 Op.getOperand(0).hasOneUse();
8215 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
8216 bool addTest = true;
8217 SDValue Chain = Op.getOperand(0);
8218 SDValue Cond = Op.getOperand(1);
8219 SDValue Dest = Op.getOperand(2);
8220 DebugLoc dl = Op.getDebugLoc();
8223 if (Cond.getOpcode() == ISD::SETCC) {
8224 SDValue NewCond = LowerSETCC(Cond, DAG);
8225 if (NewCond.getNode())
8229 // FIXME: LowerXALUO doesn't handle these!!
8230 else if (Cond.getOpcode() == X86ISD::ADD ||
8231 Cond.getOpcode() == X86ISD::SUB ||
8232 Cond.getOpcode() == X86ISD::SMUL ||
8233 Cond.getOpcode() == X86ISD::UMUL)
8234 Cond = LowerXALUO(Cond, DAG);
8237 // Look pass (and (setcc_carry (cmp ...)), 1).
8238 if (Cond.getOpcode() == ISD::AND &&
8239 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
8240 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
8241 if (C && C->getAPIntValue() == 1)
8242 Cond = Cond.getOperand(0);
8245 // If condition flag is set by a X86ISD::CMP, then use it as the condition
8246 // setting operand in place of the X86ISD::SETCC.
8247 if (Cond.getOpcode() == X86ISD::SETCC ||
8248 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
8249 CC = Cond.getOperand(0);
8251 SDValue Cmp = Cond.getOperand(1);
8252 unsigned Opc = Cmp.getOpcode();
8253 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
8254 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
8258 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
8262 // These can only come from an arithmetic instruction with overflow,
8263 // e.g. SADDO, UADDO.
8264 Cond = Cond.getNode()->getOperand(1);
8271 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
8272 SDValue Cmp = Cond.getOperand(0).getOperand(1);
8273 if (CondOpc == ISD::OR) {
8274 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
8275 // two branches instead of an explicit OR instruction with a
8277 if (Cmp == Cond.getOperand(1).getOperand(1) &&
8278 isX86LogicalCmp(Cmp)) {
8279 CC = Cond.getOperand(0).getOperand(0);
8280 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8281 Chain, Dest, CC, Cmp);
8282 CC = Cond.getOperand(1).getOperand(0);
8286 } else { // ISD::AND
8287 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
8288 // two branches instead of an explicit AND instruction with a
8289 // separate test. However, we only do this if this block doesn't
8290 // have a fall-through edge, because this requires an explicit
8291 // jmp when the condition is false.
8292 if (Cmp == Cond.getOperand(1).getOperand(1) &&
8293 isX86LogicalCmp(Cmp) &&
8294 Op.getNode()->hasOneUse()) {
8295 X86::CondCode CCode =
8296 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
8297 CCode = X86::GetOppositeBranchCondition(CCode);
8298 CC = DAG.getConstant(CCode, MVT::i8);
8299 SDNode *User = *Op.getNode()->use_begin();
8300 // Look for an unconditional branch following this conditional branch.
8301 // We need this because we need to reverse the successors in order
8302 // to implement FCMP_OEQ.
8303 if (User->getOpcode() == ISD::BR) {
8304 SDValue FalseBB = User->getOperand(1);
8306 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
8307 assert(NewBR == User);
8311 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8312 Chain, Dest, CC, Cmp);
8313 X86::CondCode CCode =
8314 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
8315 CCode = X86::GetOppositeBranchCondition(CCode);
8316 CC = DAG.getConstant(CCode, MVT::i8);
8322 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
8323 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
8324 // It should be transformed during dag combiner except when the condition
8325 // is set by a arithmetics with overflow node.
8326 X86::CondCode CCode =
8327 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
8328 CCode = X86::GetOppositeBranchCondition(CCode);
8329 CC = DAG.getConstant(CCode, MVT::i8);
8330 Cond = Cond.getOperand(0).getOperand(1);
8336 // Look pass the truncate.
8337 if (Cond.getOpcode() == ISD::TRUNCATE)
8338 Cond = Cond.getOperand(0);
8340 // We know the result of AND is compared against zero. Try to match
8342 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
8343 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
8344 if (NewSetCC.getNode()) {
8345 CC = NewSetCC.getOperand(0);
8346 Cond = NewSetCC.getOperand(1);
8353 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
8354 Cond = EmitTest(Cond, X86::COND_NE, DAG);
8356 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8357 Chain, Dest, CC, Cond);
8361 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
8362 // Calls to _alloca is needed to probe the stack when allocating more than 4k
8363 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
8364 // that the guard pages used by the OS virtual memory manager are allocated in
8365 // correct sequence.
8367 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
8368 SelectionDAG &DAG) const {
8369 assert((Subtarget->isTargetCygMing() || Subtarget->isTargetWindows()) &&
8370 "This should be used only on Windows targets");
8371 assert(!Subtarget->isTargetEnvMacho());
8372 DebugLoc dl = Op.getDebugLoc();
8375 SDValue Chain = Op.getOperand(0);
8376 SDValue Size = Op.getOperand(1);
8377 // FIXME: Ensure alignment here
8381 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
8382 unsigned Reg = (Subtarget->is64Bit() ? X86::RAX : X86::EAX);
8384 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
8385 Flag = Chain.getValue(1);
8387 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
8389 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
8390 Flag = Chain.getValue(1);
8392 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
8394 SDValue Ops1[2] = { Chain.getValue(0), Chain };
8395 return DAG.getMergeValues(Ops1, 2, dl);
8398 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
8399 MachineFunction &MF = DAG.getMachineFunction();
8400 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
8402 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
8403 DebugLoc DL = Op.getDebugLoc();
8405 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
8406 // vastart just stores the address of the VarArgsFrameIndex slot into the
8407 // memory location argument.
8408 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
8410 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
8411 MachinePointerInfo(SV), false, false, 0);
8415 // gp_offset (0 - 6 * 8)
8416 // fp_offset (48 - 48 + 8 * 16)
8417 // overflow_arg_area (point to parameters coming in memory).
8419 SmallVector<SDValue, 8> MemOps;
8420 SDValue FIN = Op.getOperand(1);
8422 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
8423 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
8425 FIN, MachinePointerInfo(SV), false, false, 0);
8426 MemOps.push_back(Store);
8429 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8430 FIN, DAG.getIntPtrConstant(4));
8431 Store = DAG.getStore(Op.getOperand(0), DL,
8432 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
8434 FIN, MachinePointerInfo(SV, 4), false, false, 0);
8435 MemOps.push_back(Store);
8437 // Store ptr to overflow_arg_area
8438 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8439 FIN, DAG.getIntPtrConstant(4));
8440 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
8442 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
8443 MachinePointerInfo(SV, 8),
8445 MemOps.push_back(Store);
8447 // Store ptr to reg_save_area.
8448 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8449 FIN, DAG.getIntPtrConstant(8));
8450 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
8452 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
8453 MachinePointerInfo(SV, 16), false, false, 0);
8454 MemOps.push_back(Store);
8455 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
8456 &MemOps[0], MemOps.size());
8459 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
8460 assert(Subtarget->is64Bit() &&
8461 "LowerVAARG only handles 64-bit va_arg!");
8462 assert((Subtarget->isTargetLinux() ||
8463 Subtarget->isTargetDarwin()) &&
8464 "Unhandled target in LowerVAARG");
8465 assert(Op.getNode()->getNumOperands() == 4);
8466 SDValue Chain = Op.getOperand(0);
8467 SDValue SrcPtr = Op.getOperand(1);
8468 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
8469 unsigned Align = Op.getConstantOperandVal(3);
8470 DebugLoc dl = Op.getDebugLoc();
8472 EVT ArgVT = Op.getNode()->getValueType(0);
8473 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
8474 uint32_t ArgSize = getTargetData()->getTypeAllocSize(ArgTy);
8477 // Decide which area this value should be read from.
8478 // TODO: Implement the AMD64 ABI in its entirety. This simple
8479 // selection mechanism works only for the basic types.
8480 if (ArgVT == MVT::f80) {
8481 llvm_unreachable("va_arg for f80 not yet implemented");
8482 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
8483 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
8484 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
8485 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
8487 llvm_unreachable("Unhandled argument type in LowerVAARG");
8491 // Sanity Check: Make sure using fp_offset makes sense.
8492 assert(!UseSoftFloat &&
8493 !(DAG.getMachineFunction()
8494 .getFunction()->hasFnAttr(Attribute::NoImplicitFloat)) &&
8495 Subtarget->hasXMM());
8498 // Insert VAARG_64 node into the DAG
8499 // VAARG_64 returns two values: Variable Argument Address, Chain
8500 SmallVector<SDValue, 11> InstOps;
8501 InstOps.push_back(Chain);
8502 InstOps.push_back(SrcPtr);
8503 InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
8504 InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
8505 InstOps.push_back(DAG.getConstant(Align, MVT::i32));
8506 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
8507 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
8508 VTs, &InstOps[0], InstOps.size(),
8510 MachinePointerInfo(SV),
8515 Chain = VAARG.getValue(1);
8517 // Load the next argument and return it
8518 return DAG.getLoad(ArgVT, dl,
8521 MachinePointerInfo(),
8525 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
8526 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
8527 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
8528 SDValue Chain = Op.getOperand(0);
8529 SDValue DstPtr = Op.getOperand(1);
8530 SDValue SrcPtr = Op.getOperand(2);
8531 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
8532 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
8533 DebugLoc DL = Op.getDebugLoc();
8535 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
8536 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
8538 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
8542 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
8543 DebugLoc dl = Op.getDebugLoc();
8544 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8546 default: return SDValue(); // Don't custom lower most intrinsics.
8547 // Comparison intrinsics.
8548 case Intrinsic::x86_sse_comieq_ss:
8549 case Intrinsic::x86_sse_comilt_ss:
8550 case Intrinsic::x86_sse_comile_ss:
8551 case Intrinsic::x86_sse_comigt_ss:
8552 case Intrinsic::x86_sse_comige_ss:
8553 case Intrinsic::x86_sse_comineq_ss:
8554 case Intrinsic::x86_sse_ucomieq_ss:
8555 case Intrinsic::x86_sse_ucomilt_ss:
8556 case Intrinsic::x86_sse_ucomile_ss:
8557 case Intrinsic::x86_sse_ucomigt_ss:
8558 case Intrinsic::x86_sse_ucomige_ss:
8559 case Intrinsic::x86_sse_ucomineq_ss:
8560 case Intrinsic::x86_sse2_comieq_sd:
8561 case Intrinsic::x86_sse2_comilt_sd:
8562 case Intrinsic::x86_sse2_comile_sd:
8563 case Intrinsic::x86_sse2_comigt_sd:
8564 case Intrinsic::x86_sse2_comige_sd:
8565 case Intrinsic::x86_sse2_comineq_sd:
8566 case Intrinsic::x86_sse2_ucomieq_sd:
8567 case Intrinsic::x86_sse2_ucomilt_sd:
8568 case Intrinsic::x86_sse2_ucomile_sd:
8569 case Intrinsic::x86_sse2_ucomigt_sd:
8570 case Intrinsic::x86_sse2_ucomige_sd:
8571 case Intrinsic::x86_sse2_ucomineq_sd: {
8573 ISD::CondCode CC = ISD::SETCC_INVALID;
8576 case Intrinsic::x86_sse_comieq_ss:
8577 case Intrinsic::x86_sse2_comieq_sd:
8581 case Intrinsic::x86_sse_comilt_ss:
8582 case Intrinsic::x86_sse2_comilt_sd:
8586 case Intrinsic::x86_sse_comile_ss:
8587 case Intrinsic::x86_sse2_comile_sd:
8591 case Intrinsic::x86_sse_comigt_ss:
8592 case Intrinsic::x86_sse2_comigt_sd:
8596 case Intrinsic::x86_sse_comige_ss:
8597 case Intrinsic::x86_sse2_comige_sd:
8601 case Intrinsic::x86_sse_comineq_ss:
8602 case Intrinsic::x86_sse2_comineq_sd:
8606 case Intrinsic::x86_sse_ucomieq_ss:
8607 case Intrinsic::x86_sse2_ucomieq_sd:
8608 Opc = X86ISD::UCOMI;
8611 case Intrinsic::x86_sse_ucomilt_ss:
8612 case Intrinsic::x86_sse2_ucomilt_sd:
8613 Opc = X86ISD::UCOMI;
8616 case Intrinsic::x86_sse_ucomile_ss:
8617 case Intrinsic::x86_sse2_ucomile_sd:
8618 Opc = X86ISD::UCOMI;
8621 case Intrinsic::x86_sse_ucomigt_ss:
8622 case Intrinsic::x86_sse2_ucomigt_sd:
8623 Opc = X86ISD::UCOMI;
8626 case Intrinsic::x86_sse_ucomige_ss:
8627 case Intrinsic::x86_sse2_ucomige_sd:
8628 Opc = X86ISD::UCOMI;
8631 case Intrinsic::x86_sse_ucomineq_ss:
8632 case Intrinsic::x86_sse2_ucomineq_sd:
8633 Opc = X86ISD::UCOMI;
8638 SDValue LHS = Op.getOperand(1);
8639 SDValue RHS = Op.getOperand(2);
8640 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
8641 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
8642 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
8643 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
8644 DAG.getConstant(X86CC, MVT::i8), Cond);
8645 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
8647 // ptest and testp intrinsics. The intrinsic these come from are designed to
8648 // return an integer value, not just an instruction so lower it to the ptest
8649 // or testp pattern and a setcc for the result.
8650 case Intrinsic::x86_sse41_ptestz:
8651 case Intrinsic::x86_sse41_ptestc:
8652 case Intrinsic::x86_sse41_ptestnzc:
8653 case Intrinsic::x86_avx_ptestz_256:
8654 case Intrinsic::x86_avx_ptestc_256:
8655 case Intrinsic::x86_avx_ptestnzc_256:
8656 case Intrinsic::x86_avx_vtestz_ps:
8657 case Intrinsic::x86_avx_vtestc_ps:
8658 case Intrinsic::x86_avx_vtestnzc_ps:
8659 case Intrinsic::x86_avx_vtestz_pd:
8660 case Intrinsic::x86_avx_vtestc_pd:
8661 case Intrinsic::x86_avx_vtestnzc_pd:
8662 case Intrinsic::x86_avx_vtestz_ps_256:
8663 case Intrinsic::x86_avx_vtestc_ps_256:
8664 case Intrinsic::x86_avx_vtestnzc_ps_256:
8665 case Intrinsic::x86_avx_vtestz_pd_256:
8666 case Intrinsic::x86_avx_vtestc_pd_256:
8667 case Intrinsic::x86_avx_vtestnzc_pd_256: {
8668 bool IsTestPacked = false;
8671 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
8672 case Intrinsic::x86_avx_vtestz_ps:
8673 case Intrinsic::x86_avx_vtestz_pd:
8674 case Intrinsic::x86_avx_vtestz_ps_256:
8675 case Intrinsic::x86_avx_vtestz_pd_256:
8676 IsTestPacked = true; // Fallthrough
8677 case Intrinsic::x86_sse41_ptestz:
8678 case Intrinsic::x86_avx_ptestz_256:
8680 X86CC = X86::COND_E;
8682 case Intrinsic::x86_avx_vtestc_ps:
8683 case Intrinsic::x86_avx_vtestc_pd:
8684 case Intrinsic::x86_avx_vtestc_ps_256:
8685 case Intrinsic::x86_avx_vtestc_pd_256:
8686 IsTestPacked = true; // Fallthrough
8687 case Intrinsic::x86_sse41_ptestc:
8688 case Intrinsic::x86_avx_ptestc_256:
8690 X86CC = X86::COND_B;
8692 case Intrinsic::x86_avx_vtestnzc_ps:
8693 case Intrinsic::x86_avx_vtestnzc_pd:
8694 case Intrinsic::x86_avx_vtestnzc_ps_256:
8695 case Intrinsic::x86_avx_vtestnzc_pd_256:
8696 IsTestPacked = true; // Fallthrough
8697 case Intrinsic::x86_sse41_ptestnzc:
8698 case Intrinsic::x86_avx_ptestnzc_256:
8700 X86CC = X86::COND_A;
8704 SDValue LHS = Op.getOperand(1);
8705 SDValue RHS = Op.getOperand(2);
8706 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
8707 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
8708 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
8709 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
8710 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
8713 // Fix vector shift instructions where the last operand is a non-immediate
8715 case Intrinsic::x86_sse2_pslli_w:
8716 case Intrinsic::x86_sse2_pslli_d:
8717 case Intrinsic::x86_sse2_pslli_q:
8718 case Intrinsic::x86_sse2_psrli_w:
8719 case Intrinsic::x86_sse2_psrli_d:
8720 case Intrinsic::x86_sse2_psrli_q:
8721 case Intrinsic::x86_sse2_psrai_w:
8722 case Intrinsic::x86_sse2_psrai_d:
8723 case Intrinsic::x86_mmx_pslli_w:
8724 case Intrinsic::x86_mmx_pslli_d:
8725 case Intrinsic::x86_mmx_pslli_q:
8726 case Intrinsic::x86_mmx_psrli_w:
8727 case Intrinsic::x86_mmx_psrli_d:
8728 case Intrinsic::x86_mmx_psrli_q:
8729 case Intrinsic::x86_mmx_psrai_w:
8730 case Intrinsic::x86_mmx_psrai_d: {
8731 SDValue ShAmt = Op.getOperand(2);
8732 if (isa<ConstantSDNode>(ShAmt))
8735 unsigned NewIntNo = 0;
8736 EVT ShAmtVT = MVT::v4i32;
8738 case Intrinsic::x86_sse2_pslli_w:
8739 NewIntNo = Intrinsic::x86_sse2_psll_w;
8741 case Intrinsic::x86_sse2_pslli_d:
8742 NewIntNo = Intrinsic::x86_sse2_psll_d;
8744 case Intrinsic::x86_sse2_pslli_q:
8745 NewIntNo = Intrinsic::x86_sse2_psll_q;
8747 case Intrinsic::x86_sse2_psrli_w:
8748 NewIntNo = Intrinsic::x86_sse2_psrl_w;
8750 case Intrinsic::x86_sse2_psrli_d:
8751 NewIntNo = Intrinsic::x86_sse2_psrl_d;
8753 case Intrinsic::x86_sse2_psrli_q:
8754 NewIntNo = Intrinsic::x86_sse2_psrl_q;
8756 case Intrinsic::x86_sse2_psrai_w:
8757 NewIntNo = Intrinsic::x86_sse2_psra_w;
8759 case Intrinsic::x86_sse2_psrai_d:
8760 NewIntNo = Intrinsic::x86_sse2_psra_d;
8763 ShAmtVT = MVT::v2i32;
8765 case Intrinsic::x86_mmx_pslli_w:
8766 NewIntNo = Intrinsic::x86_mmx_psll_w;
8768 case Intrinsic::x86_mmx_pslli_d:
8769 NewIntNo = Intrinsic::x86_mmx_psll_d;
8771 case Intrinsic::x86_mmx_pslli_q:
8772 NewIntNo = Intrinsic::x86_mmx_psll_q;
8774 case Intrinsic::x86_mmx_psrli_w:
8775 NewIntNo = Intrinsic::x86_mmx_psrl_w;
8777 case Intrinsic::x86_mmx_psrli_d:
8778 NewIntNo = Intrinsic::x86_mmx_psrl_d;
8780 case Intrinsic::x86_mmx_psrli_q:
8781 NewIntNo = Intrinsic::x86_mmx_psrl_q;
8783 case Intrinsic::x86_mmx_psrai_w:
8784 NewIntNo = Intrinsic::x86_mmx_psra_w;
8786 case Intrinsic::x86_mmx_psrai_d:
8787 NewIntNo = Intrinsic::x86_mmx_psra_d;
8789 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
8795 // The vector shift intrinsics with scalars uses 32b shift amounts but
8796 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
8800 ShOps[1] = DAG.getConstant(0, MVT::i32);
8801 if (ShAmtVT == MVT::v4i32) {
8802 ShOps[2] = DAG.getUNDEF(MVT::i32);
8803 ShOps[3] = DAG.getUNDEF(MVT::i32);
8804 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
8806 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
8807 // FIXME this must be lowered to get rid of the invalid type.
8810 EVT VT = Op.getValueType();
8811 ShAmt = DAG.getNode(ISD::BITCAST, dl, VT, ShAmt);
8812 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8813 DAG.getConstant(NewIntNo, MVT::i32),
8814 Op.getOperand(1), ShAmt);
8819 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
8820 SelectionDAG &DAG) const {
8821 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8822 MFI->setReturnAddressIsTaken(true);
8824 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8825 DebugLoc dl = Op.getDebugLoc();
8828 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
8830 DAG.getConstant(TD->getPointerSize(),
8831 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
8832 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
8833 DAG.getNode(ISD::ADD, dl, getPointerTy(),
8835 MachinePointerInfo(), false, false, 0);
8838 // Just load the return address.
8839 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
8840 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
8841 RetAddrFI, MachinePointerInfo(), false, false, 0);
8844 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
8845 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8846 MFI->setFrameAddressIsTaken(true);
8848 EVT VT = Op.getValueType();
8849 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
8850 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8851 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
8852 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
8854 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
8855 MachinePointerInfo(),
8860 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
8861 SelectionDAG &DAG) const {
8862 return DAG.getIntPtrConstant(2*TD->getPointerSize());
8865 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
8866 MachineFunction &MF = DAG.getMachineFunction();
8867 SDValue Chain = Op.getOperand(0);
8868 SDValue Offset = Op.getOperand(1);
8869 SDValue Handler = Op.getOperand(2);
8870 DebugLoc dl = Op.getDebugLoc();
8872 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
8873 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
8875 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
8877 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
8878 DAG.getIntPtrConstant(TD->getPointerSize()));
8879 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
8880 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
8882 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
8883 MF.getRegInfo().addLiveOut(StoreAddrReg);
8885 return DAG.getNode(X86ISD::EH_RETURN, dl,
8887 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
8890 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
8891 SelectionDAG &DAG) const {
8892 SDValue Root = Op.getOperand(0);
8893 SDValue Trmp = Op.getOperand(1); // trampoline
8894 SDValue FPtr = Op.getOperand(2); // nested function
8895 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
8896 DebugLoc dl = Op.getDebugLoc();
8898 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
8900 if (Subtarget->is64Bit()) {
8901 SDValue OutChains[6];
8903 // Large code-model.
8904 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
8905 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
8907 const unsigned char N86R10 = X86_MC::getX86RegNum(X86::R10);
8908 const unsigned char N86R11 = X86_MC::getX86RegNum(X86::R11);
8910 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
8912 // Load the pointer to the nested function into R11.
8913 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
8914 SDValue Addr = Trmp;
8915 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8916 Addr, MachinePointerInfo(TrmpAddr),
8919 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8920 DAG.getConstant(2, MVT::i64));
8921 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
8922 MachinePointerInfo(TrmpAddr, 2),
8925 // Load the 'nest' parameter value into R10.
8926 // R10 is specified in X86CallingConv.td
8927 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
8928 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8929 DAG.getConstant(10, MVT::i64));
8930 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8931 Addr, MachinePointerInfo(TrmpAddr, 10),
8934 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8935 DAG.getConstant(12, MVT::i64));
8936 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
8937 MachinePointerInfo(TrmpAddr, 12),
8940 // Jump to the nested function.
8941 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
8942 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8943 DAG.getConstant(20, MVT::i64));
8944 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8945 Addr, MachinePointerInfo(TrmpAddr, 20),
8948 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
8949 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8950 DAG.getConstant(22, MVT::i64));
8951 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
8952 MachinePointerInfo(TrmpAddr, 22),
8956 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
8957 return DAG.getMergeValues(Ops, 2, dl);
8959 const Function *Func =
8960 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
8961 CallingConv::ID CC = Func->getCallingConv();
8966 llvm_unreachable("Unsupported calling convention");
8967 case CallingConv::C:
8968 case CallingConv::X86_StdCall: {
8969 // Pass 'nest' parameter in ECX.
8970 // Must be kept in sync with X86CallingConv.td
8973 // Check that ECX wasn't needed by an 'inreg' parameter.
8974 FunctionType *FTy = Func->getFunctionType();
8975 const AttrListPtr &Attrs = Func->getAttributes();
8977 if (!Attrs.isEmpty() && !Func->isVarArg()) {
8978 unsigned InRegCount = 0;
8981 for (FunctionType::param_iterator I = FTy->param_begin(),
8982 E = FTy->param_end(); I != E; ++I, ++Idx)
8983 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
8984 // FIXME: should only count parameters that are lowered to integers.
8985 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
8987 if (InRegCount > 2) {
8988 report_fatal_error("Nest register in use - reduce number of inreg"
8994 case CallingConv::X86_FastCall:
8995 case CallingConv::X86_ThisCall:
8996 case CallingConv::Fast:
8997 // Pass 'nest' parameter in EAX.
8998 // Must be kept in sync with X86CallingConv.td
9003 SDValue OutChains[4];
9006 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
9007 DAG.getConstant(10, MVT::i32));
9008 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
9010 // This is storing the opcode for MOV32ri.
9011 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
9012 const unsigned char N86Reg = X86_MC::getX86RegNum(NestReg);
9013 OutChains[0] = DAG.getStore(Root, dl,
9014 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
9015 Trmp, MachinePointerInfo(TrmpAddr),
9018 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
9019 DAG.getConstant(1, MVT::i32));
9020 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
9021 MachinePointerInfo(TrmpAddr, 1),
9024 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
9025 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
9026 DAG.getConstant(5, MVT::i32));
9027 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
9028 MachinePointerInfo(TrmpAddr, 5),
9031 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
9032 DAG.getConstant(6, MVT::i32));
9033 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
9034 MachinePointerInfo(TrmpAddr, 6),
9038 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
9039 return DAG.getMergeValues(Ops, 2, dl);
9043 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
9044 SelectionDAG &DAG) const {
9046 The rounding mode is in bits 11:10 of FPSR, and has the following
9053 FLT_ROUNDS, on the other hand, expects the following:
9060 To perform the conversion, we do:
9061 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
9064 MachineFunction &MF = DAG.getMachineFunction();
9065 const TargetMachine &TM = MF.getTarget();
9066 const TargetFrameLowering &TFI = *TM.getFrameLowering();
9067 unsigned StackAlignment = TFI.getStackAlignment();
9068 EVT VT = Op.getValueType();
9069 DebugLoc DL = Op.getDebugLoc();
9071 // Save FP Control Word to stack slot
9072 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
9073 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
9076 MachineMemOperand *MMO =
9077 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
9078 MachineMemOperand::MOStore, 2, 2);
9080 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
9081 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
9082 DAG.getVTList(MVT::Other),
9083 Ops, 2, MVT::i16, MMO);
9085 // Load FP Control Word from stack slot
9086 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
9087 MachinePointerInfo(), false, false, 0);
9089 // Transform as necessary
9091 DAG.getNode(ISD::SRL, DL, MVT::i16,
9092 DAG.getNode(ISD::AND, DL, MVT::i16,
9093 CWD, DAG.getConstant(0x800, MVT::i16)),
9094 DAG.getConstant(11, MVT::i8));
9096 DAG.getNode(ISD::SRL, DL, MVT::i16,
9097 DAG.getNode(ISD::AND, DL, MVT::i16,
9098 CWD, DAG.getConstant(0x400, MVT::i16)),
9099 DAG.getConstant(9, MVT::i8));
9102 DAG.getNode(ISD::AND, DL, MVT::i16,
9103 DAG.getNode(ISD::ADD, DL, MVT::i16,
9104 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
9105 DAG.getConstant(1, MVT::i16)),
9106 DAG.getConstant(3, MVT::i16));
9109 return DAG.getNode((VT.getSizeInBits() < 16 ?
9110 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
9113 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
9114 EVT VT = Op.getValueType();
9116 unsigned NumBits = VT.getSizeInBits();
9117 DebugLoc dl = Op.getDebugLoc();
9119 Op = Op.getOperand(0);
9120 if (VT == MVT::i8) {
9121 // Zero extend to i32 since there is not an i8 bsr.
9123 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
9126 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
9127 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
9128 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
9130 // If src is zero (i.e. bsr sets ZF), returns NumBits.
9133 DAG.getConstant(NumBits+NumBits-1, OpVT),
9134 DAG.getConstant(X86::COND_E, MVT::i8),
9137 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
9139 // Finally xor with NumBits-1.
9140 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
9143 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
9147 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
9148 EVT VT = Op.getValueType();
9150 unsigned NumBits = VT.getSizeInBits();
9151 DebugLoc dl = Op.getDebugLoc();
9153 Op = Op.getOperand(0);
9154 if (VT == MVT::i8) {
9156 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
9159 // Issue a bsf (scan bits forward) which also sets EFLAGS.
9160 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
9161 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
9163 // If src is zero (i.e. bsf sets ZF), returns NumBits.
9166 DAG.getConstant(NumBits, OpVT),
9167 DAG.getConstant(X86::COND_E, MVT::i8),
9170 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
9173 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
9177 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
9178 EVT VT = Op.getValueType();
9179 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
9180 DebugLoc dl = Op.getDebugLoc();
9182 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
9183 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
9184 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
9185 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
9186 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
9188 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
9189 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
9190 // return AloBlo + AloBhi + AhiBlo;
9192 SDValue A = Op.getOperand(0);
9193 SDValue B = Op.getOperand(1);
9195 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9196 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9197 A, DAG.getConstant(32, MVT::i32));
9198 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9199 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9200 B, DAG.getConstant(32, MVT::i32));
9201 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9202 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
9204 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9205 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
9207 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9208 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
9210 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9211 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9212 AloBhi, DAG.getConstant(32, MVT::i32));
9213 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9214 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9215 AhiBlo, DAG.getConstant(32, MVT::i32));
9216 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
9217 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
9221 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
9223 EVT VT = Op.getValueType();
9224 DebugLoc dl = Op.getDebugLoc();
9225 SDValue R = Op.getOperand(0);
9226 SDValue Amt = Op.getOperand(1);
9227 LLVMContext *Context = DAG.getContext();
9229 if (!(Subtarget->hasSSE2() || Subtarget->hasAVX()))
9232 // Decompose 256-bit shifts into smaller 128-bit shifts.
9233 if (VT.getSizeInBits() == 256) {
9234 int NumElems = VT.getVectorNumElements();
9235 MVT EltVT = VT.getVectorElementType().getSimpleVT();
9236 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
9238 // Extract the two vectors
9239 SDValue V1 = Extract128BitVector(R, DAG.getConstant(0, MVT::i32), DAG, dl);
9240 SDValue V2 = Extract128BitVector(R, DAG.getConstant(NumElems/2, MVT::i32),
9243 // Recreate the shift amount vectors
9244 SmallVector<SDValue, 4> Amt1Csts;
9245 SmallVector<SDValue, 4> Amt2Csts;
9246 for (int i = 0; i < NumElems/2; ++i)
9247 Amt1Csts.push_back(Amt->getOperand(i));
9248 for (int i = NumElems/2; i < NumElems; ++i)
9249 Amt2Csts.push_back(Amt->getOperand(i));
9251 SDValue Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT,
9252 &Amt1Csts[0], NumElems/2);
9253 SDValue Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT,
9254 &Amt2Csts[0], NumElems/2);
9256 // Issue new vector shifts for the smaller types
9257 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
9258 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
9260 // Concatenate the result back
9261 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
9264 // Optimize shl/srl/sra with constant shift amount.
9265 if (isSplatVector(Amt.getNode())) {
9266 SDValue SclrAmt = Amt->getOperand(0);
9267 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(SclrAmt)) {
9268 uint64_t ShiftAmt = C->getZExtValue();
9270 if (VT == MVT::v2i64 && Op.getOpcode() == ISD::SHL)
9271 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9272 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9273 R, DAG.getConstant(ShiftAmt, MVT::i32));
9275 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SHL)
9276 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9277 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9278 R, DAG.getConstant(ShiftAmt, MVT::i32));
9280 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SHL)
9281 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9282 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9283 R, DAG.getConstant(ShiftAmt, MVT::i32));
9285 if (VT == MVT::v2i64 && Op.getOpcode() == ISD::SRL)
9286 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9287 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9288 R, DAG.getConstant(ShiftAmt, MVT::i32));
9290 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SRL)
9291 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9292 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9293 R, DAG.getConstant(ShiftAmt, MVT::i32));
9295 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SRL)
9296 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9297 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9298 R, DAG.getConstant(ShiftAmt, MVT::i32));
9300 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SRA)
9301 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9302 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9303 R, DAG.getConstant(ShiftAmt, MVT::i32));
9305 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SRA)
9306 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9307 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9308 R, DAG.getConstant(ShiftAmt, MVT::i32));
9312 // Lower SHL with variable shift amount.
9313 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
9314 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9315 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9316 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
9318 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
9320 std::vector<Constant*> CV(4, CI);
9321 Constant *C = ConstantVector::get(CV);
9322 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9323 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9324 MachinePointerInfo::getConstantPool(),
9327 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
9328 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
9329 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
9330 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
9332 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
9334 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9335 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9336 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
9338 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
9339 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
9341 std::vector<Constant*> CVM1(16, CM1);
9342 std::vector<Constant*> CVM2(16, CM2);
9343 Constant *C = ConstantVector::get(CVM1);
9344 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9345 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9346 MachinePointerInfo::getConstantPool(),
9349 // r = pblendv(r, psllw(r & (char16)15, 4), a);
9350 M = DAG.getNode(ISD::AND, dl, VT, R, M);
9351 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9352 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
9353 DAG.getConstant(4, MVT::i32));
9354 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
9356 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
9358 C = ConstantVector::get(CVM2);
9359 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9360 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9361 MachinePointerInfo::getConstantPool(),
9364 // r = pblendv(r, psllw(r & (char16)63, 2), a);
9365 M = DAG.getNode(ISD::AND, dl, VT, R, M);
9366 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9367 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
9368 DAG.getConstant(2, MVT::i32));
9369 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
9371 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
9373 // return pblendv(r, r+r, a);
9374 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT,
9375 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
9381 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
9382 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
9383 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
9384 // looks for this combo and may remove the "setcc" instruction if the "setcc"
9385 // has only one use.
9386 SDNode *N = Op.getNode();
9387 SDValue LHS = N->getOperand(0);
9388 SDValue RHS = N->getOperand(1);
9389 unsigned BaseOp = 0;
9391 DebugLoc DL = Op.getDebugLoc();
9392 switch (Op.getOpcode()) {
9393 default: llvm_unreachable("Unknown ovf instruction!");
9395 // A subtract of one will be selected as a INC. Note that INC doesn't
9396 // set CF, so we can't do this for UADDO.
9397 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
9399 BaseOp = X86ISD::INC;
9403 BaseOp = X86ISD::ADD;
9407 BaseOp = X86ISD::ADD;
9411 // A subtract of one will be selected as a DEC. Note that DEC doesn't
9412 // set CF, so we can't do this for USUBO.
9413 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
9415 BaseOp = X86ISD::DEC;
9419 BaseOp = X86ISD::SUB;
9423 BaseOp = X86ISD::SUB;
9427 BaseOp = X86ISD::SMUL;
9430 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
9431 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
9433 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
9436 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9437 DAG.getConstant(X86::COND_O, MVT::i32),
9438 SDValue(Sum.getNode(), 2));
9440 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
9444 // Also sets EFLAGS.
9445 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
9446 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
9449 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
9450 DAG.getConstant(Cond, MVT::i32),
9451 SDValue(Sum.getNode(), 1));
9453 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
9456 SDValue X86TargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const{
9457 DebugLoc dl = Op.getDebugLoc();
9458 SDNode* Node = Op.getNode();
9459 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
9460 EVT VT = Node->getValueType(0);
9462 if (Subtarget->hasSSE2() && VT.isVector()) {
9463 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
9464 ExtraVT.getScalarType().getSizeInBits();
9465 SDValue ShAmt = DAG.getConstant(BitsDiff, MVT::i32);
9467 unsigned SHLIntrinsicsID = 0;
9468 unsigned SRAIntrinsicsID = 0;
9469 switch (VT.getSimpleVT().SimpleTy) {
9473 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_q;
9474 SRAIntrinsicsID = 0;
9478 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_d;
9479 SRAIntrinsicsID = Intrinsic::x86_sse2_psrai_d;
9483 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_w;
9484 SRAIntrinsicsID = Intrinsic::x86_sse2_psrai_w;
9489 SDValue Tmp1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9490 DAG.getConstant(SHLIntrinsicsID, MVT::i32),
9491 Node->getOperand(0), ShAmt);
9493 // In case of 1 bit sext, no need to shr
9494 if (ExtraVT.getScalarType().getSizeInBits() == 1) return Tmp1;
9496 if (SRAIntrinsicsID) {
9497 Tmp1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9498 DAG.getConstant(SRAIntrinsicsID, MVT::i32),
9508 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
9509 DebugLoc dl = Op.getDebugLoc();
9511 // Go ahead and emit the fence on x86-64 even if we asked for no-sse2.
9512 // There isn't any reason to disable it if the target processor supports it.
9513 if (!Subtarget->hasSSE2() && !Subtarget->is64Bit()) {
9514 SDValue Chain = Op.getOperand(0);
9515 SDValue Zero = DAG.getConstant(0, MVT::i32);
9517 DAG.getRegister(X86::ESP, MVT::i32), // Base
9518 DAG.getTargetConstant(1, MVT::i8), // Scale
9519 DAG.getRegister(0, MVT::i32), // Index
9520 DAG.getTargetConstant(0, MVT::i32), // Disp
9521 DAG.getRegister(0, MVT::i32), // Segment.
9526 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
9527 array_lengthof(Ops));
9528 return SDValue(Res, 0);
9531 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
9533 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
9535 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
9536 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
9537 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
9538 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
9540 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
9541 if (!Op1 && !Op2 && !Op3 && Op4)
9542 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
9544 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
9545 if (Op1 && !Op2 && !Op3 && !Op4)
9546 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
9548 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
9550 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
9553 SDValue X86TargetLowering::LowerATOMIC_FENCE(SDValue Op,
9554 SelectionDAG &DAG) const {
9555 DebugLoc dl = Op.getDebugLoc();
9556 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
9557 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
9558 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
9559 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
9561 // The only fence that needs an instruction is a sequentially-consistent
9562 // cross-thread fence.
9563 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
9564 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
9565 // no-sse2). There isn't any reason to disable it if the target processor
9567 if (Subtarget->hasSSE2() || Subtarget->is64Bit())
9568 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
9570 SDValue Chain = Op.getOperand(0);
9571 SDValue Zero = DAG.getConstant(0, MVT::i32);
9573 DAG.getRegister(X86::ESP, MVT::i32), // Base
9574 DAG.getTargetConstant(1, MVT::i8), // Scale
9575 DAG.getRegister(0, MVT::i32), // Index
9576 DAG.getTargetConstant(0, MVT::i32), // Disp
9577 DAG.getRegister(0, MVT::i32), // Segment.
9582 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
9583 array_lengthof(Ops));
9584 return SDValue(Res, 0);
9587 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
9588 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
9592 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
9593 EVT T = Op.getValueType();
9594 DebugLoc DL = Op.getDebugLoc();
9597 switch(T.getSimpleVT().SimpleTy) {
9599 assert(false && "Invalid value type!");
9600 case MVT::i8: Reg = X86::AL; size = 1; break;
9601 case MVT::i16: Reg = X86::AX; size = 2; break;
9602 case MVT::i32: Reg = X86::EAX; size = 4; break;
9604 assert(Subtarget->is64Bit() && "Node not type legal!");
9605 Reg = X86::RAX; size = 8;
9608 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
9609 Op.getOperand(2), SDValue());
9610 SDValue Ops[] = { cpIn.getValue(0),
9613 DAG.getTargetConstant(size, MVT::i8),
9615 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9616 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
9617 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
9620 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
9624 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
9625 SelectionDAG &DAG) const {
9626 assert(Subtarget->is64Bit() && "Result not type legalized?");
9627 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9628 SDValue TheChain = Op.getOperand(0);
9629 DebugLoc dl = Op.getDebugLoc();
9630 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
9631 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
9632 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
9634 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
9635 DAG.getConstant(32, MVT::i8));
9637 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
9640 return DAG.getMergeValues(Ops, 2, dl);
9643 SDValue X86TargetLowering::LowerBITCAST(SDValue Op,
9644 SelectionDAG &DAG) const {
9645 EVT SrcVT = Op.getOperand(0).getValueType();
9646 EVT DstVT = Op.getValueType();
9647 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
9648 Subtarget->hasMMX() && "Unexpected custom BITCAST");
9649 assert((DstVT == MVT::i64 ||
9650 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
9651 "Unexpected custom BITCAST");
9652 // i64 <=> MMX conversions are Legal.
9653 if (SrcVT==MVT::i64 && DstVT.isVector())
9655 if (DstVT==MVT::i64 && SrcVT.isVector())
9657 // MMX <=> MMX conversions are Legal.
9658 if (SrcVT.isVector() && DstVT.isVector())
9660 // All other conversions need to be expanded.
9664 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
9665 SDNode *Node = Op.getNode();
9666 DebugLoc dl = Node->getDebugLoc();
9667 EVT T = Node->getValueType(0);
9668 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
9669 DAG.getConstant(0, T), Node->getOperand(2));
9670 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
9671 cast<AtomicSDNode>(Node)->getMemoryVT(),
9672 Node->getOperand(0),
9673 Node->getOperand(1), negOp,
9674 cast<AtomicSDNode>(Node)->getSrcValue(),
9675 cast<AtomicSDNode>(Node)->getAlignment(),
9676 cast<AtomicSDNode>(Node)->getOrdering(),
9677 cast<AtomicSDNode>(Node)->getSynchScope());
9680 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
9681 EVT VT = Op.getNode()->getValueType(0);
9683 // Let legalize expand this if it isn't a legal type yet.
9684 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
9687 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
9690 bool ExtraOp = false;
9691 switch (Op.getOpcode()) {
9692 default: assert(0 && "Invalid code");
9693 case ISD::ADDC: Opc = X86ISD::ADD; break;
9694 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
9695 case ISD::SUBC: Opc = X86ISD::SUB; break;
9696 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
9700 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
9702 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
9703 Op.getOperand(1), Op.getOperand(2));
9706 /// LowerOperation - Provide custom lowering hooks for some operations.
9708 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
9709 switch (Op.getOpcode()) {
9710 default: llvm_unreachable("Should not custom lower this!");
9711 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op,DAG);
9712 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
9713 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op,DAG);
9714 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
9715 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
9716 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
9717 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
9718 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
9719 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
9720 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
9721 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
9722 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, DAG);
9723 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
9724 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
9725 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
9726 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
9727 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
9728 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
9729 case ISD::SHL_PARTS:
9730 case ISD::SRA_PARTS:
9731 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
9732 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
9733 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
9734 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
9735 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
9736 case ISD::FABS: return LowerFABS(Op, DAG);
9737 case ISD::FNEG: return LowerFNEG(Op, DAG);
9738 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
9739 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
9740 case ISD::SETCC: return LowerSETCC(Op, DAG);
9741 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
9742 case ISD::SELECT: return LowerSELECT(Op, DAG);
9743 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
9744 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
9745 case ISD::VASTART: return LowerVASTART(Op, DAG);
9746 case ISD::VAARG: return LowerVAARG(Op, DAG);
9747 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
9748 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
9749 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
9750 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
9751 case ISD::FRAME_TO_ARGS_OFFSET:
9752 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
9753 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
9754 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
9755 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
9756 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
9757 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
9758 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
9759 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
9762 case ISD::SHL: return LowerShift(Op, DAG);
9768 case ISD::UMULO: return LowerXALUO(Op, DAG);
9769 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
9770 case ISD::BITCAST: return LowerBITCAST(Op, DAG);
9774 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
9778 void X86TargetLowering::
9779 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
9780 SelectionDAG &DAG, unsigned NewOp) const {
9781 EVT T = Node->getValueType(0);
9782 DebugLoc dl = Node->getDebugLoc();
9783 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
9785 SDValue Chain = Node->getOperand(0);
9786 SDValue In1 = Node->getOperand(1);
9787 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
9788 Node->getOperand(2), DAG.getIntPtrConstant(0));
9789 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
9790 Node->getOperand(2), DAG.getIntPtrConstant(1));
9791 SDValue Ops[] = { Chain, In1, In2L, In2H };
9792 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
9794 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
9795 cast<MemSDNode>(Node)->getMemOperand());
9796 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
9797 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
9798 Results.push_back(Result.getValue(2));
9801 /// ReplaceNodeResults - Replace a node with an illegal result type
9802 /// with a new node built out of custom code.
9803 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
9804 SmallVectorImpl<SDValue>&Results,
9805 SelectionDAG &DAG) const {
9806 DebugLoc dl = N->getDebugLoc();
9807 switch (N->getOpcode()) {
9809 assert(false && "Do not know how to custom type legalize this operation!");
9811 case ISD::SIGN_EXTEND_INREG:
9816 // We don't want to expand or promote these.
9818 case ISD::FP_TO_SINT: {
9819 std::pair<SDValue,SDValue> Vals =
9820 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
9821 SDValue FIST = Vals.first, StackSlot = Vals.second;
9822 if (FIST.getNode() != 0) {
9823 EVT VT = N->getValueType(0);
9824 // Return a load from the stack slot.
9825 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
9826 MachinePointerInfo(), false, false, 0));
9830 case ISD::READCYCLECOUNTER: {
9831 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9832 SDValue TheChain = N->getOperand(0);
9833 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
9834 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
9836 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
9838 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
9839 SDValue Ops[] = { eax, edx };
9840 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
9841 Results.push_back(edx.getValue(1));
9844 case ISD::ATOMIC_CMP_SWAP: {
9845 EVT T = N->getValueType(0);
9846 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
9847 SDValue cpInL, cpInH;
9848 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
9849 DAG.getConstant(0, MVT::i32));
9850 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
9851 DAG.getConstant(1, MVT::i32));
9852 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
9853 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
9855 SDValue swapInL, swapInH;
9856 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
9857 DAG.getConstant(0, MVT::i32));
9858 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
9859 DAG.getConstant(1, MVT::i32));
9860 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
9862 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
9863 swapInL.getValue(1));
9864 SDValue Ops[] = { swapInH.getValue(0),
9866 swapInH.getValue(1) };
9867 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9868 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
9869 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG8_DAG, dl, Tys,
9871 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
9872 MVT::i32, Result.getValue(1));
9873 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
9874 MVT::i32, cpOutL.getValue(2));
9875 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
9876 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
9877 Results.push_back(cpOutH.getValue(1));
9880 case ISD::ATOMIC_LOAD_ADD:
9881 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
9883 case ISD::ATOMIC_LOAD_AND:
9884 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
9886 case ISD::ATOMIC_LOAD_NAND:
9887 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
9889 case ISD::ATOMIC_LOAD_OR:
9890 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
9892 case ISD::ATOMIC_LOAD_SUB:
9893 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
9895 case ISD::ATOMIC_LOAD_XOR:
9896 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
9898 case ISD::ATOMIC_SWAP:
9899 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
9904 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
9906 default: return NULL;
9907 case X86ISD::BSF: return "X86ISD::BSF";
9908 case X86ISD::BSR: return "X86ISD::BSR";
9909 case X86ISD::SHLD: return "X86ISD::SHLD";
9910 case X86ISD::SHRD: return "X86ISD::SHRD";
9911 case X86ISD::FAND: return "X86ISD::FAND";
9912 case X86ISD::FOR: return "X86ISD::FOR";
9913 case X86ISD::FXOR: return "X86ISD::FXOR";
9914 case X86ISD::FSRL: return "X86ISD::FSRL";
9915 case X86ISD::FILD: return "X86ISD::FILD";
9916 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
9917 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
9918 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
9919 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
9920 case X86ISD::FLD: return "X86ISD::FLD";
9921 case X86ISD::FST: return "X86ISD::FST";
9922 case X86ISD::CALL: return "X86ISD::CALL";
9923 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
9924 case X86ISD::BT: return "X86ISD::BT";
9925 case X86ISD::CMP: return "X86ISD::CMP";
9926 case X86ISD::COMI: return "X86ISD::COMI";
9927 case X86ISD::UCOMI: return "X86ISD::UCOMI";
9928 case X86ISD::SETCC: return "X86ISD::SETCC";
9929 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
9930 case X86ISD::FSETCCsd: return "X86ISD::FSETCCsd";
9931 case X86ISD::FSETCCss: return "X86ISD::FSETCCss";
9932 case X86ISD::CMOV: return "X86ISD::CMOV";
9933 case X86ISD::BRCOND: return "X86ISD::BRCOND";
9934 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
9935 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
9936 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
9937 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
9938 case X86ISD::Wrapper: return "X86ISD::Wrapper";
9939 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
9940 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
9941 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
9942 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
9943 case X86ISD::PINSRB: return "X86ISD::PINSRB";
9944 case X86ISD::PINSRW: return "X86ISD::PINSRW";
9945 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
9946 case X86ISD::ANDNP: return "X86ISD::ANDNP";
9947 case X86ISD::PSIGNB: return "X86ISD::PSIGNB";
9948 case X86ISD::PSIGNW: return "X86ISD::PSIGNW";
9949 case X86ISD::PSIGND: return "X86ISD::PSIGND";
9950 case X86ISD::PBLENDVB: return "X86ISD::PBLENDVB";
9951 case X86ISD::FMAX: return "X86ISD::FMAX";
9952 case X86ISD::FMIN: return "X86ISD::FMIN";
9953 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
9954 case X86ISD::FRCP: return "X86ISD::FRCP";
9955 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
9956 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
9957 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
9958 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
9959 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
9960 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
9961 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
9962 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
9963 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
9964 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
9965 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
9966 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
9967 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
9968 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
9969 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
9970 case X86ISD::VSHL: return "X86ISD::VSHL";
9971 case X86ISD::VSRL: return "X86ISD::VSRL";
9972 case X86ISD::CMPPD: return "X86ISD::CMPPD";
9973 case X86ISD::CMPPS: return "X86ISD::CMPPS";
9974 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
9975 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
9976 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
9977 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
9978 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
9979 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
9980 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
9981 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
9982 case X86ISD::ADD: return "X86ISD::ADD";
9983 case X86ISD::SUB: return "X86ISD::SUB";
9984 case X86ISD::ADC: return "X86ISD::ADC";
9985 case X86ISD::SBB: return "X86ISD::SBB";
9986 case X86ISD::SMUL: return "X86ISD::SMUL";
9987 case X86ISD::UMUL: return "X86ISD::UMUL";
9988 case X86ISD::INC: return "X86ISD::INC";
9989 case X86ISD::DEC: return "X86ISD::DEC";
9990 case X86ISD::OR: return "X86ISD::OR";
9991 case X86ISD::XOR: return "X86ISD::XOR";
9992 case X86ISD::AND: return "X86ISD::AND";
9993 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
9994 case X86ISD::PTEST: return "X86ISD::PTEST";
9995 case X86ISD::TESTP: return "X86ISD::TESTP";
9996 case X86ISD::PALIGN: return "X86ISD::PALIGN";
9997 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
9998 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
9999 case X86ISD::PSHUFHW_LD: return "X86ISD::PSHUFHW_LD";
10000 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
10001 case X86ISD::PSHUFLW_LD: return "X86ISD::PSHUFLW_LD";
10002 case X86ISD::SHUFPS: return "X86ISD::SHUFPS";
10003 case X86ISD::SHUFPD: return "X86ISD::SHUFPD";
10004 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
10005 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
10006 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
10007 case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
10008 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
10009 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
10010 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
10011 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
10012 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
10013 case X86ISD::MOVSHDUP_LD: return "X86ISD::MOVSHDUP_LD";
10014 case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
10015 case X86ISD::MOVSD: return "X86ISD::MOVSD";
10016 case X86ISD::MOVSS: return "X86ISD::MOVSS";
10017 case X86ISD::UNPCKLPS: return "X86ISD::UNPCKLPS";
10018 case X86ISD::UNPCKLPD: return "X86ISD::UNPCKLPD";
10019 case X86ISD::VUNPCKLPDY: return "X86ISD::VUNPCKLPDY";
10020 case X86ISD::UNPCKHPS: return "X86ISD::UNPCKHPS";
10021 case X86ISD::UNPCKHPD: return "X86ISD::UNPCKHPD";
10022 case X86ISD::PUNPCKLBW: return "X86ISD::PUNPCKLBW";
10023 case X86ISD::PUNPCKLWD: return "X86ISD::PUNPCKLWD";
10024 case X86ISD::PUNPCKLDQ: return "X86ISD::PUNPCKLDQ";
10025 case X86ISD::PUNPCKLQDQ: return "X86ISD::PUNPCKLQDQ";
10026 case X86ISD::PUNPCKHBW: return "X86ISD::PUNPCKHBW";
10027 case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
10028 case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
10029 case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
10030 case X86ISD::VPERMILPS: return "X86ISD::VPERMILPS";
10031 case X86ISD::VPERMILPSY: return "X86ISD::VPERMILPSY";
10032 case X86ISD::VPERMILPD: return "X86ISD::VPERMILPD";
10033 case X86ISD::VPERMILPDY: return "X86ISD::VPERMILPDY";
10034 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
10035 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
10036 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
10037 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
10041 // isLegalAddressingMode - Return true if the addressing mode represented
10042 // by AM is legal for this target, for a load/store of the specified type.
10043 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
10045 // X86 supports extremely general addressing modes.
10046 CodeModel::Model M = getTargetMachine().getCodeModel();
10047 Reloc::Model R = getTargetMachine().getRelocationModel();
10049 // X86 allows a sign-extended 32-bit immediate field as a displacement.
10050 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
10055 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
10057 // If a reference to this global requires an extra load, we can't fold it.
10058 if (isGlobalStubReference(GVFlags))
10061 // If BaseGV requires a register for the PIC base, we cannot also have a
10062 // BaseReg specified.
10063 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
10066 // If lower 4G is not available, then we must use rip-relative addressing.
10067 if ((M != CodeModel::Small || R != Reloc::Static) &&
10068 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
10072 switch (AM.Scale) {
10078 // These scales always work.
10083 // These scales are formed with basereg+scalereg. Only accept if there is
10088 default: // Other stuff never works.
10096 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
10097 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10099 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
10100 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
10101 if (NumBits1 <= NumBits2)
10106 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
10107 if (!VT1.isInteger() || !VT2.isInteger())
10109 unsigned NumBits1 = VT1.getSizeInBits();
10110 unsigned NumBits2 = VT2.getSizeInBits();
10111 if (NumBits1 <= NumBits2)
10116 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
10117 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
10118 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
10121 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
10122 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
10123 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
10126 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
10127 // i16 instructions are longer (0x66 prefix) and potentially slower.
10128 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
10131 /// isShuffleMaskLegal - Targets can use this to indicate that they only
10132 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
10133 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
10134 /// are assumed to be legal.
10136 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
10138 // Very little shuffling can be done for 64-bit vectors right now.
10139 if (VT.getSizeInBits() == 64)
10140 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
10142 // FIXME: pshufb, blends, shifts.
10143 return (VT.getVectorNumElements() == 2 ||
10144 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
10145 isMOVLMask(M, VT) ||
10146 isSHUFPMask(M, VT) ||
10147 isPSHUFDMask(M, VT) ||
10148 isPSHUFHWMask(M, VT) ||
10149 isPSHUFLWMask(M, VT) ||
10150 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
10151 isUNPCKLMask(M, VT) ||
10152 isUNPCKHMask(M, VT) ||
10153 isUNPCKL_v_undef_Mask(M, VT) ||
10154 isUNPCKH_v_undef_Mask(M, VT));
10158 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
10160 unsigned NumElts = VT.getVectorNumElements();
10161 // FIXME: This collection of masks seems suspect.
10164 if (NumElts == 4 && VT.getSizeInBits() == 128) {
10165 return (isMOVLMask(Mask, VT) ||
10166 isCommutedMOVLMask(Mask, VT, true) ||
10167 isSHUFPMask(Mask, VT) ||
10168 isCommutedSHUFPMask(Mask, VT));
10173 //===----------------------------------------------------------------------===//
10174 // X86 Scheduler Hooks
10175 //===----------------------------------------------------------------------===//
10177 // private utility function
10178 MachineBasicBlock *
10179 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
10180 MachineBasicBlock *MBB,
10187 TargetRegisterClass *RC,
10188 bool invSrc) const {
10189 // For the atomic bitwise operator, we generate
10192 // ld t1 = [bitinstr.addr]
10193 // op t2 = t1, [bitinstr.val]
10195 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
10197 // fallthrough -->nextMBB
10198 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10199 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10200 MachineFunction::iterator MBBIter = MBB;
10203 /// First build the CFG
10204 MachineFunction *F = MBB->getParent();
10205 MachineBasicBlock *thisMBB = MBB;
10206 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
10207 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
10208 F->insert(MBBIter, newMBB);
10209 F->insert(MBBIter, nextMBB);
10211 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
10212 nextMBB->splice(nextMBB->begin(), thisMBB,
10213 llvm::next(MachineBasicBlock::iterator(bInstr)),
10215 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10217 // Update thisMBB to fall through to newMBB
10218 thisMBB->addSuccessor(newMBB);
10220 // newMBB jumps to itself and fall through to nextMBB
10221 newMBB->addSuccessor(nextMBB);
10222 newMBB->addSuccessor(newMBB);
10224 // Insert instructions into newMBB based on incoming instruction
10225 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
10226 "unexpected number of operands");
10227 DebugLoc dl = bInstr->getDebugLoc();
10228 MachineOperand& destOper = bInstr->getOperand(0);
10229 MachineOperand* argOpers[2 + X86::AddrNumOperands];
10230 int numArgs = bInstr->getNumOperands() - 1;
10231 for (int i=0; i < numArgs; ++i)
10232 argOpers[i] = &bInstr->getOperand(i+1);
10234 // x86 address has 4 operands: base, index, scale, and displacement
10235 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
10236 int valArgIndx = lastAddrIndx + 1;
10238 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
10239 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
10240 for (int i=0; i <= lastAddrIndx; ++i)
10241 (*MIB).addOperand(*argOpers[i]);
10243 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
10245 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
10250 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
10251 assert((argOpers[valArgIndx]->isReg() ||
10252 argOpers[valArgIndx]->isImm()) &&
10253 "invalid operand");
10254 if (argOpers[valArgIndx]->isReg())
10255 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
10257 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
10259 (*MIB).addOperand(*argOpers[valArgIndx]);
10261 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
10264 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
10265 for (int i=0; i <= lastAddrIndx; ++i)
10266 (*MIB).addOperand(*argOpers[i]);
10268 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
10269 (*MIB).setMemRefs(bInstr->memoperands_begin(),
10270 bInstr->memoperands_end());
10272 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
10273 MIB.addReg(EAXreg);
10276 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
10278 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
10282 // private utility function: 64 bit atomics on 32 bit host.
10283 MachineBasicBlock *
10284 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
10285 MachineBasicBlock *MBB,
10290 bool invSrc) const {
10291 // For the atomic bitwise operator, we generate
10292 // thisMBB (instructions are in pairs, except cmpxchg8b)
10293 // ld t1,t2 = [bitinstr.addr]
10295 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
10296 // op t5, t6 <- out1, out2, [bitinstr.val]
10297 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
10298 // mov ECX, EBX <- t5, t6
10299 // mov EAX, EDX <- t1, t2
10300 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
10301 // mov t3, t4 <- EAX, EDX
10303 // result in out1, out2
10304 // fallthrough -->nextMBB
10306 const TargetRegisterClass *RC = X86::GR32RegisterClass;
10307 const unsigned LoadOpc = X86::MOV32rm;
10308 const unsigned NotOpc = X86::NOT32r;
10309 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10310 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10311 MachineFunction::iterator MBBIter = MBB;
10314 /// First build the CFG
10315 MachineFunction *F = MBB->getParent();
10316 MachineBasicBlock *thisMBB = MBB;
10317 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
10318 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
10319 F->insert(MBBIter, newMBB);
10320 F->insert(MBBIter, nextMBB);
10322 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
10323 nextMBB->splice(nextMBB->begin(), thisMBB,
10324 llvm::next(MachineBasicBlock::iterator(bInstr)),
10326 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10328 // Update thisMBB to fall through to newMBB
10329 thisMBB->addSuccessor(newMBB);
10331 // newMBB jumps to itself and fall through to nextMBB
10332 newMBB->addSuccessor(nextMBB);
10333 newMBB->addSuccessor(newMBB);
10335 DebugLoc dl = bInstr->getDebugLoc();
10336 // Insert instructions into newMBB based on incoming instruction
10337 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
10338 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
10339 "unexpected number of operands");
10340 MachineOperand& dest1Oper = bInstr->getOperand(0);
10341 MachineOperand& dest2Oper = bInstr->getOperand(1);
10342 MachineOperand* argOpers[2 + X86::AddrNumOperands];
10343 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
10344 argOpers[i] = &bInstr->getOperand(i+2);
10346 // We use some of the operands multiple times, so conservatively just
10347 // clear any kill flags that might be present.
10348 if (argOpers[i]->isReg() && argOpers[i]->isUse())
10349 argOpers[i]->setIsKill(false);
10352 // x86 address has 5 operands: base, index, scale, displacement, and segment.
10353 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
10355 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
10356 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
10357 for (int i=0; i <= lastAddrIndx; ++i)
10358 (*MIB).addOperand(*argOpers[i]);
10359 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
10360 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
10361 // add 4 to displacement.
10362 for (int i=0; i <= lastAddrIndx-2; ++i)
10363 (*MIB).addOperand(*argOpers[i]);
10364 MachineOperand newOp3 = *(argOpers[3]);
10365 if (newOp3.isImm())
10366 newOp3.setImm(newOp3.getImm()+4);
10368 newOp3.setOffset(newOp3.getOffset()+4);
10369 (*MIB).addOperand(newOp3);
10370 (*MIB).addOperand(*argOpers[lastAddrIndx]);
10372 // t3/4 are defined later, at the bottom of the loop
10373 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
10374 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
10375 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
10376 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
10377 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
10378 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
10380 // The subsequent operations should be using the destination registers of
10381 //the PHI instructions.
10383 t1 = F->getRegInfo().createVirtualRegister(RC);
10384 t2 = F->getRegInfo().createVirtualRegister(RC);
10385 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
10386 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
10388 t1 = dest1Oper.getReg();
10389 t2 = dest2Oper.getReg();
10392 int valArgIndx = lastAddrIndx + 1;
10393 assert((argOpers[valArgIndx]->isReg() ||
10394 argOpers[valArgIndx]->isImm()) &&
10395 "invalid operand");
10396 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
10397 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
10398 if (argOpers[valArgIndx]->isReg())
10399 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
10401 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
10402 if (regOpcL != X86::MOV32rr)
10404 (*MIB).addOperand(*argOpers[valArgIndx]);
10405 assert(argOpers[valArgIndx + 1]->isReg() ==
10406 argOpers[valArgIndx]->isReg());
10407 assert(argOpers[valArgIndx + 1]->isImm() ==
10408 argOpers[valArgIndx]->isImm());
10409 if (argOpers[valArgIndx + 1]->isReg())
10410 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
10412 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
10413 if (regOpcH != X86::MOV32rr)
10415 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
10417 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
10419 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
10422 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
10424 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
10427 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
10428 for (int i=0; i <= lastAddrIndx; ++i)
10429 (*MIB).addOperand(*argOpers[i]);
10431 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
10432 (*MIB).setMemRefs(bInstr->memoperands_begin(),
10433 bInstr->memoperands_end());
10435 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
10436 MIB.addReg(X86::EAX);
10437 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
10438 MIB.addReg(X86::EDX);
10441 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
10443 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
10447 // private utility function
10448 MachineBasicBlock *
10449 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
10450 MachineBasicBlock *MBB,
10451 unsigned cmovOpc) const {
10452 // For the atomic min/max operator, we generate
10455 // ld t1 = [min/max.addr]
10456 // mov t2 = [min/max.val]
10458 // cmov[cond] t2 = t1
10460 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
10462 // fallthrough -->nextMBB
10464 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10465 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10466 MachineFunction::iterator MBBIter = MBB;
10469 /// First build the CFG
10470 MachineFunction *F = MBB->getParent();
10471 MachineBasicBlock *thisMBB = MBB;
10472 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
10473 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
10474 F->insert(MBBIter, newMBB);
10475 F->insert(MBBIter, nextMBB);
10477 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
10478 nextMBB->splice(nextMBB->begin(), thisMBB,
10479 llvm::next(MachineBasicBlock::iterator(mInstr)),
10481 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10483 // Update thisMBB to fall through to newMBB
10484 thisMBB->addSuccessor(newMBB);
10486 // newMBB jumps to newMBB and fall through to nextMBB
10487 newMBB->addSuccessor(nextMBB);
10488 newMBB->addSuccessor(newMBB);
10490 DebugLoc dl = mInstr->getDebugLoc();
10491 // Insert instructions into newMBB based on incoming instruction
10492 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
10493 "unexpected number of operands");
10494 MachineOperand& destOper = mInstr->getOperand(0);
10495 MachineOperand* argOpers[2 + X86::AddrNumOperands];
10496 int numArgs = mInstr->getNumOperands() - 1;
10497 for (int i=0; i < numArgs; ++i)
10498 argOpers[i] = &mInstr->getOperand(i+1);
10500 // x86 address has 4 operands: base, index, scale, and displacement
10501 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
10502 int valArgIndx = lastAddrIndx + 1;
10504 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
10505 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
10506 for (int i=0; i <= lastAddrIndx; ++i)
10507 (*MIB).addOperand(*argOpers[i]);
10509 // We only support register and immediate values
10510 assert((argOpers[valArgIndx]->isReg() ||
10511 argOpers[valArgIndx]->isImm()) &&
10512 "invalid operand");
10514 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
10515 if (argOpers[valArgIndx]->isReg())
10516 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
10518 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
10519 (*MIB).addOperand(*argOpers[valArgIndx]);
10521 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
10524 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
10529 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
10530 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
10534 // Cmp and exchange if none has modified the memory location
10535 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
10536 for (int i=0; i <= lastAddrIndx; ++i)
10537 (*MIB).addOperand(*argOpers[i]);
10539 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
10540 (*MIB).setMemRefs(mInstr->memoperands_begin(),
10541 mInstr->memoperands_end());
10543 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
10544 MIB.addReg(X86::EAX);
10547 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
10549 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
10553 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
10554 // or XMM0_V32I8 in AVX all of this code can be replaced with that
10555 // in the .td file.
10556 MachineBasicBlock *
10557 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
10558 unsigned numArgs, bool memArg) const {
10559 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
10560 "Target must have SSE4.2 or AVX features enabled");
10562 DebugLoc dl = MI->getDebugLoc();
10563 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10565 if (!Subtarget->hasAVX()) {
10567 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
10569 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
10572 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
10574 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
10577 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
10578 for (unsigned i = 0; i < numArgs; ++i) {
10579 MachineOperand &Op = MI->getOperand(i+1);
10580 if (!(Op.isReg() && Op.isImplicit()))
10581 MIB.addOperand(Op);
10583 BuildMI(*BB, MI, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
10584 .addReg(X86::XMM0);
10586 MI->eraseFromParent();
10590 MachineBasicBlock *
10591 X86TargetLowering::EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB) const {
10592 DebugLoc dl = MI->getDebugLoc();
10593 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10595 // Address into RAX/EAX, other two args into ECX, EDX.
10596 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
10597 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
10598 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
10599 for (int i = 0; i < X86::AddrNumOperands; ++i)
10600 MIB.addOperand(MI->getOperand(i));
10602 unsigned ValOps = X86::AddrNumOperands;
10603 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
10604 .addReg(MI->getOperand(ValOps).getReg());
10605 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
10606 .addReg(MI->getOperand(ValOps+1).getReg());
10608 // The instruction doesn't actually take any operands though.
10609 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
10611 MI->eraseFromParent(); // The pseudo is gone now.
10615 MachineBasicBlock *
10616 X86TargetLowering::EmitMwait(MachineInstr *MI, MachineBasicBlock *BB) const {
10617 DebugLoc dl = MI->getDebugLoc();
10618 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10620 // First arg in ECX, the second in EAX.
10621 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
10622 .addReg(MI->getOperand(0).getReg());
10623 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EAX)
10624 .addReg(MI->getOperand(1).getReg());
10626 // The instruction doesn't actually take any operands though.
10627 BuildMI(*BB, MI, dl, TII->get(X86::MWAITrr));
10629 MI->eraseFromParent(); // The pseudo is gone now.
10633 MachineBasicBlock *
10634 X86TargetLowering::EmitVAARG64WithCustomInserter(
10636 MachineBasicBlock *MBB) const {
10637 // Emit va_arg instruction on X86-64.
10639 // Operands to this pseudo-instruction:
10640 // 0 ) Output : destination address (reg)
10641 // 1-5) Input : va_list address (addr, i64mem)
10642 // 6 ) ArgSize : Size (in bytes) of vararg type
10643 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
10644 // 8 ) Align : Alignment of type
10645 // 9 ) EFLAGS (implicit-def)
10647 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
10648 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
10650 unsigned DestReg = MI->getOperand(0).getReg();
10651 MachineOperand &Base = MI->getOperand(1);
10652 MachineOperand &Scale = MI->getOperand(2);
10653 MachineOperand &Index = MI->getOperand(3);
10654 MachineOperand &Disp = MI->getOperand(4);
10655 MachineOperand &Segment = MI->getOperand(5);
10656 unsigned ArgSize = MI->getOperand(6).getImm();
10657 unsigned ArgMode = MI->getOperand(7).getImm();
10658 unsigned Align = MI->getOperand(8).getImm();
10660 // Memory Reference
10661 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
10662 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
10663 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
10665 // Machine Information
10666 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10667 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
10668 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
10669 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
10670 DebugLoc DL = MI->getDebugLoc();
10672 // struct va_list {
10675 // i64 overflow_area (address)
10676 // i64 reg_save_area (address)
10678 // sizeof(va_list) = 24
10679 // alignment(va_list) = 8
10681 unsigned TotalNumIntRegs = 6;
10682 unsigned TotalNumXMMRegs = 8;
10683 bool UseGPOffset = (ArgMode == 1);
10684 bool UseFPOffset = (ArgMode == 2);
10685 unsigned MaxOffset = TotalNumIntRegs * 8 +
10686 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
10688 /* Align ArgSize to a multiple of 8 */
10689 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
10690 bool NeedsAlign = (Align > 8);
10692 MachineBasicBlock *thisMBB = MBB;
10693 MachineBasicBlock *overflowMBB;
10694 MachineBasicBlock *offsetMBB;
10695 MachineBasicBlock *endMBB;
10697 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
10698 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
10699 unsigned OffsetReg = 0;
10701 if (!UseGPOffset && !UseFPOffset) {
10702 // If we only pull from the overflow region, we don't create a branch.
10703 // We don't need to alter control flow.
10704 OffsetDestReg = 0; // unused
10705 OverflowDestReg = DestReg;
10708 overflowMBB = thisMBB;
10711 // First emit code to check if gp_offset (or fp_offset) is below the bound.
10712 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
10713 // If not, pull from overflow_area. (branch to overflowMBB)
10718 // offsetMBB overflowMBB
10723 // Registers for the PHI in endMBB
10724 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
10725 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
10727 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10728 MachineFunction *MF = MBB->getParent();
10729 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10730 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10731 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10733 MachineFunction::iterator MBBIter = MBB;
10736 // Insert the new basic blocks
10737 MF->insert(MBBIter, offsetMBB);
10738 MF->insert(MBBIter, overflowMBB);
10739 MF->insert(MBBIter, endMBB);
10741 // Transfer the remainder of MBB and its successor edges to endMBB.
10742 endMBB->splice(endMBB->begin(), thisMBB,
10743 llvm::next(MachineBasicBlock::iterator(MI)),
10745 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10747 // Make offsetMBB and overflowMBB successors of thisMBB
10748 thisMBB->addSuccessor(offsetMBB);
10749 thisMBB->addSuccessor(overflowMBB);
10751 // endMBB is a successor of both offsetMBB and overflowMBB
10752 offsetMBB->addSuccessor(endMBB);
10753 overflowMBB->addSuccessor(endMBB);
10755 // Load the offset value into a register
10756 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
10757 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
10761 .addDisp(Disp, UseFPOffset ? 4 : 0)
10762 .addOperand(Segment)
10763 .setMemRefs(MMOBegin, MMOEnd);
10765 // Check if there is enough room left to pull this argument.
10766 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
10768 .addImm(MaxOffset + 8 - ArgSizeA8);
10770 // Branch to "overflowMBB" if offset >= max
10771 // Fall through to "offsetMBB" otherwise
10772 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
10773 .addMBB(overflowMBB);
10776 // In offsetMBB, emit code to use the reg_save_area.
10778 assert(OffsetReg != 0);
10780 // Read the reg_save_area address.
10781 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
10782 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
10787 .addOperand(Segment)
10788 .setMemRefs(MMOBegin, MMOEnd);
10790 // Zero-extend the offset
10791 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
10792 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
10795 .addImm(X86::sub_32bit);
10797 // Add the offset to the reg_save_area to get the final address.
10798 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
10799 .addReg(OffsetReg64)
10800 .addReg(RegSaveReg);
10802 // Compute the offset for the next argument
10803 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
10804 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
10806 .addImm(UseFPOffset ? 16 : 8);
10808 // Store it back into the va_list.
10809 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
10813 .addDisp(Disp, UseFPOffset ? 4 : 0)
10814 .addOperand(Segment)
10815 .addReg(NextOffsetReg)
10816 .setMemRefs(MMOBegin, MMOEnd);
10819 BuildMI(offsetMBB, DL, TII->get(X86::JMP_4))
10824 // Emit code to use overflow area
10827 // Load the overflow_area address into a register.
10828 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
10829 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
10834 .addOperand(Segment)
10835 .setMemRefs(MMOBegin, MMOEnd);
10837 // If we need to align it, do so. Otherwise, just copy the address
10838 // to OverflowDestReg.
10840 // Align the overflow address
10841 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
10842 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
10844 // aligned_addr = (addr + (align-1)) & ~(align-1)
10845 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
10846 .addReg(OverflowAddrReg)
10849 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
10851 .addImm(~(uint64_t)(Align-1));
10853 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
10854 .addReg(OverflowAddrReg);
10857 // Compute the next overflow address after this argument.
10858 // (the overflow address should be kept 8-byte aligned)
10859 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
10860 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
10861 .addReg(OverflowDestReg)
10862 .addImm(ArgSizeA8);
10864 // Store the new overflow address.
10865 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
10870 .addOperand(Segment)
10871 .addReg(NextAddrReg)
10872 .setMemRefs(MMOBegin, MMOEnd);
10874 // If we branched, emit the PHI to the front of endMBB.
10876 BuildMI(*endMBB, endMBB->begin(), DL,
10877 TII->get(X86::PHI), DestReg)
10878 .addReg(OffsetDestReg).addMBB(offsetMBB)
10879 .addReg(OverflowDestReg).addMBB(overflowMBB);
10882 // Erase the pseudo instruction
10883 MI->eraseFromParent();
10888 MachineBasicBlock *
10889 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
10891 MachineBasicBlock *MBB) const {
10892 // Emit code to save XMM registers to the stack. The ABI says that the
10893 // number of registers to save is given in %al, so it's theoretically
10894 // possible to do an indirect jump trick to avoid saving all of them,
10895 // however this code takes a simpler approach and just executes all
10896 // of the stores if %al is non-zero. It's less code, and it's probably
10897 // easier on the hardware branch predictor, and stores aren't all that
10898 // expensive anyway.
10900 // Create the new basic blocks. One block contains all the XMM stores,
10901 // and one block is the final destination regardless of whether any
10902 // stores were performed.
10903 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10904 MachineFunction *F = MBB->getParent();
10905 MachineFunction::iterator MBBIter = MBB;
10907 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
10908 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
10909 F->insert(MBBIter, XMMSaveMBB);
10910 F->insert(MBBIter, EndMBB);
10912 // Transfer the remainder of MBB and its successor edges to EndMBB.
10913 EndMBB->splice(EndMBB->begin(), MBB,
10914 llvm::next(MachineBasicBlock::iterator(MI)),
10916 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
10918 // The original block will now fall through to the XMM save block.
10919 MBB->addSuccessor(XMMSaveMBB);
10920 // The XMMSaveMBB will fall through to the end block.
10921 XMMSaveMBB->addSuccessor(EndMBB);
10923 // Now add the instructions.
10924 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10925 DebugLoc DL = MI->getDebugLoc();
10927 unsigned CountReg = MI->getOperand(0).getReg();
10928 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
10929 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
10931 if (!Subtarget->isTargetWin64()) {
10932 // If %al is 0, branch around the XMM save block.
10933 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
10934 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
10935 MBB->addSuccessor(EndMBB);
10938 // In the XMM save block, save all the XMM argument registers.
10939 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
10940 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
10941 MachineMemOperand *MMO =
10942 F->getMachineMemOperand(
10943 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
10944 MachineMemOperand::MOStore,
10945 /*Size=*/16, /*Align=*/16);
10946 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
10947 .addFrameIndex(RegSaveFrameIndex)
10948 .addImm(/*Scale=*/1)
10949 .addReg(/*IndexReg=*/0)
10950 .addImm(/*Disp=*/Offset)
10951 .addReg(/*Segment=*/0)
10952 .addReg(MI->getOperand(i).getReg())
10953 .addMemOperand(MMO);
10956 MI->eraseFromParent(); // The pseudo instruction is gone now.
10961 MachineBasicBlock *
10962 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
10963 MachineBasicBlock *BB) const {
10964 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10965 DebugLoc DL = MI->getDebugLoc();
10967 // To "insert" a SELECT_CC instruction, we actually have to insert the
10968 // diamond control-flow pattern. The incoming instruction knows the
10969 // destination vreg to set, the condition code register to branch on, the
10970 // true/false values to select between, and a branch opcode to use.
10971 const BasicBlock *LLVM_BB = BB->getBasicBlock();
10972 MachineFunction::iterator It = BB;
10978 // cmpTY ccX, r1, r2
10980 // fallthrough --> copy0MBB
10981 MachineBasicBlock *thisMBB = BB;
10982 MachineFunction *F = BB->getParent();
10983 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
10984 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
10985 F->insert(It, copy0MBB);
10986 F->insert(It, sinkMBB);
10988 // If the EFLAGS register isn't dead in the terminator, then claim that it's
10989 // live into the sink and copy blocks.
10990 const MachineFunction *MF = BB->getParent();
10991 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
10992 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
10994 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
10995 const MachineOperand &MO = MI->getOperand(I);
10996 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
10997 unsigned Reg = MO.getReg();
10998 if (Reg != X86::EFLAGS) continue;
10999 copy0MBB->addLiveIn(Reg);
11000 sinkMBB->addLiveIn(Reg);
11003 // Transfer the remainder of BB and its successor edges to sinkMBB.
11004 sinkMBB->splice(sinkMBB->begin(), BB,
11005 llvm::next(MachineBasicBlock::iterator(MI)),
11007 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
11009 // Add the true and fallthrough blocks as its successors.
11010 BB->addSuccessor(copy0MBB);
11011 BB->addSuccessor(sinkMBB);
11013 // Create the conditional branch instruction.
11015 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
11016 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
11019 // %FalseValue = ...
11020 // # fallthrough to sinkMBB
11021 copy0MBB->addSuccessor(sinkMBB);
11024 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
11026 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
11027 TII->get(X86::PHI), MI->getOperand(0).getReg())
11028 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
11029 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
11031 MI->eraseFromParent(); // The pseudo instruction is gone now.
11035 MachineBasicBlock *
11036 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
11037 MachineBasicBlock *BB) const {
11038 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11039 DebugLoc DL = MI->getDebugLoc();
11041 assert(!Subtarget->isTargetEnvMacho());
11043 // The lowering is pretty easy: we're just emitting the call to _alloca. The
11044 // non-trivial part is impdef of ESP.
11046 if (Subtarget->isTargetWin64()) {
11047 if (Subtarget->isTargetCygMing()) {
11048 // ___chkstk(Mingw64):
11049 // Clobbers R10, R11, RAX and EFLAGS.
11051 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
11052 .addExternalSymbol("___chkstk")
11053 .addReg(X86::RAX, RegState::Implicit)
11054 .addReg(X86::RSP, RegState::Implicit)
11055 .addReg(X86::RAX, RegState::Define | RegState::Implicit)
11056 .addReg(X86::RSP, RegState::Define | RegState::Implicit)
11057 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
11059 // __chkstk(MSVCRT): does not update stack pointer.
11060 // Clobbers R10, R11 and EFLAGS.
11061 // FIXME: RAX(allocated size) might be reused and not killed.
11062 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
11063 .addExternalSymbol("__chkstk")
11064 .addReg(X86::RAX, RegState::Implicit)
11065 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
11066 // RAX has the offset to subtracted from RSP.
11067 BuildMI(*BB, MI, DL, TII->get(X86::SUB64rr), X86::RSP)
11072 const char *StackProbeSymbol =
11073 Subtarget->isTargetWindows() ? "_chkstk" : "_alloca";
11075 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
11076 .addExternalSymbol(StackProbeSymbol)
11077 .addReg(X86::EAX, RegState::Implicit)
11078 .addReg(X86::ESP, RegState::Implicit)
11079 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
11080 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
11081 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
11084 MI->eraseFromParent(); // The pseudo instruction is gone now.
11088 MachineBasicBlock *
11089 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
11090 MachineBasicBlock *BB) const {
11091 // This is pretty easy. We're taking the value that we received from
11092 // our load from the relocation, sticking it in either RDI (x86-64)
11093 // or EAX and doing an indirect call. The return value will then
11094 // be in the normal return register.
11095 const X86InstrInfo *TII
11096 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
11097 DebugLoc DL = MI->getDebugLoc();
11098 MachineFunction *F = BB->getParent();
11100 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
11101 assert(MI->getOperand(3).isGlobal() && "This should be a global");
11103 if (Subtarget->is64Bit()) {
11104 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
11105 TII->get(X86::MOV64rm), X86::RDI)
11107 .addImm(0).addReg(0)
11108 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
11109 MI->getOperand(3).getTargetFlags())
11111 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
11112 addDirectMem(MIB, X86::RDI);
11113 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
11114 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
11115 TII->get(X86::MOV32rm), X86::EAX)
11117 .addImm(0).addReg(0)
11118 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
11119 MI->getOperand(3).getTargetFlags())
11121 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
11122 addDirectMem(MIB, X86::EAX);
11124 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
11125 TII->get(X86::MOV32rm), X86::EAX)
11126 .addReg(TII->getGlobalBaseReg(F))
11127 .addImm(0).addReg(0)
11128 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
11129 MI->getOperand(3).getTargetFlags())
11131 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
11132 addDirectMem(MIB, X86::EAX);
11135 MI->eraseFromParent(); // The pseudo instruction is gone now.
11139 MachineBasicBlock *
11140 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
11141 MachineBasicBlock *BB) const {
11142 switch (MI->getOpcode()) {
11143 default: assert(false && "Unexpected instr type to insert");
11144 case X86::TAILJMPd64:
11145 case X86::TAILJMPr64:
11146 case X86::TAILJMPm64:
11147 assert(!"TAILJMP64 would not be touched here.");
11148 case X86::TCRETURNdi64:
11149 case X86::TCRETURNri64:
11150 case X86::TCRETURNmi64:
11151 // Defs of TCRETURNxx64 has Win64's callee-saved registers, as subset.
11152 // On AMD64, additional defs should be added before register allocation.
11153 if (!Subtarget->isTargetWin64()) {
11154 MI->addRegisterDefined(X86::RSI);
11155 MI->addRegisterDefined(X86::RDI);
11156 MI->addRegisterDefined(X86::XMM6);
11157 MI->addRegisterDefined(X86::XMM7);
11158 MI->addRegisterDefined(X86::XMM8);
11159 MI->addRegisterDefined(X86::XMM9);
11160 MI->addRegisterDefined(X86::XMM10);
11161 MI->addRegisterDefined(X86::XMM11);
11162 MI->addRegisterDefined(X86::XMM12);
11163 MI->addRegisterDefined(X86::XMM13);
11164 MI->addRegisterDefined(X86::XMM14);
11165 MI->addRegisterDefined(X86::XMM15);
11168 case X86::WIN_ALLOCA:
11169 return EmitLoweredWinAlloca(MI, BB);
11170 case X86::TLSCall_32:
11171 case X86::TLSCall_64:
11172 return EmitLoweredTLSCall(MI, BB);
11173 case X86::CMOV_GR8:
11174 case X86::CMOV_FR32:
11175 case X86::CMOV_FR64:
11176 case X86::CMOV_V4F32:
11177 case X86::CMOV_V2F64:
11178 case X86::CMOV_V2I64:
11179 case X86::CMOV_V8F32:
11180 case X86::CMOV_V4F64:
11181 case X86::CMOV_V4I64:
11182 case X86::CMOV_GR16:
11183 case X86::CMOV_GR32:
11184 case X86::CMOV_RFP32:
11185 case X86::CMOV_RFP64:
11186 case X86::CMOV_RFP80:
11187 return EmitLoweredSelect(MI, BB);
11189 case X86::FP32_TO_INT16_IN_MEM:
11190 case X86::FP32_TO_INT32_IN_MEM:
11191 case X86::FP32_TO_INT64_IN_MEM:
11192 case X86::FP64_TO_INT16_IN_MEM:
11193 case X86::FP64_TO_INT32_IN_MEM:
11194 case X86::FP64_TO_INT64_IN_MEM:
11195 case X86::FP80_TO_INT16_IN_MEM:
11196 case X86::FP80_TO_INT32_IN_MEM:
11197 case X86::FP80_TO_INT64_IN_MEM: {
11198 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11199 DebugLoc DL = MI->getDebugLoc();
11201 // Change the floating point control register to use "round towards zero"
11202 // mode when truncating to an integer value.
11203 MachineFunction *F = BB->getParent();
11204 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
11205 addFrameReference(BuildMI(*BB, MI, DL,
11206 TII->get(X86::FNSTCW16m)), CWFrameIdx);
11208 // Load the old value of the high byte of the control word...
11210 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
11211 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
11214 // Set the high part to be round to zero...
11215 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
11218 // Reload the modified control word now...
11219 addFrameReference(BuildMI(*BB, MI, DL,
11220 TII->get(X86::FLDCW16m)), CWFrameIdx);
11222 // Restore the memory image of control word to original value
11223 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
11226 // Get the X86 opcode to use.
11228 switch (MI->getOpcode()) {
11229 default: llvm_unreachable("illegal opcode!");
11230 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
11231 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
11232 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
11233 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
11234 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
11235 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
11236 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
11237 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
11238 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
11242 MachineOperand &Op = MI->getOperand(0);
11244 AM.BaseType = X86AddressMode::RegBase;
11245 AM.Base.Reg = Op.getReg();
11247 AM.BaseType = X86AddressMode::FrameIndexBase;
11248 AM.Base.FrameIndex = Op.getIndex();
11250 Op = MI->getOperand(1);
11252 AM.Scale = Op.getImm();
11253 Op = MI->getOperand(2);
11255 AM.IndexReg = Op.getImm();
11256 Op = MI->getOperand(3);
11257 if (Op.isGlobal()) {
11258 AM.GV = Op.getGlobal();
11260 AM.Disp = Op.getImm();
11262 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
11263 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
11265 // Reload the original control word now.
11266 addFrameReference(BuildMI(*BB, MI, DL,
11267 TII->get(X86::FLDCW16m)), CWFrameIdx);
11269 MI->eraseFromParent(); // The pseudo instruction is gone now.
11272 // String/text processing lowering.
11273 case X86::PCMPISTRM128REG:
11274 case X86::VPCMPISTRM128REG:
11275 return EmitPCMP(MI, BB, 3, false /* in-mem */);
11276 case X86::PCMPISTRM128MEM:
11277 case X86::VPCMPISTRM128MEM:
11278 return EmitPCMP(MI, BB, 3, true /* in-mem */);
11279 case X86::PCMPESTRM128REG:
11280 case X86::VPCMPESTRM128REG:
11281 return EmitPCMP(MI, BB, 5, false /* in mem */);
11282 case X86::PCMPESTRM128MEM:
11283 case X86::VPCMPESTRM128MEM:
11284 return EmitPCMP(MI, BB, 5, true /* in mem */);
11286 // Thread synchronization.
11288 return EmitMonitor(MI, BB);
11290 return EmitMwait(MI, BB);
11292 // Atomic Lowering.
11293 case X86::ATOMAND32:
11294 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
11295 X86::AND32ri, X86::MOV32rm,
11297 X86::NOT32r, X86::EAX,
11298 X86::GR32RegisterClass);
11299 case X86::ATOMOR32:
11300 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
11301 X86::OR32ri, X86::MOV32rm,
11303 X86::NOT32r, X86::EAX,
11304 X86::GR32RegisterClass);
11305 case X86::ATOMXOR32:
11306 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
11307 X86::XOR32ri, X86::MOV32rm,
11309 X86::NOT32r, X86::EAX,
11310 X86::GR32RegisterClass);
11311 case X86::ATOMNAND32:
11312 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
11313 X86::AND32ri, X86::MOV32rm,
11315 X86::NOT32r, X86::EAX,
11316 X86::GR32RegisterClass, true);
11317 case X86::ATOMMIN32:
11318 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
11319 case X86::ATOMMAX32:
11320 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
11321 case X86::ATOMUMIN32:
11322 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
11323 case X86::ATOMUMAX32:
11324 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
11326 case X86::ATOMAND16:
11327 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
11328 X86::AND16ri, X86::MOV16rm,
11330 X86::NOT16r, X86::AX,
11331 X86::GR16RegisterClass);
11332 case X86::ATOMOR16:
11333 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
11334 X86::OR16ri, X86::MOV16rm,
11336 X86::NOT16r, X86::AX,
11337 X86::GR16RegisterClass);
11338 case X86::ATOMXOR16:
11339 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
11340 X86::XOR16ri, X86::MOV16rm,
11342 X86::NOT16r, X86::AX,
11343 X86::GR16RegisterClass);
11344 case X86::ATOMNAND16:
11345 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
11346 X86::AND16ri, X86::MOV16rm,
11348 X86::NOT16r, X86::AX,
11349 X86::GR16RegisterClass, true);
11350 case X86::ATOMMIN16:
11351 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
11352 case X86::ATOMMAX16:
11353 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
11354 case X86::ATOMUMIN16:
11355 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
11356 case X86::ATOMUMAX16:
11357 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
11359 case X86::ATOMAND8:
11360 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
11361 X86::AND8ri, X86::MOV8rm,
11363 X86::NOT8r, X86::AL,
11364 X86::GR8RegisterClass);
11366 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
11367 X86::OR8ri, X86::MOV8rm,
11369 X86::NOT8r, X86::AL,
11370 X86::GR8RegisterClass);
11371 case X86::ATOMXOR8:
11372 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
11373 X86::XOR8ri, X86::MOV8rm,
11375 X86::NOT8r, X86::AL,
11376 X86::GR8RegisterClass);
11377 case X86::ATOMNAND8:
11378 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
11379 X86::AND8ri, X86::MOV8rm,
11381 X86::NOT8r, X86::AL,
11382 X86::GR8RegisterClass, true);
11383 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
11384 // This group is for 64-bit host.
11385 case X86::ATOMAND64:
11386 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
11387 X86::AND64ri32, X86::MOV64rm,
11389 X86::NOT64r, X86::RAX,
11390 X86::GR64RegisterClass);
11391 case X86::ATOMOR64:
11392 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
11393 X86::OR64ri32, X86::MOV64rm,
11395 X86::NOT64r, X86::RAX,
11396 X86::GR64RegisterClass);
11397 case X86::ATOMXOR64:
11398 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
11399 X86::XOR64ri32, X86::MOV64rm,
11401 X86::NOT64r, X86::RAX,
11402 X86::GR64RegisterClass);
11403 case X86::ATOMNAND64:
11404 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
11405 X86::AND64ri32, X86::MOV64rm,
11407 X86::NOT64r, X86::RAX,
11408 X86::GR64RegisterClass, true);
11409 case X86::ATOMMIN64:
11410 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
11411 case X86::ATOMMAX64:
11412 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
11413 case X86::ATOMUMIN64:
11414 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
11415 case X86::ATOMUMAX64:
11416 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
11418 // This group does 64-bit operations on a 32-bit host.
11419 case X86::ATOMAND6432:
11420 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11421 X86::AND32rr, X86::AND32rr,
11422 X86::AND32ri, X86::AND32ri,
11424 case X86::ATOMOR6432:
11425 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11426 X86::OR32rr, X86::OR32rr,
11427 X86::OR32ri, X86::OR32ri,
11429 case X86::ATOMXOR6432:
11430 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11431 X86::XOR32rr, X86::XOR32rr,
11432 X86::XOR32ri, X86::XOR32ri,
11434 case X86::ATOMNAND6432:
11435 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11436 X86::AND32rr, X86::AND32rr,
11437 X86::AND32ri, X86::AND32ri,
11439 case X86::ATOMADD6432:
11440 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11441 X86::ADD32rr, X86::ADC32rr,
11442 X86::ADD32ri, X86::ADC32ri,
11444 case X86::ATOMSUB6432:
11445 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11446 X86::SUB32rr, X86::SBB32rr,
11447 X86::SUB32ri, X86::SBB32ri,
11449 case X86::ATOMSWAP6432:
11450 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11451 X86::MOV32rr, X86::MOV32rr,
11452 X86::MOV32ri, X86::MOV32ri,
11454 case X86::VASTART_SAVE_XMM_REGS:
11455 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
11457 case X86::VAARG_64:
11458 return EmitVAARG64WithCustomInserter(MI, BB);
11462 //===----------------------------------------------------------------------===//
11463 // X86 Optimization Hooks
11464 //===----------------------------------------------------------------------===//
11466 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
11470 const SelectionDAG &DAG,
11471 unsigned Depth) const {
11472 unsigned Opc = Op.getOpcode();
11473 assert((Opc >= ISD::BUILTIN_OP_END ||
11474 Opc == ISD::INTRINSIC_WO_CHAIN ||
11475 Opc == ISD::INTRINSIC_W_CHAIN ||
11476 Opc == ISD::INTRINSIC_VOID) &&
11477 "Should use MaskedValueIsZero if you don't know whether Op"
11478 " is a target node!");
11480 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
11494 // These nodes' second result is a boolean.
11495 if (Op.getResNo() == 0)
11498 case X86ISD::SETCC:
11499 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
11500 Mask.getBitWidth() - 1);
11505 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
11506 unsigned Depth) const {
11507 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
11508 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
11509 return Op.getValueType().getScalarType().getSizeInBits();
11515 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
11516 /// node is a GlobalAddress + offset.
11517 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
11518 const GlobalValue* &GA,
11519 int64_t &Offset) const {
11520 if (N->getOpcode() == X86ISD::Wrapper) {
11521 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
11522 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
11523 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
11527 return TargetLowering::isGAPlusOffset(N, GA, Offset);
11530 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
11531 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
11532 TargetLowering::DAGCombinerInfo &DCI) {
11533 DebugLoc dl = N->getDebugLoc();
11534 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
11535 SDValue V1 = SVOp->getOperand(0);
11536 SDValue V2 = SVOp->getOperand(1);
11537 EVT VT = SVOp->getValueType(0);
11539 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
11540 V2.getOpcode() == ISD::CONCAT_VECTORS) {
11544 // V UNDEF BUILD_VECTOR UNDEF
11546 // CONCAT_VECTOR CONCAT_VECTOR
11549 // RESULT: V + zero extended
11551 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
11552 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
11553 V1.getOperand(1).getOpcode() != ISD::UNDEF)
11556 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
11559 // To match the shuffle mask, the first half of the mask should
11560 // be exactly the first vector, and all the rest a splat with the
11561 // first element of the second one.
11562 int NumElems = VT.getVectorNumElements();
11563 for (int i = 0; i < NumElems/2; ++i)
11564 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
11565 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
11568 // Emit a zeroed vector and insert the desired subvector on its
11570 SDValue Zeros = getZeroVector(VT, true /* HasSSE2 */, DAG, dl);
11571 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0),
11572 DAG.getConstant(0, MVT::i32), DAG, dl);
11573 return DCI.CombineTo(N, InsV);
11579 /// PerformShuffleCombine - Performs several different shuffle combines.
11580 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
11581 TargetLowering::DAGCombinerInfo &DCI) {
11582 DebugLoc dl = N->getDebugLoc();
11583 EVT VT = N->getValueType(0);
11585 // Don't create instructions with illegal types after legalize types has run.
11586 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11587 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
11590 // Only handle pure VECTOR_SHUFFLE nodes.
11591 if (VT.getSizeInBits() == 256 && N->getOpcode() == ISD::VECTOR_SHUFFLE)
11592 return PerformShuffleCombine256(N, DAG, DCI);
11594 // Only handle 128 wide vector from here on.
11595 if (VT.getSizeInBits() != 128)
11598 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
11599 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
11600 // consecutive, non-overlapping, and in the right order.
11601 SmallVector<SDValue, 16> Elts;
11602 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
11603 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
11605 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
11608 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
11609 /// generation and convert it from being a bunch of shuffles and extracts
11610 /// to a simple store and scalar loads to extract the elements.
11611 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
11612 const TargetLowering &TLI) {
11613 SDValue InputVector = N->getOperand(0);
11615 // Only operate on vectors of 4 elements, where the alternative shuffling
11616 // gets to be more expensive.
11617 if (InputVector.getValueType() != MVT::v4i32)
11620 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
11621 // single use which is a sign-extend or zero-extend, and all elements are
11623 SmallVector<SDNode *, 4> Uses;
11624 unsigned ExtractedElements = 0;
11625 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
11626 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
11627 if (UI.getUse().getResNo() != InputVector.getResNo())
11630 SDNode *Extract = *UI;
11631 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
11634 if (Extract->getValueType(0) != MVT::i32)
11636 if (!Extract->hasOneUse())
11638 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
11639 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
11641 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
11644 // Record which element was extracted.
11645 ExtractedElements |=
11646 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
11648 Uses.push_back(Extract);
11651 // If not all the elements were used, this may not be worthwhile.
11652 if (ExtractedElements != 15)
11655 // Ok, we've now decided to do the transformation.
11656 DebugLoc dl = InputVector.getDebugLoc();
11658 // Store the value to a temporary stack slot.
11659 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
11660 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
11661 MachinePointerInfo(), false, false, 0);
11663 // Replace each use (extract) with a load of the appropriate element.
11664 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
11665 UE = Uses.end(); UI != UE; ++UI) {
11666 SDNode *Extract = *UI;
11668 // cOMpute the element's address.
11669 SDValue Idx = Extract->getOperand(1);
11671 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
11672 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
11673 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
11675 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
11676 StackPtr, OffsetVal);
11678 // Load the scalar.
11679 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
11680 ScalarAddr, MachinePointerInfo(),
11683 // Replace the exact with the load.
11684 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
11687 // The replacement was made in place; don't return anything.
11691 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
11692 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
11693 const X86Subtarget *Subtarget) {
11694 DebugLoc DL = N->getDebugLoc();
11695 SDValue Cond = N->getOperand(0);
11696 // Get the LHS/RHS of the select.
11697 SDValue LHS = N->getOperand(1);
11698 SDValue RHS = N->getOperand(2);
11700 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
11701 // instructions match the semantics of the common C idiom x<y?x:y but not
11702 // x<=y?x:y, because of how they handle negative zero (which can be
11703 // ignored in unsafe-math mode).
11704 if (Subtarget->hasSSE2() &&
11705 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
11706 Cond.getOpcode() == ISD::SETCC) {
11707 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
11709 unsigned Opcode = 0;
11710 // Check for x CC y ? x : y.
11711 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
11712 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
11716 // Converting this to a min would handle NaNs incorrectly, and swapping
11717 // the operands would cause it to handle comparisons between positive
11718 // and negative zero incorrectly.
11719 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
11720 if (!UnsafeFPMath &&
11721 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
11723 std::swap(LHS, RHS);
11725 Opcode = X86ISD::FMIN;
11728 // Converting this to a min would handle comparisons between positive
11729 // and negative zero incorrectly.
11730 if (!UnsafeFPMath &&
11731 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
11733 Opcode = X86ISD::FMIN;
11736 // Converting this to a min would handle both negative zeros and NaNs
11737 // incorrectly, but we can swap the operands to fix both.
11738 std::swap(LHS, RHS);
11742 Opcode = X86ISD::FMIN;
11746 // Converting this to a max would handle comparisons between positive
11747 // and negative zero incorrectly.
11748 if (!UnsafeFPMath &&
11749 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
11751 Opcode = X86ISD::FMAX;
11754 // Converting this to a max would handle NaNs incorrectly, and swapping
11755 // the operands would cause it to handle comparisons between positive
11756 // and negative zero incorrectly.
11757 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
11758 if (!UnsafeFPMath &&
11759 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
11761 std::swap(LHS, RHS);
11763 Opcode = X86ISD::FMAX;
11766 // Converting this to a max would handle both negative zeros and NaNs
11767 // incorrectly, but we can swap the operands to fix both.
11768 std::swap(LHS, RHS);
11772 Opcode = X86ISD::FMAX;
11775 // Check for x CC y ? y : x -- a min/max with reversed arms.
11776 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
11777 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
11781 // Converting this to a min would handle comparisons between positive
11782 // and negative zero incorrectly, and swapping the operands would
11783 // cause it to handle NaNs incorrectly.
11784 if (!UnsafeFPMath &&
11785 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
11786 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11788 std::swap(LHS, RHS);
11790 Opcode = X86ISD::FMIN;
11793 // Converting this to a min would handle NaNs incorrectly.
11794 if (!UnsafeFPMath &&
11795 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
11797 Opcode = X86ISD::FMIN;
11800 // Converting this to a min would handle both negative zeros and NaNs
11801 // incorrectly, but we can swap the operands to fix both.
11802 std::swap(LHS, RHS);
11806 Opcode = X86ISD::FMIN;
11810 // Converting this to a max would handle NaNs incorrectly.
11811 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11813 Opcode = X86ISD::FMAX;
11816 // Converting this to a max would handle comparisons between positive
11817 // and negative zero incorrectly, and swapping the operands would
11818 // cause it to handle NaNs incorrectly.
11819 if (!UnsafeFPMath &&
11820 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
11821 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11823 std::swap(LHS, RHS);
11825 Opcode = X86ISD::FMAX;
11828 // Converting this to a max would handle both negative zeros and NaNs
11829 // incorrectly, but we can swap the operands to fix both.
11830 std::swap(LHS, RHS);
11834 Opcode = X86ISD::FMAX;
11840 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
11843 // If this is a select between two integer constants, try to do some
11845 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
11846 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
11847 // Don't do this for crazy integer types.
11848 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
11849 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
11850 // so that TrueC (the true value) is larger than FalseC.
11851 bool NeedsCondInvert = false;
11853 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
11854 // Efficiently invertible.
11855 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
11856 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
11857 isa<ConstantSDNode>(Cond.getOperand(1))))) {
11858 NeedsCondInvert = true;
11859 std::swap(TrueC, FalseC);
11862 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
11863 if (FalseC->getAPIntValue() == 0 &&
11864 TrueC->getAPIntValue().isPowerOf2()) {
11865 if (NeedsCondInvert) // Invert the condition if needed.
11866 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11867 DAG.getConstant(1, Cond.getValueType()));
11869 // Zero extend the condition if needed.
11870 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
11872 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
11873 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
11874 DAG.getConstant(ShAmt, MVT::i8));
11877 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
11878 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
11879 if (NeedsCondInvert) // Invert the condition if needed.
11880 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11881 DAG.getConstant(1, Cond.getValueType()));
11883 // Zero extend the condition if needed.
11884 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
11885 FalseC->getValueType(0), Cond);
11886 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11887 SDValue(FalseC, 0));
11890 // Optimize cases that will turn into an LEA instruction. This requires
11891 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
11892 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
11893 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
11894 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
11896 bool isFastMultiplier = false;
11898 switch ((unsigned char)Diff) {
11900 case 1: // result = add base, cond
11901 case 2: // result = lea base( , cond*2)
11902 case 3: // result = lea base(cond, cond*2)
11903 case 4: // result = lea base( , cond*4)
11904 case 5: // result = lea base(cond, cond*4)
11905 case 8: // result = lea base( , cond*8)
11906 case 9: // result = lea base(cond, cond*8)
11907 isFastMultiplier = true;
11912 if (isFastMultiplier) {
11913 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
11914 if (NeedsCondInvert) // Invert the condition if needed.
11915 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11916 DAG.getConstant(1, Cond.getValueType()));
11918 // Zero extend the condition if needed.
11919 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
11921 // Scale the condition by the difference.
11923 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
11924 DAG.getConstant(Diff, Cond.getValueType()));
11926 // Add the base if non-zero.
11927 if (FalseC->getAPIntValue() != 0)
11928 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11929 SDValue(FalseC, 0));
11939 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
11940 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
11941 TargetLowering::DAGCombinerInfo &DCI) {
11942 DebugLoc DL = N->getDebugLoc();
11944 // If the flag operand isn't dead, don't touch this CMOV.
11945 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
11948 SDValue FalseOp = N->getOperand(0);
11949 SDValue TrueOp = N->getOperand(1);
11950 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
11951 SDValue Cond = N->getOperand(3);
11952 if (CC == X86::COND_E || CC == X86::COND_NE) {
11953 switch (Cond.getOpcode()) {
11957 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
11958 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
11959 return (CC == X86::COND_E) ? FalseOp : TrueOp;
11963 // If this is a select between two integer constants, try to do some
11964 // optimizations. Note that the operands are ordered the opposite of SELECT
11966 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
11967 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
11968 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
11969 // larger than FalseC (the false value).
11970 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
11971 CC = X86::GetOppositeBranchCondition(CC);
11972 std::swap(TrueC, FalseC);
11975 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
11976 // This is efficient for any integer data type (including i8/i16) and
11978 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
11979 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11980 DAG.getConstant(CC, MVT::i8), Cond);
11982 // Zero extend the condition if needed.
11983 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
11985 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
11986 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
11987 DAG.getConstant(ShAmt, MVT::i8));
11988 if (N->getNumValues() == 2) // Dead flag value?
11989 return DCI.CombineTo(N, Cond, SDValue());
11993 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
11994 // for any integer data type, including i8/i16.
11995 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
11996 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11997 DAG.getConstant(CC, MVT::i8), Cond);
11999 // Zero extend the condition if needed.
12000 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
12001 FalseC->getValueType(0), Cond);
12002 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
12003 SDValue(FalseC, 0));
12005 if (N->getNumValues() == 2) // Dead flag value?
12006 return DCI.CombineTo(N, Cond, SDValue());
12010 // Optimize cases that will turn into an LEA instruction. This requires
12011 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
12012 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
12013 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
12014 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
12016 bool isFastMultiplier = false;
12018 switch ((unsigned char)Diff) {
12020 case 1: // result = add base, cond
12021 case 2: // result = lea base( , cond*2)
12022 case 3: // result = lea base(cond, cond*2)
12023 case 4: // result = lea base( , cond*4)
12024 case 5: // result = lea base(cond, cond*4)
12025 case 8: // result = lea base( , cond*8)
12026 case 9: // result = lea base(cond, cond*8)
12027 isFastMultiplier = true;
12032 if (isFastMultiplier) {
12033 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
12034 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
12035 DAG.getConstant(CC, MVT::i8), Cond);
12036 // Zero extend the condition if needed.
12037 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
12039 // Scale the condition by the difference.
12041 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
12042 DAG.getConstant(Diff, Cond.getValueType()));
12044 // Add the base if non-zero.
12045 if (FalseC->getAPIntValue() != 0)
12046 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
12047 SDValue(FalseC, 0));
12048 if (N->getNumValues() == 2) // Dead flag value?
12049 return DCI.CombineTo(N, Cond, SDValue());
12059 /// PerformMulCombine - Optimize a single multiply with constant into two
12060 /// in order to implement it with two cheaper instructions, e.g.
12061 /// LEA + SHL, LEA + LEA.
12062 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
12063 TargetLowering::DAGCombinerInfo &DCI) {
12064 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
12067 EVT VT = N->getValueType(0);
12068 if (VT != MVT::i64)
12071 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
12074 uint64_t MulAmt = C->getZExtValue();
12075 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
12078 uint64_t MulAmt1 = 0;
12079 uint64_t MulAmt2 = 0;
12080 if ((MulAmt % 9) == 0) {
12082 MulAmt2 = MulAmt / 9;
12083 } else if ((MulAmt % 5) == 0) {
12085 MulAmt2 = MulAmt / 5;
12086 } else if ((MulAmt % 3) == 0) {
12088 MulAmt2 = MulAmt / 3;
12091 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
12092 DebugLoc DL = N->getDebugLoc();
12094 if (isPowerOf2_64(MulAmt2) &&
12095 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
12096 // If second multiplifer is pow2, issue it first. We want the multiply by
12097 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
12099 std::swap(MulAmt1, MulAmt2);
12102 if (isPowerOf2_64(MulAmt1))
12103 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
12104 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
12106 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
12107 DAG.getConstant(MulAmt1, VT));
12109 if (isPowerOf2_64(MulAmt2))
12110 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
12111 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
12113 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
12114 DAG.getConstant(MulAmt2, VT));
12116 // Do not add new nodes to DAG combiner worklist.
12117 DCI.CombineTo(N, NewMul, false);
12122 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
12123 SDValue N0 = N->getOperand(0);
12124 SDValue N1 = N->getOperand(1);
12125 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
12126 EVT VT = N0.getValueType();
12128 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
12129 // since the result of setcc_c is all zero's or all ones.
12130 if (N1C && N0.getOpcode() == ISD::AND &&
12131 N0.getOperand(1).getOpcode() == ISD::Constant) {
12132 SDValue N00 = N0.getOperand(0);
12133 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
12134 ((N00.getOpcode() == ISD::ANY_EXTEND ||
12135 N00.getOpcode() == ISD::ZERO_EXTEND) &&
12136 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
12137 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
12138 APInt ShAmt = N1C->getAPIntValue();
12139 Mask = Mask.shl(ShAmt);
12141 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
12142 N00, DAG.getConstant(Mask, VT));
12149 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
12151 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
12152 const X86Subtarget *Subtarget) {
12153 EVT VT = N->getValueType(0);
12154 if (!VT.isVector() && VT.isInteger() &&
12155 N->getOpcode() == ISD::SHL)
12156 return PerformSHLCombine(N, DAG);
12158 // On X86 with SSE2 support, we can transform this to a vector shift if
12159 // all elements are shifted by the same amount. We can't do this in legalize
12160 // because the a constant vector is typically transformed to a constant pool
12161 // so we have no knowledge of the shift amount.
12162 if (!(Subtarget->hasSSE2() || Subtarget->hasAVX()))
12165 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
12168 SDValue ShAmtOp = N->getOperand(1);
12169 EVT EltVT = VT.getVectorElementType();
12170 DebugLoc DL = N->getDebugLoc();
12171 SDValue BaseShAmt = SDValue();
12172 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
12173 unsigned NumElts = VT.getVectorNumElements();
12175 for (; i != NumElts; ++i) {
12176 SDValue Arg = ShAmtOp.getOperand(i);
12177 if (Arg.getOpcode() == ISD::UNDEF) continue;
12181 for (; i != NumElts; ++i) {
12182 SDValue Arg = ShAmtOp.getOperand(i);
12183 if (Arg.getOpcode() == ISD::UNDEF) continue;
12184 if (Arg != BaseShAmt) {
12188 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
12189 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
12190 SDValue InVec = ShAmtOp.getOperand(0);
12191 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
12192 unsigned NumElts = InVec.getValueType().getVectorNumElements();
12194 for (; i != NumElts; ++i) {
12195 SDValue Arg = InVec.getOperand(i);
12196 if (Arg.getOpcode() == ISD::UNDEF) continue;
12200 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
12201 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
12202 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
12203 if (C->getZExtValue() == SplatIdx)
12204 BaseShAmt = InVec.getOperand(1);
12207 if (BaseShAmt.getNode() == 0)
12208 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
12209 DAG.getIntPtrConstant(0));
12213 // The shift amount is an i32.
12214 if (EltVT.bitsGT(MVT::i32))
12215 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
12216 else if (EltVT.bitsLT(MVT::i32))
12217 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
12219 // The shift amount is identical so we can do a vector shift.
12220 SDValue ValOp = N->getOperand(0);
12221 switch (N->getOpcode()) {
12223 llvm_unreachable("Unknown shift opcode!");
12226 if (VT == MVT::v2i64)
12227 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12228 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
12230 if (VT == MVT::v4i32)
12231 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12232 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
12234 if (VT == MVT::v8i16)
12235 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12236 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
12240 if (VT == MVT::v4i32)
12241 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12242 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
12244 if (VT == MVT::v8i16)
12245 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12246 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
12250 if (VT == MVT::v2i64)
12251 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12252 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
12254 if (VT == MVT::v4i32)
12255 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12256 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
12258 if (VT == MVT::v8i16)
12259 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12260 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
12268 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
12269 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
12270 // and friends. Likewise for OR -> CMPNEQSS.
12271 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
12272 TargetLowering::DAGCombinerInfo &DCI,
12273 const X86Subtarget *Subtarget) {
12276 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
12277 // we're requiring SSE2 for both.
12278 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
12279 SDValue N0 = N->getOperand(0);
12280 SDValue N1 = N->getOperand(1);
12281 SDValue CMP0 = N0->getOperand(1);
12282 SDValue CMP1 = N1->getOperand(1);
12283 DebugLoc DL = N->getDebugLoc();
12285 // The SETCCs should both refer to the same CMP.
12286 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
12289 SDValue CMP00 = CMP0->getOperand(0);
12290 SDValue CMP01 = CMP0->getOperand(1);
12291 EVT VT = CMP00.getValueType();
12293 if (VT == MVT::f32 || VT == MVT::f64) {
12294 bool ExpectingFlags = false;
12295 // Check for any users that want flags:
12296 for (SDNode::use_iterator UI = N->use_begin(),
12298 !ExpectingFlags && UI != UE; ++UI)
12299 switch (UI->getOpcode()) {
12304 ExpectingFlags = true;
12306 case ISD::CopyToReg:
12307 case ISD::SIGN_EXTEND:
12308 case ISD::ZERO_EXTEND:
12309 case ISD::ANY_EXTEND:
12313 if (!ExpectingFlags) {
12314 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
12315 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
12317 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
12318 X86::CondCode tmp = cc0;
12323 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
12324 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
12325 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
12326 X86ISD::NodeType NTOperator = is64BitFP ?
12327 X86ISD::FSETCCsd : X86ISD::FSETCCss;
12328 // FIXME: need symbolic constants for these magic numbers.
12329 // See X86ATTInstPrinter.cpp:printSSECC().
12330 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
12331 SDValue OnesOrZeroesF = DAG.getNode(NTOperator, DL, MVT::f32, CMP00, CMP01,
12332 DAG.getConstant(x86cc, MVT::i8));
12333 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, MVT::i32,
12335 SDValue ANDed = DAG.getNode(ISD::AND, DL, MVT::i32, OnesOrZeroesI,
12336 DAG.getConstant(1, MVT::i32));
12337 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
12338 return OneBitOfTruth;
12346 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
12347 /// so it can be folded inside ANDNP.
12348 static bool CanFoldXORWithAllOnes(const SDNode *N) {
12349 EVT VT = N->getValueType(0);
12351 // Match direct AllOnes for 128 and 256-bit vectors
12352 if (ISD::isBuildVectorAllOnes(N))
12355 // Look through a bit convert.
12356 if (N->getOpcode() == ISD::BITCAST)
12357 N = N->getOperand(0).getNode();
12359 // Sometimes the operand may come from a insert_subvector building a 256-bit
12361 if (VT.getSizeInBits() == 256 &&
12362 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
12363 SDValue V1 = N->getOperand(0);
12364 SDValue V2 = N->getOperand(1);
12366 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
12367 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
12368 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
12369 ISD::isBuildVectorAllOnes(V2.getNode()))
12376 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
12377 TargetLowering::DAGCombinerInfo &DCI,
12378 const X86Subtarget *Subtarget) {
12379 if (DCI.isBeforeLegalizeOps())
12382 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
12386 // Want to form ANDNP nodes:
12387 // 1) In the hopes of then easily combining them with OR and AND nodes
12388 // to form PBLEND/PSIGN.
12389 // 2) To match ANDN packed intrinsics
12390 EVT VT = N->getValueType(0);
12391 if (VT != MVT::v2i64 && VT != MVT::v4i64)
12394 SDValue N0 = N->getOperand(0);
12395 SDValue N1 = N->getOperand(1);
12396 DebugLoc DL = N->getDebugLoc();
12398 // Check LHS for vnot
12399 if (N0.getOpcode() == ISD::XOR &&
12400 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
12401 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
12402 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
12404 // Check RHS for vnot
12405 if (N1.getOpcode() == ISD::XOR &&
12406 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
12407 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
12408 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
12413 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
12414 TargetLowering::DAGCombinerInfo &DCI,
12415 const X86Subtarget *Subtarget) {
12416 if (DCI.isBeforeLegalizeOps())
12419 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
12423 EVT VT = N->getValueType(0);
12424 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64 && VT != MVT::v2i64)
12427 SDValue N0 = N->getOperand(0);
12428 SDValue N1 = N->getOperand(1);
12430 // look for psign/blend
12431 if (Subtarget->hasSSSE3()) {
12432 if (VT == MVT::v2i64) {
12433 // Canonicalize pandn to RHS
12434 if (N0.getOpcode() == X86ISD::ANDNP)
12436 // or (and (m, x), (pandn m, y))
12437 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
12438 SDValue Mask = N1.getOperand(0);
12439 SDValue X = N1.getOperand(1);
12441 if (N0.getOperand(0) == Mask)
12442 Y = N0.getOperand(1);
12443 if (N0.getOperand(1) == Mask)
12444 Y = N0.getOperand(0);
12446 // Check to see if the mask appeared in both the AND and ANDNP and
12450 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
12451 if (Mask.getOpcode() != ISD::BITCAST ||
12452 X.getOpcode() != ISD::BITCAST ||
12453 Y.getOpcode() != ISD::BITCAST)
12456 // Look through mask bitcast.
12457 Mask = Mask.getOperand(0);
12458 EVT MaskVT = Mask.getValueType();
12460 // Validate that the Mask operand is a vector sra node. The sra node
12461 // will be an intrinsic.
12462 if (Mask.getOpcode() != ISD::INTRINSIC_WO_CHAIN)
12465 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
12466 // there is no psrai.b
12467 switch (cast<ConstantSDNode>(Mask.getOperand(0))->getZExtValue()) {
12468 case Intrinsic::x86_sse2_psrai_w:
12469 case Intrinsic::x86_sse2_psrai_d:
12471 default: return SDValue();
12474 // Check that the SRA is all signbits.
12475 SDValue SraC = Mask.getOperand(2);
12476 unsigned SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
12477 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
12478 if ((SraAmt + 1) != EltBits)
12481 DebugLoc DL = N->getDebugLoc();
12483 // Now we know we at least have a plendvb with the mask val. See if
12484 // we can form a psignb/w/d.
12485 // psign = x.type == y.type == mask.type && y = sub(0, x);
12486 X = X.getOperand(0);
12487 Y = Y.getOperand(0);
12488 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
12489 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
12490 X.getValueType() == MaskVT && X.getValueType() == Y.getValueType()){
12493 case 8: Opc = X86ISD::PSIGNB; break;
12494 case 16: Opc = X86ISD::PSIGNW; break;
12495 case 32: Opc = X86ISD::PSIGND; break;
12499 SDValue Sign = DAG.getNode(Opc, DL, MaskVT, X, Mask.getOperand(1));
12500 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Sign);
12503 // PBLENDVB only available on SSE 4.1
12504 if (!Subtarget->hasSSE41())
12507 X = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, X);
12508 Y = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Y);
12509 Mask = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Mask);
12510 Mask = DAG.getNode(X86ISD::PBLENDVB, DL, MVT::v16i8, X, Y, Mask);
12511 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Mask);
12516 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
12517 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
12519 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
12521 if (!N0.hasOneUse() || !N1.hasOneUse())
12524 SDValue ShAmt0 = N0.getOperand(1);
12525 if (ShAmt0.getValueType() != MVT::i8)
12527 SDValue ShAmt1 = N1.getOperand(1);
12528 if (ShAmt1.getValueType() != MVT::i8)
12530 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
12531 ShAmt0 = ShAmt0.getOperand(0);
12532 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
12533 ShAmt1 = ShAmt1.getOperand(0);
12535 DebugLoc DL = N->getDebugLoc();
12536 unsigned Opc = X86ISD::SHLD;
12537 SDValue Op0 = N0.getOperand(0);
12538 SDValue Op1 = N1.getOperand(0);
12539 if (ShAmt0.getOpcode() == ISD::SUB) {
12540 Opc = X86ISD::SHRD;
12541 std::swap(Op0, Op1);
12542 std::swap(ShAmt0, ShAmt1);
12545 unsigned Bits = VT.getSizeInBits();
12546 if (ShAmt1.getOpcode() == ISD::SUB) {
12547 SDValue Sum = ShAmt1.getOperand(0);
12548 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
12549 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
12550 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
12551 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
12552 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
12553 return DAG.getNode(Opc, DL, VT,
12555 DAG.getNode(ISD::TRUNCATE, DL,
12558 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
12559 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
12561 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
12562 return DAG.getNode(Opc, DL, VT,
12563 N0.getOperand(0), N1.getOperand(0),
12564 DAG.getNode(ISD::TRUNCATE, DL,
12571 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
12572 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
12573 const X86Subtarget *Subtarget) {
12574 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
12575 // the FP state in cases where an emms may be missing.
12576 // A preferable solution to the general problem is to figure out the right
12577 // places to insert EMMS. This qualifies as a quick hack.
12579 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
12580 StoreSDNode *St = cast<StoreSDNode>(N);
12581 EVT VT = St->getValue().getValueType();
12582 if (VT.getSizeInBits() != 64)
12585 const Function *F = DAG.getMachineFunction().getFunction();
12586 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
12587 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
12588 && Subtarget->hasSSE2();
12589 if ((VT.isVector() ||
12590 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
12591 isa<LoadSDNode>(St->getValue()) &&
12592 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
12593 St->getChain().hasOneUse() && !St->isVolatile()) {
12594 SDNode* LdVal = St->getValue().getNode();
12595 LoadSDNode *Ld = 0;
12596 int TokenFactorIndex = -1;
12597 SmallVector<SDValue, 8> Ops;
12598 SDNode* ChainVal = St->getChain().getNode();
12599 // Must be a store of a load. We currently handle two cases: the load
12600 // is a direct child, and it's under an intervening TokenFactor. It is
12601 // possible to dig deeper under nested TokenFactors.
12602 if (ChainVal == LdVal)
12603 Ld = cast<LoadSDNode>(St->getChain());
12604 else if (St->getValue().hasOneUse() &&
12605 ChainVal->getOpcode() == ISD::TokenFactor) {
12606 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
12607 if (ChainVal->getOperand(i).getNode() == LdVal) {
12608 TokenFactorIndex = i;
12609 Ld = cast<LoadSDNode>(St->getValue());
12611 Ops.push_back(ChainVal->getOperand(i));
12615 if (!Ld || !ISD::isNormalLoad(Ld))
12618 // If this is not the MMX case, i.e. we are just turning i64 load/store
12619 // into f64 load/store, avoid the transformation if there are multiple
12620 // uses of the loaded value.
12621 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
12624 DebugLoc LdDL = Ld->getDebugLoc();
12625 DebugLoc StDL = N->getDebugLoc();
12626 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
12627 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
12629 if (Subtarget->is64Bit() || F64IsLegal) {
12630 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
12631 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
12632 Ld->getPointerInfo(), Ld->isVolatile(),
12633 Ld->isNonTemporal(), Ld->getAlignment());
12634 SDValue NewChain = NewLd.getValue(1);
12635 if (TokenFactorIndex != -1) {
12636 Ops.push_back(NewChain);
12637 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
12640 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
12641 St->getPointerInfo(),
12642 St->isVolatile(), St->isNonTemporal(),
12643 St->getAlignment());
12646 // Otherwise, lower to two pairs of 32-bit loads / stores.
12647 SDValue LoAddr = Ld->getBasePtr();
12648 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
12649 DAG.getConstant(4, MVT::i32));
12651 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
12652 Ld->getPointerInfo(),
12653 Ld->isVolatile(), Ld->isNonTemporal(),
12654 Ld->getAlignment());
12655 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
12656 Ld->getPointerInfo().getWithOffset(4),
12657 Ld->isVolatile(), Ld->isNonTemporal(),
12658 MinAlign(Ld->getAlignment(), 4));
12660 SDValue NewChain = LoLd.getValue(1);
12661 if (TokenFactorIndex != -1) {
12662 Ops.push_back(LoLd);
12663 Ops.push_back(HiLd);
12664 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
12668 LoAddr = St->getBasePtr();
12669 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
12670 DAG.getConstant(4, MVT::i32));
12672 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
12673 St->getPointerInfo(),
12674 St->isVolatile(), St->isNonTemporal(),
12675 St->getAlignment());
12676 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
12677 St->getPointerInfo().getWithOffset(4),
12679 St->isNonTemporal(),
12680 MinAlign(St->getAlignment(), 4));
12681 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
12686 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
12687 /// X86ISD::FXOR nodes.
12688 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
12689 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
12690 // F[X]OR(0.0, x) -> x
12691 // F[X]OR(x, 0.0) -> x
12692 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
12693 if (C->getValueAPF().isPosZero())
12694 return N->getOperand(1);
12695 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
12696 if (C->getValueAPF().isPosZero())
12697 return N->getOperand(0);
12701 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
12702 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
12703 // FAND(0.0, x) -> 0.0
12704 // FAND(x, 0.0) -> 0.0
12705 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
12706 if (C->getValueAPF().isPosZero())
12707 return N->getOperand(0);
12708 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
12709 if (C->getValueAPF().isPosZero())
12710 return N->getOperand(1);
12714 static SDValue PerformBTCombine(SDNode *N,
12716 TargetLowering::DAGCombinerInfo &DCI) {
12717 // BT ignores high bits in the bit index operand.
12718 SDValue Op1 = N->getOperand(1);
12719 if (Op1.hasOneUse()) {
12720 unsigned BitWidth = Op1.getValueSizeInBits();
12721 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
12722 APInt KnownZero, KnownOne;
12723 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
12724 !DCI.isBeforeLegalizeOps());
12725 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12726 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
12727 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
12728 DCI.CommitTargetLoweringOpt(TLO);
12733 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
12734 SDValue Op = N->getOperand(0);
12735 if (Op.getOpcode() == ISD::BITCAST)
12736 Op = Op.getOperand(0);
12737 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
12738 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
12739 VT.getVectorElementType().getSizeInBits() ==
12740 OpVT.getVectorElementType().getSizeInBits()) {
12741 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
12746 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
12747 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
12748 // (and (i32 x86isd::setcc_carry), 1)
12749 // This eliminates the zext. This transformation is necessary because
12750 // ISD::SETCC is always legalized to i8.
12751 DebugLoc dl = N->getDebugLoc();
12752 SDValue N0 = N->getOperand(0);
12753 EVT VT = N->getValueType(0);
12754 if (N0.getOpcode() == ISD::AND &&
12756 N0.getOperand(0).hasOneUse()) {
12757 SDValue N00 = N0.getOperand(0);
12758 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
12760 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
12761 if (!C || C->getZExtValue() != 1)
12763 return DAG.getNode(ISD::AND, dl, VT,
12764 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
12765 N00.getOperand(0), N00.getOperand(1)),
12766 DAG.getConstant(1, VT));
12772 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
12773 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG) {
12774 unsigned X86CC = N->getConstantOperandVal(0);
12775 SDValue EFLAG = N->getOperand(1);
12776 DebugLoc DL = N->getDebugLoc();
12778 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
12779 // a zext and produces an all-ones bit which is more useful than 0/1 in some
12781 if (X86CC == X86::COND_B)
12782 return DAG.getNode(ISD::AND, DL, MVT::i8,
12783 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
12784 DAG.getConstant(X86CC, MVT::i8), EFLAG),
12785 DAG.getConstant(1, MVT::i8));
12790 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
12791 const X86TargetLowering *XTLI) {
12792 SDValue Op0 = N->getOperand(0);
12793 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
12794 // a 32-bit target where SSE doesn't support i64->FP operations.
12795 if (Op0.getOpcode() == ISD::LOAD) {
12796 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
12797 EVT VT = Ld->getValueType(0);
12798 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
12799 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
12800 !XTLI->getSubtarget()->is64Bit() &&
12801 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
12802 SDValue FILDChain = XTLI->BuildFILD(SDValue(N, 0), Ld->getValueType(0),
12803 Ld->getChain(), Op0, DAG);
12804 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
12811 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
12812 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
12813 X86TargetLowering::DAGCombinerInfo &DCI) {
12814 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
12815 // the result is either zero or one (depending on the input carry bit).
12816 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
12817 if (X86::isZeroNode(N->getOperand(0)) &&
12818 X86::isZeroNode(N->getOperand(1)) &&
12819 // We don't have a good way to replace an EFLAGS use, so only do this when
12821 SDValue(N, 1).use_empty()) {
12822 DebugLoc DL = N->getDebugLoc();
12823 EVT VT = N->getValueType(0);
12824 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
12825 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
12826 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
12827 DAG.getConstant(X86::COND_B,MVT::i8),
12829 DAG.getConstant(1, VT));
12830 return DCI.CombineTo(N, Res1, CarryOut);
12836 // fold (add Y, (sete X, 0)) -> adc 0, Y
12837 // (add Y, (setne X, 0)) -> sbb -1, Y
12838 // (sub (sete X, 0), Y) -> sbb 0, Y
12839 // (sub (setne X, 0), Y) -> adc -1, Y
12840 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
12841 DebugLoc DL = N->getDebugLoc();
12843 // Look through ZExts.
12844 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
12845 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
12848 SDValue SetCC = Ext.getOperand(0);
12849 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
12852 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
12853 if (CC != X86::COND_E && CC != X86::COND_NE)
12856 SDValue Cmp = SetCC.getOperand(1);
12857 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
12858 !X86::isZeroNode(Cmp.getOperand(1)) ||
12859 !Cmp.getOperand(0).getValueType().isInteger())
12862 SDValue CmpOp0 = Cmp.getOperand(0);
12863 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
12864 DAG.getConstant(1, CmpOp0.getValueType()));
12866 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
12867 if (CC == X86::COND_NE)
12868 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
12869 DL, OtherVal.getValueType(), OtherVal,
12870 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
12871 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
12872 DL, OtherVal.getValueType(), OtherVal,
12873 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
12876 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG) {
12877 SDValue Op0 = N->getOperand(0);
12878 SDValue Op1 = N->getOperand(1);
12880 // X86 can't encode an immediate LHS of a sub. See if we can push the
12881 // negation into a preceding instruction.
12882 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
12883 uint64_t Op0C = C->getSExtValue();
12885 // If the RHS of the sub is a XOR with one use and a constant, invert the
12886 // immediate. Then add one to the LHS of the sub so we can turn
12887 // X-Y -> X+~Y+1, saving one register.
12888 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
12889 isa<ConstantSDNode>(Op1.getOperand(1))) {
12890 uint64_t XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getSExtValue();
12891 EVT VT = Op0.getValueType();
12892 SDValue NewXor = DAG.getNode(ISD::XOR, Op1.getDebugLoc(), VT,
12894 DAG.getConstant(~XorC, VT));
12895 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, NewXor,
12896 DAG.getConstant(Op0C+1, VT));
12900 return OptimizeConditionalInDecrement(N, DAG);
12903 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
12904 DAGCombinerInfo &DCI) const {
12905 SelectionDAG &DAG = DCI.DAG;
12906 switch (N->getOpcode()) {
12908 case ISD::EXTRACT_VECTOR_ELT:
12909 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
12910 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
12911 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
12912 case ISD::ADD: return OptimizeConditionalInDecrement(N, DAG);
12913 case ISD::SUB: return PerformSubCombine(N, DAG);
12914 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
12915 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
12918 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
12919 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
12920 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
12921 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
12922 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, this);
12924 case X86ISD::FOR: return PerformFORCombine(N, DAG);
12925 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
12926 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
12927 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
12928 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
12929 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG);
12930 case X86ISD::SHUFPS: // Handle all target specific shuffles
12931 case X86ISD::SHUFPD:
12932 case X86ISD::PALIGN:
12933 case X86ISD::PUNPCKHBW:
12934 case X86ISD::PUNPCKHWD:
12935 case X86ISD::PUNPCKHDQ:
12936 case X86ISD::PUNPCKHQDQ:
12937 case X86ISD::UNPCKHPS:
12938 case X86ISD::UNPCKHPD:
12939 case X86ISD::VUNPCKHPSY:
12940 case X86ISD::VUNPCKHPDY:
12941 case X86ISD::PUNPCKLBW:
12942 case X86ISD::PUNPCKLWD:
12943 case X86ISD::PUNPCKLDQ:
12944 case X86ISD::PUNPCKLQDQ:
12945 case X86ISD::UNPCKLPS:
12946 case X86ISD::UNPCKLPD:
12947 case X86ISD::VUNPCKLPSY:
12948 case X86ISD::VUNPCKLPDY:
12949 case X86ISD::MOVHLPS:
12950 case X86ISD::MOVLHPS:
12951 case X86ISD::PSHUFD:
12952 case X86ISD::PSHUFHW:
12953 case X86ISD::PSHUFLW:
12954 case X86ISD::MOVSS:
12955 case X86ISD::MOVSD:
12956 case X86ISD::VPERMILPS:
12957 case X86ISD::VPERMILPSY:
12958 case X86ISD::VPERMILPD:
12959 case X86ISD::VPERMILPDY:
12960 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI);
12966 /// isTypeDesirableForOp - Return true if the target has native support for
12967 /// the specified value type and it is 'desirable' to use the type for the
12968 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
12969 /// instruction encodings are longer and some i16 instructions are slow.
12970 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
12971 if (!isTypeLegal(VT))
12973 if (VT != MVT::i16)
12980 case ISD::SIGN_EXTEND:
12981 case ISD::ZERO_EXTEND:
12982 case ISD::ANY_EXTEND:
12995 /// IsDesirableToPromoteOp - This method query the target whether it is
12996 /// beneficial for dag combiner to promote the specified node. If true, it
12997 /// should return the desired promotion type by reference.
12998 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
12999 EVT VT = Op.getValueType();
13000 if (VT != MVT::i16)
13003 bool Promote = false;
13004 bool Commute = false;
13005 switch (Op.getOpcode()) {
13008 LoadSDNode *LD = cast<LoadSDNode>(Op);
13009 // If the non-extending load has a single use and it's not live out, then it
13010 // might be folded.
13011 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
13012 Op.hasOneUse()*/) {
13013 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13014 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
13015 // The only case where we'd want to promote LOAD (rather then it being
13016 // promoted as an operand is when it's only use is liveout.
13017 if (UI->getOpcode() != ISD::CopyToReg)
13024 case ISD::SIGN_EXTEND:
13025 case ISD::ZERO_EXTEND:
13026 case ISD::ANY_EXTEND:
13031 SDValue N0 = Op.getOperand(0);
13032 // Look out for (store (shl (load), x)).
13033 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
13046 SDValue N0 = Op.getOperand(0);
13047 SDValue N1 = Op.getOperand(1);
13048 if (!Commute && MayFoldLoad(N1))
13050 // Avoid disabling potential load folding opportunities.
13051 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
13053 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
13063 //===----------------------------------------------------------------------===//
13064 // X86 Inline Assembly Support
13065 //===----------------------------------------------------------------------===//
13067 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
13068 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
13070 std::string AsmStr = IA->getAsmString();
13072 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
13073 SmallVector<StringRef, 4> AsmPieces;
13074 SplitString(AsmStr, AsmPieces, ";\n");
13076 switch (AsmPieces.size()) {
13077 default: return false;
13079 AsmStr = AsmPieces[0];
13081 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
13083 // FIXME: this should verify that we are targeting a 486 or better. If not,
13084 // we will turn this bswap into something that will be lowered to logical ops
13085 // instead of emitting the bswap asm. For now, we don't support 486 or lower
13086 // so don't worry about this.
13088 if (AsmPieces.size() == 2 &&
13089 (AsmPieces[0] == "bswap" ||
13090 AsmPieces[0] == "bswapq" ||
13091 AsmPieces[0] == "bswapl") &&
13092 (AsmPieces[1] == "$0" ||
13093 AsmPieces[1] == "${0:q}")) {
13094 // No need to check constraints, nothing other than the equivalent of
13095 // "=r,0" would be valid here.
13096 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
13097 if (!Ty || Ty->getBitWidth() % 16 != 0)
13099 return IntrinsicLowering::LowerToByteSwap(CI);
13101 // rorw $$8, ${0:w} --> llvm.bswap.i16
13102 if (CI->getType()->isIntegerTy(16) &&
13103 AsmPieces.size() == 3 &&
13104 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
13105 AsmPieces[1] == "$$8," &&
13106 AsmPieces[2] == "${0:w}" &&
13107 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
13109 const std::string &ConstraintsStr = IA->getConstraintString();
13110 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
13111 std::sort(AsmPieces.begin(), AsmPieces.end());
13112 if (AsmPieces.size() == 4 &&
13113 AsmPieces[0] == "~{cc}" &&
13114 AsmPieces[1] == "~{dirflag}" &&
13115 AsmPieces[2] == "~{flags}" &&
13116 AsmPieces[3] == "~{fpsr}") {
13117 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
13118 if (!Ty || Ty->getBitWidth() % 16 != 0)
13120 return IntrinsicLowering::LowerToByteSwap(CI);
13125 if (CI->getType()->isIntegerTy(32) &&
13126 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
13127 SmallVector<StringRef, 4> Words;
13128 SplitString(AsmPieces[0], Words, " \t,");
13129 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
13130 Words[2] == "${0:w}") {
13132 SplitString(AsmPieces[1], Words, " \t,");
13133 if (Words.size() == 3 && Words[0] == "rorl" && Words[1] == "$$16" &&
13134 Words[2] == "$0") {
13136 SplitString(AsmPieces[2], Words, " \t,");
13137 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
13138 Words[2] == "${0:w}") {
13140 const std::string &ConstraintsStr = IA->getConstraintString();
13141 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
13142 std::sort(AsmPieces.begin(), AsmPieces.end());
13143 if (AsmPieces.size() == 4 &&
13144 AsmPieces[0] == "~{cc}" &&
13145 AsmPieces[1] == "~{dirflag}" &&
13146 AsmPieces[2] == "~{flags}" &&
13147 AsmPieces[3] == "~{fpsr}") {
13148 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
13149 if (!Ty || Ty->getBitWidth() % 16 != 0)
13151 return IntrinsicLowering::LowerToByteSwap(CI);
13158 if (CI->getType()->isIntegerTy(64)) {
13159 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
13160 if (Constraints.size() >= 2 &&
13161 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
13162 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
13163 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
13164 SmallVector<StringRef, 4> Words;
13165 SplitString(AsmPieces[0], Words, " \t");
13166 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
13168 SplitString(AsmPieces[1], Words, " \t");
13169 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
13171 SplitString(AsmPieces[2], Words, " \t,");
13172 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
13173 Words[2] == "%edx") {
13174 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
13175 if (!Ty || Ty->getBitWidth() % 16 != 0)
13177 return IntrinsicLowering::LowerToByteSwap(CI);
13190 /// getConstraintType - Given a constraint letter, return the type of
13191 /// constraint it is for this target.
13192 X86TargetLowering::ConstraintType
13193 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
13194 if (Constraint.size() == 1) {
13195 switch (Constraint[0]) {
13206 return C_RegisterClass;
13230 return TargetLowering::getConstraintType(Constraint);
13233 /// Examine constraint type and operand type and determine a weight value.
13234 /// This object must already have been set up with the operand type
13235 /// and the current alternative constraint selected.
13236 TargetLowering::ConstraintWeight
13237 X86TargetLowering::getSingleConstraintMatchWeight(
13238 AsmOperandInfo &info, const char *constraint) const {
13239 ConstraintWeight weight = CW_Invalid;
13240 Value *CallOperandVal = info.CallOperandVal;
13241 // If we don't have a value, we can't do a match,
13242 // but allow it at the lowest weight.
13243 if (CallOperandVal == NULL)
13245 Type *type = CallOperandVal->getType();
13246 // Look at the constraint type.
13247 switch (*constraint) {
13249 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
13260 if (CallOperandVal->getType()->isIntegerTy())
13261 weight = CW_SpecificReg;
13266 if (type->isFloatingPointTy())
13267 weight = CW_SpecificReg;
13270 if (type->isX86_MMXTy() && Subtarget->hasMMX())
13271 weight = CW_SpecificReg;
13275 if ((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasXMM())
13276 weight = CW_Register;
13279 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
13280 if (C->getZExtValue() <= 31)
13281 weight = CW_Constant;
13285 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
13286 if (C->getZExtValue() <= 63)
13287 weight = CW_Constant;
13291 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
13292 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
13293 weight = CW_Constant;
13297 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
13298 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
13299 weight = CW_Constant;
13303 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
13304 if (C->getZExtValue() <= 3)
13305 weight = CW_Constant;
13309 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
13310 if (C->getZExtValue() <= 0xff)
13311 weight = CW_Constant;
13316 if (dyn_cast<ConstantFP>(CallOperandVal)) {
13317 weight = CW_Constant;
13321 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
13322 if ((C->getSExtValue() >= -0x80000000LL) &&
13323 (C->getSExtValue() <= 0x7fffffffLL))
13324 weight = CW_Constant;
13328 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
13329 if (C->getZExtValue() <= 0xffffffff)
13330 weight = CW_Constant;
13337 /// LowerXConstraint - try to replace an X constraint, which matches anything,
13338 /// with another that has more specific requirements based on the type of the
13339 /// corresponding operand.
13340 const char *X86TargetLowering::
13341 LowerXConstraint(EVT ConstraintVT) const {
13342 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
13343 // 'f' like normal targets.
13344 if (ConstraintVT.isFloatingPoint()) {
13345 if (Subtarget->hasXMMInt())
13347 if (Subtarget->hasXMM())
13351 return TargetLowering::LowerXConstraint(ConstraintVT);
13354 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
13355 /// vector. If it is invalid, don't add anything to Ops.
13356 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
13357 std::string &Constraint,
13358 std::vector<SDValue>&Ops,
13359 SelectionDAG &DAG) const {
13360 SDValue Result(0, 0);
13362 // Only support length 1 constraints for now.
13363 if (Constraint.length() > 1) return;
13365 char ConstraintLetter = Constraint[0];
13366 switch (ConstraintLetter) {
13369 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13370 if (C->getZExtValue() <= 31) {
13371 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
13377 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13378 if (C->getZExtValue() <= 63) {
13379 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
13385 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13386 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
13387 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
13393 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13394 if (C->getZExtValue() <= 255) {
13395 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
13401 // 32-bit signed value
13402 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13403 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
13404 C->getSExtValue())) {
13405 // Widen to 64 bits here to get it sign extended.
13406 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
13409 // FIXME gcc accepts some relocatable values here too, but only in certain
13410 // memory models; it's complicated.
13415 // 32-bit unsigned value
13416 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13417 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
13418 C->getZExtValue())) {
13419 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
13423 // FIXME gcc accepts some relocatable values here too, but only in certain
13424 // memory models; it's complicated.
13428 // Literal immediates are always ok.
13429 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
13430 // Widen to 64 bits here to get it sign extended.
13431 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
13435 // In any sort of PIC mode addresses need to be computed at runtime by
13436 // adding in a register or some sort of table lookup. These can't
13437 // be used as immediates.
13438 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
13441 // If we are in non-pic codegen mode, we allow the address of a global (with
13442 // an optional displacement) to be used with 'i'.
13443 GlobalAddressSDNode *GA = 0;
13444 int64_t Offset = 0;
13446 // Match either (GA), (GA+C), (GA+C1+C2), etc.
13448 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
13449 Offset += GA->getOffset();
13451 } else if (Op.getOpcode() == ISD::ADD) {
13452 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
13453 Offset += C->getZExtValue();
13454 Op = Op.getOperand(0);
13457 } else if (Op.getOpcode() == ISD::SUB) {
13458 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
13459 Offset += -C->getZExtValue();
13460 Op = Op.getOperand(0);
13465 // Otherwise, this isn't something we can handle, reject it.
13469 const GlobalValue *GV = GA->getGlobal();
13470 // If we require an extra load to get this address, as in PIC mode, we
13471 // can't accept it.
13472 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
13473 getTargetMachine())))
13476 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
13477 GA->getValueType(0), Offset);
13482 if (Result.getNode()) {
13483 Ops.push_back(Result);
13486 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
13489 std::pair<unsigned, const TargetRegisterClass*>
13490 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
13492 // First, see if this is a constraint that directly corresponds to an LLVM
13494 if (Constraint.size() == 1) {
13495 // GCC Constraint Letters
13496 switch (Constraint[0]) {
13498 // TODO: Slight differences here in allocation order and leaving
13499 // RIP in the class. Do they matter any more here than they do
13500 // in the normal allocation?
13501 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
13502 if (Subtarget->is64Bit()) {
13503 if (VT == MVT::i32 || VT == MVT::f32)
13504 return std::make_pair(0U, X86::GR32RegisterClass);
13505 else if (VT == MVT::i16)
13506 return std::make_pair(0U, X86::GR16RegisterClass);
13507 else if (VT == MVT::i8 || VT == MVT::i1)
13508 return std::make_pair(0U, X86::GR8RegisterClass);
13509 else if (VT == MVT::i64 || VT == MVT::f64)
13510 return std::make_pair(0U, X86::GR64RegisterClass);
13513 // 32-bit fallthrough
13514 case 'Q': // Q_REGS
13515 if (VT == MVT::i32 || VT == MVT::f32)
13516 return std::make_pair(0U, X86::GR32_ABCDRegisterClass);
13517 else if (VT == MVT::i16)
13518 return std::make_pair(0U, X86::GR16_ABCDRegisterClass);
13519 else if (VT == MVT::i8 || VT == MVT::i1)
13520 return std::make_pair(0U, X86::GR8_ABCD_LRegisterClass);
13521 else if (VT == MVT::i64)
13522 return std::make_pair(0U, X86::GR64_ABCDRegisterClass);
13524 case 'r': // GENERAL_REGS
13525 case 'l': // INDEX_REGS
13526 if (VT == MVT::i8 || VT == MVT::i1)
13527 return std::make_pair(0U, X86::GR8RegisterClass);
13528 if (VT == MVT::i16)
13529 return std::make_pair(0U, X86::GR16RegisterClass);
13530 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
13531 return std::make_pair(0U, X86::GR32RegisterClass);
13532 return std::make_pair(0U, X86::GR64RegisterClass);
13533 case 'R': // LEGACY_REGS
13534 if (VT == MVT::i8 || VT == MVT::i1)
13535 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
13536 if (VT == MVT::i16)
13537 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
13538 if (VT == MVT::i32 || !Subtarget->is64Bit())
13539 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
13540 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
13541 case 'f': // FP Stack registers.
13542 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
13543 // value to the correct fpstack register class.
13544 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
13545 return std::make_pair(0U, X86::RFP32RegisterClass);
13546 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
13547 return std::make_pair(0U, X86::RFP64RegisterClass);
13548 return std::make_pair(0U, X86::RFP80RegisterClass);
13549 case 'y': // MMX_REGS if MMX allowed.
13550 if (!Subtarget->hasMMX()) break;
13551 return std::make_pair(0U, X86::VR64RegisterClass);
13552 case 'Y': // SSE_REGS if SSE2 allowed
13553 if (!Subtarget->hasXMMInt()) break;
13555 case 'x': // SSE_REGS if SSE1 allowed
13556 if (!Subtarget->hasXMM()) break;
13558 switch (VT.getSimpleVT().SimpleTy) {
13560 // Scalar SSE types.
13563 return std::make_pair(0U, X86::FR32RegisterClass);
13566 return std::make_pair(0U, X86::FR64RegisterClass);
13574 return std::make_pair(0U, X86::VR128RegisterClass);
13580 // Use the default implementation in TargetLowering to convert the register
13581 // constraint into a member of a register class.
13582 std::pair<unsigned, const TargetRegisterClass*> Res;
13583 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
13585 // Not found as a standard register?
13586 if (Res.second == 0) {
13587 // Map st(0) -> st(7) -> ST0
13588 if (Constraint.size() == 7 && Constraint[0] == '{' &&
13589 tolower(Constraint[1]) == 's' &&
13590 tolower(Constraint[2]) == 't' &&
13591 Constraint[3] == '(' &&
13592 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
13593 Constraint[5] == ')' &&
13594 Constraint[6] == '}') {
13596 Res.first = X86::ST0+Constraint[4]-'0';
13597 Res.second = X86::RFP80RegisterClass;
13601 // GCC allows "st(0)" to be called just plain "st".
13602 if (StringRef("{st}").equals_lower(Constraint)) {
13603 Res.first = X86::ST0;
13604 Res.second = X86::RFP80RegisterClass;
13609 if (StringRef("{flags}").equals_lower(Constraint)) {
13610 Res.first = X86::EFLAGS;
13611 Res.second = X86::CCRRegisterClass;
13615 // 'A' means EAX + EDX.
13616 if (Constraint == "A") {
13617 Res.first = X86::EAX;
13618 Res.second = X86::GR32_ADRegisterClass;
13624 // Otherwise, check to see if this is a register class of the wrong value
13625 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
13626 // turn into {ax},{dx}.
13627 if (Res.second->hasType(VT))
13628 return Res; // Correct type already, nothing to do.
13630 // All of the single-register GCC register classes map their values onto
13631 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
13632 // really want an 8-bit or 32-bit register, map to the appropriate register
13633 // class and return the appropriate register.
13634 if (Res.second == X86::GR16RegisterClass) {
13635 if (VT == MVT::i8) {
13636 unsigned DestReg = 0;
13637 switch (Res.first) {
13639 case X86::AX: DestReg = X86::AL; break;
13640 case X86::DX: DestReg = X86::DL; break;
13641 case X86::CX: DestReg = X86::CL; break;
13642 case X86::BX: DestReg = X86::BL; break;
13645 Res.first = DestReg;
13646 Res.second = X86::GR8RegisterClass;
13648 } else if (VT == MVT::i32) {
13649 unsigned DestReg = 0;
13650 switch (Res.first) {
13652 case X86::AX: DestReg = X86::EAX; break;
13653 case X86::DX: DestReg = X86::EDX; break;
13654 case X86::CX: DestReg = X86::ECX; break;
13655 case X86::BX: DestReg = X86::EBX; break;
13656 case X86::SI: DestReg = X86::ESI; break;
13657 case X86::DI: DestReg = X86::EDI; break;
13658 case X86::BP: DestReg = X86::EBP; break;
13659 case X86::SP: DestReg = X86::ESP; break;
13662 Res.first = DestReg;
13663 Res.second = X86::GR32RegisterClass;
13665 } else if (VT == MVT::i64) {
13666 unsigned DestReg = 0;
13667 switch (Res.first) {
13669 case X86::AX: DestReg = X86::RAX; break;
13670 case X86::DX: DestReg = X86::RDX; break;
13671 case X86::CX: DestReg = X86::RCX; break;
13672 case X86::BX: DestReg = X86::RBX; break;
13673 case X86::SI: DestReg = X86::RSI; break;
13674 case X86::DI: DestReg = X86::RDI; break;
13675 case X86::BP: DestReg = X86::RBP; break;
13676 case X86::SP: DestReg = X86::RSP; break;
13679 Res.first = DestReg;
13680 Res.second = X86::GR64RegisterClass;
13683 } else if (Res.second == X86::FR32RegisterClass ||
13684 Res.second == X86::FR64RegisterClass ||
13685 Res.second == X86::VR128RegisterClass) {
13686 // Handle references to XMM physical registers that got mapped into the
13687 // wrong class. This can happen with constraints like {xmm0} where the
13688 // target independent register mapper will just pick the first match it can
13689 // find, ignoring the required type.
13690 if (VT == MVT::f32)
13691 Res.second = X86::FR32RegisterClass;
13692 else if (VT == MVT::f64)
13693 Res.second = X86::FR64RegisterClass;
13694 else if (X86::VR128RegisterClass->hasType(VT))
13695 Res.second = X86::VR128RegisterClass;