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();
172 X86ScalarSSEf32 = Subtarget->hasXMM();
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 // We may not have a libcall for MEMBARRIER so we should lower this.
453 setOperationAction(ISD::MEMBARRIER , 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()) {
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()) {
926 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
927 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
928 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
930 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
931 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
932 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
934 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
935 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
938 if (Subtarget->hasSSE42())
939 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
941 if (!UseSoftFloat && Subtarget->hasAVX()) {
942 addRegisterClass(MVT::v32i8, X86::VR256RegisterClass);
943 addRegisterClass(MVT::v16i16, X86::VR256RegisterClass);
944 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
945 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
946 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
947 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
949 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
950 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
951 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
953 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
954 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
955 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
956 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
957 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
958 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
960 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
961 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
962 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
963 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
964 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
965 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
967 // Custom lower several nodes for 256-bit types.
968 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
969 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
970 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
973 // Extract subvector is special because the value type
974 // (result) is 128-bit but the source is 256-bit wide.
975 if (VT.is128BitVector())
976 setOperationAction(ISD::EXTRACT_SUBVECTOR, SVT, Custom);
978 // Do not attempt to custom lower other non-256-bit vectors
979 if (!VT.is256BitVector())
982 setOperationAction(ISD::BUILD_VECTOR, SVT, Custom);
983 setOperationAction(ISD::VECTOR_SHUFFLE, SVT, Custom);
984 setOperationAction(ISD::INSERT_VECTOR_ELT, SVT, Custom);
985 setOperationAction(ISD::EXTRACT_VECTOR_ELT, SVT, Custom);
986 setOperationAction(ISD::SCALAR_TO_VECTOR, SVT, Custom);
987 setOperationAction(ISD::INSERT_SUBVECTOR, SVT, Custom);
990 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
991 for (unsigned i = (unsigned)MVT::v32i8; i != (unsigned)MVT::v4i64; ++i) {
992 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
995 // Do not attempt to promote non-256-bit vectors
996 if (!VT.is256BitVector())
999 setOperationAction(ISD::AND, SVT, Promote);
1000 AddPromotedToType (ISD::AND, SVT, MVT::v4i64);
1001 setOperationAction(ISD::OR, SVT, Promote);
1002 AddPromotedToType (ISD::OR, SVT, MVT::v4i64);
1003 setOperationAction(ISD::XOR, SVT, Promote);
1004 AddPromotedToType (ISD::XOR, SVT, MVT::v4i64);
1005 setOperationAction(ISD::LOAD, SVT, Promote);
1006 AddPromotedToType (ISD::LOAD, SVT, MVT::v4i64);
1007 setOperationAction(ISD::SELECT, SVT, Promote);
1008 AddPromotedToType (ISD::SELECT, SVT, MVT::v4i64);
1012 // SIGN_EXTEND_INREGs are evaluated by the extend type. Handle the expansion
1013 // of this type with custom code.
1014 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1015 VT != (unsigned)MVT::LAST_VECTOR_VALUETYPE; VT++) {
1016 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT, Custom);
1019 // We want to custom lower some of our intrinsics.
1020 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1023 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1024 // handle type legalization for these operations here.
1026 // FIXME: We really should do custom legalization for addition and
1027 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1028 // than generic legalization for 64-bit multiplication-with-overflow, though.
1029 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1030 // Add/Sub/Mul with overflow operations are custom lowered.
1032 setOperationAction(ISD::SADDO, VT, Custom);
1033 setOperationAction(ISD::UADDO, VT, Custom);
1034 setOperationAction(ISD::SSUBO, VT, Custom);
1035 setOperationAction(ISD::USUBO, VT, Custom);
1036 setOperationAction(ISD::SMULO, VT, Custom);
1037 setOperationAction(ISD::UMULO, VT, Custom);
1040 // There are no 8-bit 3-address imul/mul instructions
1041 setOperationAction(ISD::SMULO, MVT::i8, Expand);
1042 setOperationAction(ISD::UMULO, MVT::i8, Expand);
1044 if (!Subtarget->is64Bit()) {
1045 // These libcalls are not available in 32-bit.
1046 setLibcallName(RTLIB::SHL_I128, 0);
1047 setLibcallName(RTLIB::SRL_I128, 0);
1048 setLibcallName(RTLIB::SRA_I128, 0);
1051 // We have target-specific dag combine patterns for the following nodes:
1052 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1053 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1054 setTargetDAGCombine(ISD::BUILD_VECTOR);
1055 setTargetDAGCombine(ISD::SELECT);
1056 setTargetDAGCombine(ISD::SHL);
1057 setTargetDAGCombine(ISD::SRA);
1058 setTargetDAGCombine(ISD::SRL);
1059 setTargetDAGCombine(ISD::OR);
1060 setTargetDAGCombine(ISD::AND);
1061 setTargetDAGCombine(ISD::ADD);
1062 setTargetDAGCombine(ISD::SUB);
1063 setTargetDAGCombine(ISD::STORE);
1064 setTargetDAGCombine(ISD::ZERO_EXTEND);
1065 setTargetDAGCombine(ISD::SINT_TO_FP);
1066 if (Subtarget->is64Bit())
1067 setTargetDAGCombine(ISD::MUL);
1069 computeRegisterProperties();
1071 // On Darwin, -Os means optimize for size without hurting performance,
1072 // do not reduce the limit.
1073 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1074 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1075 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1076 maxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1077 maxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1078 maxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1079 setPrefLoopAlignment(16);
1080 benefitFromCodePlacementOpt = true;
1082 setPrefFunctionAlignment(4);
1086 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1091 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1092 /// the desired ByVal argument alignment.
1093 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1096 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1097 if (VTy->getBitWidth() == 128)
1099 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1100 unsigned EltAlign = 0;
1101 getMaxByValAlign(ATy->getElementType(), EltAlign);
1102 if (EltAlign > MaxAlign)
1103 MaxAlign = EltAlign;
1104 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1105 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1106 unsigned EltAlign = 0;
1107 getMaxByValAlign(STy->getElementType(i), EltAlign);
1108 if (EltAlign > MaxAlign)
1109 MaxAlign = EltAlign;
1117 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1118 /// function arguments in the caller parameter area. For X86, aggregates
1119 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1120 /// are at 4-byte boundaries.
1121 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1122 if (Subtarget->is64Bit()) {
1123 // Max of 8 and alignment of type.
1124 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1131 if (Subtarget->hasXMM())
1132 getMaxByValAlign(Ty, Align);
1136 /// getOptimalMemOpType - Returns the target specific optimal type for load
1137 /// and store operations as a result of memset, memcpy, and memmove
1138 /// lowering. If DstAlign is zero that means it's safe to destination
1139 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1140 /// means there isn't a need to check it against alignment requirement,
1141 /// probably because the source does not need to be loaded. If
1142 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1143 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1144 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1145 /// constant so it does not need to be loaded.
1146 /// It returns EVT::Other if the type should be determined using generic
1147 /// target-independent logic.
1149 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1150 unsigned DstAlign, unsigned SrcAlign,
1151 bool NonScalarIntSafe,
1153 MachineFunction &MF) const {
1154 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1155 // linux. This is because the stack realignment code can't handle certain
1156 // cases like PR2962. This should be removed when PR2962 is fixed.
1157 const Function *F = MF.getFunction();
1158 if (NonScalarIntSafe &&
1159 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1161 (Subtarget->isUnalignedMemAccessFast() ||
1162 ((DstAlign == 0 || DstAlign >= 16) &&
1163 (SrcAlign == 0 || SrcAlign >= 16))) &&
1164 Subtarget->getStackAlignment() >= 16) {
1165 if (Subtarget->hasSSE2())
1167 if (Subtarget->hasSSE1())
1169 } else if (!MemcpyStrSrc && Size >= 8 &&
1170 !Subtarget->is64Bit() &&
1171 Subtarget->getStackAlignment() >= 8 &&
1172 Subtarget->hasXMMInt()) {
1173 // Do not use f64 to lower memcpy if source is string constant. It's
1174 // better to use i32 to avoid the loads.
1178 if (Subtarget->is64Bit() && Size >= 8)
1183 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1184 /// current function. The returned value is a member of the
1185 /// MachineJumpTableInfo::JTEntryKind enum.
1186 unsigned X86TargetLowering::getJumpTableEncoding() const {
1187 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1189 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1190 Subtarget->isPICStyleGOT())
1191 return MachineJumpTableInfo::EK_Custom32;
1193 // Otherwise, use the normal jump table encoding heuristics.
1194 return TargetLowering::getJumpTableEncoding();
1198 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1199 const MachineBasicBlock *MBB,
1200 unsigned uid,MCContext &Ctx) const{
1201 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1202 Subtarget->isPICStyleGOT());
1203 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1205 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1206 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1209 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1211 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1212 SelectionDAG &DAG) const {
1213 if (!Subtarget->is64Bit())
1214 // This doesn't have DebugLoc associated with it, but is not really the
1215 // same as a Register.
1216 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1220 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1221 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1223 const MCExpr *X86TargetLowering::
1224 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1225 MCContext &Ctx) const {
1226 // X86-64 uses RIP relative addressing based on the jump table label.
1227 if (Subtarget->isPICStyleRIPRel())
1228 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1230 // Otherwise, the reference is relative to the PIC base.
1231 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1234 // FIXME: Why this routine is here? Move to RegInfo!
1235 std::pair<const TargetRegisterClass*, uint8_t>
1236 X86TargetLowering::findRepresentativeClass(EVT VT) const{
1237 const TargetRegisterClass *RRC = 0;
1239 switch (VT.getSimpleVT().SimpleTy) {
1241 return TargetLowering::findRepresentativeClass(VT);
1242 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1243 RRC = (Subtarget->is64Bit()
1244 ? X86::GR64RegisterClass : X86::GR32RegisterClass);
1247 RRC = X86::VR64RegisterClass;
1249 case MVT::f32: case MVT::f64:
1250 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1251 case MVT::v4f32: case MVT::v2f64:
1252 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1254 RRC = X86::VR128RegisterClass;
1257 return std::make_pair(RRC, Cost);
1260 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1261 unsigned &Offset) const {
1262 if (!Subtarget->isTargetLinux())
1265 if (Subtarget->is64Bit()) {
1266 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1268 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1281 //===----------------------------------------------------------------------===//
1282 // Return Value Calling Convention Implementation
1283 //===----------------------------------------------------------------------===//
1285 #include "X86GenCallingConv.inc"
1288 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1289 MachineFunction &MF, bool isVarArg,
1290 const SmallVectorImpl<ISD::OutputArg> &Outs,
1291 LLVMContext &Context) const {
1292 SmallVector<CCValAssign, 16> RVLocs;
1293 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1295 return CCInfo.CheckReturn(Outs, RetCC_X86);
1299 X86TargetLowering::LowerReturn(SDValue Chain,
1300 CallingConv::ID CallConv, bool isVarArg,
1301 const SmallVectorImpl<ISD::OutputArg> &Outs,
1302 const SmallVectorImpl<SDValue> &OutVals,
1303 DebugLoc dl, SelectionDAG &DAG) const {
1304 MachineFunction &MF = DAG.getMachineFunction();
1305 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1307 SmallVector<CCValAssign, 16> RVLocs;
1308 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1309 RVLocs, *DAG.getContext());
1310 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1312 // Add the regs to the liveout set for the function.
1313 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1314 for (unsigned i = 0; i != RVLocs.size(); ++i)
1315 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1316 MRI.addLiveOut(RVLocs[i].getLocReg());
1320 SmallVector<SDValue, 6> RetOps;
1321 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1322 // Operand #1 = Bytes To Pop
1323 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1326 // Copy the result values into the output registers.
1327 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1328 CCValAssign &VA = RVLocs[i];
1329 assert(VA.isRegLoc() && "Can only return in registers!");
1330 SDValue ValToCopy = OutVals[i];
1331 EVT ValVT = ValToCopy.getValueType();
1333 // If this is x86-64, and we disabled SSE, we can't return FP values,
1334 // or SSE or MMX vectors.
1335 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1336 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1337 (Subtarget->is64Bit() && !Subtarget->hasXMM())) {
1338 report_fatal_error("SSE register return with SSE disabled");
1340 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1341 // llvm-gcc has never done it right and no one has noticed, so this
1342 // should be OK for now.
1343 if (ValVT == MVT::f64 &&
1344 (Subtarget->is64Bit() && !Subtarget->hasXMMInt()))
1345 report_fatal_error("SSE2 register return with SSE2 disabled");
1347 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1348 // the RET instruction and handled by the FP Stackifier.
1349 if (VA.getLocReg() == X86::ST0 ||
1350 VA.getLocReg() == X86::ST1) {
1351 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1352 // change the value to the FP stack register class.
1353 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1354 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1355 RetOps.push_back(ValToCopy);
1356 // Don't emit a copytoreg.
1360 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1361 // which is returned in RAX / RDX.
1362 if (Subtarget->is64Bit()) {
1363 if (ValVT == MVT::x86mmx) {
1364 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1365 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
1366 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1368 // If we don't have SSE2 available, convert to v4f32 so the generated
1369 // register is legal.
1370 if (!Subtarget->hasSSE2())
1371 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
1376 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1377 Flag = Chain.getValue(1);
1380 // The x86-64 ABI for returning structs by value requires that we copy
1381 // the sret argument into %rax for the return. We saved the argument into
1382 // a virtual register in the entry block, so now we copy the value out
1384 if (Subtarget->is64Bit() &&
1385 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1386 MachineFunction &MF = DAG.getMachineFunction();
1387 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1388 unsigned Reg = FuncInfo->getSRetReturnReg();
1390 "SRetReturnReg should have been set in LowerFormalArguments().");
1391 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1393 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1394 Flag = Chain.getValue(1);
1396 // RAX now acts like a return value.
1397 MRI.addLiveOut(X86::RAX);
1400 RetOps[0] = Chain; // Update chain.
1402 // Add the flag if we have it.
1404 RetOps.push_back(Flag);
1406 return DAG.getNode(X86ISD::RET_FLAG, dl,
1407 MVT::Other, &RetOps[0], RetOps.size());
1410 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N) const {
1411 if (N->getNumValues() != 1)
1413 if (!N->hasNUsesOfValue(1, 0))
1416 SDNode *Copy = *N->use_begin();
1417 if (Copy->getOpcode() != ISD::CopyToReg &&
1418 Copy->getOpcode() != ISD::FP_EXTEND)
1421 bool HasRet = false;
1422 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1424 if (UI->getOpcode() != X86ISD::RET_FLAG)
1433 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
1434 ISD::NodeType ExtendKind) const {
1436 // TODO: Is this also valid on 32-bit?
1437 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
1438 ReturnMVT = MVT::i8;
1440 ReturnMVT = MVT::i32;
1442 EVT MinVT = getRegisterType(Context, ReturnMVT);
1443 return VT.bitsLT(MinVT) ? MinVT : VT;
1446 /// LowerCallResult - Lower the result values of a call into the
1447 /// appropriate copies out of appropriate physical registers.
1450 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1451 CallingConv::ID CallConv, bool isVarArg,
1452 const SmallVectorImpl<ISD::InputArg> &Ins,
1453 DebugLoc dl, SelectionDAG &DAG,
1454 SmallVectorImpl<SDValue> &InVals) const {
1456 // Assign locations to each value returned by this call.
1457 SmallVector<CCValAssign, 16> RVLocs;
1458 bool Is64Bit = Subtarget->is64Bit();
1459 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1460 getTargetMachine(), RVLocs, *DAG.getContext());
1461 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1463 // Copy all of the result registers out of their specified physreg.
1464 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1465 CCValAssign &VA = RVLocs[i];
1466 EVT CopyVT = VA.getValVT();
1468 // If this is x86-64, and we disabled SSE, we can't return FP values
1469 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1470 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasXMM())) {
1471 report_fatal_error("SSE register return with SSE disabled");
1476 // If this is a call to a function that returns an fp value on the floating
1477 // point stack, we must guarantee the the value is popped from the stack, so
1478 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1479 // if the return value is not used. We use the FpPOP_RETVAL instruction
1481 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1482 // If we prefer to use the value in xmm registers, copy it out as f80 and
1483 // use a truncate to move it from fp stack reg to xmm reg.
1484 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1485 SDValue Ops[] = { Chain, InFlag };
1486 Chain = SDValue(DAG.getMachineNode(X86::FpPOP_RETVAL, dl, CopyVT,
1487 MVT::Other, MVT::Glue, Ops, 2), 1);
1488 Val = Chain.getValue(0);
1490 // Round the f80 to the right size, which also moves it to the appropriate
1492 if (CopyVT != VA.getValVT())
1493 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1494 // This truncation won't change the value.
1495 DAG.getIntPtrConstant(1));
1497 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1498 CopyVT, InFlag).getValue(1);
1499 Val = Chain.getValue(0);
1501 InFlag = Chain.getValue(2);
1502 InVals.push_back(Val);
1509 //===----------------------------------------------------------------------===//
1510 // C & StdCall & Fast Calling Convention implementation
1511 //===----------------------------------------------------------------------===//
1512 // StdCall calling convention seems to be standard for many Windows' API
1513 // routines and around. It differs from C calling convention just a little:
1514 // callee should clean up the stack, not caller. Symbols should be also
1515 // decorated in some fancy way :) It doesn't support any vector arguments.
1516 // For info on fast calling convention see Fast Calling Convention (tail call)
1517 // implementation LowerX86_32FastCCCallTo.
1519 /// CallIsStructReturn - Determines whether a call uses struct return
1521 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1525 return Outs[0].Flags.isSRet();
1528 /// ArgsAreStructReturn - Determines whether a function uses struct
1529 /// return semantics.
1531 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1535 return Ins[0].Flags.isSRet();
1538 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1539 /// by "Src" to address "Dst" with size and alignment information specified by
1540 /// the specific parameter attribute. The copy will be passed as a byval
1541 /// function parameter.
1543 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1544 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1546 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1548 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1549 /*isVolatile*/false, /*AlwaysInline=*/true,
1550 MachinePointerInfo(), MachinePointerInfo());
1553 /// IsTailCallConvention - Return true if the calling convention is one that
1554 /// supports tail call optimization.
1555 static bool IsTailCallConvention(CallingConv::ID CC) {
1556 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1559 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1560 if (!CI->isTailCall())
1564 CallingConv::ID CalleeCC = CS.getCallingConv();
1565 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
1571 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1572 /// a tailcall target by changing its ABI.
1573 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1574 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1578 X86TargetLowering::LowerMemArgument(SDValue Chain,
1579 CallingConv::ID CallConv,
1580 const SmallVectorImpl<ISD::InputArg> &Ins,
1581 DebugLoc dl, SelectionDAG &DAG,
1582 const CCValAssign &VA,
1583 MachineFrameInfo *MFI,
1585 // Create the nodes corresponding to a load from this parameter slot.
1586 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1587 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1588 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1591 // If value is passed by pointer we have address passed instead of the value
1593 if (VA.getLocInfo() == CCValAssign::Indirect)
1594 ValVT = VA.getLocVT();
1596 ValVT = VA.getValVT();
1598 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1599 // changed with more analysis.
1600 // In case of tail call optimization mark all arguments mutable. Since they
1601 // could be overwritten by lowering of arguments in case of a tail call.
1602 if (Flags.isByVal()) {
1603 unsigned Bytes = Flags.getByValSize();
1604 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
1605 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
1606 return DAG.getFrameIndex(FI, getPointerTy());
1608 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1609 VA.getLocMemOffset(), isImmutable);
1610 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1611 return DAG.getLoad(ValVT, dl, Chain, FIN,
1612 MachinePointerInfo::getFixedStack(FI),
1618 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1619 CallingConv::ID CallConv,
1621 const SmallVectorImpl<ISD::InputArg> &Ins,
1624 SmallVectorImpl<SDValue> &InVals)
1626 MachineFunction &MF = DAG.getMachineFunction();
1627 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1629 const Function* Fn = MF.getFunction();
1630 if (Fn->hasExternalLinkage() &&
1631 Subtarget->isTargetCygMing() &&
1632 Fn->getName() == "main")
1633 FuncInfo->setForceFramePointer(true);
1635 MachineFrameInfo *MFI = MF.getFrameInfo();
1636 bool Is64Bit = Subtarget->is64Bit();
1637 bool IsWin64 = Subtarget->isTargetWin64();
1639 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1640 "Var args not supported with calling convention fastcc or ghc");
1642 // Assign locations to all of the incoming arguments.
1643 SmallVector<CCValAssign, 16> ArgLocs;
1644 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1645 ArgLocs, *DAG.getContext());
1647 // Allocate shadow area for Win64
1649 CCInfo.AllocateStack(32, 8);
1652 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
1654 unsigned LastVal = ~0U;
1656 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1657 CCValAssign &VA = ArgLocs[i];
1658 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1660 assert(VA.getValNo() != LastVal &&
1661 "Don't support value assigned to multiple locs yet");
1662 LastVal = VA.getValNo();
1664 if (VA.isRegLoc()) {
1665 EVT RegVT = VA.getLocVT();
1666 TargetRegisterClass *RC = NULL;
1667 if (RegVT == MVT::i32)
1668 RC = X86::GR32RegisterClass;
1669 else if (Is64Bit && RegVT == MVT::i64)
1670 RC = X86::GR64RegisterClass;
1671 else if (RegVT == MVT::f32)
1672 RC = X86::FR32RegisterClass;
1673 else if (RegVT == MVT::f64)
1674 RC = X86::FR64RegisterClass;
1675 else if (RegVT.isVector() && RegVT.getSizeInBits() == 256)
1676 RC = X86::VR256RegisterClass;
1677 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1678 RC = X86::VR128RegisterClass;
1679 else if (RegVT == MVT::x86mmx)
1680 RC = X86::VR64RegisterClass;
1682 llvm_unreachable("Unknown argument type!");
1684 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1685 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1687 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1688 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1690 if (VA.getLocInfo() == CCValAssign::SExt)
1691 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1692 DAG.getValueType(VA.getValVT()));
1693 else if (VA.getLocInfo() == CCValAssign::ZExt)
1694 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1695 DAG.getValueType(VA.getValVT()));
1696 else if (VA.getLocInfo() == CCValAssign::BCvt)
1697 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
1699 if (VA.isExtInLoc()) {
1700 // Handle MMX values passed in XMM regs.
1701 if (RegVT.isVector()) {
1702 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(),
1705 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1708 assert(VA.isMemLoc());
1709 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1712 // If value is passed via pointer - do a load.
1713 if (VA.getLocInfo() == CCValAssign::Indirect)
1714 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
1715 MachinePointerInfo(), false, false, 0);
1717 InVals.push_back(ArgValue);
1720 // The x86-64 ABI for returning structs by value requires that we copy
1721 // the sret argument into %rax for the return. Save the argument into
1722 // a virtual register so that we can access it from the return points.
1723 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1724 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1725 unsigned Reg = FuncInfo->getSRetReturnReg();
1727 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1728 FuncInfo->setSRetReturnReg(Reg);
1730 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1731 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1734 unsigned StackSize = CCInfo.getNextStackOffset();
1735 // Align stack specially for tail calls.
1736 if (FuncIsMadeTailCallSafe(CallConv))
1737 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1739 // If the function takes variable number of arguments, make a frame index for
1740 // the start of the first vararg value... for expansion of llvm.va_start.
1742 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1743 CallConv != CallingConv::X86_ThisCall)) {
1744 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1747 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1749 // FIXME: We should really autogenerate these arrays
1750 static const unsigned GPR64ArgRegsWin64[] = {
1751 X86::RCX, X86::RDX, X86::R8, X86::R9
1753 static const unsigned GPR64ArgRegs64Bit[] = {
1754 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1756 static const unsigned XMMArgRegs64Bit[] = {
1757 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1758 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1760 const unsigned *GPR64ArgRegs;
1761 unsigned NumXMMRegs = 0;
1764 // The XMM registers which might contain var arg parameters are shadowed
1765 // in their paired GPR. So we only need to save the GPR to their home
1767 TotalNumIntRegs = 4;
1768 GPR64ArgRegs = GPR64ArgRegsWin64;
1770 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1771 GPR64ArgRegs = GPR64ArgRegs64Bit;
1773 NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs64Bit, TotalNumXMMRegs);
1775 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1778 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1779 assert(!(NumXMMRegs && !Subtarget->hasXMM()) &&
1780 "SSE register cannot be used when SSE is disabled!");
1781 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1782 "SSE register cannot be used when SSE is disabled!");
1783 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasXMM())
1784 // Kernel mode asks for SSE to be disabled, so don't push them
1786 TotalNumXMMRegs = 0;
1789 const TargetFrameLowering &TFI = *getTargetMachine().getFrameLowering();
1790 // Get to the caller-allocated home save location. Add 8 to account
1791 // for the return address.
1792 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
1793 FuncInfo->setRegSaveFrameIndex(
1794 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
1795 // Fixup to set vararg frame on shadow area (4 x i64).
1797 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
1799 // For X86-64, if there are vararg parameters that are passed via
1800 // registers, then we must store them to their spots on the stack so they
1801 // may be loaded by deferencing the result of va_next.
1802 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1803 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1804 FuncInfo->setRegSaveFrameIndex(
1805 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1809 // Store the integer parameter registers.
1810 SmallVector<SDValue, 8> MemOps;
1811 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1813 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1814 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1815 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1816 DAG.getIntPtrConstant(Offset));
1817 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1818 X86::GR64RegisterClass);
1819 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1821 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1822 MachinePointerInfo::getFixedStack(
1823 FuncInfo->getRegSaveFrameIndex(), Offset),
1825 MemOps.push_back(Store);
1829 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1830 // Now store the XMM (fp + vector) parameter registers.
1831 SmallVector<SDValue, 11> SaveXMMOps;
1832 SaveXMMOps.push_back(Chain);
1834 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1835 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1836 SaveXMMOps.push_back(ALVal);
1838 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1839 FuncInfo->getRegSaveFrameIndex()));
1840 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1841 FuncInfo->getVarArgsFPOffset()));
1843 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1844 unsigned VReg = MF.addLiveIn(XMMArgRegs64Bit[NumXMMRegs],
1845 X86::VR128RegisterClass);
1846 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1847 SaveXMMOps.push_back(Val);
1849 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1851 &SaveXMMOps[0], SaveXMMOps.size()));
1854 if (!MemOps.empty())
1855 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1856 &MemOps[0], MemOps.size());
1860 // Some CCs need callee pop.
1861 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, GuaranteedTailCallOpt)) {
1862 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1864 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1865 // If this is an sret function, the return should pop the hidden pointer.
1866 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1867 FuncInfo->setBytesToPopOnReturn(4);
1871 // RegSaveFrameIndex is X86-64 only.
1872 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1873 if (CallConv == CallingConv::X86_FastCall ||
1874 CallConv == CallingConv::X86_ThisCall)
1875 // fastcc functions can't have varargs.
1876 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1883 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1884 SDValue StackPtr, SDValue Arg,
1885 DebugLoc dl, SelectionDAG &DAG,
1886 const CCValAssign &VA,
1887 ISD::ArgFlagsTy Flags) const {
1888 unsigned LocMemOffset = VA.getLocMemOffset();
1889 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1890 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1891 if (Flags.isByVal())
1892 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1894 return DAG.getStore(Chain, dl, Arg, PtrOff,
1895 MachinePointerInfo::getStack(LocMemOffset),
1899 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1900 /// optimization is performed and it is required.
1902 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1903 SDValue &OutRetAddr, SDValue Chain,
1904 bool IsTailCall, bool Is64Bit,
1905 int FPDiff, DebugLoc dl) const {
1906 // Adjust the Return address stack slot.
1907 EVT VT = getPointerTy();
1908 OutRetAddr = getReturnAddressFrameIndex(DAG);
1910 // Load the "old" Return address.
1911 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
1913 return SDValue(OutRetAddr.getNode(), 1);
1916 /// EmitTailCallStoreRetAddr - Emit a store of the return address if tail call
1917 /// optimization is performed and it is required (FPDiff!=0).
1919 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1920 SDValue Chain, SDValue RetAddrFrIdx,
1921 bool Is64Bit, int FPDiff, DebugLoc dl) {
1922 // Store the return address to the appropriate stack slot.
1923 if (!FPDiff) return Chain;
1924 // Calculate the new stack slot for the return address.
1925 int SlotSize = Is64Bit ? 8 : 4;
1926 int NewReturnAddrFI =
1927 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1928 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1929 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1930 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1931 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
1937 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1938 CallingConv::ID CallConv, bool isVarArg,
1940 const SmallVectorImpl<ISD::OutputArg> &Outs,
1941 const SmallVectorImpl<SDValue> &OutVals,
1942 const SmallVectorImpl<ISD::InputArg> &Ins,
1943 DebugLoc dl, SelectionDAG &DAG,
1944 SmallVectorImpl<SDValue> &InVals) const {
1945 MachineFunction &MF = DAG.getMachineFunction();
1946 bool Is64Bit = Subtarget->is64Bit();
1947 bool IsWin64 = Subtarget->isTargetWin64();
1948 bool IsStructRet = CallIsStructReturn(Outs);
1949 bool IsSibcall = false;
1952 // Check if it's really possible to do a tail call.
1953 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1954 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1955 Outs, OutVals, Ins, DAG);
1957 // Sibcalls are automatically detected tailcalls which do not require
1959 if (!GuaranteedTailCallOpt && isTailCall)
1966 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1967 "Var args not supported with calling convention fastcc or ghc");
1969 // Analyze operands of the call, assigning locations to each operand.
1970 SmallVector<CCValAssign, 16> ArgLocs;
1971 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1972 ArgLocs, *DAG.getContext());
1974 // Allocate shadow area for Win64
1976 CCInfo.AllocateStack(32, 8);
1979 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
1981 // Get a count of how many bytes are to be pushed on the stack.
1982 unsigned NumBytes = CCInfo.getNextStackOffset();
1984 // This is a sibcall. The memory operands are available in caller's
1985 // own caller's stack.
1987 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1988 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1991 if (isTailCall && !IsSibcall) {
1992 // Lower arguments at fp - stackoffset + fpdiff.
1993 unsigned NumBytesCallerPushed =
1994 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1995 FPDiff = NumBytesCallerPushed - NumBytes;
1997 // Set the delta of movement of the returnaddr stackslot.
1998 // But only set if delta is greater than previous delta.
1999 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
2000 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
2004 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2006 SDValue RetAddrFrIdx;
2007 // Load return address for tail calls.
2008 if (isTailCall && FPDiff)
2009 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2010 Is64Bit, FPDiff, dl);
2012 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2013 SmallVector<SDValue, 8> MemOpChains;
2016 // Walk the register/memloc assignments, inserting copies/loads. In the case
2017 // of tail call optimization arguments are handle later.
2018 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2019 CCValAssign &VA = ArgLocs[i];
2020 EVT RegVT = VA.getLocVT();
2021 SDValue Arg = OutVals[i];
2022 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2023 bool isByVal = Flags.isByVal();
2025 // Promote the value if needed.
2026 switch (VA.getLocInfo()) {
2027 default: llvm_unreachable("Unknown loc info!");
2028 case CCValAssign::Full: break;
2029 case CCValAssign::SExt:
2030 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2032 case CCValAssign::ZExt:
2033 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2035 case CCValAssign::AExt:
2036 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
2037 // Special case: passing MMX values in XMM registers.
2038 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2039 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2040 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2042 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2044 case CCValAssign::BCvt:
2045 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2047 case CCValAssign::Indirect: {
2048 // Store the argument.
2049 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2050 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2051 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2052 MachinePointerInfo::getFixedStack(FI),
2059 if (VA.isRegLoc()) {
2060 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2061 if (isVarArg && IsWin64) {
2062 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2063 // shadow reg if callee is a varargs function.
2064 unsigned ShadowReg = 0;
2065 switch (VA.getLocReg()) {
2066 case X86::XMM0: ShadowReg = X86::RCX; break;
2067 case X86::XMM1: ShadowReg = X86::RDX; break;
2068 case X86::XMM2: ShadowReg = X86::R8; break;
2069 case X86::XMM3: ShadowReg = X86::R9; break;
2072 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2074 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2075 assert(VA.isMemLoc());
2076 if (StackPtr.getNode() == 0)
2077 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
2078 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2079 dl, DAG, VA, Flags));
2083 if (!MemOpChains.empty())
2084 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2085 &MemOpChains[0], MemOpChains.size());
2087 // Build a sequence of copy-to-reg nodes chained together with token chain
2088 // and flag operands which copy the outgoing args into registers.
2090 // Tail call byval lowering might overwrite argument registers so in case of
2091 // tail call optimization the copies to registers are lowered later.
2093 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2094 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2095 RegsToPass[i].second, InFlag);
2096 InFlag = Chain.getValue(1);
2099 if (Subtarget->isPICStyleGOT()) {
2100 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2103 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
2104 DAG.getNode(X86ISD::GlobalBaseReg,
2105 DebugLoc(), getPointerTy()),
2107 InFlag = Chain.getValue(1);
2109 // If we are tail calling and generating PIC/GOT style code load the
2110 // address of the callee into ECX. The value in ecx is used as target of
2111 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2112 // for tail calls on PIC/GOT architectures. Normally we would just put the
2113 // address of GOT into ebx and then call target@PLT. But for tail calls
2114 // ebx would be restored (since ebx is callee saved) before jumping to the
2117 // Note: The actual moving to ECX is done further down.
2118 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2119 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2120 !G->getGlobal()->hasProtectedVisibility())
2121 Callee = LowerGlobalAddress(Callee, DAG);
2122 else if (isa<ExternalSymbolSDNode>(Callee))
2123 Callee = LowerExternalSymbol(Callee, DAG);
2127 if (Is64Bit && isVarArg && !IsWin64) {
2128 // From AMD64 ABI document:
2129 // For calls that may call functions that use varargs or stdargs
2130 // (prototype-less calls or calls to functions containing ellipsis (...) in
2131 // the declaration) %al is used as hidden argument to specify the number
2132 // of SSE registers used. The contents of %al do not need to match exactly
2133 // the number of registers, but must be an ubound on the number of SSE
2134 // registers used and is in the range 0 - 8 inclusive.
2136 // Count the number of XMM registers allocated.
2137 static const unsigned XMMArgRegs[] = {
2138 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2139 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2141 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2142 assert((Subtarget->hasXMM() || !NumXMMRegs)
2143 && "SSE registers cannot be used when SSE is disabled");
2145 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2146 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2147 InFlag = Chain.getValue(1);
2151 // For tail calls lower the arguments to the 'real' stack slot.
2153 // Force all the incoming stack arguments to be loaded from the stack
2154 // before any new outgoing arguments are stored to the stack, because the
2155 // outgoing stack slots may alias the incoming argument stack slots, and
2156 // the alias isn't otherwise explicit. This is slightly more conservative
2157 // than necessary, because it means that each store effectively depends
2158 // on every argument instead of just those arguments it would clobber.
2159 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2161 SmallVector<SDValue, 8> MemOpChains2;
2164 // Do not flag preceding copytoreg stuff together with the following stuff.
2166 if (GuaranteedTailCallOpt) {
2167 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2168 CCValAssign &VA = ArgLocs[i];
2171 assert(VA.isMemLoc());
2172 SDValue Arg = OutVals[i];
2173 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2174 // Create frame index.
2175 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2176 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2177 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2178 FIN = DAG.getFrameIndex(FI, getPointerTy());
2180 if (Flags.isByVal()) {
2181 // Copy relative to framepointer.
2182 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2183 if (StackPtr.getNode() == 0)
2184 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2186 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2188 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2192 // Store relative to framepointer.
2193 MemOpChains2.push_back(
2194 DAG.getStore(ArgChain, dl, Arg, FIN,
2195 MachinePointerInfo::getFixedStack(FI),
2201 if (!MemOpChains2.empty())
2202 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2203 &MemOpChains2[0], MemOpChains2.size());
2205 // Copy arguments to their registers.
2206 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2207 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2208 RegsToPass[i].second, InFlag);
2209 InFlag = Chain.getValue(1);
2213 // Store the return address to the appropriate stack slot.
2214 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2218 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2219 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2220 // In the 64-bit large code model, we have to make all calls
2221 // through a register, since the call instruction's 32-bit
2222 // pc-relative offset may not be large enough to hold the whole
2224 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2225 // If the callee is a GlobalAddress node (quite common, every direct call
2226 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2229 // We should use extra load for direct calls to dllimported functions in
2231 const GlobalValue *GV = G->getGlobal();
2232 if (!GV->hasDLLImportLinkage()) {
2233 unsigned char OpFlags = 0;
2234 bool ExtraLoad = false;
2235 unsigned WrapperKind = ISD::DELETED_NODE;
2237 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2238 // external symbols most go through the PLT in PIC mode. If the symbol
2239 // has hidden or protected visibility, or if it is static or local, then
2240 // we don't need to use the PLT - we can directly call it.
2241 if (Subtarget->isTargetELF() &&
2242 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2243 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2244 OpFlags = X86II::MO_PLT;
2245 } else if (Subtarget->isPICStyleStubAny() &&
2246 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2247 (!Subtarget->getTargetTriple().isMacOSX() ||
2248 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2249 // PC-relative references to external symbols should go through $stub,
2250 // unless we're building with the leopard linker or later, which
2251 // automatically synthesizes these stubs.
2252 OpFlags = X86II::MO_DARWIN_STUB;
2253 } else if (Subtarget->isPICStyleRIPRel() &&
2254 isa<Function>(GV) &&
2255 cast<Function>(GV)->hasFnAttr(Attribute::NonLazyBind)) {
2256 // If the function is marked as non-lazy, generate an indirect call
2257 // which loads from the GOT directly. This avoids runtime overhead
2258 // at the cost of eager binding (and one extra byte of encoding).
2259 OpFlags = X86II::MO_GOTPCREL;
2260 WrapperKind = X86ISD::WrapperRIP;
2264 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2265 G->getOffset(), OpFlags);
2267 // Add a wrapper if needed.
2268 if (WrapperKind != ISD::DELETED_NODE)
2269 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
2270 // Add extra indirection if needed.
2272 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
2273 MachinePointerInfo::getGOT(),
2276 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2277 unsigned char OpFlags = 0;
2279 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
2280 // external symbols should go through the PLT.
2281 if (Subtarget->isTargetELF() &&
2282 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2283 OpFlags = X86II::MO_PLT;
2284 } else if (Subtarget->isPICStyleStubAny() &&
2285 (!Subtarget->getTargetTriple().isMacOSX() ||
2286 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2287 // PC-relative references to external symbols should go through $stub,
2288 // unless we're building with the leopard linker or later, which
2289 // automatically synthesizes these stubs.
2290 OpFlags = X86II::MO_DARWIN_STUB;
2293 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2297 // Returns a chain & a flag for retval copy to use.
2298 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2299 SmallVector<SDValue, 8> Ops;
2301 if (!IsSibcall && isTailCall) {
2302 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2303 DAG.getIntPtrConstant(0, true), InFlag);
2304 InFlag = Chain.getValue(1);
2307 Ops.push_back(Chain);
2308 Ops.push_back(Callee);
2311 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2313 // Add argument registers to the end of the list so that they are known live
2315 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2316 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2317 RegsToPass[i].second.getValueType()));
2319 // Add an implicit use GOT pointer in EBX.
2320 if (!isTailCall && Subtarget->isPICStyleGOT())
2321 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2323 // Add an implicit use of AL for non-Windows x86 64-bit vararg functions.
2324 if (Is64Bit && isVarArg && !IsWin64)
2325 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2327 if (InFlag.getNode())
2328 Ops.push_back(InFlag);
2332 //// If this is the first return lowered for this function, add the regs
2333 //// to the liveout set for the function.
2334 // This isn't right, although it's probably harmless on x86; liveouts
2335 // should be computed from returns not tail calls. Consider a void
2336 // function making a tail call to a function returning int.
2337 return DAG.getNode(X86ISD::TC_RETURN, dl,
2338 NodeTys, &Ops[0], Ops.size());
2341 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2342 InFlag = Chain.getValue(1);
2344 // Create the CALLSEQ_END node.
2345 unsigned NumBytesForCalleeToPush;
2346 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, GuaranteedTailCallOpt))
2347 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2348 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2349 // If this is a call to a struct-return function, the callee
2350 // pops the hidden struct pointer, so we have to push it back.
2351 // This is common for Darwin/X86, Linux & Mingw32 targets.
2352 NumBytesForCalleeToPush = 4;
2354 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2356 // Returns a flag for retval copy to use.
2358 Chain = DAG.getCALLSEQ_END(Chain,
2359 DAG.getIntPtrConstant(NumBytes, true),
2360 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2363 InFlag = Chain.getValue(1);
2366 // Handle result values, copying them out of physregs into vregs that we
2368 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2369 Ins, dl, DAG, InVals);
2373 //===----------------------------------------------------------------------===//
2374 // Fast Calling Convention (tail call) implementation
2375 //===----------------------------------------------------------------------===//
2377 // Like std call, callee cleans arguments, convention except that ECX is
2378 // reserved for storing the tail called function address. Only 2 registers are
2379 // free for argument passing (inreg). Tail call optimization is performed
2381 // * tailcallopt is enabled
2382 // * caller/callee are fastcc
2383 // On X86_64 architecture with GOT-style position independent code only local
2384 // (within module) calls are supported at the moment.
2385 // To keep the stack aligned according to platform abi the function
2386 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2387 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2388 // If a tail called function callee has more arguments than the caller the
2389 // caller needs to make sure that there is room to move the RETADDR to. This is
2390 // achieved by reserving an area the size of the argument delta right after the
2391 // original REtADDR, but before the saved framepointer or the spilled registers
2392 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2404 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2405 /// for a 16 byte align requirement.
2407 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2408 SelectionDAG& DAG) const {
2409 MachineFunction &MF = DAG.getMachineFunction();
2410 const TargetMachine &TM = MF.getTarget();
2411 const TargetFrameLowering &TFI = *TM.getFrameLowering();
2412 unsigned StackAlignment = TFI.getStackAlignment();
2413 uint64_t AlignMask = StackAlignment - 1;
2414 int64_t Offset = StackSize;
2415 uint64_t SlotSize = TD->getPointerSize();
2416 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2417 // Number smaller than 12 so just add the difference.
2418 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2420 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2421 Offset = ((~AlignMask) & Offset) + StackAlignment +
2422 (StackAlignment-SlotSize);
2427 /// MatchingStackOffset - Return true if the given stack call argument is
2428 /// already available in the same position (relatively) of the caller's
2429 /// incoming argument stack.
2431 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2432 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2433 const X86InstrInfo *TII) {
2434 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2436 if (Arg.getOpcode() == ISD::CopyFromReg) {
2437 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2438 if (!TargetRegisterInfo::isVirtualRegister(VR))
2440 MachineInstr *Def = MRI->getVRegDef(VR);
2443 if (!Flags.isByVal()) {
2444 if (!TII->isLoadFromStackSlot(Def, FI))
2447 unsigned Opcode = Def->getOpcode();
2448 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2449 Def->getOperand(1).isFI()) {
2450 FI = Def->getOperand(1).getIndex();
2451 Bytes = Flags.getByValSize();
2455 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2456 if (Flags.isByVal())
2457 // ByVal argument is passed in as a pointer but it's now being
2458 // dereferenced. e.g.
2459 // define @foo(%struct.X* %A) {
2460 // tail call @bar(%struct.X* byval %A)
2463 SDValue Ptr = Ld->getBasePtr();
2464 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2467 FI = FINode->getIndex();
2468 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
2469 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
2470 FI = FINode->getIndex();
2471 Bytes = Flags.getByValSize();
2475 assert(FI != INT_MAX);
2476 if (!MFI->isFixedObjectIndex(FI))
2478 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2481 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2482 /// for tail call optimization. Targets which want to do tail call
2483 /// optimization should implement this function.
2485 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2486 CallingConv::ID CalleeCC,
2488 bool isCalleeStructRet,
2489 bool isCallerStructRet,
2490 const SmallVectorImpl<ISD::OutputArg> &Outs,
2491 const SmallVectorImpl<SDValue> &OutVals,
2492 const SmallVectorImpl<ISD::InputArg> &Ins,
2493 SelectionDAG& DAG) const {
2494 if (!IsTailCallConvention(CalleeCC) &&
2495 CalleeCC != CallingConv::C)
2498 // If -tailcallopt is specified, make fastcc functions tail-callable.
2499 const MachineFunction &MF = DAG.getMachineFunction();
2500 const Function *CallerF = DAG.getMachineFunction().getFunction();
2501 CallingConv::ID CallerCC = CallerF->getCallingConv();
2502 bool CCMatch = CallerCC == CalleeCC;
2504 if (GuaranteedTailCallOpt) {
2505 if (IsTailCallConvention(CalleeCC) && CCMatch)
2510 // Look for obvious safe cases to perform tail call optimization that do not
2511 // require ABI changes. This is what gcc calls sibcall.
2513 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2514 // emit a special epilogue.
2515 if (RegInfo->needsStackRealignment(MF))
2518 // Also avoid sibcall optimization if either caller or callee uses struct
2519 // return semantics.
2520 if (isCalleeStructRet || isCallerStructRet)
2523 // An stdcall caller is expected to clean up its arguments; the callee
2524 // isn't going to do that.
2525 if (!CCMatch && CallerCC==CallingConv::X86_StdCall)
2528 // Do not sibcall optimize vararg calls unless all arguments are passed via
2530 if (isVarArg && !Outs.empty()) {
2532 // Optimizing for varargs on Win64 is unlikely to be safe without
2533 // additional testing.
2534 if (Subtarget->isTargetWin64())
2537 SmallVector<CCValAssign, 16> ArgLocs;
2538 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2539 getTargetMachine(), ArgLocs, *DAG.getContext());
2541 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2542 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
2543 if (!ArgLocs[i].isRegLoc())
2547 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2548 // Therefore if it's not used by the call it is not safe to optimize this into
2550 bool Unused = false;
2551 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2558 SmallVector<CCValAssign, 16> RVLocs;
2559 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(),
2560 getTargetMachine(), RVLocs, *DAG.getContext());
2561 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2562 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2563 CCValAssign &VA = RVLocs[i];
2564 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2569 // If the calling conventions do not match, then we'd better make sure the
2570 // results are returned in the same way as what the caller expects.
2572 SmallVector<CCValAssign, 16> RVLocs1;
2573 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
2574 getTargetMachine(), RVLocs1, *DAG.getContext());
2575 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2577 SmallVector<CCValAssign, 16> RVLocs2;
2578 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
2579 getTargetMachine(), RVLocs2, *DAG.getContext());
2580 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2582 if (RVLocs1.size() != RVLocs2.size())
2584 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2585 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2587 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2589 if (RVLocs1[i].isRegLoc()) {
2590 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2593 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2599 // If the callee takes no arguments then go on to check the results of the
2601 if (!Outs.empty()) {
2602 // Check if stack adjustment is needed. For now, do not do this if any
2603 // argument is passed on the stack.
2604 SmallVector<CCValAssign, 16> ArgLocs;
2605 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2606 getTargetMachine(), ArgLocs, *DAG.getContext());
2608 // Allocate shadow area for Win64
2609 if (Subtarget->isTargetWin64()) {
2610 CCInfo.AllocateStack(32, 8);
2613 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2614 if (CCInfo.getNextStackOffset()) {
2615 MachineFunction &MF = DAG.getMachineFunction();
2616 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2619 // Check if the arguments are already laid out in the right way as
2620 // the caller's fixed stack objects.
2621 MachineFrameInfo *MFI = MF.getFrameInfo();
2622 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2623 const X86InstrInfo *TII =
2624 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2625 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2626 CCValAssign &VA = ArgLocs[i];
2627 SDValue Arg = OutVals[i];
2628 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2629 if (VA.getLocInfo() == CCValAssign::Indirect)
2631 if (!VA.isRegLoc()) {
2632 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2639 // If the tailcall address may be in a register, then make sure it's
2640 // possible to register allocate for it. In 32-bit, the call address can
2641 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2642 // callee-saved registers are restored. These happen to be the same
2643 // registers used to pass 'inreg' arguments so watch out for those.
2644 if (!Subtarget->is64Bit() &&
2645 !isa<GlobalAddressSDNode>(Callee) &&
2646 !isa<ExternalSymbolSDNode>(Callee)) {
2647 unsigned NumInRegs = 0;
2648 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2649 CCValAssign &VA = ArgLocs[i];
2652 unsigned Reg = VA.getLocReg();
2655 case X86::EAX: case X86::EDX: case X86::ECX:
2656 if (++NumInRegs == 3)
2668 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2669 return X86::createFastISel(funcInfo);
2673 //===----------------------------------------------------------------------===//
2674 // Other Lowering Hooks
2675 //===----------------------------------------------------------------------===//
2677 static bool MayFoldLoad(SDValue Op) {
2678 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
2681 static bool MayFoldIntoStore(SDValue Op) {
2682 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
2685 static bool isTargetShuffle(unsigned Opcode) {
2687 default: return false;
2688 case X86ISD::PSHUFD:
2689 case X86ISD::PSHUFHW:
2690 case X86ISD::PSHUFLW:
2691 case X86ISD::SHUFPD:
2692 case X86ISD::PALIGN:
2693 case X86ISD::SHUFPS:
2694 case X86ISD::MOVLHPS:
2695 case X86ISD::MOVLHPD:
2696 case X86ISD::MOVHLPS:
2697 case X86ISD::MOVLPS:
2698 case X86ISD::MOVLPD:
2699 case X86ISD::MOVSHDUP:
2700 case X86ISD::MOVSLDUP:
2701 case X86ISD::MOVDDUP:
2704 case X86ISD::UNPCKLPS:
2705 case X86ISD::UNPCKLPD:
2706 case X86ISD::VUNPCKLPSY:
2707 case X86ISD::VUNPCKLPDY:
2708 case X86ISD::PUNPCKLWD:
2709 case X86ISD::PUNPCKLBW:
2710 case X86ISD::PUNPCKLDQ:
2711 case X86ISD::PUNPCKLQDQ:
2712 case X86ISD::UNPCKHPS:
2713 case X86ISD::UNPCKHPD:
2714 case X86ISD::PUNPCKHWD:
2715 case X86ISD::PUNPCKHBW:
2716 case X86ISD::PUNPCKHDQ:
2717 case X86ISD::PUNPCKHQDQ:
2718 case X86ISD::VPERMIL:
2724 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2725 SDValue V1, SelectionDAG &DAG) {
2727 default: llvm_unreachable("Unknown x86 shuffle node");
2728 case X86ISD::MOVSHDUP:
2729 case X86ISD::MOVSLDUP:
2730 case X86ISD::MOVDDUP:
2731 return DAG.getNode(Opc, dl, VT, V1);
2737 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2738 SDValue V1, unsigned TargetMask, SelectionDAG &DAG) {
2740 default: llvm_unreachable("Unknown x86 shuffle node");
2741 case X86ISD::PSHUFD:
2742 case X86ISD::PSHUFHW:
2743 case X86ISD::PSHUFLW:
2744 case X86ISD::VPERMIL:
2745 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
2751 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2752 SDValue V1, SDValue V2, unsigned TargetMask, SelectionDAG &DAG) {
2754 default: llvm_unreachable("Unknown x86 shuffle node");
2755 case X86ISD::PALIGN:
2756 case X86ISD::SHUFPD:
2757 case X86ISD::SHUFPS:
2758 return DAG.getNode(Opc, dl, VT, V1, V2,
2759 DAG.getConstant(TargetMask, MVT::i8));
2764 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2765 SDValue V1, SDValue V2, SelectionDAG &DAG) {
2767 default: llvm_unreachable("Unknown x86 shuffle node");
2768 case X86ISD::MOVLHPS:
2769 case X86ISD::MOVLHPD:
2770 case X86ISD::MOVHLPS:
2771 case X86ISD::MOVLPS:
2772 case X86ISD::MOVLPD:
2775 case X86ISD::UNPCKLPS:
2776 case X86ISD::UNPCKLPD:
2777 case X86ISD::VUNPCKLPSY:
2778 case X86ISD::VUNPCKLPDY:
2779 case X86ISD::PUNPCKLWD:
2780 case X86ISD::PUNPCKLBW:
2781 case X86ISD::PUNPCKLDQ:
2782 case X86ISD::PUNPCKLQDQ:
2783 case X86ISD::UNPCKHPS:
2784 case X86ISD::UNPCKHPD:
2785 case X86ISD::PUNPCKHWD:
2786 case X86ISD::PUNPCKHBW:
2787 case X86ISD::PUNPCKHDQ:
2788 case X86ISD::PUNPCKHQDQ:
2789 return DAG.getNode(Opc, dl, VT, V1, V2);
2794 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2795 MachineFunction &MF = DAG.getMachineFunction();
2796 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2797 int ReturnAddrIndex = FuncInfo->getRAIndex();
2799 if (ReturnAddrIndex == 0) {
2800 // Set up a frame object for the return address.
2801 uint64_t SlotSize = TD->getPointerSize();
2802 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2804 FuncInfo->setRAIndex(ReturnAddrIndex);
2807 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2811 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2812 bool hasSymbolicDisplacement) {
2813 // Offset should fit into 32 bit immediate field.
2814 if (!isInt<32>(Offset))
2817 // If we don't have a symbolic displacement - we don't have any extra
2819 if (!hasSymbolicDisplacement)
2822 // FIXME: Some tweaks might be needed for medium code model.
2823 if (M != CodeModel::Small && M != CodeModel::Kernel)
2826 // For small code model we assume that latest object is 16MB before end of 31
2827 // bits boundary. We may also accept pretty large negative constants knowing
2828 // that all objects are in the positive half of address space.
2829 if (M == CodeModel::Small && Offset < 16*1024*1024)
2832 // For kernel code model we know that all object resist in the negative half
2833 // of 32bits address space. We may not accept negative offsets, since they may
2834 // be just off and we may accept pretty large positive ones.
2835 if (M == CodeModel::Kernel && Offset > 0)
2841 /// isCalleePop - Determines whether the callee is required to pop its
2842 /// own arguments. Callee pop is necessary to support tail calls.
2843 bool X86::isCalleePop(CallingConv::ID CallingConv,
2844 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
2848 switch (CallingConv) {
2851 case CallingConv::X86_StdCall:
2853 case CallingConv::X86_FastCall:
2855 case CallingConv::X86_ThisCall:
2857 case CallingConv::Fast:
2859 case CallingConv::GHC:
2864 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2865 /// specific condition code, returning the condition code and the LHS/RHS of the
2866 /// comparison to make.
2867 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2868 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2870 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2871 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2872 // X > -1 -> X == 0, jump !sign.
2873 RHS = DAG.getConstant(0, RHS.getValueType());
2874 return X86::COND_NS;
2875 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2876 // X < 0 -> X == 0, jump on sign.
2878 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2880 RHS = DAG.getConstant(0, RHS.getValueType());
2881 return X86::COND_LE;
2885 switch (SetCCOpcode) {
2886 default: llvm_unreachable("Invalid integer condition!");
2887 case ISD::SETEQ: return X86::COND_E;
2888 case ISD::SETGT: return X86::COND_G;
2889 case ISD::SETGE: return X86::COND_GE;
2890 case ISD::SETLT: return X86::COND_L;
2891 case ISD::SETLE: return X86::COND_LE;
2892 case ISD::SETNE: return X86::COND_NE;
2893 case ISD::SETULT: return X86::COND_B;
2894 case ISD::SETUGT: return X86::COND_A;
2895 case ISD::SETULE: return X86::COND_BE;
2896 case ISD::SETUGE: return X86::COND_AE;
2900 // First determine if it is required or is profitable to flip the operands.
2902 // If LHS is a foldable load, but RHS is not, flip the condition.
2903 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
2904 !ISD::isNON_EXTLoad(RHS.getNode())) {
2905 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2906 std::swap(LHS, RHS);
2909 switch (SetCCOpcode) {
2915 std::swap(LHS, RHS);
2919 // On a floating point condition, the flags are set as follows:
2921 // 0 | 0 | 0 | X > Y
2922 // 0 | 0 | 1 | X < Y
2923 // 1 | 0 | 0 | X == Y
2924 // 1 | 1 | 1 | unordered
2925 switch (SetCCOpcode) {
2926 default: llvm_unreachable("Condcode should be pre-legalized away");
2928 case ISD::SETEQ: return X86::COND_E;
2929 case ISD::SETOLT: // flipped
2931 case ISD::SETGT: return X86::COND_A;
2932 case ISD::SETOLE: // flipped
2934 case ISD::SETGE: return X86::COND_AE;
2935 case ISD::SETUGT: // flipped
2937 case ISD::SETLT: return X86::COND_B;
2938 case ISD::SETUGE: // flipped
2940 case ISD::SETLE: return X86::COND_BE;
2942 case ISD::SETNE: return X86::COND_NE;
2943 case ISD::SETUO: return X86::COND_P;
2944 case ISD::SETO: return X86::COND_NP;
2946 case ISD::SETUNE: return X86::COND_INVALID;
2950 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2951 /// code. Current x86 isa includes the following FP cmov instructions:
2952 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2953 static bool hasFPCMov(unsigned X86CC) {
2969 /// isFPImmLegal - Returns true if the target can instruction select the
2970 /// specified FP immediate natively. If false, the legalizer will
2971 /// materialize the FP immediate as a load from a constant pool.
2972 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2973 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2974 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2980 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2981 /// the specified range (L, H].
2982 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2983 return (Val < 0) || (Val >= Low && Val < Hi);
2986 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2987 /// specified value.
2988 static bool isUndefOrEqual(int Val, int CmpVal) {
2989 if (Val < 0 || Val == CmpVal)
2994 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2995 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2996 /// the second operand.
2997 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2998 if (VT == MVT::v4f32 || VT == MVT::v4i32 )
2999 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
3000 if (VT == MVT::v2f64 || VT == MVT::v2i64)
3001 return (Mask[0] < 2 && Mask[1] < 2);
3005 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
3006 SmallVector<int, 8> M;
3008 return ::isPSHUFDMask(M, N->getValueType(0));
3011 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
3012 /// is suitable for input to PSHUFHW.
3013 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3014 if (VT != MVT::v8i16)
3017 // Lower quadword copied in order or undef.
3018 for (int i = 0; i != 4; ++i)
3019 if (Mask[i] >= 0 && Mask[i] != i)
3022 // Upper quadword shuffled.
3023 for (int i = 4; i != 8; ++i)
3024 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
3030 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
3031 SmallVector<int, 8> M;
3033 return ::isPSHUFHWMask(M, N->getValueType(0));
3036 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
3037 /// is suitable for input to PSHUFLW.
3038 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3039 if (VT != MVT::v8i16)
3042 // Upper quadword copied in order.
3043 for (int i = 4; i != 8; ++i)
3044 if (Mask[i] >= 0 && Mask[i] != i)
3047 // Lower quadword shuffled.
3048 for (int i = 0; i != 4; ++i)
3055 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
3056 SmallVector<int, 8> M;
3058 return ::isPSHUFLWMask(M, N->getValueType(0));
3061 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
3062 /// is suitable for input to PALIGNR.
3063 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
3065 int i, e = VT.getVectorNumElements();
3067 // Do not handle v2i64 / v2f64 shuffles with palignr.
3068 if (e < 4 || !hasSSSE3)
3071 for (i = 0; i != e; ++i)
3075 // All undef, not a palignr.
3079 // Make sure we're shifting in the right direction.
3083 int s = Mask[i] - i;
3085 // Check the rest of the elements to see if they are consecutive.
3086 for (++i; i != e; ++i) {
3088 if (m >= 0 && m != s+i)
3094 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
3095 SmallVector<int, 8> M;
3097 return ::isPALIGNRMask(M, N->getValueType(0), true);
3100 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
3101 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
3102 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3103 int NumElems = VT.getVectorNumElements();
3104 if (NumElems != 2 && NumElems != 4)
3107 int Half = NumElems / 2;
3108 for (int i = 0; i < Half; ++i)
3109 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3111 for (int i = Half; i < NumElems; ++i)
3112 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3118 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
3119 SmallVector<int, 8> M;
3121 return ::isSHUFPMask(M, N->getValueType(0));
3124 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
3125 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
3126 /// half elements to come from vector 1 (which would equal the dest.) and
3127 /// the upper half to come from vector 2.
3128 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3129 int NumElems = VT.getVectorNumElements();
3131 if (NumElems != 2 && NumElems != 4)
3134 int Half = NumElems / 2;
3135 for (int i = 0; i < Half; ++i)
3136 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3138 for (int i = Half; i < NumElems; ++i)
3139 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3144 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
3145 SmallVector<int, 8> M;
3147 return isCommutedSHUFPMask(M, N->getValueType(0));
3150 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3151 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3152 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
3153 if (N->getValueType(0).getVectorNumElements() != 4)
3156 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3157 return isUndefOrEqual(N->getMaskElt(0), 6) &&
3158 isUndefOrEqual(N->getMaskElt(1), 7) &&
3159 isUndefOrEqual(N->getMaskElt(2), 2) &&
3160 isUndefOrEqual(N->getMaskElt(3), 3);
3163 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3164 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3166 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
3167 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3172 return isUndefOrEqual(N->getMaskElt(0), 2) &&
3173 isUndefOrEqual(N->getMaskElt(1), 3) &&
3174 isUndefOrEqual(N->getMaskElt(2), 2) &&
3175 isUndefOrEqual(N->getMaskElt(3), 3);
3178 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3179 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3180 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
3181 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3183 if (NumElems != 2 && NumElems != 4)
3186 for (unsigned i = 0; i < NumElems/2; ++i)
3187 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
3190 for (unsigned i = NumElems/2; i < NumElems; ++i)
3191 if (!isUndefOrEqual(N->getMaskElt(i), i))
3197 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3198 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3199 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
3200 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3202 if ((NumElems != 2 && NumElems != 4)
3203 || N->getValueType(0).getSizeInBits() > 128)
3206 for (unsigned i = 0; i < NumElems/2; ++i)
3207 if (!isUndefOrEqual(N->getMaskElt(i), i))
3210 for (unsigned i = 0; i < NumElems/2; ++i)
3211 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
3217 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3218 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3219 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3220 bool V2IsSplat = false) {
3221 int NumElts = VT.getVectorNumElements();
3222 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3225 // Handle vector lengths > 128 bits. Define a "section" as a set of
3226 // 128 bits. AVX defines UNPCK* to operate independently on 128-bit
3228 unsigned NumSections = VT.getSizeInBits() / 128;
3229 if (NumSections == 0 ) NumSections = 1; // Handle MMX
3230 unsigned NumSectionElts = NumElts / NumSections;
3233 unsigned End = NumSectionElts;
3234 for (unsigned s = 0; s < NumSections; ++s) {
3235 for (unsigned i = Start, j = s * NumSectionElts;
3239 int BitI1 = Mask[i+1];
3240 if (!isUndefOrEqual(BitI, j))
3243 if (!isUndefOrEqual(BitI1, NumElts))
3246 if (!isUndefOrEqual(BitI1, j + NumElts))
3250 // Process the next 128 bits.
3251 Start += NumSectionElts;
3252 End += NumSectionElts;
3258 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3259 SmallVector<int, 8> M;
3261 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
3264 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3265 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3266 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3267 bool V2IsSplat = false) {
3268 int NumElts = VT.getVectorNumElements();
3269 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3272 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3274 int BitI1 = Mask[i+1];
3275 if (!isUndefOrEqual(BitI, j + NumElts/2))
3278 if (isUndefOrEqual(BitI1, NumElts))
3281 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
3288 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3289 SmallVector<int, 8> M;
3291 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3294 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3295 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3297 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3298 int NumElems = VT.getVectorNumElements();
3299 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3302 // Handle vector lengths > 128 bits. Define a "section" as a set of
3303 // 128 bits. AVX defines UNPCK* to operate independently on 128-bit
3305 unsigned NumSections = VT.getSizeInBits() / 128;
3306 if (NumSections == 0 ) NumSections = 1; // Handle MMX
3307 unsigned NumSectionElts = NumElems / NumSections;
3309 for (unsigned s = 0; s < NumSections; ++s) {
3310 for (unsigned i = s * NumSectionElts, j = s * NumSectionElts;
3311 i != NumSectionElts * (s + 1);
3314 int BitI1 = Mask[i+1];
3316 if (!isUndefOrEqual(BitI, j))
3318 if (!isUndefOrEqual(BitI1, j))
3326 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3327 SmallVector<int, 8> M;
3329 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3332 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3333 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3335 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3336 int NumElems = VT.getVectorNumElements();
3337 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3340 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3342 int BitI1 = Mask[i+1];
3343 if (!isUndefOrEqual(BitI, j))
3345 if (!isUndefOrEqual(BitI1, j))
3351 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3352 SmallVector<int, 8> M;
3354 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3357 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3358 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3359 /// MOVSD, and MOVD, i.e. setting the lowest element.
3360 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3361 if (VT.getVectorElementType().getSizeInBits() < 32)
3364 int NumElts = VT.getVectorNumElements();
3366 if (!isUndefOrEqual(Mask[0], NumElts))
3369 for (int i = 1; i < NumElts; ++i)
3370 if (!isUndefOrEqual(Mask[i], i))
3376 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3377 SmallVector<int, 8> M;
3379 return ::isMOVLMask(M, N->getValueType(0));
3382 /// isVPERMILMask - Return true if the specified VECTOR_SHUFFLE operand
3383 /// specifies a shuffle of elements that is suitable for input to VPERMIL*.
3384 static bool isVPERMILMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3385 unsigned NumElts = VT.getVectorNumElements();
3386 unsigned NumLanes = VT.getSizeInBits()/128;
3388 // Match any permutation of 128-bit vector with 32/64-bit types
3389 if (NumLanes == 1) {
3390 if (NumElts == 4 || NumElts == 2)
3395 // Only match 256-bit with 32/64-bit types
3396 if (NumElts != 8 && NumElts != 4)
3399 // The mask on the high lane should be the same as the low. Actually,
3400 // they can differ if any of the corresponding index in a lane is undef.
3401 int LaneSize = NumElts/NumLanes;
3402 for (int i = 0; i < LaneSize; ++i) {
3403 int HighElt = i+LaneSize;
3404 if (Mask[i] < 0 || Mask[HighElt] < 0)
3407 if (Mask[HighElt]-Mask[i] != LaneSize)
3414 /// getShuffleVPERMILImmediateediate - Return the appropriate immediate to shuffle
3415 /// the specified VECTOR_MASK mask with VPERMIL* instructions.
3416 static unsigned getShuffleVPERMILImmediate(SDNode *N) {
3417 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3418 EVT VT = SVOp->getValueType(0);
3420 int NumElts = VT.getVectorNumElements();
3421 int NumLanes = VT.getSizeInBits()/128;
3424 for (int i = 0; i < NumElts/NumLanes /* lane size */; ++i)
3425 Mask |= SVOp->getMaskElt(i) << (i*2);
3430 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3431 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3432 /// element of vector 2 and the other elements to come from vector 1 in order.
3433 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3434 bool V2IsSplat = false, bool V2IsUndef = false) {
3435 int NumOps = VT.getVectorNumElements();
3436 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3439 if (!isUndefOrEqual(Mask[0], 0))
3442 for (int i = 1; i < NumOps; ++i)
3443 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3444 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3445 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3451 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3452 bool V2IsUndef = false) {
3453 SmallVector<int, 8> M;
3455 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3458 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3459 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3460 /// Masks to match: <1, 1, 3, 3> or <1, 1, 3, 3, 5, 5, 7, 7>
3461 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N,
3462 const X86Subtarget *Subtarget) {
3463 if (!Subtarget->hasSSE3() && !Subtarget->hasAVX())
3466 // The second vector must be undef
3467 if (N->getOperand(1).getOpcode() != ISD::UNDEF)
3470 EVT VT = N->getValueType(0);
3471 unsigned NumElems = VT.getVectorNumElements();
3473 if ((VT.getSizeInBits() == 128 && NumElems != 4) ||
3474 (VT.getSizeInBits() == 256 && NumElems != 8))
3477 // "i+1" is the value the indexed mask element must have
3478 for (unsigned i = 0; i < NumElems; i += 2)
3479 if (!isUndefOrEqual(N->getMaskElt(i), i+1) ||
3480 !isUndefOrEqual(N->getMaskElt(i+1), i+1))
3486 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3487 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3488 /// Masks to match: <0, 0, 2, 2> or <0, 0, 2, 2, 4, 4, 6, 6>
3489 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N,
3490 const X86Subtarget *Subtarget) {
3491 if (!Subtarget->hasSSE3() && !Subtarget->hasAVX())
3494 // The second vector must be undef
3495 if (N->getOperand(1).getOpcode() != ISD::UNDEF)
3498 EVT VT = N->getValueType(0);
3499 unsigned NumElems = VT.getVectorNumElements();
3501 if ((VT.getSizeInBits() == 128 && NumElems != 4) ||
3502 (VT.getSizeInBits() == 256 && NumElems != 8))
3505 // "i" is the value the indexed mask element must have
3506 for (unsigned i = 0; i < NumElems; i += 2)
3507 if (!isUndefOrEqual(N->getMaskElt(i), i) ||
3508 !isUndefOrEqual(N->getMaskElt(i+1), i))
3514 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3515 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3516 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3517 int e = N->getValueType(0).getVectorNumElements() / 2;
3519 for (int i = 0; i < e; ++i)
3520 if (!isUndefOrEqual(N->getMaskElt(i), i))
3522 for (int i = 0; i < e; ++i)
3523 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3528 /// isVEXTRACTF128Index - Return true if the specified
3529 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
3530 /// suitable for input to VEXTRACTF128.
3531 bool X86::isVEXTRACTF128Index(SDNode *N) {
3532 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3535 // The index should be aligned on a 128-bit boundary.
3537 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3539 unsigned VL = N->getValueType(0).getVectorNumElements();
3540 unsigned VBits = N->getValueType(0).getSizeInBits();
3541 unsigned ElSize = VBits / VL;
3542 bool Result = (Index * ElSize) % 128 == 0;
3547 /// isVINSERTF128Index - Return true if the specified INSERT_SUBVECTOR
3548 /// operand specifies a subvector insert that is suitable for input to
3550 bool X86::isVINSERTF128Index(SDNode *N) {
3551 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3554 // The index should be aligned on a 128-bit boundary.
3556 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3558 unsigned VL = N->getValueType(0).getVectorNumElements();
3559 unsigned VBits = N->getValueType(0).getSizeInBits();
3560 unsigned ElSize = VBits / VL;
3561 bool Result = (Index * ElSize) % 128 == 0;
3566 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3567 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3568 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3569 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3570 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3572 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3574 for (int i = 0; i < NumOperands; ++i) {
3575 int Val = SVOp->getMaskElt(NumOperands-i-1);
3576 if (Val < 0) Val = 0;
3577 if (Val >= NumOperands) Val -= NumOperands;
3579 if (i != NumOperands - 1)
3585 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3586 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3587 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3588 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3590 // 8 nodes, but we only care about the last 4.
3591 for (unsigned i = 7; i >= 4; --i) {
3592 int Val = SVOp->getMaskElt(i);
3601 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3602 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3603 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3604 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3606 // 8 nodes, but we only care about the first 4.
3607 for (int i = 3; i >= 0; --i) {
3608 int Val = SVOp->getMaskElt(i);
3617 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3618 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3619 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3620 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3621 EVT VVT = N->getValueType(0);
3622 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3626 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3627 Val = SVOp->getMaskElt(i);
3631 assert(Val - i > 0 && "PALIGNR imm should be positive");
3632 return (Val - i) * EltSize;
3635 /// getExtractVEXTRACTF128Immediate - Return the appropriate immediate
3636 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
3638 unsigned X86::getExtractVEXTRACTF128Immediate(SDNode *N) {
3639 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3640 llvm_unreachable("Illegal extract subvector for VEXTRACTF128");
3643 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3645 EVT VecVT = N->getOperand(0).getValueType();
3646 EVT ElVT = VecVT.getVectorElementType();
3648 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
3649 return Index / NumElemsPerChunk;
3652 /// getInsertVINSERTF128Immediate - Return the appropriate immediate
3653 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
3655 unsigned X86::getInsertVINSERTF128Immediate(SDNode *N) {
3656 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3657 llvm_unreachable("Illegal insert subvector for VINSERTF128");
3660 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3662 EVT VecVT = N->getValueType(0);
3663 EVT ElVT = VecVT.getVectorElementType();
3665 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
3666 return Index / NumElemsPerChunk;
3669 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3671 bool X86::isZeroNode(SDValue Elt) {
3672 return ((isa<ConstantSDNode>(Elt) &&
3673 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3674 (isa<ConstantFPSDNode>(Elt) &&
3675 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3678 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3679 /// their permute mask.
3680 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3681 SelectionDAG &DAG) {
3682 EVT VT = SVOp->getValueType(0);
3683 unsigned NumElems = VT.getVectorNumElements();
3684 SmallVector<int, 8> MaskVec;
3686 for (unsigned i = 0; i != NumElems; ++i) {
3687 int idx = SVOp->getMaskElt(i);
3689 MaskVec.push_back(idx);
3690 else if (idx < (int)NumElems)
3691 MaskVec.push_back(idx + NumElems);
3693 MaskVec.push_back(idx - NumElems);
3695 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3696 SVOp->getOperand(0), &MaskVec[0]);
3699 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3700 /// the two vector operands have swapped position.
3701 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3702 unsigned NumElems = VT.getVectorNumElements();
3703 for (unsigned i = 0; i != NumElems; ++i) {
3707 else if (idx < (int)NumElems)
3708 Mask[i] = idx + NumElems;
3710 Mask[i] = idx - NumElems;
3714 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3715 /// match movhlps. The lower half elements should come from upper half of
3716 /// V1 (and in order), and the upper half elements should come from the upper
3717 /// half of V2 (and in order).
3718 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3719 if (Op->getValueType(0).getVectorNumElements() != 4)
3721 for (unsigned i = 0, e = 2; i != e; ++i)
3722 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3724 for (unsigned i = 2; i != 4; ++i)
3725 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3730 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3731 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3733 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3734 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3736 N = N->getOperand(0).getNode();
3737 if (!ISD::isNON_EXTLoad(N))
3740 *LD = cast<LoadSDNode>(N);
3744 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3745 /// match movlp{s|d}. The lower half elements should come from lower half of
3746 /// V1 (and in order), and the upper half elements should come from the upper
3747 /// half of V2 (and in order). And since V1 will become the source of the
3748 /// MOVLP, it must be either a vector load or a scalar load to vector.
3749 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3750 ShuffleVectorSDNode *Op) {
3751 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3753 // Is V2 is a vector load, don't do this transformation. We will try to use
3754 // load folding shufps op.
3755 if (ISD::isNON_EXTLoad(V2))
3758 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3760 if (NumElems != 2 && NumElems != 4)
3762 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3763 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3765 for (unsigned i = NumElems/2; i != NumElems; ++i)
3766 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3771 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3773 static bool isSplatVector(SDNode *N) {
3774 if (N->getOpcode() != ISD::BUILD_VECTOR)
3777 SDValue SplatValue = N->getOperand(0);
3778 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3779 if (N->getOperand(i) != SplatValue)
3784 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3785 /// to an zero vector.
3786 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3787 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3788 SDValue V1 = N->getOperand(0);
3789 SDValue V2 = N->getOperand(1);
3790 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3791 for (unsigned i = 0; i != NumElems; ++i) {
3792 int Idx = N->getMaskElt(i);
3793 if (Idx >= (int)NumElems) {
3794 unsigned Opc = V2.getOpcode();
3795 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3797 if (Opc != ISD::BUILD_VECTOR ||
3798 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3800 } else if (Idx >= 0) {
3801 unsigned Opc = V1.getOpcode();
3802 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3804 if (Opc != ISD::BUILD_VECTOR ||
3805 !X86::isZeroNode(V1.getOperand(Idx)))
3812 /// getZeroVector - Returns a vector of specified type with all zero elements.
3814 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3816 assert(VT.isVector() && "Expected a vector type");
3818 // Always build SSE zero vectors as <4 x i32> bitcasted
3819 // to their dest type. This ensures they get CSE'd.
3821 if (VT.getSizeInBits() == 128) { // SSE
3822 if (HasSSE2) { // SSE2
3823 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3824 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3826 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3827 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3829 } else if (VT.getSizeInBits() == 256) { // AVX
3830 // 256-bit logic and arithmetic instructions in AVX are
3831 // all floating-point, no support for integer ops. Default
3832 // to emitting fp zeroed vectors then.
3833 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3834 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3835 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
3837 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
3840 /// getOnesVector - Returns a vector of specified type with all bits set.
3841 /// Always build ones vectors as <4 x i32>. For 256-bit types, use two
3842 /// <4 x i32> inserted in a <8 x i32> appropriately. Then bitcast to their
3843 /// original type, ensuring they get CSE'd.
3844 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3845 assert(VT.isVector() && "Expected a vector type");
3846 assert((VT.is128BitVector() || VT.is256BitVector())
3847 && "Expected a 128-bit or 256-bit vector type");
3849 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3850 SDValue Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
3851 Cst, Cst, Cst, Cst);
3853 if (VT.is256BitVector()) {
3854 SDValue InsV = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, MVT::v8i32),
3855 Vec, DAG.getConstant(0, MVT::i32), DAG, dl);
3856 Vec = Insert128BitVector(InsV, Vec,
3857 DAG.getConstant(4 /* NumElems/2 */, MVT::i32), DAG, dl);
3860 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
3863 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3864 /// that point to V2 points to its first element.
3865 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3866 EVT VT = SVOp->getValueType(0);
3867 unsigned NumElems = VT.getVectorNumElements();
3869 bool Changed = false;
3870 SmallVector<int, 8> MaskVec;
3871 SVOp->getMask(MaskVec);
3873 for (unsigned i = 0; i != NumElems; ++i) {
3874 if (MaskVec[i] > (int)NumElems) {
3875 MaskVec[i] = NumElems;
3880 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3881 SVOp->getOperand(1), &MaskVec[0]);
3882 return SDValue(SVOp, 0);
3885 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3886 /// operation of specified width.
3887 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3889 unsigned NumElems = VT.getVectorNumElements();
3890 SmallVector<int, 8> Mask;
3891 Mask.push_back(NumElems);
3892 for (unsigned i = 1; i != NumElems; ++i)
3894 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3897 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3898 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3900 unsigned NumElems = VT.getVectorNumElements();
3901 SmallVector<int, 8> Mask;
3902 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3904 Mask.push_back(i + NumElems);
3906 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3909 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
3910 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3912 unsigned NumElems = VT.getVectorNumElements();
3913 unsigned Half = NumElems/2;
3914 SmallVector<int, 8> Mask;
3915 for (unsigned i = 0; i != Half; ++i) {
3916 Mask.push_back(i + Half);
3917 Mask.push_back(i + NumElems + Half);
3919 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3922 // PromoteSplatv8v16 - All i16 and i8 vector types can't be used directly by
3923 // a generic shuffle instruction because the target has no such instructions.
3924 // Generate shuffles which repeat i16 and i8 several times until they can be
3925 // represented by v4f32 and then be manipulated by target suported shuffles.
3926 static SDValue PromoteSplatv8v16(SDValue V, SelectionDAG &DAG, int &EltNo) {
3927 EVT VT = V.getValueType();
3928 int NumElems = VT.getVectorNumElements();
3929 DebugLoc dl = V.getDebugLoc();
3931 while (NumElems > 4) {
3932 if (EltNo < NumElems/2) {
3933 V = getUnpackl(DAG, dl, VT, V, V);
3935 V = getUnpackh(DAG, dl, VT, V, V);
3936 EltNo -= NumElems/2;
3943 /// getLegalSplat - Generate a legal splat with supported x86 shuffles
3944 static SDValue getLegalSplat(SelectionDAG &DAG, SDValue V, int EltNo) {
3945 EVT VT = V.getValueType();
3946 DebugLoc dl = V.getDebugLoc();
3947 assert((VT.getSizeInBits() == 128 || VT.getSizeInBits() == 256)
3948 && "Vector size not supported");
3950 bool Is128 = VT.getSizeInBits() == 128;
3951 EVT NVT = Is128 ? MVT::v4f32 : MVT::v8f32;
3952 V = DAG.getNode(ISD::BITCAST, dl, NVT, V);
3955 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3956 V = DAG.getVectorShuffle(NVT, dl, V, DAG.getUNDEF(NVT), &SplatMask[0]);
3958 // The second half of indicies refer to the higher part, which is a
3959 // duplication of the lower one. This makes this shuffle a perfect match
3960 // for the VPERM instruction.
3961 int SplatMask[8] = { EltNo, EltNo, EltNo, EltNo,
3962 EltNo+4, EltNo+4, EltNo+4, EltNo+4 };
3963 V = DAG.getVectorShuffle(NVT, dl, V, DAG.getUNDEF(NVT), &SplatMask[0]);
3966 return DAG.getNode(ISD::BITCAST, dl, VT, V);
3969 /// PromoteVectorToScalarSplat - Since there's no native support for
3970 /// scalar_to_vector for 256-bit AVX, a 128-bit scalar_to_vector +
3971 /// INSERT_SUBVECTOR is generated. Recognize this idiom and do the
3972 /// shuffle before the insertion, this yields less instructions in the end.
3973 static SDValue PromoteVectorToScalarSplat(ShuffleVectorSDNode *SV,
3974 SelectionDAG &DAG) {
3975 EVT SrcVT = SV->getValueType(0);
3976 SDValue V1 = SV->getOperand(0);
3977 DebugLoc dl = SV->getDebugLoc();
3978 int NumElems = SrcVT.getVectorNumElements();
3980 assert(SrcVT.is256BitVector() && "unknown howto handle vector type");
3982 SmallVector<int, 4> Mask;
3983 for (int i = 0; i < NumElems/2; ++i)
3984 Mask.push_back(SV->getMaskElt(i));
3986 EVT SVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
3988 SDValue SV1 = DAG.getVectorShuffle(SVT, dl, V1.getOperand(1),
3989 DAG.getUNDEF(SVT), &Mask[0]);
3990 SDValue InsV = Insert128BitVector(DAG.getUNDEF(SrcVT), SV1,
3991 DAG.getConstant(0, MVT::i32), DAG, dl);
3993 return Insert128BitVector(InsV, SV1,
3994 DAG.getConstant(NumElems/2, MVT::i32), DAG, dl);
3997 /// PromoteSplat - Promote a splat of v4i32, v8i16 or v16i8 to v4f32 and
3998 /// v8i32, v16i16 or v32i8 to v8f32.
3999 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
4000 EVT SrcVT = SV->getValueType(0);
4001 SDValue V1 = SV->getOperand(0);
4002 DebugLoc dl = SV->getDebugLoc();
4004 int EltNo = SV->getSplatIndex();
4005 int NumElems = SrcVT.getVectorNumElements();
4006 unsigned Size = SrcVT.getSizeInBits();
4008 // Extract the 128-bit part containing the splat element and update
4009 // the splat element index when it refers to the higher register.
4011 unsigned Idx = (EltNo > NumElems/2) ? NumElems/2 : 0;
4012 V1 = Extract128BitVector(V1, DAG.getConstant(Idx, MVT::i32), DAG, dl);
4014 EltNo -= NumElems/2;
4017 // Make this 128-bit vector duplicate i8 and i16 elements
4019 V1 = PromoteSplatv8v16(V1, DAG, EltNo);
4021 // Recreate the 256-bit vector and place the same 128-bit vector
4022 // into the low and high part. This is necessary because we want
4023 // to use VPERM to shuffle the v8f32 vector, and VPERM only shuffles
4024 // inside each separate v4f32 lane.
4026 SDValue InsV = Insert128BitVector(DAG.getUNDEF(SrcVT), V1,
4027 DAG.getConstant(0, MVT::i32), DAG, dl);
4028 V1 = Insert128BitVector(InsV, V1,
4029 DAG.getConstant(NumElems/2, MVT::i32), DAG, dl);
4032 return getLegalSplat(DAG, V1, EltNo);
4035 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
4036 /// vector of zero or undef vector. This produces a shuffle where the low
4037 /// element of V2 is swizzled into the zero/undef vector, landing at element
4038 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4039 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4040 bool isZero, bool HasSSE2,
4041 SelectionDAG &DAG) {
4042 EVT VT = V2.getValueType();
4044 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
4045 unsigned NumElems = VT.getVectorNumElements();
4046 SmallVector<int, 16> MaskVec;
4047 for (unsigned i = 0; i != NumElems; ++i)
4048 // If this is the insertion idx, put the low elt of V2 here.
4049 MaskVec.push_back(i == Idx ? NumElems : i);
4050 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
4053 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4054 /// element of the result of the vector shuffle.
4055 static SDValue getShuffleScalarElt(SDNode *N, int Index, SelectionDAG &DAG,
4058 return SDValue(); // Limit search depth.
4060 SDValue V = SDValue(N, 0);
4061 EVT VT = V.getValueType();
4062 unsigned Opcode = V.getOpcode();
4064 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4065 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4066 Index = SV->getMaskElt(Index);
4069 return DAG.getUNDEF(VT.getVectorElementType());
4071 int NumElems = VT.getVectorNumElements();
4072 SDValue NewV = (Index < NumElems) ? SV->getOperand(0) : SV->getOperand(1);
4073 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG, Depth+1);
4076 // Recurse into target specific vector shuffles to find scalars.
4077 if (isTargetShuffle(Opcode)) {
4078 int NumElems = VT.getVectorNumElements();
4079 SmallVector<unsigned, 16> ShuffleMask;
4083 case X86ISD::SHUFPS:
4084 case X86ISD::SHUFPD:
4085 ImmN = N->getOperand(N->getNumOperands()-1);
4086 DecodeSHUFPSMask(NumElems,
4087 cast<ConstantSDNode>(ImmN)->getZExtValue(),
4090 case X86ISD::PUNPCKHBW:
4091 case X86ISD::PUNPCKHWD:
4092 case X86ISD::PUNPCKHDQ:
4093 case X86ISD::PUNPCKHQDQ:
4094 DecodePUNPCKHMask(NumElems, ShuffleMask);
4096 case X86ISD::UNPCKHPS:
4097 case X86ISD::UNPCKHPD:
4098 DecodeUNPCKHPMask(NumElems, ShuffleMask);
4100 case X86ISD::PUNPCKLBW:
4101 case X86ISD::PUNPCKLWD:
4102 case X86ISD::PUNPCKLDQ:
4103 case X86ISD::PUNPCKLQDQ:
4104 DecodePUNPCKLMask(VT, ShuffleMask);
4106 case X86ISD::UNPCKLPS:
4107 case X86ISD::UNPCKLPD:
4108 case X86ISD::VUNPCKLPSY:
4109 case X86ISD::VUNPCKLPDY:
4110 DecodeUNPCKLPMask(VT, ShuffleMask);
4112 case X86ISD::MOVHLPS:
4113 DecodeMOVHLPSMask(NumElems, ShuffleMask);
4115 case X86ISD::MOVLHPS:
4116 DecodeMOVLHPSMask(NumElems, ShuffleMask);
4118 case X86ISD::PSHUFD:
4119 ImmN = N->getOperand(N->getNumOperands()-1);
4120 DecodePSHUFMask(NumElems,
4121 cast<ConstantSDNode>(ImmN)->getZExtValue(),
4124 case X86ISD::PSHUFHW:
4125 ImmN = N->getOperand(N->getNumOperands()-1);
4126 DecodePSHUFHWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
4129 case X86ISD::PSHUFLW:
4130 ImmN = N->getOperand(N->getNumOperands()-1);
4131 DecodePSHUFLWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
4135 case X86ISD::MOVSD: {
4136 // The index 0 always comes from the first element of the second source,
4137 // this is why MOVSS and MOVSD are used in the first place. The other
4138 // elements come from the other positions of the first source vector.
4139 unsigned OpNum = (Index == 0) ? 1 : 0;
4140 return getShuffleScalarElt(V.getOperand(OpNum).getNode(), Index, DAG,
4143 case X86ISD::VPERMIL:
4144 ImmN = N->getOperand(N->getNumOperands()-1);
4145 DecodeVPERMILMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4148 assert("not implemented for target shuffle node");
4152 Index = ShuffleMask[Index];
4154 return DAG.getUNDEF(VT.getVectorElementType());
4156 SDValue NewV = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
4157 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG,
4161 // Actual nodes that may contain scalar elements
4162 if (Opcode == ISD::BITCAST) {
4163 V = V.getOperand(0);
4164 EVT SrcVT = V.getValueType();
4165 unsigned NumElems = VT.getVectorNumElements();
4167 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4171 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4172 return (Index == 0) ? V.getOperand(0)
4173 : DAG.getUNDEF(VT.getVectorElementType());
4175 if (V.getOpcode() == ISD::BUILD_VECTOR)
4176 return V.getOperand(Index);
4181 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
4182 /// shuffle operation which come from a consecutively from a zero. The
4183 /// search can start in two different directions, from left or right.
4185 unsigned getNumOfConsecutiveZeros(SDNode *N, int NumElems,
4186 bool ZerosFromLeft, SelectionDAG &DAG) {
4189 while (i < NumElems) {
4190 unsigned Index = ZerosFromLeft ? i : NumElems-i-1;
4191 SDValue Elt = getShuffleScalarElt(N, Index, DAG, 0);
4192 if (!(Elt.getNode() &&
4193 (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt))))
4201 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies from MaskI to
4202 /// MaskE correspond consecutively to elements from one of the vector operands,
4203 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
4205 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp, int MaskI, int MaskE,
4206 int OpIdx, int NumElems, unsigned &OpNum) {
4207 bool SeenV1 = false;
4208 bool SeenV2 = false;
4210 for (int i = MaskI; i <= MaskE; ++i, ++OpIdx) {
4211 int Idx = SVOp->getMaskElt(i);
4212 // Ignore undef indicies
4221 // Only accept consecutive elements from the same vector
4222 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
4226 OpNum = SeenV1 ? 0 : 1;
4230 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
4231 /// logical left shift of a vector.
4232 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4233 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4234 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4235 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4236 false /* check zeros from right */, DAG);
4242 // Considering the elements in the mask that are not consecutive zeros,
4243 // check if they consecutively come from only one of the source vectors.
4245 // V1 = {X, A, B, C} 0
4247 // vector_shuffle V1, V2 <1, 2, 3, X>
4249 if (!isShuffleMaskConsecutive(SVOp,
4250 0, // Mask Start Index
4251 NumElems-NumZeros-1, // Mask End Index
4252 NumZeros, // Where to start looking in the src vector
4253 NumElems, // Number of elements in vector
4254 OpSrc)) // Which source operand ?
4259 ShVal = SVOp->getOperand(OpSrc);
4263 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
4264 /// logical left shift of a vector.
4265 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4266 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4267 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4268 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4269 true /* check zeros from left */, DAG);
4275 // Considering the elements in the mask that are not consecutive zeros,
4276 // check if they consecutively come from only one of the source vectors.
4278 // 0 { A, B, X, X } = V2
4280 // vector_shuffle V1, V2 <X, X, 4, 5>
4282 if (!isShuffleMaskConsecutive(SVOp,
4283 NumZeros, // Mask Start Index
4284 NumElems-1, // Mask End Index
4285 0, // Where to start looking in the src vector
4286 NumElems, // Number of elements in vector
4287 OpSrc)) // Which source operand ?
4292 ShVal = SVOp->getOperand(OpSrc);
4296 /// isVectorShift - Returns true if the shuffle can be implemented as a
4297 /// logical left or right shift of a vector.
4298 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4299 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4300 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
4301 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
4307 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
4309 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4310 unsigned NumNonZero, unsigned NumZero,
4312 const TargetLowering &TLI) {
4316 DebugLoc dl = Op.getDebugLoc();
4319 for (unsigned i = 0; i < 16; ++i) {
4320 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4321 if (ThisIsNonZero && First) {
4323 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4325 V = DAG.getUNDEF(MVT::v8i16);
4330 SDValue ThisElt(0, 0), LastElt(0, 0);
4331 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4332 if (LastIsNonZero) {
4333 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4334 MVT::i16, Op.getOperand(i-1));
4336 if (ThisIsNonZero) {
4337 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4338 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4339 ThisElt, DAG.getConstant(8, MVT::i8));
4341 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4345 if (ThisElt.getNode())
4346 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4347 DAG.getIntPtrConstant(i/2));
4351 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
4354 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
4356 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4357 unsigned NumNonZero, unsigned NumZero,
4359 const TargetLowering &TLI) {
4363 DebugLoc dl = Op.getDebugLoc();
4366 for (unsigned i = 0; i < 8; ++i) {
4367 bool isNonZero = (NonZeros & (1 << i)) != 0;
4371 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4373 V = DAG.getUNDEF(MVT::v8i16);
4376 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4377 MVT::v8i16, V, Op.getOperand(i),
4378 DAG.getIntPtrConstant(i));
4385 /// getVShift - Return a vector logical shift node.
4387 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4388 unsigned NumBits, SelectionDAG &DAG,
4389 const TargetLowering &TLI, DebugLoc dl) {
4390 EVT ShVT = MVT::v2i64;
4391 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
4392 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
4393 return DAG.getNode(ISD::BITCAST, dl, VT,
4394 DAG.getNode(Opc, dl, ShVT, SrcOp,
4395 DAG.getConstant(NumBits,
4396 TLI.getShiftAmountTy(SrcOp.getValueType()))));
4400 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
4401 SelectionDAG &DAG) const {
4403 // Check if the scalar load can be widened into a vector load. And if
4404 // the address is "base + cst" see if the cst can be "absorbed" into
4405 // the shuffle mask.
4406 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4407 SDValue Ptr = LD->getBasePtr();
4408 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4410 EVT PVT = LD->getValueType(0);
4411 if (PVT != MVT::i32 && PVT != MVT::f32)
4416 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4417 FI = FINode->getIndex();
4419 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4420 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4421 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4422 Offset = Ptr.getConstantOperandVal(1);
4423 Ptr = Ptr.getOperand(0);
4428 SDValue Chain = LD->getChain();
4429 // Make sure the stack object alignment is at least 16.
4430 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4431 if (DAG.InferPtrAlignment(Ptr) < 16) {
4432 if (MFI->isFixedObjectIndex(FI)) {
4433 // Can't change the alignment. FIXME: It's possible to compute
4434 // the exact stack offset and reference FI + adjust offset instead.
4435 // If someone *really* cares about this. That's the way to implement it.
4438 MFI->setObjectAlignment(FI, 16);
4442 // (Offset % 16) must be multiple of 4. Then address is then
4443 // Ptr + (Offset & ~15).
4446 if ((Offset % 16) & 3)
4448 int64_t StartOffset = Offset & ~15;
4450 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
4451 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4453 int EltNo = (Offset - StartOffset) >> 2;
4454 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
4455 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
4456 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,
4457 LD->getPointerInfo().getWithOffset(StartOffset),
4459 // Canonicalize it to a v4i32 shuffle.
4460 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1);
4461 return DAG.getNode(ISD::BITCAST, dl, VT,
4462 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
4463 DAG.getUNDEF(MVT::v4i32),&Mask[0]));
4469 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
4470 /// vector of type 'VT', see if the elements can be replaced by a single large
4471 /// load which has the same value as a build_vector whose operands are 'elts'.
4473 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4475 /// FIXME: we'd also like to handle the case where the last elements are zero
4476 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4477 /// There's even a handy isZeroNode for that purpose.
4478 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
4479 DebugLoc &DL, SelectionDAG &DAG) {
4480 EVT EltVT = VT.getVectorElementType();
4481 unsigned NumElems = Elts.size();
4483 LoadSDNode *LDBase = NULL;
4484 unsigned LastLoadedElt = -1U;
4486 // For each element in the initializer, see if we've found a load or an undef.
4487 // If we don't find an initial load element, or later load elements are
4488 // non-consecutive, bail out.
4489 for (unsigned i = 0; i < NumElems; ++i) {
4490 SDValue Elt = Elts[i];
4492 if (!Elt.getNode() ||
4493 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4496 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4498 LDBase = cast<LoadSDNode>(Elt.getNode());
4502 if (Elt.getOpcode() == ISD::UNDEF)
4505 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4506 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
4511 // If we have found an entire vector of loads and undefs, then return a large
4512 // load of the entire vector width starting at the base pointer. If we found
4513 // consecutive loads for the low half, generate a vzext_load node.
4514 if (LastLoadedElt == NumElems - 1) {
4515 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
4516 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4517 LDBase->getPointerInfo(),
4518 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
4519 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4520 LDBase->getPointerInfo(),
4521 LDBase->isVolatile(), LDBase->isNonTemporal(),
4522 LDBase->getAlignment());
4523 } else if (NumElems == 4 && LastLoadedElt == 1 &&
4524 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
4525 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4526 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4527 SDValue ResNode = DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys,
4529 LDBase->getMemOperand());
4530 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
4536 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
4537 DebugLoc dl = Op.getDebugLoc();
4539 EVT VT = Op.getValueType();
4540 EVT ExtVT = VT.getVectorElementType();
4541 unsigned NumElems = Op.getNumOperands();
4544 // - pxor (SSE2), xorps (SSE1), vpxor (128 AVX), xorp[s|d] (256 AVX)
4546 // - pcmpeqd (SSE2 and 128 AVX), fallback to constant pools (256 AVX)
4547 if (ISD::isBuildVectorAllZeros(Op.getNode()) ||
4548 ISD::isBuildVectorAllOnes(Op.getNode())) {
4549 // Canonicalize this to <4 x i32> or <8 x 32> (SSE) to
4550 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
4551 // eliminated on x86-32 hosts.
4552 if (Op.getValueType() == MVT::v4i32 ||
4553 Op.getValueType() == MVT::v8i32)
4556 if (ISD::isBuildVectorAllOnes(Op.getNode()))
4557 return getOnesVector(Op.getValueType(), DAG, dl);
4558 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
4561 unsigned EVTBits = ExtVT.getSizeInBits();
4563 unsigned NumZero = 0;
4564 unsigned NumNonZero = 0;
4565 unsigned NonZeros = 0;
4566 bool IsAllConstants = true;
4567 SmallSet<SDValue, 8> Values;
4568 for (unsigned i = 0; i < NumElems; ++i) {
4569 SDValue Elt = Op.getOperand(i);
4570 if (Elt.getOpcode() == ISD::UNDEF)
4573 if (Elt.getOpcode() != ISD::Constant &&
4574 Elt.getOpcode() != ISD::ConstantFP)
4575 IsAllConstants = false;
4576 if (X86::isZeroNode(Elt))
4579 NonZeros |= (1 << i);
4584 // All undef vector. Return an UNDEF. All zero vectors were handled above.
4585 if (NumNonZero == 0)
4586 return DAG.getUNDEF(VT);
4588 // Special case for single non-zero, non-undef, element.
4589 if (NumNonZero == 1) {
4590 unsigned Idx = CountTrailingZeros_32(NonZeros);
4591 SDValue Item = Op.getOperand(Idx);
4593 // If this is an insertion of an i64 value on x86-32, and if the top bits of
4594 // the value are obviously zero, truncate the value to i32 and do the
4595 // insertion that way. Only do this if the value is non-constant or if the
4596 // value is a constant being inserted into element 0. It is cheaper to do
4597 // a constant pool load than it is to do a movd + shuffle.
4598 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
4599 (!IsAllConstants || Idx == 0)) {
4600 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
4602 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
4603 EVT VecVT = MVT::v4i32;
4604 unsigned VecElts = 4;
4606 // Truncate the value (which may itself be a constant) to i32, and
4607 // convert it to a vector with movd (S2V+shuffle to zero extend).
4608 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
4609 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
4610 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4611 Subtarget->hasSSE2(), DAG);
4613 // Now we have our 32-bit value zero extended in the low element of
4614 // a vector. If Idx != 0, swizzle it into place.
4616 SmallVector<int, 4> Mask;
4617 Mask.push_back(Idx);
4618 for (unsigned i = 1; i != VecElts; ++i)
4620 Item = DAG.getVectorShuffle(VecVT, dl, Item,
4621 DAG.getUNDEF(Item.getValueType()),
4624 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Item);
4628 // If we have a constant or non-constant insertion into the low element of
4629 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
4630 // the rest of the elements. This will be matched as movd/movq/movss/movsd
4631 // depending on what the source datatype is.
4634 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4635 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
4636 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
4637 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4638 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
4639 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
4641 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
4642 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
4643 assert(VT.getSizeInBits() == 128 && "Expected an SSE value type!");
4644 EVT MiddleVT = MVT::v4i32;
4645 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
4646 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4647 Subtarget->hasSSE2(), DAG);
4648 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
4652 // Is it a vector logical left shift?
4653 if (NumElems == 2 && Idx == 1 &&
4654 X86::isZeroNode(Op.getOperand(0)) &&
4655 !X86::isZeroNode(Op.getOperand(1))) {
4656 unsigned NumBits = VT.getSizeInBits();
4657 return getVShift(true, VT,
4658 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4659 VT, Op.getOperand(1)),
4660 NumBits/2, DAG, *this, dl);
4663 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
4666 // Otherwise, if this is a vector with i32 or f32 elements, and the element
4667 // is a non-constant being inserted into an element other than the low one,
4668 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
4669 // movd/movss) to move this into the low element, then shuffle it into
4671 if (EVTBits == 32) {
4672 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4674 // Turn it into a shuffle of zero and zero-extended scalar to vector.
4675 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
4676 Subtarget->hasSSE2(), DAG);
4677 SmallVector<int, 8> MaskVec;
4678 for (unsigned i = 0; i < NumElems; i++)
4679 MaskVec.push_back(i == Idx ? 0 : 1);
4680 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
4684 // Splat is obviously ok. Let legalizer expand it to a shuffle.
4685 if (Values.size() == 1) {
4686 if (EVTBits == 32) {
4687 // Instead of a shuffle like this:
4688 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
4689 // Check if it's possible to issue this instead.
4690 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
4691 unsigned Idx = CountTrailingZeros_32(NonZeros);
4692 SDValue Item = Op.getOperand(Idx);
4693 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
4694 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
4699 // A vector full of immediates; various special cases are already
4700 // handled, so this is best done with a single constant-pool load.
4704 // For AVX-length vectors, build the individual 128-bit pieces and use
4705 // shuffles to put them in place.
4706 if (VT.getSizeInBits() == 256 && !ISD::isBuildVectorAllZeros(Op.getNode())) {
4707 SmallVector<SDValue, 32> V;
4708 for (unsigned i = 0; i < NumElems; ++i)
4709 V.push_back(Op.getOperand(i));
4711 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
4713 // Build both the lower and upper subvector.
4714 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[0], NumElems/2);
4715 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[NumElems / 2],
4718 // Recreate the wider vector with the lower and upper part.
4719 SDValue Vec = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT), Upper,
4720 DAG.getConstant(NumElems/2, MVT::i32), DAG, dl);
4721 return Insert128BitVector(Vec, Lower, DAG.getConstant(0, MVT::i32),
4725 // Let legalizer expand 2-wide build_vectors.
4726 if (EVTBits == 64) {
4727 if (NumNonZero == 1) {
4728 // One half is zero or undef.
4729 unsigned Idx = CountTrailingZeros_32(NonZeros);
4730 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
4731 Op.getOperand(Idx));
4732 return getShuffleVectorZeroOrUndef(V2, Idx, true,
4733 Subtarget->hasSSE2(), DAG);
4738 // If element VT is < 32 bits, convert it to inserts into a zero vector.
4739 if (EVTBits == 8 && NumElems == 16) {
4740 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
4742 if (V.getNode()) return V;
4745 if (EVTBits == 16 && NumElems == 8) {
4746 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
4748 if (V.getNode()) return V;
4751 // If element VT is == 32 bits, turn it into a number of shuffles.
4752 SmallVector<SDValue, 8> V;
4754 if (NumElems == 4 && NumZero > 0) {
4755 for (unsigned i = 0; i < 4; ++i) {
4756 bool isZero = !(NonZeros & (1 << i));
4758 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4760 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4763 for (unsigned i = 0; i < 2; ++i) {
4764 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
4767 V[i] = V[i*2]; // Must be a zero vector.
4770 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
4773 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
4776 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
4781 SmallVector<int, 8> MaskVec;
4782 bool Reverse = (NonZeros & 0x3) == 2;
4783 for (unsigned i = 0; i < 2; ++i)
4784 MaskVec.push_back(Reverse ? 1-i : i);
4785 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4786 for (unsigned i = 0; i < 2; ++i)
4787 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4788 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4791 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4792 // Check for a build vector of consecutive loads.
4793 for (unsigned i = 0; i < NumElems; ++i)
4794 V[i] = Op.getOperand(i);
4796 // Check for elements which are consecutive loads.
4797 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4801 // For SSE 4.1, use insertps to put the high elements into the low element.
4802 if (getSubtarget()->hasSSE41()) {
4804 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
4805 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
4807 Result = DAG.getUNDEF(VT);
4809 for (unsigned i = 1; i < NumElems; ++i) {
4810 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
4811 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
4812 Op.getOperand(i), DAG.getIntPtrConstant(i));
4817 // Otherwise, expand into a number of unpckl*, start by extending each of
4818 // our (non-undef) elements to the full vector width with the element in the
4819 // bottom slot of the vector (which generates no code for SSE).
4820 for (unsigned i = 0; i < NumElems; ++i) {
4821 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4822 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4824 V[i] = DAG.getUNDEF(VT);
4827 // Next, we iteratively mix elements, e.g. for v4f32:
4828 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4829 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4830 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4831 unsigned EltStride = NumElems >> 1;
4832 while (EltStride != 0) {
4833 for (unsigned i = 0; i < EltStride; ++i) {
4834 // If V[i+EltStride] is undef and this is the first round of mixing,
4835 // then it is safe to just drop this shuffle: V[i] is already in the
4836 // right place, the one element (since it's the first round) being
4837 // inserted as undef can be dropped. This isn't safe for successive
4838 // rounds because they will permute elements within both vectors.
4839 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
4840 EltStride == NumElems/2)
4843 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
4853 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4854 // We support concatenate two MMX registers and place them in a MMX
4855 // register. This is better than doing a stack convert.
4856 DebugLoc dl = Op.getDebugLoc();
4857 EVT ResVT = Op.getValueType();
4858 assert(Op.getNumOperands() == 2);
4859 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4860 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4862 SDValue InVec = DAG.getNode(ISD::BITCAST,dl, MVT::v1i64, Op.getOperand(0));
4863 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4864 InVec = Op.getOperand(1);
4865 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4866 unsigned NumElts = ResVT.getVectorNumElements();
4867 VecOp = DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
4868 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4869 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4871 InVec = DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, InVec);
4872 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4873 Mask[0] = 0; Mask[1] = 2;
4874 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4876 return DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
4879 // v8i16 shuffles - Prefer shuffles in the following order:
4880 // 1. [all] pshuflw, pshufhw, optional move
4881 // 2. [ssse3] 1 x pshufb
4882 // 3. [ssse3] 2 x pshufb + 1 x por
4883 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4885 X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
4886 SelectionDAG &DAG) const {
4887 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4888 SDValue V1 = SVOp->getOperand(0);
4889 SDValue V2 = SVOp->getOperand(1);
4890 DebugLoc dl = SVOp->getDebugLoc();
4891 SmallVector<int, 8> MaskVals;
4893 // Determine if more than 1 of the words in each of the low and high quadwords
4894 // of the result come from the same quadword of one of the two inputs. Undef
4895 // mask values count as coming from any quadword, for better codegen.
4896 SmallVector<unsigned, 4> LoQuad(4);
4897 SmallVector<unsigned, 4> HiQuad(4);
4898 BitVector InputQuads(4);
4899 for (unsigned i = 0; i < 8; ++i) {
4900 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4901 int EltIdx = SVOp->getMaskElt(i);
4902 MaskVals.push_back(EltIdx);
4911 InputQuads.set(EltIdx / 4);
4914 int BestLoQuad = -1;
4915 unsigned MaxQuad = 1;
4916 for (unsigned i = 0; i < 4; ++i) {
4917 if (LoQuad[i] > MaxQuad) {
4919 MaxQuad = LoQuad[i];
4923 int BestHiQuad = -1;
4925 for (unsigned i = 0; i < 4; ++i) {
4926 if (HiQuad[i] > MaxQuad) {
4928 MaxQuad = HiQuad[i];
4932 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4933 // of the two input vectors, shuffle them into one input vector so only a
4934 // single pshufb instruction is necessary. If There are more than 2 input
4935 // quads, disable the next transformation since it does not help SSSE3.
4936 bool V1Used = InputQuads[0] || InputQuads[1];
4937 bool V2Used = InputQuads[2] || InputQuads[3];
4938 if (Subtarget->hasSSSE3()) {
4939 if (InputQuads.count() == 2 && V1Used && V2Used) {
4940 BestLoQuad = InputQuads.find_first();
4941 BestHiQuad = InputQuads.find_next(BestLoQuad);
4943 if (InputQuads.count() > 2) {
4949 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4950 // the shuffle mask. If a quad is scored as -1, that means that it contains
4951 // words from all 4 input quadwords.
4953 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4954 SmallVector<int, 8> MaskV;
4955 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4956 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4957 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4958 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
4959 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
4960 NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
4962 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4963 // source words for the shuffle, to aid later transformations.
4964 bool AllWordsInNewV = true;
4965 bool InOrder[2] = { true, true };
4966 for (unsigned i = 0; i != 8; ++i) {
4967 int idx = MaskVals[i];
4969 InOrder[i/4] = false;
4970 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4972 AllWordsInNewV = false;
4976 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4977 if (AllWordsInNewV) {
4978 for (int i = 0; i != 8; ++i) {
4979 int idx = MaskVals[i];
4982 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4983 if ((idx != i) && idx < 4)
4985 if ((idx != i) && idx > 3)
4994 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4995 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4996 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4997 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
4998 unsigned TargetMask = 0;
4999 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
5000 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
5001 TargetMask = pshufhw ? X86::getShufflePSHUFHWImmediate(NewV.getNode()):
5002 X86::getShufflePSHUFLWImmediate(NewV.getNode());
5003 V1 = NewV.getOperand(0);
5004 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
5008 // If we have SSSE3, and all words of the result are from 1 input vector,
5009 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
5010 // is present, fall back to case 4.
5011 if (Subtarget->hasSSSE3()) {
5012 SmallVector<SDValue,16> pshufbMask;
5014 // If we have elements from both input vectors, set the high bit of the
5015 // shuffle mask element to zero out elements that come from V2 in the V1
5016 // mask, and elements that come from V1 in the V2 mask, so that the two
5017 // results can be OR'd together.
5018 bool TwoInputs = V1Used && V2Used;
5019 for (unsigned i = 0; i != 8; ++i) {
5020 int EltIdx = MaskVals[i] * 2;
5021 if (TwoInputs && (EltIdx >= 16)) {
5022 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5023 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5026 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
5027 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
5029 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V1);
5030 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
5031 DAG.getNode(ISD::BUILD_VECTOR, dl,
5032 MVT::v16i8, &pshufbMask[0], 16));
5034 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5036 // Calculate the shuffle mask for the second input, shuffle it, and
5037 // OR it with the first shuffled input.
5039 for (unsigned i = 0; i != 8; ++i) {
5040 int EltIdx = MaskVals[i] * 2;
5042 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5043 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5046 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
5047 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
5049 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V2);
5050 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
5051 DAG.getNode(ISD::BUILD_VECTOR, dl,
5052 MVT::v16i8, &pshufbMask[0], 16));
5053 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
5054 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5057 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
5058 // and update MaskVals with new element order.
5059 BitVector InOrder(8);
5060 if (BestLoQuad >= 0) {
5061 SmallVector<int, 8> MaskV;
5062 for (int i = 0; i != 4; ++i) {
5063 int idx = MaskVals[i];
5065 MaskV.push_back(-1);
5067 } else if ((idx / 4) == BestLoQuad) {
5068 MaskV.push_back(idx & 3);
5071 MaskV.push_back(-1);
5074 for (unsigned i = 4; i != 8; ++i)
5076 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
5079 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
5080 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
5082 X86::getShufflePSHUFLWImmediate(NewV.getNode()),
5086 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
5087 // and update MaskVals with the new element order.
5088 if (BestHiQuad >= 0) {
5089 SmallVector<int, 8> MaskV;
5090 for (unsigned i = 0; i != 4; ++i)
5092 for (unsigned i = 4; i != 8; ++i) {
5093 int idx = MaskVals[i];
5095 MaskV.push_back(-1);
5097 } else if ((idx / 4) == BestHiQuad) {
5098 MaskV.push_back((idx & 3) + 4);
5101 MaskV.push_back(-1);
5104 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
5107 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
5108 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
5110 X86::getShufflePSHUFHWImmediate(NewV.getNode()),
5114 // In case BestHi & BestLo were both -1, which means each quadword has a word
5115 // from each of the four input quadwords, calculate the InOrder bitvector now
5116 // before falling through to the insert/extract cleanup.
5117 if (BestLoQuad == -1 && BestHiQuad == -1) {
5119 for (int i = 0; i != 8; ++i)
5120 if (MaskVals[i] < 0 || MaskVals[i] == i)
5124 // The other elements are put in the right place using pextrw and pinsrw.
5125 for (unsigned i = 0; i != 8; ++i) {
5128 int EltIdx = MaskVals[i];
5131 SDValue ExtOp = (EltIdx < 8)
5132 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
5133 DAG.getIntPtrConstant(EltIdx))
5134 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
5135 DAG.getIntPtrConstant(EltIdx - 8));
5136 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
5137 DAG.getIntPtrConstant(i));
5142 // v16i8 shuffles - Prefer shuffles in the following order:
5143 // 1. [ssse3] 1 x pshufb
5144 // 2. [ssse3] 2 x pshufb + 1 x por
5145 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
5147 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
5149 const X86TargetLowering &TLI) {
5150 SDValue V1 = SVOp->getOperand(0);
5151 SDValue V2 = SVOp->getOperand(1);
5152 DebugLoc dl = SVOp->getDebugLoc();
5153 SmallVector<int, 16> MaskVals;
5154 SVOp->getMask(MaskVals);
5156 // If we have SSSE3, case 1 is generated when all result bytes come from
5157 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
5158 // present, fall back to case 3.
5159 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
5162 for (unsigned i = 0; i < 16; ++i) {
5163 int EltIdx = MaskVals[i];
5172 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
5173 if (TLI.getSubtarget()->hasSSSE3()) {
5174 SmallVector<SDValue,16> pshufbMask;
5176 // If all result elements are from one input vector, then only translate
5177 // undef mask values to 0x80 (zero out result) in the pshufb mask.
5179 // Otherwise, we have elements from both input vectors, and must zero out
5180 // elements that come from V2 in the first mask, and V1 in the second mask
5181 // so that we can OR them together.
5182 bool TwoInputs = !(V1Only || V2Only);
5183 for (unsigned i = 0; i != 16; ++i) {
5184 int EltIdx = MaskVals[i];
5185 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
5186 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5189 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
5191 // If all the elements are from V2, assign it to V1 and return after
5192 // building the first pshufb.
5195 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
5196 DAG.getNode(ISD::BUILD_VECTOR, dl,
5197 MVT::v16i8, &pshufbMask[0], 16));
5201 // Calculate the shuffle mask for the second input, shuffle it, and
5202 // OR it with the first shuffled input.
5204 for (unsigned i = 0; i != 16; ++i) {
5205 int EltIdx = MaskVals[i];
5207 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5210 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
5212 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
5213 DAG.getNode(ISD::BUILD_VECTOR, dl,
5214 MVT::v16i8, &pshufbMask[0], 16));
5215 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
5218 // No SSSE3 - Calculate in place words and then fix all out of place words
5219 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
5220 // the 16 different words that comprise the two doublequadword input vectors.
5221 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5222 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
5223 SDValue NewV = V2Only ? V2 : V1;
5224 for (int i = 0; i != 8; ++i) {
5225 int Elt0 = MaskVals[i*2];
5226 int Elt1 = MaskVals[i*2+1];
5228 // This word of the result is all undef, skip it.
5229 if (Elt0 < 0 && Elt1 < 0)
5232 // This word of the result is already in the correct place, skip it.
5233 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
5235 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
5238 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
5239 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
5242 // If Elt0 and Elt1 are defined, are consecutive, and can be load
5243 // using a single extract together, load it and store it.
5244 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
5245 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
5246 DAG.getIntPtrConstant(Elt1 / 2));
5247 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5248 DAG.getIntPtrConstant(i));
5252 // If Elt1 is defined, extract it from the appropriate source. If the
5253 // source byte is not also odd, shift the extracted word left 8 bits
5254 // otherwise clear the bottom 8 bits if we need to do an or.
5256 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
5257 DAG.getIntPtrConstant(Elt1 / 2));
5258 if ((Elt1 & 1) == 0)
5259 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
5261 TLI.getShiftAmountTy(InsElt.getValueType())));
5263 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
5264 DAG.getConstant(0xFF00, MVT::i16));
5266 // If Elt0 is defined, extract it from the appropriate source. If the
5267 // source byte is not also even, shift the extracted word right 8 bits. If
5268 // Elt1 was also defined, OR the extracted values together before
5269 // inserting them in the result.
5271 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
5272 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
5273 if ((Elt0 & 1) != 0)
5274 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
5276 TLI.getShiftAmountTy(InsElt0.getValueType())));
5278 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
5279 DAG.getConstant(0x00FF, MVT::i16));
5280 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
5283 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5284 DAG.getIntPtrConstant(i));
5286 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
5289 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
5290 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
5291 /// done when every pair / quad of shuffle mask elements point to elements in
5292 /// the right sequence. e.g.
5293 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
5295 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
5296 SelectionDAG &DAG, DebugLoc dl) {
5297 EVT VT = SVOp->getValueType(0);
5298 SDValue V1 = SVOp->getOperand(0);
5299 SDValue V2 = SVOp->getOperand(1);
5300 unsigned NumElems = VT.getVectorNumElements();
5301 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
5303 switch (VT.getSimpleVT().SimpleTy) {
5304 default: assert(false && "Unexpected!");
5305 case MVT::v4f32: NewVT = MVT::v2f64; break;
5306 case MVT::v4i32: NewVT = MVT::v2i64; break;
5307 case MVT::v8i16: NewVT = MVT::v4i32; break;
5308 case MVT::v16i8: NewVT = MVT::v4i32; break;
5311 int Scale = NumElems / NewWidth;
5312 SmallVector<int, 8> MaskVec;
5313 for (unsigned i = 0; i < NumElems; i += Scale) {
5315 for (int j = 0; j < Scale; ++j) {
5316 int EltIdx = SVOp->getMaskElt(i+j);
5320 StartIdx = EltIdx - (EltIdx % Scale);
5321 if (EltIdx != StartIdx + j)
5325 MaskVec.push_back(-1);
5327 MaskVec.push_back(StartIdx / Scale);
5330 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
5331 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
5332 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
5335 /// getVZextMovL - Return a zero-extending vector move low node.
5337 static SDValue getVZextMovL(EVT VT, EVT OpVT,
5338 SDValue SrcOp, SelectionDAG &DAG,
5339 const X86Subtarget *Subtarget, DebugLoc dl) {
5340 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
5341 LoadSDNode *LD = NULL;
5342 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
5343 LD = dyn_cast<LoadSDNode>(SrcOp);
5345 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
5347 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
5348 if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
5349 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
5350 SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
5351 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
5353 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
5354 return DAG.getNode(ISD::BITCAST, dl, VT,
5355 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5356 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5364 return DAG.getNode(ISD::BITCAST, dl, VT,
5365 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5366 DAG.getNode(ISD::BITCAST, dl,
5370 /// LowerVECTOR_SHUFFLE_256 - Handle all 256-bit wide vectors shuffles
5371 /// which could not be matched by any known target speficic shuffle
5373 LowerVECTOR_SHUFFLE_256(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
5377 /// LowerVECTOR_SHUFFLE_128v4 - Handle all 128-bit wide vectors with
5378 /// 4 elements, and match them with several different shuffle types.
5380 LowerVECTOR_SHUFFLE_128v4(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
5381 SDValue V1 = SVOp->getOperand(0);
5382 SDValue V2 = SVOp->getOperand(1);
5383 DebugLoc dl = SVOp->getDebugLoc();
5384 EVT VT = SVOp->getValueType(0);
5386 assert(VT.getSizeInBits() == 128 && "Unsupported vector size");
5388 SmallVector<std::pair<int, int>, 8> Locs;
5390 SmallVector<int, 8> Mask1(4U, -1);
5391 SmallVector<int, 8> PermMask;
5392 SVOp->getMask(PermMask);
5396 for (unsigned i = 0; i != 4; ++i) {
5397 int Idx = PermMask[i];
5399 Locs[i] = std::make_pair(-1, -1);
5401 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
5403 Locs[i] = std::make_pair(0, NumLo);
5407 Locs[i] = std::make_pair(1, NumHi);
5409 Mask1[2+NumHi] = Idx;
5415 if (NumLo <= 2 && NumHi <= 2) {
5416 // If no more than two elements come from either vector. This can be
5417 // implemented with two shuffles. First shuffle gather the elements.
5418 // The second shuffle, which takes the first shuffle as both of its
5419 // vector operands, put the elements into the right order.
5420 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5422 SmallVector<int, 8> Mask2(4U, -1);
5424 for (unsigned i = 0; i != 4; ++i) {
5425 if (Locs[i].first == -1)
5428 unsigned Idx = (i < 2) ? 0 : 4;
5429 Idx += Locs[i].first * 2 + Locs[i].second;
5434 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
5435 } else if (NumLo == 3 || NumHi == 3) {
5436 // Otherwise, we must have three elements from one vector, call it X, and
5437 // one element from the other, call it Y. First, use a shufps to build an
5438 // intermediate vector with the one element from Y and the element from X
5439 // that will be in the same half in the final destination (the indexes don't
5440 // matter). Then, use a shufps to build the final vector, taking the half
5441 // containing the element from Y from the intermediate, and the other half
5444 // Normalize it so the 3 elements come from V1.
5445 CommuteVectorShuffleMask(PermMask, VT);
5449 // Find the element from V2.
5451 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
5452 int Val = PermMask[HiIndex];
5459 Mask1[0] = PermMask[HiIndex];
5461 Mask1[2] = PermMask[HiIndex^1];
5463 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5466 Mask1[0] = PermMask[0];
5467 Mask1[1] = PermMask[1];
5468 Mask1[2] = HiIndex & 1 ? 6 : 4;
5469 Mask1[3] = HiIndex & 1 ? 4 : 6;
5470 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5472 Mask1[0] = HiIndex & 1 ? 2 : 0;
5473 Mask1[1] = HiIndex & 1 ? 0 : 2;
5474 Mask1[2] = PermMask[2];
5475 Mask1[3] = PermMask[3];
5480 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
5484 // Break it into (shuffle shuffle_hi, shuffle_lo).
5487 SmallVector<int,8> LoMask(4U, -1);
5488 SmallVector<int,8> HiMask(4U, -1);
5490 SmallVector<int,8> *MaskPtr = &LoMask;
5491 unsigned MaskIdx = 0;
5494 for (unsigned i = 0; i != 4; ++i) {
5501 int Idx = PermMask[i];
5503 Locs[i] = std::make_pair(-1, -1);
5504 } else if (Idx < 4) {
5505 Locs[i] = std::make_pair(MaskIdx, LoIdx);
5506 (*MaskPtr)[LoIdx] = Idx;
5509 Locs[i] = std::make_pair(MaskIdx, HiIdx);
5510 (*MaskPtr)[HiIdx] = Idx;
5515 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
5516 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
5517 SmallVector<int, 8> MaskOps;
5518 for (unsigned i = 0; i != 4; ++i) {
5519 if (Locs[i].first == -1) {
5520 MaskOps.push_back(-1);
5522 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
5523 MaskOps.push_back(Idx);
5526 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
5529 static bool MayFoldVectorLoad(SDValue V) {
5530 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
5531 V = V.getOperand(0);
5532 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5533 V = V.getOperand(0);
5539 // FIXME: the version above should always be used. Since there's
5540 // a bug where several vector shuffles can't be folded because the
5541 // DAG is not updated during lowering and a node claims to have two
5542 // uses while it only has one, use this version, and let isel match
5543 // another instruction if the load really happens to have more than
5544 // one use. Remove this version after this bug get fixed.
5545 // rdar://8434668, PR8156
5546 static bool RelaxedMayFoldVectorLoad(SDValue V) {
5547 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
5548 V = V.getOperand(0);
5549 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5550 V = V.getOperand(0);
5551 if (ISD::isNormalLoad(V.getNode()))
5556 /// CanFoldShuffleIntoVExtract - Check if the current shuffle is used by
5557 /// a vector extract, and if both can be later optimized into a single load.
5558 /// This is done in visitEXTRACT_VECTOR_ELT and the conditions are checked
5559 /// here because otherwise a target specific shuffle node is going to be
5560 /// emitted for this shuffle, and the optimization not done.
5561 /// FIXME: This is probably not the best approach, but fix the problem
5562 /// until the right path is decided.
5564 bool CanXFormVExtractWithShuffleIntoLoad(SDValue V, SelectionDAG &DAG,
5565 const TargetLowering &TLI) {
5566 EVT VT = V.getValueType();
5567 ShuffleVectorSDNode *SVOp = dyn_cast<ShuffleVectorSDNode>(V);
5569 // Be sure that the vector shuffle is present in a pattern like this:
5570 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), c) -> (f32 load $addr)
5574 SDNode *N = *V.getNode()->use_begin();
5575 if (N->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5578 SDValue EltNo = N->getOperand(1);
5579 if (!isa<ConstantSDNode>(EltNo))
5582 // If the bit convert changed the number of elements, it is unsafe
5583 // to examine the mask.
5584 bool HasShuffleIntoBitcast = false;
5585 if (V.getOpcode() == ISD::BITCAST) {
5586 EVT SrcVT = V.getOperand(0).getValueType();
5587 if (SrcVT.getVectorNumElements() != VT.getVectorNumElements())
5589 V = V.getOperand(0);
5590 HasShuffleIntoBitcast = true;
5593 // Select the input vector, guarding against out of range extract vector.
5594 unsigned NumElems = VT.getVectorNumElements();
5595 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5596 int Idx = (Elt > NumElems) ? -1 : SVOp->getMaskElt(Elt);
5597 V = (Idx < (int)NumElems) ? V.getOperand(0) : V.getOperand(1);
5599 // Skip one more bit_convert if necessary
5600 if (V.getOpcode() == ISD::BITCAST)
5601 V = V.getOperand(0);
5603 if (ISD::isNormalLoad(V.getNode())) {
5604 // Is the original load suitable?
5605 LoadSDNode *LN0 = cast<LoadSDNode>(V);
5607 // FIXME: avoid the multi-use bug that is preventing lots of
5608 // of foldings to be detected, this is still wrong of course, but
5609 // give the temporary desired behavior, and if it happens that
5610 // the load has real more uses, during isel it will not fold, and
5611 // will generate poor code.
5612 if (!LN0 || LN0->isVolatile()) // || !LN0->hasOneUse()
5615 if (!HasShuffleIntoBitcast)
5618 // If there's a bitcast before the shuffle, check if the load type and
5619 // alignment is valid.
5620 unsigned Align = LN0->getAlignment();
5622 TLI.getTargetData()->getABITypeAlignment(
5623 VT.getTypeForEVT(*DAG.getContext()));
5625 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
5633 SDValue getMOVDDup(SDValue &Op, DebugLoc &dl, SDValue V1, SelectionDAG &DAG) {
5634 EVT VT = Op.getValueType();
5636 // Canonizalize to v2f64.
5637 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
5638 return DAG.getNode(ISD::BITCAST, dl, VT,
5639 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
5644 SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
5646 SDValue V1 = Op.getOperand(0);
5647 SDValue V2 = Op.getOperand(1);
5648 EVT VT = Op.getValueType();
5650 assert(VT != MVT::v2i64 && "unsupported shuffle type");
5652 if (HasSSE2 && VT == MVT::v2f64)
5653 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
5656 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V2, DAG);
5660 SDValue getMOVHighToLow(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG) {
5661 SDValue V1 = Op.getOperand(0);
5662 SDValue V2 = Op.getOperand(1);
5663 EVT VT = Op.getValueType();
5665 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
5666 "unsupported shuffle type");
5668 if (V2.getOpcode() == ISD::UNDEF)
5672 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
5676 SDValue getMOVLP(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
5677 SDValue V1 = Op.getOperand(0);
5678 SDValue V2 = Op.getOperand(1);
5679 EVT VT = Op.getValueType();
5680 unsigned NumElems = VT.getVectorNumElements();
5682 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
5683 // operand of these instructions is only memory, so check if there's a
5684 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
5686 bool CanFoldLoad = false;
5688 // Trivial case, when V2 comes from a load.
5689 if (MayFoldVectorLoad(V2))
5692 // When V1 is a load, it can be folded later into a store in isel, example:
5693 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
5695 // (MOVLPSmr addr:$src1, VR128:$src2)
5696 // So, recognize this potential and also use MOVLPS or MOVLPD
5697 if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
5700 // Both of them can't be memory operations though.
5701 if (MayFoldVectorLoad(V1) && MayFoldVectorLoad(V2))
5702 CanFoldLoad = false;
5705 if (HasSSE2 && NumElems == 2)
5706 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
5709 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
5712 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5713 // movl and movlp will both match v2i64, but v2i64 is never matched by
5714 // movl earlier because we make it strict to avoid messing with the movlp load
5715 // folding logic (see the code above getMOVLP call). Match it here then,
5716 // this is horrible, but will stay like this until we move all shuffle
5717 // matching to x86 specific nodes. Note that for the 1st condition all
5718 // types are matched with movsd.
5719 if ((HasSSE2 && NumElems == 2) || !X86::isMOVLMask(SVOp))
5720 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5722 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5725 assert(VT != MVT::v4i32 && "unsupported shuffle type");
5727 // Invert the operand order and use SHUFPS to match it.
5728 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V2, V1,
5729 X86::getShuffleSHUFImmediate(SVOp), DAG);
5732 static inline unsigned getUNPCKLOpcode(EVT VT) {
5733 switch(VT.getSimpleVT().SimpleTy) {
5734 case MVT::v4i32: return X86ISD::PUNPCKLDQ;
5735 case MVT::v2i64: return X86ISD::PUNPCKLQDQ;
5736 case MVT::v4f32: return X86ISD::UNPCKLPS;
5737 case MVT::v2f64: return X86ISD::UNPCKLPD;
5738 case MVT::v8f32: return X86ISD::VUNPCKLPSY;
5739 case MVT::v4f64: return X86ISD::VUNPCKLPDY;
5740 case MVT::v16i8: return X86ISD::PUNPCKLBW;
5741 case MVT::v8i16: return X86ISD::PUNPCKLWD;
5743 llvm_unreachable("Unknown type for unpckl");
5748 static inline unsigned getUNPCKHOpcode(EVT VT) {
5749 switch(VT.getSimpleVT().SimpleTy) {
5750 case MVT::v4i32: return X86ISD::PUNPCKHDQ;
5751 case MVT::v2i64: return X86ISD::PUNPCKHQDQ;
5752 case MVT::v4f32: return X86ISD::UNPCKHPS;
5753 case MVT::v2f64: return X86ISD::UNPCKHPD;
5754 case MVT::v16i8: return X86ISD::PUNPCKHBW;
5755 case MVT::v8i16: return X86ISD::PUNPCKHWD;
5757 llvm_unreachable("Unknown type for unpckh");
5763 SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG,
5764 const TargetLowering &TLI,
5765 const X86Subtarget *Subtarget) {
5766 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5767 EVT VT = Op.getValueType();
5768 DebugLoc dl = Op.getDebugLoc();
5769 SDValue V1 = Op.getOperand(0);
5770 SDValue V2 = Op.getOperand(1);
5772 if (isZeroShuffle(SVOp))
5773 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
5775 // Handle splat operations
5776 if (SVOp->isSplat()) {
5777 unsigned NumElem = VT.getVectorNumElements();
5778 // Special case, this is the only place now where it's allowed to return
5779 // a vector_shuffle operation without using a target specific node, because
5780 // *hopefully* it will be optimized away by the dag combiner. FIXME: should
5781 // this be moved to DAGCombine instead?
5782 if (NumElem <= 4 && CanXFormVExtractWithShuffleIntoLoad(Op, DAG, TLI))
5785 // Since there's no native support for scalar_to_vector for 256-bit AVX, a
5786 // 128-bit scalar_to_vector + INSERT_SUBVECTOR is generated. Recognize this
5787 // idiom and do the shuffle before the insertion, this yields less
5788 // instructions in the end.
5789 if (VT.is256BitVector() &&
5790 V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
5791 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
5792 V1.getOperand(1).getOpcode() == ISD::SCALAR_TO_VECTOR)
5793 return PromoteVectorToScalarSplat(SVOp, DAG);
5795 // Handle splats by matching through known shuffle masks
5796 if ((VT.is128BitVector() && NumElem <= 4) ||
5797 (VT.is256BitVector() && NumElem <= 8))
5800 // All i16 and i8 vector types can't be used directly by a generic shuffle
5801 // instruction because the target has no such instruction. Generate shuffles
5802 // which repeat i16 and i8 several times until they fit in i32, and then can
5803 // be manipulated by target suported shuffles. After the insertion of the
5804 // necessary shuffles, the result is bitcasted back to v4f32 or v8f32.
5805 return PromoteSplat(SVOp, DAG);
5808 // If the shuffle can be profitably rewritten as a narrower shuffle, then
5810 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
5811 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5812 if (NewOp.getNode())
5813 return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
5814 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
5815 // FIXME: Figure out a cleaner way to do this.
5816 // Try to make use of movq to zero out the top part.
5817 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
5818 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5819 if (NewOp.getNode()) {
5820 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
5821 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
5822 DAG, Subtarget, dl);
5824 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
5825 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5826 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
5827 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
5828 DAG, Subtarget, dl);
5835 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
5836 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5837 SDValue V1 = Op.getOperand(0);
5838 SDValue V2 = Op.getOperand(1);
5839 EVT VT = Op.getValueType();
5840 DebugLoc dl = Op.getDebugLoc();
5841 unsigned NumElems = VT.getVectorNumElements();
5842 bool isMMX = VT.getSizeInBits() == 64;
5843 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
5844 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
5845 bool V1IsSplat = false;
5846 bool V2IsSplat = false;
5847 bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
5848 bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
5849 bool HasSSSE3 = Subtarget->hasSSSE3() || Subtarget->hasAVX();
5850 MachineFunction &MF = DAG.getMachineFunction();
5851 bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
5853 // Shuffle operations on MMX not supported.
5857 // Vector shuffle lowering takes 3 steps:
5859 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
5860 // narrowing and commutation of operands should be handled.
5861 // 2) Matching of shuffles with known shuffle masks to x86 target specific
5863 // 3) Rewriting of unmatched masks into new generic shuffle operations,
5864 // so the shuffle can be broken into other shuffles and the legalizer can
5865 // try the lowering again.
5867 // The general ideia is that no vector_shuffle operation should be left to
5868 // be matched during isel, all of them must be converted to a target specific
5871 // Normalize the input vectors. Here splats, zeroed vectors, profitable
5872 // narrowing and commutation of operands should be handled. The actual code
5873 // doesn't include all of those, work in progress...
5874 SDValue NewOp = NormalizeVectorShuffle(Op, DAG, *this, Subtarget);
5875 if (NewOp.getNode())
5878 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
5879 // unpckh_undef). Only use pshufd if speed is more important than size.
5880 if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp))
5881 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
5882 if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp))
5883 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5885 if (X86::isMOVDDUPMask(SVOp) && HasSSE3 && V2IsUndef &&
5886 RelaxedMayFoldVectorLoad(V1))
5887 return getMOVDDup(Op, dl, V1, DAG);
5889 if (X86::isMOVHLPS_v_undef_Mask(SVOp))
5890 return getMOVHighToLow(Op, dl, DAG);
5892 // Use to match splats
5893 if (HasSSE2 && X86::isUNPCKHMask(SVOp) && V2IsUndef &&
5894 (VT == MVT::v2f64 || VT == MVT::v2i64))
5895 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5897 if (X86::isPSHUFDMask(SVOp)) {
5898 // The actual implementation will match the mask in the if above and then
5899 // during isel it can match several different instructions, not only pshufd
5900 // as its name says, sad but true, emulate the behavior for now...
5901 if (X86::isMOVDDUPMask(SVOp) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
5902 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
5904 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5906 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
5907 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
5909 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5910 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V1,
5913 if (VT == MVT::v4f32)
5914 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V1,
5918 // Check if this can be converted into a logical shift.
5919 bool isLeft = false;
5922 bool isShift = getSubtarget()->hasSSE2() &&
5923 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
5924 if (isShift && ShVal.hasOneUse()) {
5925 // If the shifted value has multiple uses, it may be cheaper to use
5926 // v_set0 + movlhps or movhlps, etc.
5927 EVT EltVT = VT.getVectorElementType();
5928 ShAmt *= EltVT.getSizeInBits();
5929 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5932 if (X86::isMOVLMask(SVOp)) {
5935 if (ISD::isBuildVectorAllZeros(V1.getNode()))
5936 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
5937 if (!X86::isMOVLPMask(SVOp)) {
5938 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5939 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5941 if (VT == MVT::v4i32 || VT == MVT::v4f32)
5942 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5946 // FIXME: fold these into legal mask.
5947 if (X86::isMOVLHPSMask(SVOp) && !X86::isUNPCKLMask(SVOp))
5948 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
5950 if (X86::isMOVHLPSMask(SVOp))
5951 return getMOVHighToLow(Op, dl, DAG);
5953 if (X86::isMOVSHDUPMask(SVOp, Subtarget))
5954 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
5956 if (X86::isMOVSLDUPMask(SVOp, Subtarget))
5957 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
5959 if (X86::isMOVLPMask(SVOp))
5960 return getMOVLP(Op, dl, DAG, HasSSE2);
5962 if (ShouldXformToMOVHLPS(SVOp) ||
5963 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
5964 return CommuteVectorShuffle(SVOp, DAG);
5967 // No better options. Use a vshl / vsrl.
5968 EVT EltVT = VT.getVectorElementType();
5969 ShAmt *= EltVT.getSizeInBits();
5970 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5973 bool Commuted = false;
5974 // FIXME: This should also accept a bitcast of a splat? Be careful, not
5975 // 1,1,1,1 -> v8i16 though.
5976 V1IsSplat = isSplatVector(V1.getNode());
5977 V2IsSplat = isSplatVector(V2.getNode());
5979 // Canonicalize the splat or undef, if present, to be on the RHS.
5980 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
5981 Op = CommuteVectorShuffle(SVOp, DAG);
5982 SVOp = cast<ShuffleVectorSDNode>(Op);
5983 V1 = SVOp->getOperand(0);
5984 V2 = SVOp->getOperand(1);
5985 std::swap(V1IsSplat, V2IsSplat);
5986 std::swap(V1IsUndef, V2IsUndef);
5990 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
5991 // Shuffling low element of v1 into undef, just return v1.
5994 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
5995 // the instruction selector will not match, so get a canonical MOVL with
5996 // swapped operands to undo the commute.
5997 return getMOVL(DAG, dl, VT, V2, V1);
6000 if (X86::isUNPCKLMask(SVOp))
6001 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V2, DAG);
6003 if (X86::isUNPCKHMask(SVOp))
6004 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V2, DAG);
6007 // Normalize mask so all entries that point to V2 points to its first
6008 // element then try to match unpck{h|l} again. If match, return a
6009 // new vector_shuffle with the corrected mask.
6010 SDValue NewMask = NormalizeMask(SVOp, DAG);
6011 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
6012 if (NSVOp != SVOp) {
6013 if (X86::isUNPCKLMask(NSVOp, true)) {
6015 } else if (X86::isUNPCKHMask(NSVOp, true)) {
6022 // Commute is back and try unpck* again.
6023 // FIXME: this seems wrong.
6024 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
6025 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
6027 if (X86::isUNPCKLMask(NewSVOp))
6028 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V2, V1, DAG);
6030 if (X86::isUNPCKHMask(NewSVOp))
6031 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V2, V1, DAG);
6034 // Normalize the node to match x86 shuffle ops if needed
6035 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
6036 return CommuteVectorShuffle(SVOp, DAG);
6038 // The checks below are all present in isShuffleMaskLegal, but they are
6039 // inlined here right now to enable us to directly emit target specific
6040 // nodes, and remove one by one until they don't return Op anymore.
6041 SmallVector<int, 16> M;
6044 if (isPALIGNRMask(M, VT, HasSSSE3))
6045 return getTargetShuffleNode(X86ISD::PALIGN, dl, VT, V1, V2,
6046 X86::getShufflePALIGNRImmediate(SVOp),
6049 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
6050 SVOp->getSplatIndex() == 0 && V2IsUndef) {
6051 if (VT == MVT::v2f64)
6052 return getTargetShuffleNode(X86ISD::UNPCKLPD, dl, VT, V1, V1, DAG);
6053 if (VT == MVT::v2i64)
6054 return getTargetShuffleNode(X86ISD::PUNPCKLQDQ, dl, VT, V1, V1, DAG);
6057 if (isPSHUFHWMask(M, VT))
6058 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
6059 X86::getShufflePSHUFHWImmediate(SVOp),
6062 if (isPSHUFLWMask(M, VT))
6063 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
6064 X86::getShufflePSHUFLWImmediate(SVOp),
6067 if (isSHUFPMask(M, VT)) {
6068 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
6069 if (VT == MVT::v4f32 || VT == MVT::v4i32)
6070 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V2,
6072 if (VT == MVT::v2f64 || VT == MVT::v2i64)
6073 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V2,
6077 if (X86::isUNPCKL_v_undef_Mask(SVOp))
6078 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
6079 if (X86::isUNPCKH_v_undef_Mask(SVOp))
6080 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
6082 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
6083 if (VT == MVT::v8i16) {
6084 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
6085 if (NewOp.getNode())
6089 if (VT == MVT::v16i8) {
6090 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
6091 if (NewOp.getNode())
6095 // Handle all 128-bit wide vectors with 4 elements, and match them with
6096 // several different shuffle types.
6097 if (NumElems == 4 && VT.getSizeInBits() == 128)
6098 return LowerVECTOR_SHUFFLE_128v4(SVOp, DAG);
6100 //===--------------------------------------------------------------------===//
6101 // Custom lower or generate target specific nodes for 256-bit shuffles.
6103 // Handle VPERMIL permutations
6104 if (isVPERMILMask(M, VT)) {
6105 unsigned TargetMask = getShuffleVPERMILImmediate(SVOp);
6106 if (VT == MVT::v8f32)
6107 return getTargetShuffleNode(X86ISD::VPERMIL, dl, VT, V1, TargetMask, DAG);
6110 // Handle general 256-bit shuffles
6111 if (VT.is256BitVector())
6112 return LowerVECTOR_SHUFFLE_256(SVOp, DAG);
6118 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
6119 SelectionDAG &DAG) const {
6120 EVT VT = Op.getValueType();
6121 DebugLoc dl = Op.getDebugLoc();
6122 if (VT.getSizeInBits() == 8) {
6123 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
6124 Op.getOperand(0), Op.getOperand(1));
6125 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
6126 DAG.getValueType(VT));
6127 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6128 } else if (VT.getSizeInBits() == 16) {
6129 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6130 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
6132 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
6133 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6134 DAG.getNode(ISD::BITCAST, dl,
6138 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
6139 Op.getOperand(0), Op.getOperand(1));
6140 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
6141 DAG.getValueType(VT));
6142 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6143 } else if (VT == MVT::f32) {
6144 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
6145 // the result back to FR32 register. It's only worth matching if the
6146 // result has a single use which is a store or a bitcast to i32. And in
6147 // the case of a store, it's not worth it if the index is a constant 0,
6148 // because a MOVSSmr can be used instead, which is smaller and faster.
6149 if (!Op.hasOneUse())
6151 SDNode *User = *Op.getNode()->use_begin();
6152 if ((User->getOpcode() != ISD::STORE ||
6153 (isa<ConstantSDNode>(Op.getOperand(1)) &&
6154 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
6155 (User->getOpcode() != ISD::BITCAST ||
6156 User->getValueType(0) != MVT::i32))
6158 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6159 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
6162 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
6163 } else if (VT == MVT::i32) {
6164 // ExtractPS works with constant index.
6165 if (isa<ConstantSDNode>(Op.getOperand(1)))
6173 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
6174 SelectionDAG &DAG) const {
6175 if (!isa<ConstantSDNode>(Op.getOperand(1)))
6178 SDValue Vec = Op.getOperand(0);
6179 EVT VecVT = Vec.getValueType();
6181 // If this is a 256-bit vector result, first extract the 128-bit
6182 // vector and then extract from the 128-bit vector.
6183 if (VecVT.getSizeInBits() > 128) {
6184 DebugLoc dl = Op.getNode()->getDebugLoc();
6185 unsigned NumElems = VecVT.getVectorNumElements();
6186 SDValue Idx = Op.getOperand(1);
6188 if (!isa<ConstantSDNode>(Idx))
6191 unsigned ExtractNumElems = NumElems / (VecVT.getSizeInBits() / 128);
6192 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
6194 // Get the 128-bit vector.
6195 bool Upper = IdxVal >= ExtractNumElems;
6196 Vec = Extract128BitVector(Vec, Idx, DAG, dl);
6199 SDValue ScaledIdx = Idx;
6201 ScaledIdx = DAG.getNode(ISD::SUB, dl, Idx.getValueType(), Idx,
6202 DAG.getConstant(ExtractNumElems,
6203 Idx.getValueType()));
6204 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
6208 assert(Vec.getValueSizeInBits() <= 128 && "Unexpected vector length");
6210 if (Subtarget->hasSSE41()) {
6211 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
6216 EVT VT = Op.getValueType();
6217 DebugLoc dl = Op.getDebugLoc();
6218 // TODO: handle v16i8.
6219 if (VT.getSizeInBits() == 16) {
6220 SDValue Vec = Op.getOperand(0);
6221 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6223 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
6224 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6225 DAG.getNode(ISD::BITCAST, dl,
6228 // Transform it so it match pextrw which produces a 32-bit result.
6229 EVT EltVT = MVT::i32;
6230 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
6231 Op.getOperand(0), Op.getOperand(1));
6232 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
6233 DAG.getValueType(VT));
6234 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6235 } else if (VT.getSizeInBits() == 32) {
6236 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6240 // SHUFPS the element to the lowest double word, then movss.
6241 int Mask[4] = { Idx, -1, -1, -1 };
6242 EVT VVT = Op.getOperand(0).getValueType();
6243 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
6244 DAG.getUNDEF(VVT), Mask);
6245 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
6246 DAG.getIntPtrConstant(0));
6247 } else if (VT.getSizeInBits() == 64) {
6248 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
6249 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
6250 // to match extract_elt for f64.
6251 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6255 // UNPCKHPD the element to the lowest double word, then movsd.
6256 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
6257 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
6258 int Mask[2] = { 1, -1 };
6259 EVT VVT = Op.getOperand(0).getValueType();
6260 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
6261 DAG.getUNDEF(VVT), Mask);
6262 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
6263 DAG.getIntPtrConstant(0));
6270 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
6271 SelectionDAG &DAG) const {
6272 EVT VT = Op.getValueType();
6273 EVT EltVT = VT.getVectorElementType();
6274 DebugLoc dl = Op.getDebugLoc();
6276 SDValue N0 = Op.getOperand(0);
6277 SDValue N1 = Op.getOperand(1);
6278 SDValue N2 = Op.getOperand(2);
6280 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
6281 isa<ConstantSDNode>(N2)) {
6283 if (VT == MVT::v8i16)
6284 Opc = X86ISD::PINSRW;
6285 else if (VT == MVT::v16i8)
6286 Opc = X86ISD::PINSRB;
6288 Opc = X86ISD::PINSRB;
6290 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
6292 if (N1.getValueType() != MVT::i32)
6293 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
6294 if (N2.getValueType() != MVT::i32)
6295 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
6296 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
6297 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
6298 // Bits [7:6] of the constant are the source select. This will always be
6299 // zero here. The DAG Combiner may combine an extract_elt index into these
6300 // bits. For example (insert (extract, 3), 2) could be matched by putting
6301 // the '3' into bits [7:6] of X86ISD::INSERTPS.
6302 // Bits [5:4] of the constant are the destination select. This is the
6303 // value of the incoming immediate.
6304 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
6305 // combine either bitwise AND or insert of float 0.0 to set these bits.
6306 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
6307 // Create this as a scalar to vector..
6308 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
6309 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
6310 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
6311 // PINSR* works with constant index.
6318 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
6319 EVT VT = Op.getValueType();
6320 EVT EltVT = VT.getVectorElementType();
6322 DebugLoc dl = Op.getDebugLoc();
6323 SDValue N0 = Op.getOperand(0);
6324 SDValue N1 = Op.getOperand(1);
6325 SDValue N2 = Op.getOperand(2);
6327 // If this is a 256-bit vector result, first insert into a 128-bit
6328 // vector and then insert into the 256-bit vector.
6329 if (VT.getSizeInBits() > 128) {
6330 if (!isa<ConstantSDNode>(N2))
6333 // Get the 128-bit vector.
6334 unsigned NumElems = VT.getVectorNumElements();
6335 unsigned IdxVal = cast<ConstantSDNode>(N2)->getZExtValue();
6336 bool Upper = IdxVal >= NumElems / 2;
6338 SDValue SubN0 = Extract128BitVector(N0, N2, DAG, dl);
6341 SDValue ScaledN2 = N2;
6343 ScaledN2 = DAG.getNode(ISD::SUB, dl, N2.getValueType(), N2,
6344 DAG.getConstant(NumElems /
6345 (VT.getSizeInBits() / 128),
6346 N2.getValueType()));
6347 Op = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, SubN0.getValueType(), SubN0,
6350 // Insert the 128-bit vector
6351 // FIXME: Why UNDEF?
6352 return Insert128BitVector(N0, Op, N2, DAG, dl);
6355 if (Subtarget->hasSSE41())
6356 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
6358 if (EltVT == MVT::i8)
6361 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
6362 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
6363 // as its second argument.
6364 if (N1.getValueType() != MVT::i32)
6365 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
6366 if (N2.getValueType() != MVT::i32)
6367 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
6368 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
6374 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6375 LLVMContext *Context = DAG.getContext();
6376 DebugLoc dl = Op.getDebugLoc();
6377 EVT OpVT = Op.getValueType();
6379 // If this is a 256-bit vector result, first insert into a 128-bit
6380 // vector and then insert into the 256-bit vector.
6381 if (OpVT.getSizeInBits() > 128) {
6382 // Insert into a 128-bit vector.
6383 EVT VT128 = EVT::getVectorVT(*Context,
6384 OpVT.getVectorElementType(),
6385 OpVT.getVectorNumElements() / 2);
6387 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
6389 // Insert the 128-bit vector.
6390 return Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, OpVT), Op,
6391 DAG.getConstant(0, MVT::i32),
6395 if (Op.getValueType() == MVT::v1i64 &&
6396 Op.getOperand(0).getValueType() == MVT::i64)
6397 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
6399 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
6400 assert(Op.getValueType().getSimpleVT().getSizeInBits() == 128 &&
6401 "Expected an SSE type!");
6402 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(),
6403 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
6406 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
6407 // a simple subregister reference or explicit instructions to grab
6408 // upper bits of a vector.
6410 X86TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
6411 if (Subtarget->hasAVX()) {
6412 DebugLoc dl = Op.getNode()->getDebugLoc();
6413 SDValue Vec = Op.getNode()->getOperand(0);
6414 SDValue Idx = Op.getNode()->getOperand(1);
6416 if (Op.getNode()->getValueType(0).getSizeInBits() == 128
6417 && Vec.getNode()->getValueType(0).getSizeInBits() == 256) {
6418 return Extract128BitVector(Vec, Idx, DAG, dl);
6424 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
6425 // simple superregister reference or explicit instructions to insert
6426 // the upper bits of a vector.
6428 X86TargetLowering::LowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
6429 if (Subtarget->hasAVX()) {
6430 DebugLoc dl = Op.getNode()->getDebugLoc();
6431 SDValue Vec = Op.getNode()->getOperand(0);
6432 SDValue SubVec = Op.getNode()->getOperand(1);
6433 SDValue Idx = Op.getNode()->getOperand(2);
6435 if (Op.getNode()->getValueType(0).getSizeInBits() == 256
6436 && SubVec.getNode()->getValueType(0).getSizeInBits() == 128) {
6437 return Insert128BitVector(Vec, SubVec, Idx, DAG, dl);
6443 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
6444 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
6445 // one of the above mentioned nodes. It has to be wrapped because otherwise
6446 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
6447 // be used to form addressing mode. These wrapped nodes will be selected
6450 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
6451 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
6453 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6455 unsigned char OpFlag = 0;
6456 unsigned WrapperKind = X86ISD::Wrapper;
6457 CodeModel::Model M = getTargetMachine().getCodeModel();
6459 if (Subtarget->isPICStyleRIPRel() &&
6460 (M == CodeModel::Small || M == CodeModel::Kernel))
6461 WrapperKind = X86ISD::WrapperRIP;
6462 else if (Subtarget->isPICStyleGOT())
6463 OpFlag = X86II::MO_GOTOFF;
6464 else if (Subtarget->isPICStyleStubPIC())
6465 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6467 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
6469 CP->getOffset(), OpFlag);
6470 DebugLoc DL = CP->getDebugLoc();
6471 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6472 // With PIC, the address is actually $g + Offset.
6474 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6475 DAG.getNode(X86ISD::GlobalBaseReg,
6476 DebugLoc(), getPointerTy()),
6483 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
6484 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
6486 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6488 unsigned char OpFlag = 0;
6489 unsigned WrapperKind = X86ISD::Wrapper;
6490 CodeModel::Model M = getTargetMachine().getCodeModel();
6492 if (Subtarget->isPICStyleRIPRel() &&
6493 (M == CodeModel::Small || M == CodeModel::Kernel))
6494 WrapperKind = X86ISD::WrapperRIP;
6495 else if (Subtarget->isPICStyleGOT())
6496 OpFlag = X86II::MO_GOTOFF;
6497 else if (Subtarget->isPICStyleStubPIC())
6498 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6500 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
6502 DebugLoc DL = JT->getDebugLoc();
6503 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6505 // With PIC, the address is actually $g + Offset.
6507 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6508 DAG.getNode(X86ISD::GlobalBaseReg,
6509 DebugLoc(), getPointerTy()),
6516 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
6517 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
6519 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6521 unsigned char OpFlag = 0;
6522 unsigned WrapperKind = X86ISD::Wrapper;
6523 CodeModel::Model M = getTargetMachine().getCodeModel();
6525 if (Subtarget->isPICStyleRIPRel() &&
6526 (M == CodeModel::Small || M == CodeModel::Kernel))
6527 WrapperKind = X86ISD::WrapperRIP;
6528 else if (Subtarget->isPICStyleGOT())
6529 OpFlag = X86II::MO_GOTOFF;
6530 else if (Subtarget->isPICStyleStubPIC())
6531 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6533 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
6535 DebugLoc DL = Op.getDebugLoc();
6536 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6539 // With PIC, the address is actually $g + Offset.
6540 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
6541 !Subtarget->is64Bit()) {
6542 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6543 DAG.getNode(X86ISD::GlobalBaseReg,
6544 DebugLoc(), getPointerTy()),
6552 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
6553 // Create the TargetBlockAddressAddress node.
6554 unsigned char OpFlags =
6555 Subtarget->ClassifyBlockAddressReference();
6556 CodeModel::Model M = getTargetMachine().getCodeModel();
6557 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
6558 DebugLoc dl = Op.getDebugLoc();
6559 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
6560 /*isTarget=*/true, OpFlags);
6562 if (Subtarget->isPICStyleRIPRel() &&
6563 (M == CodeModel::Small || M == CodeModel::Kernel))
6564 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6566 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6568 // With PIC, the address is actually $g + Offset.
6569 if (isGlobalRelativeToPICBase(OpFlags)) {
6570 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6571 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6579 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
6581 SelectionDAG &DAG) const {
6582 // Create the TargetGlobalAddress node, folding in the constant
6583 // offset if it is legal.
6584 unsigned char OpFlags =
6585 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
6586 CodeModel::Model M = getTargetMachine().getCodeModel();
6588 if (OpFlags == X86II::MO_NO_FLAG &&
6589 X86::isOffsetSuitableForCodeModel(Offset, M)) {
6590 // A direct static reference to a global.
6591 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
6594 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
6597 if (Subtarget->isPICStyleRIPRel() &&
6598 (M == CodeModel::Small || M == CodeModel::Kernel))
6599 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6601 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6603 // With PIC, the address is actually $g + Offset.
6604 if (isGlobalRelativeToPICBase(OpFlags)) {
6605 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6606 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6610 // For globals that require a load from a stub to get the address, emit the
6612 if (isGlobalStubReference(OpFlags))
6613 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
6614 MachinePointerInfo::getGOT(), false, false, 0);
6616 // If there was a non-zero offset that we didn't fold, create an explicit
6619 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
6620 DAG.getConstant(Offset, getPointerTy()));
6626 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
6627 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
6628 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
6629 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
6633 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
6634 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
6635 unsigned char OperandFlags) {
6636 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6637 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6638 DebugLoc dl = GA->getDebugLoc();
6639 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6640 GA->getValueType(0),
6644 SDValue Ops[] = { Chain, TGA, *InFlag };
6645 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
6647 SDValue Ops[] = { Chain, TGA };
6648 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
6651 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
6652 MFI->setAdjustsStack(true);
6654 SDValue Flag = Chain.getValue(1);
6655 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
6658 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
6660 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6663 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
6664 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
6665 DAG.getNode(X86ISD::GlobalBaseReg,
6666 DebugLoc(), PtrVT), InFlag);
6667 InFlag = Chain.getValue(1);
6669 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
6672 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
6674 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6676 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
6677 X86::RAX, X86II::MO_TLSGD);
6680 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
6681 // "local exec" model.
6682 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6683 const EVT PtrVT, TLSModel::Model model,
6685 DebugLoc dl = GA->getDebugLoc();
6687 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
6688 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
6689 is64Bit ? 257 : 256));
6691 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
6692 DAG.getIntPtrConstant(0),
6693 MachinePointerInfo(Ptr), false, false, 0);
6695 unsigned char OperandFlags = 0;
6696 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
6698 unsigned WrapperKind = X86ISD::Wrapper;
6699 if (model == TLSModel::LocalExec) {
6700 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
6701 } else if (is64Bit) {
6702 assert(model == TLSModel::InitialExec);
6703 OperandFlags = X86II::MO_GOTTPOFF;
6704 WrapperKind = X86ISD::WrapperRIP;
6706 assert(model == TLSModel::InitialExec);
6707 OperandFlags = X86II::MO_INDNTPOFF;
6710 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
6712 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6713 GA->getValueType(0),
6714 GA->getOffset(), OperandFlags);
6715 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
6717 if (model == TLSModel::InitialExec)
6718 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
6719 MachinePointerInfo::getGOT(), false, false, 0);
6721 // The address of the thread local variable is the add of the thread
6722 // pointer with the offset of the variable.
6723 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
6727 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
6729 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
6730 const GlobalValue *GV = GA->getGlobal();
6732 if (Subtarget->isTargetELF()) {
6733 // TODO: implement the "local dynamic" model
6734 // TODO: implement the "initial exec"model for pic executables
6736 // If GV is an alias then use the aliasee for determining
6737 // thread-localness.
6738 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
6739 GV = GA->resolveAliasedGlobal(false);
6741 TLSModel::Model model
6742 = getTLSModel(GV, getTargetMachine().getRelocationModel());
6745 case TLSModel::GeneralDynamic:
6746 case TLSModel::LocalDynamic: // not implemented
6747 if (Subtarget->is64Bit())
6748 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
6749 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
6751 case TLSModel::InitialExec:
6752 case TLSModel::LocalExec:
6753 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
6754 Subtarget->is64Bit());
6756 } else if (Subtarget->isTargetDarwin()) {
6757 // Darwin only has one model of TLS. Lower to that.
6758 unsigned char OpFlag = 0;
6759 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
6760 X86ISD::WrapperRIP : X86ISD::Wrapper;
6762 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6764 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
6765 !Subtarget->is64Bit();
6767 OpFlag = X86II::MO_TLVP_PIC_BASE;
6769 OpFlag = X86II::MO_TLVP;
6770 DebugLoc DL = Op.getDebugLoc();
6771 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
6772 GA->getValueType(0),
6773 GA->getOffset(), OpFlag);
6774 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6776 // With PIC32, the address is actually $g + Offset.
6778 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6779 DAG.getNode(X86ISD::GlobalBaseReg,
6780 DebugLoc(), getPointerTy()),
6783 // Lowering the machine isd will make sure everything is in the right
6785 SDValue Chain = DAG.getEntryNode();
6786 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6787 SDValue Args[] = { Chain, Offset };
6788 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args, 2);
6790 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
6791 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6792 MFI->setAdjustsStack(true);
6794 // And our return value (tls address) is in the standard call return value
6796 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
6797 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
6801 "TLS not implemented for this target.");
6803 llvm_unreachable("Unreachable");
6808 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values and
6809 /// take a 2 x i32 value to shift plus a shift amount.
6810 SDValue X86TargetLowering::LowerShiftParts(SDValue Op, SelectionDAG &DAG) const {
6811 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
6812 EVT VT = Op.getValueType();
6813 unsigned VTBits = VT.getSizeInBits();
6814 DebugLoc dl = Op.getDebugLoc();
6815 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
6816 SDValue ShOpLo = Op.getOperand(0);
6817 SDValue ShOpHi = Op.getOperand(1);
6818 SDValue ShAmt = Op.getOperand(2);
6819 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
6820 DAG.getConstant(VTBits - 1, MVT::i8))
6821 : DAG.getConstant(0, VT);
6824 if (Op.getOpcode() == ISD::SHL_PARTS) {
6825 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
6826 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
6828 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
6829 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
6832 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
6833 DAG.getConstant(VTBits, MVT::i8));
6834 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6835 AndNode, DAG.getConstant(0, MVT::i8));
6838 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6839 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
6840 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
6842 if (Op.getOpcode() == ISD::SHL_PARTS) {
6843 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6844 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6846 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6847 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6850 SDValue Ops[2] = { Lo, Hi };
6851 return DAG.getMergeValues(Ops, 2, dl);
6854 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
6855 SelectionDAG &DAG) const {
6856 EVT SrcVT = Op.getOperand(0).getValueType();
6858 if (SrcVT.isVector())
6861 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
6862 "Unknown SINT_TO_FP to lower!");
6864 // These are really Legal; return the operand so the caller accepts it as
6866 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
6868 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
6869 Subtarget->is64Bit()) {
6873 DebugLoc dl = Op.getDebugLoc();
6874 unsigned Size = SrcVT.getSizeInBits()/8;
6875 MachineFunction &MF = DAG.getMachineFunction();
6876 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
6877 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6878 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6880 MachinePointerInfo::getFixedStack(SSFI),
6882 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
6885 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
6887 SelectionDAG &DAG) const {
6889 DebugLoc DL = Op.getDebugLoc();
6891 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
6893 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
6895 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
6897 unsigned ByteSize = SrcVT.getSizeInBits()/8;
6899 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
6900 MachineMemOperand *MMO;
6902 int SSFI = FI->getIndex();
6904 DAG.getMachineFunction()
6905 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6906 MachineMemOperand::MOLoad, ByteSize, ByteSize);
6908 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
6909 StackSlot = StackSlot.getOperand(1);
6911 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
6912 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
6914 Tys, Ops, array_lengthof(Ops),
6918 Chain = Result.getValue(1);
6919 SDValue InFlag = Result.getValue(2);
6921 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
6922 // shouldn't be necessary except that RFP cannot be live across
6923 // multiple blocks. When stackifier is fixed, they can be uncoupled.
6924 MachineFunction &MF = DAG.getMachineFunction();
6925 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
6926 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
6927 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6928 Tys = DAG.getVTList(MVT::Other);
6930 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
6932 MachineMemOperand *MMO =
6933 DAG.getMachineFunction()
6934 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6935 MachineMemOperand::MOStore, SSFISize, SSFISize);
6937 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
6938 Ops, array_lengthof(Ops),
6939 Op.getValueType(), MMO);
6940 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
6941 MachinePointerInfo::getFixedStack(SSFI),
6948 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
6949 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
6950 SelectionDAG &DAG) const {
6951 // This algorithm is not obvious. Here it is in C code, more or less:
6953 double uint64_to_double( uint32_t hi, uint32_t lo ) {
6954 static const __m128i exp = { 0x4330000045300000ULL, 0 };
6955 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
6957 // Copy ints to xmm registers.
6958 __m128i xh = _mm_cvtsi32_si128( hi );
6959 __m128i xl = _mm_cvtsi32_si128( lo );
6961 // Combine into low half of a single xmm register.
6962 __m128i x = _mm_unpacklo_epi32( xh, xl );
6966 // Merge in appropriate exponents to give the integer bits the right
6968 x = _mm_unpacklo_epi32( x, exp );
6970 // Subtract away the biases to deal with the IEEE-754 double precision
6972 d = _mm_sub_pd( (__m128d) x, bias );
6974 // All conversions up to here are exact. The correctly rounded result is
6975 // calculated using the current rounding mode using the following
6977 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
6978 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
6979 // store doesn't really need to be here (except
6980 // maybe to zero the other double)
6985 DebugLoc dl = Op.getDebugLoc();
6986 LLVMContext *Context = DAG.getContext();
6988 // Build some magic constants.
6989 std::vector<Constant*> CV0;
6990 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
6991 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
6992 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6993 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6994 Constant *C0 = ConstantVector::get(CV0);
6995 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
6997 std::vector<Constant*> CV1;
6999 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
7001 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
7002 Constant *C1 = ConstantVector::get(CV1);
7003 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
7005 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
7006 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7008 DAG.getIntPtrConstant(1)));
7009 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
7010 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7012 DAG.getIntPtrConstant(0)));
7013 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
7014 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
7015 MachinePointerInfo::getConstantPool(),
7017 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
7018 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck2);
7019 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
7020 MachinePointerInfo::getConstantPool(),
7022 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
7024 // Add the halves; easiest way is to swap them into another reg first.
7025 int ShufMask[2] = { 1, -1 };
7026 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
7027 DAG.getUNDEF(MVT::v2f64), ShufMask);
7028 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
7029 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
7030 DAG.getIntPtrConstant(0));
7033 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
7034 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
7035 SelectionDAG &DAG) const {
7036 DebugLoc dl = Op.getDebugLoc();
7037 // FP constant to bias correct the final result.
7038 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
7041 // Load the 32-bit value into an XMM register.
7042 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
7043 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7045 DAG.getIntPtrConstant(0)));
7047 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7048 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
7049 DAG.getIntPtrConstant(0));
7051 // Or the load with the bias.
7052 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
7053 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7054 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
7056 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7057 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
7058 MVT::v2f64, Bias)));
7059 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7060 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
7061 DAG.getIntPtrConstant(0));
7063 // Subtract the bias.
7064 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
7066 // Handle final rounding.
7067 EVT DestVT = Op.getValueType();
7069 if (DestVT.bitsLT(MVT::f64)) {
7070 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
7071 DAG.getIntPtrConstant(0));
7072 } else if (DestVT.bitsGT(MVT::f64)) {
7073 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
7076 // Handle final rounding.
7080 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
7081 SelectionDAG &DAG) const {
7082 SDValue N0 = Op.getOperand(0);
7083 DebugLoc dl = Op.getDebugLoc();
7085 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
7086 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
7087 // the optimization here.
7088 if (DAG.SignBitIsZero(N0))
7089 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
7091 EVT SrcVT = N0.getValueType();
7092 EVT DstVT = Op.getValueType();
7093 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
7094 return LowerUINT_TO_FP_i64(Op, DAG);
7095 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
7096 return LowerUINT_TO_FP_i32(Op, DAG);
7098 // Make a 64-bit buffer, and use it to build an FILD.
7099 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
7100 if (SrcVT == MVT::i32) {
7101 SDValue WordOff = DAG.getConstant(4, getPointerTy());
7102 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
7103 getPointerTy(), StackSlot, WordOff);
7104 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7105 StackSlot, MachinePointerInfo(),
7107 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
7108 OffsetSlot, MachinePointerInfo(),
7110 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
7114 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
7115 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7116 StackSlot, MachinePointerInfo(),
7118 // For i64 source, we need to add the appropriate power of 2 if the input
7119 // was negative. This is the same as the optimization in
7120 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
7121 // we must be careful to do the computation in x87 extended precision, not
7122 // in SSE. (The generic code can't know it's OK to do this, or how to.)
7123 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
7124 MachineMemOperand *MMO =
7125 DAG.getMachineFunction()
7126 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7127 MachineMemOperand::MOLoad, 8, 8);
7129 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
7130 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
7131 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops, 3,
7134 APInt FF(32, 0x5F800000ULL);
7136 // Check whether the sign bit is set.
7137 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
7138 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
7141 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
7142 SDValue FudgePtr = DAG.getConstantPool(
7143 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
7146 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
7147 SDValue Zero = DAG.getIntPtrConstant(0);
7148 SDValue Four = DAG.getIntPtrConstant(4);
7149 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
7151 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
7153 // Load the value out, extending it from f32 to f80.
7154 // FIXME: Avoid the extend by constructing the right constant pool?
7155 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
7156 FudgePtr, MachinePointerInfo::getConstantPool(),
7157 MVT::f32, false, false, 4);
7158 // Extend everything to 80 bits to force it to be done on x87.
7159 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
7160 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
7163 std::pair<SDValue,SDValue> X86TargetLowering::
7164 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
7165 DebugLoc DL = Op.getDebugLoc();
7167 EVT DstTy = Op.getValueType();
7170 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
7174 assert(DstTy.getSimpleVT() <= MVT::i64 &&
7175 DstTy.getSimpleVT() >= MVT::i16 &&
7176 "Unknown FP_TO_SINT to lower!");
7178 // These are really Legal.
7179 if (DstTy == MVT::i32 &&
7180 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
7181 return std::make_pair(SDValue(), SDValue());
7182 if (Subtarget->is64Bit() &&
7183 DstTy == MVT::i64 &&
7184 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
7185 return std::make_pair(SDValue(), SDValue());
7187 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
7189 MachineFunction &MF = DAG.getMachineFunction();
7190 unsigned MemSize = DstTy.getSizeInBits()/8;
7191 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
7192 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7197 switch (DstTy.getSimpleVT().SimpleTy) {
7198 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
7199 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
7200 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
7201 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
7204 SDValue Chain = DAG.getEntryNode();
7205 SDValue Value = Op.getOperand(0);
7206 EVT TheVT = Op.getOperand(0).getValueType();
7207 if (isScalarFPTypeInSSEReg(TheVT)) {
7208 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
7209 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
7210 MachinePointerInfo::getFixedStack(SSFI),
7212 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
7214 Chain, StackSlot, DAG.getValueType(TheVT)
7217 MachineMemOperand *MMO =
7218 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7219 MachineMemOperand::MOLoad, MemSize, MemSize);
7220 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, 3,
7222 Chain = Value.getValue(1);
7223 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
7224 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7227 MachineMemOperand *MMO =
7228 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7229 MachineMemOperand::MOStore, MemSize, MemSize);
7231 // Build the FP_TO_INT*_IN_MEM
7232 SDValue Ops[] = { Chain, Value, StackSlot };
7233 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
7234 Ops, 3, DstTy, MMO);
7236 return std::make_pair(FIST, StackSlot);
7239 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
7240 SelectionDAG &DAG) const {
7241 if (Op.getValueType().isVector())
7244 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
7245 SDValue FIST = Vals.first, StackSlot = Vals.second;
7246 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
7247 if (FIST.getNode() == 0) return Op;
7250 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
7251 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
7254 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
7255 SelectionDAG &DAG) const {
7256 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
7257 SDValue FIST = Vals.first, StackSlot = Vals.second;
7258 assert(FIST.getNode() && "Unexpected failure");
7261 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
7262 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
7265 SDValue X86TargetLowering::LowerFABS(SDValue Op,
7266 SelectionDAG &DAG) const {
7267 LLVMContext *Context = DAG.getContext();
7268 DebugLoc dl = Op.getDebugLoc();
7269 EVT VT = Op.getValueType();
7272 EltVT = VT.getVectorElementType();
7273 std::vector<Constant*> CV;
7274 if (EltVT == MVT::f64) {
7275 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
7279 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
7285 Constant *C = ConstantVector::get(CV);
7286 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7287 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7288 MachinePointerInfo::getConstantPool(),
7290 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
7293 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
7294 LLVMContext *Context = DAG.getContext();
7295 DebugLoc dl = Op.getDebugLoc();
7296 EVT VT = Op.getValueType();
7299 EltVT = VT.getVectorElementType();
7300 std::vector<Constant*> CV;
7301 if (EltVT == MVT::f64) {
7302 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
7306 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
7312 Constant *C = ConstantVector::get(CV);
7313 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7314 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7315 MachinePointerInfo::getConstantPool(),
7317 if (VT.isVector()) {
7318 return DAG.getNode(ISD::BITCAST, dl, VT,
7319 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
7320 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7322 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, Mask)));
7324 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
7328 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
7329 LLVMContext *Context = DAG.getContext();
7330 SDValue Op0 = Op.getOperand(0);
7331 SDValue Op1 = Op.getOperand(1);
7332 DebugLoc dl = Op.getDebugLoc();
7333 EVT VT = Op.getValueType();
7334 EVT SrcVT = Op1.getValueType();
7336 // If second operand is smaller, extend it first.
7337 if (SrcVT.bitsLT(VT)) {
7338 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
7341 // And if it is bigger, shrink it first.
7342 if (SrcVT.bitsGT(VT)) {
7343 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
7347 // At this point the operands and the result should have the same
7348 // type, and that won't be f80 since that is not custom lowered.
7350 // First get the sign bit of second operand.
7351 std::vector<Constant*> CV;
7352 if (SrcVT == MVT::f64) {
7353 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
7354 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
7356 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
7357 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7358 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7359 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7361 Constant *C = ConstantVector::get(CV);
7362 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7363 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
7364 MachinePointerInfo::getConstantPool(),
7366 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
7368 // Shift sign bit right or left if the two operands have different types.
7369 if (SrcVT.bitsGT(VT)) {
7370 // Op0 is MVT::f32, Op1 is MVT::f64.
7371 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
7372 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
7373 DAG.getConstant(32, MVT::i32));
7374 SignBit = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, SignBit);
7375 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
7376 DAG.getIntPtrConstant(0));
7379 // Clear first operand sign bit.
7381 if (VT == MVT::f64) {
7382 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
7383 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
7385 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
7386 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7387 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7388 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7390 C = ConstantVector::get(CV);
7391 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7392 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7393 MachinePointerInfo::getConstantPool(),
7395 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
7397 // Or the value with the sign bit.
7398 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
7401 SDValue X86TargetLowering::LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) const {
7402 SDValue N0 = Op.getOperand(0);
7403 DebugLoc dl = Op.getDebugLoc();
7404 EVT VT = Op.getValueType();
7406 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
7407 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
7408 DAG.getConstant(1, VT));
7409 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT));
7412 /// Emit nodes that will be selected as "test Op0,Op0", or something
7414 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
7415 SelectionDAG &DAG) const {
7416 DebugLoc dl = Op.getDebugLoc();
7418 // CF and OF aren't always set the way we want. Determine which
7419 // of these we need.
7420 bool NeedCF = false;
7421 bool NeedOF = false;
7424 case X86::COND_A: case X86::COND_AE:
7425 case X86::COND_B: case X86::COND_BE:
7428 case X86::COND_G: case X86::COND_GE:
7429 case X86::COND_L: case X86::COND_LE:
7430 case X86::COND_O: case X86::COND_NO:
7435 // See if we can use the EFLAGS value from the operand instead of
7436 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
7437 // we prove that the arithmetic won't overflow, we can't use OF or CF.
7438 if (Op.getResNo() != 0 || NeedOF || NeedCF)
7439 // Emit a CMP with 0, which is the TEST pattern.
7440 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
7441 DAG.getConstant(0, Op.getValueType()));
7443 unsigned Opcode = 0;
7444 unsigned NumOperands = 0;
7445 switch (Op.getNode()->getOpcode()) {
7447 // Due to an isel shortcoming, be conservative if this add is likely to be
7448 // selected as part of a load-modify-store instruction. When the root node
7449 // in a match is a store, isel doesn't know how to remap non-chain non-flag
7450 // uses of other nodes in the match, such as the ADD in this case. This
7451 // leads to the ADD being left around and reselected, with the result being
7452 // two adds in the output. Alas, even if none our users are stores, that
7453 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
7454 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
7455 // climbing the DAG back to the root, and it doesn't seem to be worth the
7457 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7458 UE = Op.getNode()->use_end(); UI != UE; ++UI)
7459 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
7462 if (ConstantSDNode *C =
7463 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
7464 // An add of one will be selected as an INC.
7465 if (C->getAPIntValue() == 1) {
7466 Opcode = X86ISD::INC;
7471 // An add of negative one (subtract of one) will be selected as a DEC.
7472 if (C->getAPIntValue().isAllOnesValue()) {
7473 Opcode = X86ISD::DEC;
7479 // Otherwise use a regular EFLAGS-setting add.
7480 Opcode = X86ISD::ADD;
7484 // If the primary and result isn't used, don't bother using X86ISD::AND,
7485 // because a TEST instruction will be better.
7486 bool NonFlagUse = false;
7487 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7488 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
7490 unsigned UOpNo = UI.getOperandNo();
7491 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
7492 // Look pass truncate.
7493 UOpNo = User->use_begin().getOperandNo();
7494 User = *User->use_begin();
7497 if (User->getOpcode() != ISD::BRCOND &&
7498 User->getOpcode() != ISD::SETCC &&
7499 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
7512 // Due to the ISEL shortcoming noted above, be conservative if this op is
7513 // likely to be selected as part of a load-modify-store instruction.
7514 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7515 UE = Op.getNode()->use_end(); UI != UE; ++UI)
7516 if (UI->getOpcode() == ISD::STORE)
7519 // Otherwise use a regular EFLAGS-setting instruction.
7520 switch (Op.getNode()->getOpcode()) {
7521 default: llvm_unreachable("unexpected operator!");
7522 case ISD::SUB: Opcode = X86ISD::SUB; break;
7523 case ISD::OR: Opcode = X86ISD::OR; break;
7524 case ISD::XOR: Opcode = X86ISD::XOR; break;
7525 case ISD::AND: Opcode = X86ISD::AND; break;
7537 return SDValue(Op.getNode(), 1);
7544 // Emit a CMP with 0, which is the TEST pattern.
7545 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
7546 DAG.getConstant(0, Op.getValueType()));
7548 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
7549 SmallVector<SDValue, 4> Ops;
7550 for (unsigned i = 0; i != NumOperands; ++i)
7551 Ops.push_back(Op.getOperand(i));
7553 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
7554 DAG.ReplaceAllUsesWith(Op, New);
7555 return SDValue(New.getNode(), 1);
7558 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
7560 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
7561 SelectionDAG &DAG) const {
7562 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
7563 if (C->getAPIntValue() == 0)
7564 return EmitTest(Op0, X86CC, DAG);
7566 DebugLoc dl = Op0.getDebugLoc();
7567 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
7570 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
7571 /// if it's possible.
7572 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
7573 DebugLoc dl, SelectionDAG &DAG) const {
7574 SDValue Op0 = And.getOperand(0);
7575 SDValue Op1 = And.getOperand(1);
7576 if (Op0.getOpcode() == ISD::TRUNCATE)
7577 Op0 = Op0.getOperand(0);
7578 if (Op1.getOpcode() == ISD::TRUNCATE)
7579 Op1 = Op1.getOperand(0);
7582 if (Op1.getOpcode() == ISD::SHL)
7583 std::swap(Op0, Op1);
7584 if (Op0.getOpcode() == ISD::SHL) {
7585 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
7586 if (And00C->getZExtValue() == 1) {
7587 // If we looked past a truncate, check that it's only truncating away
7589 unsigned BitWidth = Op0.getValueSizeInBits();
7590 unsigned AndBitWidth = And.getValueSizeInBits();
7591 if (BitWidth > AndBitWidth) {
7592 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
7593 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
7594 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
7598 RHS = Op0.getOperand(1);
7600 } else if (Op1.getOpcode() == ISD::Constant) {
7601 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
7602 SDValue AndLHS = Op0;
7603 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
7604 LHS = AndLHS.getOperand(0);
7605 RHS = AndLHS.getOperand(1);
7609 if (LHS.getNode()) {
7610 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
7611 // instruction. Since the shift amount is in-range-or-undefined, we know
7612 // that doing a bittest on the i32 value is ok. We extend to i32 because
7613 // the encoding for the i16 version is larger than the i32 version.
7614 // Also promote i16 to i32 for performance / code size reason.
7615 if (LHS.getValueType() == MVT::i8 ||
7616 LHS.getValueType() == MVT::i16)
7617 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
7619 // If the operand types disagree, extend the shift amount to match. Since
7620 // BT ignores high bits (like shifts) we can use anyextend.
7621 if (LHS.getValueType() != RHS.getValueType())
7622 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
7624 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
7625 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
7626 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7627 DAG.getConstant(Cond, MVT::i8), BT);
7633 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
7634 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
7635 SDValue Op0 = Op.getOperand(0);
7636 SDValue Op1 = Op.getOperand(1);
7637 DebugLoc dl = Op.getDebugLoc();
7638 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
7640 // Optimize to BT if possible.
7641 // Lower (X & (1 << N)) == 0 to BT(X, N).
7642 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
7643 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
7644 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
7645 Op1.getOpcode() == ISD::Constant &&
7646 cast<ConstantSDNode>(Op1)->isNullValue() &&
7647 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7648 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
7649 if (NewSetCC.getNode())
7653 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
7655 if (Op1.getOpcode() == ISD::Constant &&
7656 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
7657 cast<ConstantSDNode>(Op1)->isNullValue()) &&
7658 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7660 // If the input is a setcc, then reuse the input setcc or use a new one with
7661 // the inverted condition.
7662 if (Op0.getOpcode() == X86ISD::SETCC) {
7663 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
7664 bool Invert = (CC == ISD::SETNE) ^
7665 cast<ConstantSDNode>(Op1)->isNullValue();
7666 if (!Invert) return Op0;
7668 CCode = X86::GetOppositeBranchCondition(CCode);
7669 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7670 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
7674 bool isFP = Op1.getValueType().isFloatingPoint();
7675 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
7676 if (X86CC == X86::COND_INVALID)
7679 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
7680 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7681 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
7684 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
7686 SDValue Op0 = Op.getOperand(0);
7687 SDValue Op1 = Op.getOperand(1);
7688 SDValue CC = Op.getOperand(2);
7689 EVT VT = Op.getValueType();
7690 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
7691 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
7692 DebugLoc dl = Op.getDebugLoc();
7696 EVT VT0 = Op0.getValueType();
7697 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
7698 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
7701 switch (SetCCOpcode) {
7704 case ISD::SETEQ: SSECC = 0; break;
7706 case ISD::SETGT: Swap = true; // Fallthrough
7708 case ISD::SETOLT: SSECC = 1; break;
7710 case ISD::SETGE: Swap = true; // Fallthrough
7712 case ISD::SETOLE: SSECC = 2; break;
7713 case ISD::SETUO: SSECC = 3; break;
7715 case ISD::SETNE: SSECC = 4; break;
7716 case ISD::SETULE: Swap = true;
7717 case ISD::SETUGE: SSECC = 5; break;
7718 case ISD::SETULT: Swap = true;
7719 case ISD::SETUGT: SSECC = 6; break;
7720 case ISD::SETO: SSECC = 7; break;
7723 std::swap(Op0, Op1);
7725 // In the two special cases we can't handle, emit two comparisons.
7727 if (SetCCOpcode == ISD::SETUEQ) {
7729 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
7730 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
7731 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
7733 else if (SetCCOpcode == ISD::SETONE) {
7735 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
7736 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
7737 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
7739 llvm_unreachable("Illegal FP comparison");
7741 // Handle all other FP comparisons here.
7742 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
7745 // We are handling one of the integer comparisons here. Since SSE only has
7746 // GT and EQ comparisons for integer, swapping operands and multiple
7747 // operations may be required for some comparisons.
7748 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
7749 bool Swap = false, Invert = false, FlipSigns = false;
7751 switch (VT.getSimpleVT().SimpleTy) {
7753 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
7754 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
7755 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
7756 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
7759 switch (SetCCOpcode) {
7761 case ISD::SETNE: Invert = true;
7762 case ISD::SETEQ: Opc = EQOpc; break;
7763 case ISD::SETLT: Swap = true;
7764 case ISD::SETGT: Opc = GTOpc; break;
7765 case ISD::SETGE: Swap = true;
7766 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
7767 case ISD::SETULT: Swap = true;
7768 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
7769 case ISD::SETUGE: Swap = true;
7770 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
7773 std::swap(Op0, Op1);
7775 // Since SSE has no unsigned integer comparisons, we need to flip the sign
7776 // bits of the inputs before performing those operations.
7778 EVT EltVT = VT.getVectorElementType();
7779 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
7781 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
7782 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
7784 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
7785 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
7788 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
7790 // If the logical-not of the result is required, perform that now.
7792 Result = DAG.getNOT(dl, Result, VT);
7797 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
7798 static bool isX86LogicalCmp(SDValue Op) {
7799 unsigned Opc = Op.getNode()->getOpcode();
7800 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
7802 if (Op.getResNo() == 1 &&
7803 (Opc == X86ISD::ADD ||
7804 Opc == X86ISD::SUB ||
7805 Opc == X86ISD::ADC ||
7806 Opc == X86ISD::SBB ||
7807 Opc == X86ISD::SMUL ||
7808 Opc == X86ISD::UMUL ||
7809 Opc == X86ISD::INC ||
7810 Opc == X86ISD::DEC ||
7811 Opc == X86ISD::OR ||
7812 Opc == X86ISD::XOR ||
7813 Opc == X86ISD::AND))
7816 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
7822 static bool isZero(SDValue V) {
7823 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
7824 return C && C->isNullValue();
7827 static bool isAllOnes(SDValue V) {
7828 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
7829 return C && C->isAllOnesValue();
7832 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
7833 bool addTest = true;
7834 SDValue Cond = Op.getOperand(0);
7835 SDValue Op1 = Op.getOperand(1);
7836 SDValue Op2 = Op.getOperand(2);
7837 DebugLoc DL = Op.getDebugLoc();
7840 if (Cond.getOpcode() == ISD::SETCC) {
7841 SDValue NewCond = LowerSETCC(Cond, DAG);
7842 if (NewCond.getNode())
7846 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
7847 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
7848 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
7849 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
7850 if (Cond.getOpcode() == X86ISD::SETCC &&
7851 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
7852 isZero(Cond.getOperand(1).getOperand(1))) {
7853 SDValue Cmp = Cond.getOperand(1);
7855 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
7857 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
7858 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
7859 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
7861 SDValue CmpOp0 = Cmp.getOperand(0);
7862 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
7863 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
7865 SDValue Res = // Res = 0 or -1.
7866 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
7867 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
7869 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
7870 Res = DAG.getNOT(DL, Res, Res.getValueType());
7872 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
7873 if (N2C == 0 || !N2C->isNullValue())
7874 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
7879 // Look past (and (setcc_carry (cmp ...)), 1).
7880 if (Cond.getOpcode() == ISD::AND &&
7881 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7882 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7883 if (C && C->getAPIntValue() == 1)
7884 Cond = Cond.getOperand(0);
7887 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7888 // setting operand in place of the X86ISD::SETCC.
7889 if (Cond.getOpcode() == X86ISD::SETCC ||
7890 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7891 CC = Cond.getOperand(0);
7893 SDValue Cmp = Cond.getOperand(1);
7894 unsigned Opc = Cmp.getOpcode();
7895 EVT VT = Op.getValueType();
7897 bool IllegalFPCMov = false;
7898 if (VT.isFloatingPoint() && !VT.isVector() &&
7899 !isScalarFPTypeInSSEReg(VT)) // FPStack?
7900 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
7902 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
7903 Opc == X86ISD::BT) { // FIXME
7910 // Look pass the truncate.
7911 if (Cond.getOpcode() == ISD::TRUNCATE)
7912 Cond = Cond.getOperand(0);
7914 // We know the result of AND is compared against zero. Try to match
7916 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7917 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
7918 if (NewSetCC.getNode()) {
7919 CC = NewSetCC.getOperand(0);
7920 Cond = NewSetCC.getOperand(1);
7927 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7928 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7931 // a < b ? -1 : 0 -> RES = ~setcc_carry
7932 // a < b ? 0 : -1 -> RES = setcc_carry
7933 // a >= b ? -1 : 0 -> RES = setcc_carry
7934 // a >= b ? 0 : -1 -> RES = ~setcc_carry
7935 if (Cond.getOpcode() == X86ISD::CMP) {
7936 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
7938 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
7939 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
7940 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
7941 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
7942 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
7943 return DAG.getNOT(DL, Res, Res.getValueType());
7948 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
7949 // condition is true.
7950 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
7951 SDValue Ops[] = { Op2, Op1, CC, Cond };
7952 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops, array_lengthof(Ops));
7955 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
7956 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
7957 // from the AND / OR.
7958 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
7959 Opc = Op.getOpcode();
7960 if (Opc != ISD::OR && Opc != ISD::AND)
7962 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7963 Op.getOperand(0).hasOneUse() &&
7964 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
7965 Op.getOperand(1).hasOneUse());
7968 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
7969 // 1 and that the SETCC node has a single use.
7970 static bool isXor1OfSetCC(SDValue Op) {
7971 if (Op.getOpcode() != ISD::XOR)
7973 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7974 if (N1C && N1C->getAPIntValue() == 1) {
7975 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7976 Op.getOperand(0).hasOneUse();
7981 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
7982 bool addTest = true;
7983 SDValue Chain = Op.getOperand(0);
7984 SDValue Cond = Op.getOperand(1);
7985 SDValue Dest = Op.getOperand(2);
7986 DebugLoc dl = Op.getDebugLoc();
7989 if (Cond.getOpcode() == ISD::SETCC) {
7990 SDValue NewCond = LowerSETCC(Cond, DAG);
7991 if (NewCond.getNode())
7995 // FIXME: LowerXALUO doesn't handle these!!
7996 else if (Cond.getOpcode() == X86ISD::ADD ||
7997 Cond.getOpcode() == X86ISD::SUB ||
7998 Cond.getOpcode() == X86ISD::SMUL ||
7999 Cond.getOpcode() == X86ISD::UMUL)
8000 Cond = LowerXALUO(Cond, DAG);
8003 // Look pass (and (setcc_carry (cmp ...)), 1).
8004 if (Cond.getOpcode() == ISD::AND &&
8005 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
8006 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
8007 if (C && C->getAPIntValue() == 1)
8008 Cond = Cond.getOperand(0);
8011 // If condition flag is set by a X86ISD::CMP, then use it as the condition
8012 // setting operand in place of the X86ISD::SETCC.
8013 if (Cond.getOpcode() == X86ISD::SETCC ||
8014 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
8015 CC = Cond.getOperand(0);
8017 SDValue Cmp = Cond.getOperand(1);
8018 unsigned Opc = Cmp.getOpcode();
8019 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
8020 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
8024 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
8028 // These can only come from an arithmetic instruction with overflow,
8029 // e.g. SADDO, UADDO.
8030 Cond = Cond.getNode()->getOperand(1);
8037 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
8038 SDValue Cmp = Cond.getOperand(0).getOperand(1);
8039 if (CondOpc == ISD::OR) {
8040 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
8041 // two branches instead of an explicit OR instruction with a
8043 if (Cmp == Cond.getOperand(1).getOperand(1) &&
8044 isX86LogicalCmp(Cmp)) {
8045 CC = Cond.getOperand(0).getOperand(0);
8046 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8047 Chain, Dest, CC, Cmp);
8048 CC = Cond.getOperand(1).getOperand(0);
8052 } else { // ISD::AND
8053 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
8054 // two branches instead of an explicit AND instruction with a
8055 // separate test. However, we only do this if this block doesn't
8056 // have a fall-through edge, because this requires an explicit
8057 // jmp when the condition is false.
8058 if (Cmp == Cond.getOperand(1).getOperand(1) &&
8059 isX86LogicalCmp(Cmp) &&
8060 Op.getNode()->hasOneUse()) {
8061 X86::CondCode CCode =
8062 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
8063 CCode = X86::GetOppositeBranchCondition(CCode);
8064 CC = DAG.getConstant(CCode, MVT::i8);
8065 SDNode *User = *Op.getNode()->use_begin();
8066 // Look for an unconditional branch following this conditional branch.
8067 // We need this because we need to reverse the successors in order
8068 // to implement FCMP_OEQ.
8069 if (User->getOpcode() == ISD::BR) {
8070 SDValue FalseBB = User->getOperand(1);
8072 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
8073 assert(NewBR == User);
8077 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8078 Chain, Dest, CC, Cmp);
8079 X86::CondCode CCode =
8080 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
8081 CCode = X86::GetOppositeBranchCondition(CCode);
8082 CC = DAG.getConstant(CCode, MVT::i8);
8088 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
8089 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
8090 // It should be transformed during dag combiner except when the condition
8091 // is set by a arithmetics with overflow node.
8092 X86::CondCode CCode =
8093 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
8094 CCode = X86::GetOppositeBranchCondition(CCode);
8095 CC = DAG.getConstant(CCode, MVT::i8);
8096 Cond = Cond.getOperand(0).getOperand(1);
8102 // Look pass the truncate.
8103 if (Cond.getOpcode() == ISD::TRUNCATE)
8104 Cond = Cond.getOperand(0);
8106 // We know the result of AND is compared against zero. Try to match
8108 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
8109 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
8110 if (NewSetCC.getNode()) {
8111 CC = NewSetCC.getOperand(0);
8112 Cond = NewSetCC.getOperand(1);
8119 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
8120 Cond = EmitTest(Cond, X86::COND_NE, DAG);
8122 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8123 Chain, Dest, CC, Cond);
8127 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
8128 // Calls to _alloca is needed to probe the stack when allocating more than 4k
8129 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
8130 // that the guard pages used by the OS virtual memory manager are allocated in
8131 // correct sequence.
8133 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
8134 SelectionDAG &DAG) const {
8135 assert((Subtarget->isTargetCygMing() || Subtarget->isTargetWindows()) &&
8136 "This should be used only on Windows targets");
8137 assert(!Subtarget->isTargetEnvMacho());
8138 DebugLoc dl = Op.getDebugLoc();
8141 SDValue Chain = Op.getOperand(0);
8142 SDValue Size = Op.getOperand(1);
8143 // FIXME: Ensure alignment here
8147 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
8148 unsigned Reg = (Subtarget->is64Bit() ? X86::RAX : X86::EAX);
8150 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
8151 Flag = Chain.getValue(1);
8153 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
8155 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
8156 Flag = Chain.getValue(1);
8158 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
8160 SDValue Ops1[2] = { Chain.getValue(0), Chain };
8161 return DAG.getMergeValues(Ops1, 2, dl);
8164 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
8165 MachineFunction &MF = DAG.getMachineFunction();
8166 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
8168 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
8169 DebugLoc DL = Op.getDebugLoc();
8171 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
8172 // vastart just stores the address of the VarArgsFrameIndex slot into the
8173 // memory location argument.
8174 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
8176 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
8177 MachinePointerInfo(SV), false, false, 0);
8181 // gp_offset (0 - 6 * 8)
8182 // fp_offset (48 - 48 + 8 * 16)
8183 // overflow_arg_area (point to parameters coming in memory).
8185 SmallVector<SDValue, 8> MemOps;
8186 SDValue FIN = Op.getOperand(1);
8188 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
8189 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
8191 FIN, MachinePointerInfo(SV), false, false, 0);
8192 MemOps.push_back(Store);
8195 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8196 FIN, DAG.getIntPtrConstant(4));
8197 Store = DAG.getStore(Op.getOperand(0), DL,
8198 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
8200 FIN, MachinePointerInfo(SV, 4), false, false, 0);
8201 MemOps.push_back(Store);
8203 // Store ptr to overflow_arg_area
8204 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8205 FIN, DAG.getIntPtrConstant(4));
8206 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
8208 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
8209 MachinePointerInfo(SV, 8),
8211 MemOps.push_back(Store);
8213 // Store ptr to reg_save_area.
8214 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8215 FIN, DAG.getIntPtrConstant(8));
8216 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
8218 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
8219 MachinePointerInfo(SV, 16), false, false, 0);
8220 MemOps.push_back(Store);
8221 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
8222 &MemOps[0], MemOps.size());
8225 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
8226 assert(Subtarget->is64Bit() &&
8227 "LowerVAARG only handles 64-bit va_arg!");
8228 assert((Subtarget->isTargetLinux() ||
8229 Subtarget->isTargetDarwin()) &&
8230 "Unhandled target in LowerVAARG");
8231 assert(Op.getNode()->getNumOperands() == 4);
8232 SDValue Chain = Op.getOperand(0);
8233 SDValue SrcPtr = Op.getOperand(1);
8234 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
8235 unsigned Align = Op.getConstantOperandVal(3);
8236 DebugLoc dl = Op.getDebugLoc();
8238 EVT ArgVT = Op.getNode()->getValueType(0);
8239 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
8240 uint32_t ArgSize = getTargetData()->getTypeAllocSize(ArgTy);
8243 // Decide which area this value should be read from.
8244 // TODO: Implement the AMD64 ABI in its entirety. This simple
8245 // selection mechanism works only for the basic types.
8246 if (ArgVT == MVT::f80) {
8247 llvm_unreachable("va_arg for f80 not yet implemented");
8248 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
8249 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
8250 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
8251 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
8253 llvm_unreachable("Unhandled argument type in LowerVAARG");
8257 // Sanity Check: Make sure using fp_offset makes sense.
8258 assert(!UseSoftFloat &&
8259 !(DAG.getMachineFunction()
8260 .getFunction()->hasFnAttr(Attribute::NoImplicitFloat)) &&
8261 Subtarget->hasXMM());
8264 // Insert VAARG_64 node into the DAG
8265 // VAARG_64 returns two values: Variable Argument Address, Chain
8266 SmallVector<SDValue, 11> InstOps;
8267 InstOps.push_back(Chain);
8268 InstOps.push_back(SrcPtr);
8269 InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
8270 InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
8271 InstOps.push_back(DAG.getConstant(Align, MVT::i32));
8272 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
8273 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
8274 VTs, &InstOps[0], InstOps.size(),
8276 MachinePointerInfo(SV),
8281 Chain = VAARG.getValue(1);
8283 // Load the next argument and return it
8284 return DAG.getLoad(ArgVT, dl,
8287 MachinePointerInfo(),
8291 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
8292 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
8293 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
8294 SDValue Chain = Op.getOperand(0);
8295 SDValue DstPtr = Op.getOperand(1);
8296 SDValue SrcPtr = Op.getOperand(2);
8297 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
8298 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
8299 DebugLoc DL = Op.getDebugLoc();
8301 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
8302 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
8304 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
8308 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
8309 DebugLoc dl = Op.getDebugLoc();
8310 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8312 default: return SDValue(); // Don't custom lower most intrinsics.
8313 // Comparison intrinsics.
8314 case Intrinsic::x86_sse_comieq_ss:
8315 case Intrinsic::x86_sse_comilt_ss:
8316 case Intrinsic::x86_sse_comile_ss:
8317 case Intrinsic::x86_sse_comigt_ss:
8318 case Intrinsic::x86_sse_comige_ss:
8319 case Intrinsic::x86_sse_comineq_ss:
8320 case Intrinsic::x86_sse_ucomieq_ss:
8321 case Intrinsic::x86_sse_ucomilt_ss:
8322 case Intrinsic::x86_sse_ucomile_ss:
8323 case Intrinsic::x86_sse_ucomigt_ss:
8324 case Intrinsic::x86_sse_ucomige_ss:
8325 case Intrinsic::x86_sse_ucomineq_ss:
8326 case Intrinsic::x86_sse2_comieq_sd:
8327 case Intrinsic::x86_sse2_comilt_sd:
8328 case Intrinsic::x86_sse2_comile_sd:
8329 case Intrinsic::x86_sse2_comigt_sd:
8330 case Intrinsic::x86_sse2_comige_sd:
8331 case Intrinsic::x86_sse2_comineq_sd:
8332 case Intrinsic::x86_sse2_ucomieq_sd:
8333 case Intrinsic::x86_sse2_ucomilt_sd:
8334 case Intrinsic::x86_sse2_ucomile_sd:
8335 case Intrinsic::x86_sse2_ucomigt_sd:
8336 case Intrinsic::x86_sse2_ucomige_sd:
8337 case Intrinsic::x86_sse2_ucomineq_sd: {
8339 ISD::CondCode CC = ISD::SETCC_INVALID;
8342 case Intrinsic::x86_sse_comieq_ss:
8343 case Intrinsic::x86_sse2_comieq_sd:
8347 case Intrinsic::x86_sse_comilt_ss:
8348 case Intrinsic::x86_sse2_comilt_sd:
8352 case Intrinsic::x86_sse_comile_ss:
8353 case Intrinsic::x86_sse2_comile_sd:
8357 case Intrinsic::x86_sse_comigt_ss:
8358 case Intrinsic::x86_sse2_comigt_sd:
8362 case Intrinsic::x86_sse_comige_ss:
8363 case Intrinsic::x86_sse2_comige_sd:
8367 case Intrinsic::x86_sse_comineq_ss:
8368 case Intrinsic::x86_sse2_comineq_sd:
8372 case Intrinsic::x86_sse_ucomieq_ss:
8373 case Intrinsic::x86_sse2_ucomieq_sd:
8374 Opc = X86ISD::UCOMI;
8377 case Intrinsic::x86_sse_ucomilt_ss:
8378 case Intrinsic::x86_sse2_ucomilt_sd:
8379 Opc = X86ISD::UCOMI;
8382 case Intrinsic::x86_sse_ucomile_ss:
8383 case Intrinsic::x86_sse2_ucomile_sd:
8384 Opc = X86ISD::UCOMI;
8387 case Intrinsic::x86_sse_ucomigt_ss:
8388 case Intrinsic::x86_sse2_ucomigt_sd:
8389 Opc = X86ISD::UCOMI;
8392 case Intrinsic::x86_sse_ucomige_ss:
8393 case Intrinsic::x86_sse2_ucomige_sd:
8394 Opc = X86ISD::UCOMI;
8397 case Intrinsic::x86_sse_ucomineq_ss:
8398 case Intrinsic::x86_sse2_ucomineq_sd:
8399 Opc = X86ISD::UCOMI;
8404 SDValue LHS = Op.getOperand(1);
8405 SDValue RHS = Op.getOperand(2);
8406 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
8407 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
8408 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
8409 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
8410 DAG.getConstant(X86CC, MVT::i8), Cond);
8411 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
8413 // ptest and testp intrinsics. The intrinsic these come from are designed to
8414 // return an integer value, not just an instruction so lower it to the ptest
8415 // or testp pattern and a setcc for the result.
8416 case Intrinsic::x86_sse41_ptestz:
8417 case Intrinsic::x86_sse41_ptestc:
8418 case Intrinsic::x86_sse41_ptestnzc:
8419 case Intrinsic::x86_avx_ptestz_256:
8420 case Intrinsic::x86_avx_ptestc_256:
8421 case Intrinsic::x86_avx_ptestnzc_256:
8422 case Intrinsic::x86_avx_vtestz_ps:
8423 case Intrinsic::x86_avx_vtestc_ps:
8424 case Intrinsic::x86_avx_vtestnzc_ps:
8425 case Intrinsic::x86_avx_vtestz_pd:
8426 case Intrinsic::x86_avx_vtestc_pd:
8427 case Intrinsic::x86_avx_vtestnzc_pd:
8428 case Intrinsic::x86_avx_vtestz_ps_256:
8429 case Intrinsic::x86_avx_vtestc_ps_256:
8430 case Intrinsic::x86_avx_vtestnzc_ps_256:
8431 case Intrinsic::x86_avx_vtestz_pd_256:
8432 case Intrinsic::x86_avx_vtestc_pd_256:
8433 case Intrinsic::x86_avx_vtestnzc_pd_256: {
8434 bool IsTestPacked = false;
8437 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
8438 case Intrinsic::x86_avx_vtestz_ps:
8439 case Intrinsic::x86_avx_vtestz_pd:
8440 case Intrinsic::x86_avx_vtestz_ps_256:
8441 case Intrinsic::x86_avx_vtestz_pd_256:
8442 IsTestPacked = true; // Fallthrough
8443 case Intrinsic::x86_sse41_ptestz:
8444 case Intrinsic::x86_avx_ptestz_256:
8446 X86CC = X86::COND_E;
8448 case Intrinsic::x86_avx_vtestc_ps:
8449 case Intrinsic::x86_avx_vtestc_pd:
8450 case Intrinsic::x86_avx_vtestc_ps_256:
8451 case Intrinsic::x86_avx_vtestc_pd_256:
8452 IsTestPacked = true; // Fallthrough
8453 case Intrinsic::x86_sse41_ptestc:
8454 case Intrinsic::x86_avx_ptestc_256:
8456 X86CC = X86::COND_B;
8458 case Intrinsic::x86_avx_vtestnzc_ps:
8459 case Intrinsic::x86_avx_vtestnzc_pd:
8460 case Intrinsic::x86_avx_vtestnzc_ps_256:
8461 case Intrinsic::x86_avx_vtestnzc_pd_256:
8462 IsTestPacked = true; // Fallthrough
8463 case Intrinsic::x86_sse41_ptestnzc:
8464 case Intrinsic::x86_avx_ptestnzc_256:
8466 X86CC = X86::COND_A;
8470 SDValue LHS = Op.getOperand(1);
8471 SDValue RHS = Op.getOperand(2);
8472 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
8473 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
8474 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
8475 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
8476 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
8479 // Fix vector shift instructions where the last operand is a non-immediate
8481 case Intrinsic::x86_sse2_pslli_w:
8482 case Intrinsic::x86_sse2_pslli_d:
8483 case Intrinsic::x86_sse2_pslli_q:
8484 case Intrinsic::x86_sse2_psrli_w:
8485 case Intrinsic::x86_sse2_psrli_d:
8486 case Intrinsic::x86_sse2_psrli_q:
8487 case Intrinsic::x86_sse2_psrai_w:
8488 case Intrinsic::x86_sse2_psrai_d:
8489 case Intrinsic::x86_mmx_pslli_w:
8490 case Intrinsic::x86_mmx_pslli_d:
8491 case Intrinsic::x86_mmx_pslli_q:
8492 case Intrinsic::x86_mmx_psrli_w:
8493 case Intrinsic::x86_mmx_psrli_d:
8494 case Intrinsic::x86_mmx_psrli_q:
8495 case Intrinsic::x86_mmx_psrai_w:
8496 case Intrinsic::x86_mmx_psrai_d: {
8497 SDValue ShAmt = Op.getOperand(2);
8498 if (isa<ConstantSDNode>(ShAmt))
8501 unsigned NewIntNo = 0;
8502 EVT ShAmtVT = MVT::v4i32;
8504 case Intrinsic::x86_sse2_pslli_w:
8505 NewIntNo = Intrinsic::x86_sse2_psll_w;
8507 case Intrinsic::x86_sse2_pslli_d:
8508 NewIntNo = Intrinsic::x86_sse2_psll_d;
8510 case Intrinsic::x86_sse2_pslli_q:
8511 NewIntNo = Intrinsic::x86_sse2_psll_q;
8513 case Intrinsic::x86_sse2_psrli_w:
8514 NewIntNo = Intrinsic::x86_sse2_psrl_w;
8516 case Intrinsic::x86_sse2_psrli_d:
8517 NewIntNo = Intrinsic::x86_sse2_psrl_d;
8519 case Intrinsic::x86_sse2_psrli_q:
8520 NewIntNo = Intrinsic::x86_sse2_psrl_q;
8522 case Intrinsic::x86_sse2_psrai_w:
8523 NewIntNo = Intrinsic::x86_sse2_psra_w;
8525 case Intrinsic::x86_sse2_psrai_d:
8526 NewIntNo = Intrinsic::x86_sse2_psra_d;
8529 ShAmtVT = MVT::v2i32;
8531 case Intrinsic::x86_mmx_pslli_w:
8532 NewIntNo = Intrinsic::x86_mmx_psll_w;
8534 case Intrinsic::x86_mmx_pslli_d:
8535 NewIntNo = Intrinsic::x86_mmx_psll_d;
8537 case Intrinsic::x86_mmx_pslli_q:
8538 NewIntNo = Intrinsic::x86_mmx_psll_q;
8540 case Intrinsic::x86_mmx_psrli_w:
8541 NewIntNo = Intrinsic::x86_mmx_psrl_w;
8543 case Intrinsic::x86_mmx_psrli_d:
8544 NewIntNo = Intrinsic::x86_mmx_psrl_d;
8546 case Intrinsic::x86_mmx_psrli_q:
8547 NewIntNo = Intrinsic::x86_mmx_psrl_q;
8549 case Intrinsic::x86_mmx_psrai_w:
8550 NewIntNo = Intrinsic::x86_mmx_psra_w;
8552 case Intrinsic::x86_mmx_psrai_d:
8553 NewIntNo = Intrinsic::x86_mmx_psra_d;
8555 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
8561 // The vector shift intrinsics with scalars uses 32b shift amounts but
8562 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
8566 ShOps[1] = DAG.getConstant(0, MVT::i32);
8567 if (ShAmtVT == MVT::v4i32) {
8568 ShOps[2] = DAG.getUNDEF(MVT::i32);
8569 ShOps[3] = DAG.getUNDEF(MVT::i32);
8570 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
8572 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
8573 // FIXME this must be lowered to get rid of the invalid type.
8576 EVT VT = Op.getValueType();
8577 ShAmt = DAG.getNode(ISD::BITCAST, dl, VT, ShAmt);
8578 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8579 DAG.getConstant(NewIntNo, MVT::i32),
8580 Op.getOperand(1), ShAmt);
8585 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
8586 SelectionDAG &DAG) const {
8587 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8588 MFI->setReturnAddressIsTaken(true);
8590 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8591 DebugLoc dl = Op.getDebugLoc();
8594 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
8596 DAG.getConstant(TD->getPointerSize(),
8597 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
8598 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
8599 DAG.getNode(ISD::ADD, dl, getPointerTy(),
8601 MachinePointerInfo(), false, false, 0);
8604 // Just load the return address.
8605 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
8606 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
8607 RetAddrFI, MachinePointerInfo(), false, false, 0);
8610 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
8611 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8612 MFI->setFrameAddressIsTaken(true);
8614 EVT VT = Op.getValueType();
8615 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
8616 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8617 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
8618 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
8620 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
8621 MachinePointerInfo(),
8626 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
8627 SelectionDAG &DAG) const {
8628 return DAG.getIntPtrConstant(2*TD->getPointerSize());
8631 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
8632 MachineFunction &MF = DAG.getMachineFunction();
8633 SDValue Chain = Op.getOperand(0);
8634 SDValue Offset = Op.getOperand(1);
8635 SDValue Handler = Op.getOperand(2);
8636 DebugLoc dl = Op.getDebugLoc();
8638 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
8639 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
8641 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
8643 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
8644 DAG.getIntPtrConstant(TD->getPointerSize()));
8645 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
8646 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
8648 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
8649 MF.getRegInfo().addLiveOut(StoreAddrReg);
8651 return DAG.getNode(X86ISD::EH_RETURN, dl,
8653 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
8656 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
8657 SelectionDAG &DAG) const {
8658 SDValue Root = Op.getOperand(0);
8659 SDValue Trmp = Op.getOperand(1); // trampoline
8660 SDValue FPtr = Op.getOperand(2); // nested function
8661 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
8662 DebugLoc dl = Op.getDebugLoc();
8664 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
8666 if (Subtarget->is64Bit()) {
8667 SDValue OutChains[6];
8669 // Large code-model.
8670 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
8671 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
8673 const unsigned char N86R10 = X86_MC::getX86RegNum(X86::R10);
8674 const unsigned char N86R11 = X86_MC::getX86RegNum(X86::R11);
8676 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
8678 // Load the pointer to the nested function into R11.
8679 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
8680 SDValue Addr = Trmp;
8681 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8682 Addr, MachinePointerInfo(TrmpAddr),
8685 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8686 DAG.getConstant(2, MVT::i64));
8687 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
8688 MachinePointerInfo(TrmpAddr, 2),
8691 // Load the 'nest' parameter value into R10.
8692 // R10 is specified in X86CallingConv.td
8693 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
8694 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8695 DAG.getConstant(10, MVT::i64));
8696 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8697 Addr, MachinePointerInfo(TrmpAddr, 10),
8700 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8701 DAG.getConstant(12, MVT::i64));
8702 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
8703 MachinePointerInfo(TrmpAddr, 12),
8706 // Jump to the nested function.
8707 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
8708 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8709 DAG.getConstant(20, MVT::i64));
8710 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8711 Addr, MachinePointerInfo(TrmpAddr, 20),
8714 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
8715 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8716 DAG.getConstant(22, MVT::i64));
8717 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
8718 MachinePointerInfo(TrmpAddr, 22),
8722 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
8723 return DAG.getMergeValues(Ops, 2, dl);
8725 const Function *Func =
8726 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
8727 CallingConv::ID CC = Func->getCallingConv();
8732 llvm_unreachable("Unsupported calling convention");
8733 case CallingConv::C:
8734 case CallingConv::X86_StdCall: {
8735 // Pass 'nest' parameter in ECX.
8736 // Must be kept in sync with X86CallingConv.td
8739 // Check that ECX wasn't needed by an 'inreg' parameter.
8740 FunctionType *FTy = Func->getFunctionType();
8741 const AttrListPtr &Attrs = Func->getAttributes();
8743 if (!Attrs.isEmpty() && !Func->isVarArg()) {
8744 unsigned InRegCount = 0;
8747 for (FunctionType::param_iterator I = FTy->param_begin(),
8748 E = FTy->param_end(); I != E; ++I, ++Idx)
8749 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
8750 // FIXME: should only count parameters that are lowered to integers.
8751 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
8753 if (InRegCount > 2) {
8754 report_fatal_error("Nest register in use - reduce number of inreg"
8760 case CallingConv::X86_FastCall:
8761 case CallingConv::X86_ThisCall:
8762 case CallingConv::Fast:
8763 // Pass 'nest' parameter in EAX.
8764 // Must be kept in sync with X86CallingConv.td
8769 SDValue OutChains[4];
8772 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8773 DAG.getConstant(10, MVT::i32));
8774 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
8776 // This is storing the opcode for MOV32ri.
8777 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
8778 const unsigned char N86Reg = X86_MC::getX86RegNum(NestReg);
8779 OutChains[0] = DAG.getStore(Root, dl,
8780 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
8781 Trmp, MachinePointerInfo(TrmpAddr),
8784 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8785 DAG.getConstant(1, MVT::i32));
8786 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
8787 MachinePointerInfo(TrmpAddr, 1),
8790 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
8791 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8792 DAG.getConstant(5, MVT::i32));
8793 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
8794 MachinePointerInfo(TrmpAddr, 5),
8797 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8798 DAG.getConstant(6, MVT::i32));
8799 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
8800 MachinePointerInfo(TrmpAddr, 6),
8804 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
8805 return DAG.getMergeValues(Ops, 2, dl);
8809 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
8810 SelectionDAG &DAG) const {
8812 The rounding mode is in bits 11:10 of FPSR, and has the following
8819 FLT_ROUNDS, on the other hand, expects the following:
8826 To perform the conversion, we do:
8827 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
8830 MachineFunction &MF = DAG.getMachineFunction();
8831 const TargetMachine &TM = MF.getTarget();
8832 const TargetFrameLowering &TFI = *TM.getFrameLowering();
8833 unsigned StackAlignment = TFI.getStackAlignment();
8834 EVT VT = Op.getValueType();
8835 DebugLoc DL = Op.getDebugLoc();
8837 // Save FP Control Word to stack slot
8838 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
8839 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8842 MachineMemOperand *MMO =
8843 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8844 MachineMemOperand::MOStore, 2, 2);
8846 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
8847 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
8848 DAG.getVTList(MVT::Other),
8849 Ops, 2, MVT::i16, MMO);
8851 // Load FP Control Word from stack slot
8852 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
8853 MachinePointerInfo(), false, false, 0);
8855 // Transform as necessary
8857 DAG.getNode(ISD::SRL, DL, MVT::i16,
8858 DAG.getNode(ISD::AND, DL, MVT::i16,
8859 CWD, DAG.getConstant(0x800, MVT::i16)),
8860 DAG.getConstant(11, MVT::i8));
8862 DAG.getNode(ISD::SRL, DL, MVT::i16,
8863 DAG.getNode(ISD::AND, DL, MVT::i16,
8864 CWD, DAG.getConstant(0x400, MVT::i16)),
8865 DAG.getConstant(9, MVT::i8));
8868 DAG.getNode(ISD::AND, DL, MVT::i16,
8869 DAG.getNode(ISD::ADD, DL, MVT::i16,
8870 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
8871 DAG.getConstant(1, MVT::i16)),
8872 DAG.getConstant(3, MVT::i16));
8875 return DAG.getNode((VT.getSizeInBits() < 16 ?
8876 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
8879 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
8880 EVT VT = Op.getValueType();
8882 unsigned NumBits = VT.getSizeInBits();
8883 DebugLoc dl = Op.getDebugLoc();
8885 Op = Op.getOperand(0);
8886 if (VT == MVT::i8) {
8887 // Zero extend to i32 since there is not an i8 bsr.
8889 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8892 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
8893 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8894 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
8896 // If src is zero (i.e. bsr sets ZF), returns NumBits.
8899 DAG.getConstant(NumBits+NumBits-1, OpVT),
8900 DAG.getConstant(X86::COND_E, MVT::i8),
8903 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8905 // Finally xor with NumBits-1.
8906 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
8909 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8913 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
8914 EVT VT = Op.getValueType();
8916 unsigned NumBits = VT.getSizeInBits();
8917 DebugLoc dl = Op.getDebugLoc();
8919 Op = Op.getOperand(0);
8920 if (VT == MVT::i8) {
8922 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8925 // Issue a bsf (scan bits forward) which also sets EFLAGS.
8926 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8927 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
8929 // If src is zero (i.e. bsf sets ZF), returns NumBits.
8932 DAG.getConstant(NumBits, OpVT),
8933 DAG.getConstant(X86::COND_E, MVT::i8),
8936 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8939 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8943 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
8944 EVT VT = Op.getValueType();
8945 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
8946 DebugLoc dl = Op.getDebugLoc();
8948 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
8949 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
8950 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
8951 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
8952 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
8954 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
8955 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
8956 // return AloBlo + AloBhi + AhiBlo;
8958 SDValue A = Op.getOperand(0);
8959 SDValue B = Op.getOperand(1);
8961 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8962 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8963 A, DAG.getConstant(32, MVT::i32));
8964 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8965 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8966 B, DAG.getConstant(32, MVT::i32));
8967 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8968 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8970 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8971 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8973 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8974 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8976 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8977 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8978 AloBhi, DAG.getConstant(32, MVT::i32));
8979 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8980 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8981 AhiBlo, DAG.getConstant(32, MVT::i32));
8982 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
8983 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
8987 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
8989 EVT VT = Op.getValueType();
8990 DebugLoc dl = Op.getDebugLoc();
8991 SDValue R = Op.getOperand(0);
8992 SDValue Amt = Op.getOperand(1);
8994 LLVMContext *Context = DAG.getContext();
8997 if (!Subtarget->hasSSE2()) return SDValue();
8999 // Optimize shl/srl/sra with constant shift amount.
9000 if (isSplatVector(Amt.getNode())) {
9001 SDValue SclrAmt = Amt->getOperand(0);
9002 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(SclrAmt)) {
9003 uint64_t ShiftAmt = C->getZExtValue();
9005 if (VT == MVT::v2i64 && Op.getOpcode() == ISD::SHL)
9006 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9007 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9008 R, DAG.getConstant(ShiftAmt, MVT::i32));
9010 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SHL)
9011 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9012 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9013 R, DAG.getConstant(ShiftAmt, MVT::i32));
9015 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SHL)
9016 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9017 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9018 R, DAG.getConstant(ShiftAmt, MVT::i32));
9020 if (VT == MVT::v2i64 && Op.getOpcode() == ISD::SRL)
9021 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9022 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9023 R, DAG.getConstant(ShiftAmt, MVT::i32));
9025 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SRL)
9026 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9027 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9028 R, DAG.getConstant(ShiftAmt, MVT::i32));
9030 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SRL)
9031 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9032 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9033 R, DAG.getConstant(ShiftAmt, MVT::i32));
9035 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SRA)
9036 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9037 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9038 R, DAG.getConstant(ShiftAmt, MVT::i32));
9040 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SRA)
9041 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9042 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9043 R, DAG.getConstant(ShiftAmt, MVT::i32));
9047 // Lower SHL with variable shift amount.
9048 // Cannot lower SHL without SSE2 or later.
9049 if (!Subtarget->hasSSE2()) return SDValue();
9051 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
9052 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9053 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9054 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
9056 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
9058 std::vector<Constant*> CV(4, CI);
9059 Constant *C = ConstantVector::get(CV);
9060 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9061 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9062 MachinePointerInfo::getConstantPool(),
9065 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
9066 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
9067 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
9068 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
9070 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
9072 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9073 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9074 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
9076 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
9077 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
9079 std::vector<Constant*> CVM1(16, CM1);
9080 std::vector<Constant*> CVM2(16, CM2);
9081 Constant *C = ConstantVector::get(CVM1);
9082 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9083 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9084 MachinePointerInfo::getConstantPool(),
9087 // r = pblendv(r, psllw(r & (char16)15, 4), a);
9088 M = DAG.getNode(ISD::AND, dl, VT, R, M);
9089 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9090 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
9091 DAG.getConstant(4, MVT::i32));
9092 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
9094 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
9096 C = ConstantVector::get(CVM2);
9097 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9098 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9099 MachinePointerInfo::getConstantPool(),
9102 // r = pblendv(r, psllw(r & (char16)63, 2), a);
9103 M = DAG.getNode(ISD::AND, dl, VT, R, M);
9104 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9105 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
9106 DAG.getConstant(2, MVT::i32));
9107 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
9109 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
9111 // return pblendv(r, r+r, a);
9112 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT,
9113 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
9119 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
9120 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
9121 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
9122 // looks for this combo and may remove the "setcc" instruction if the "setcc"
9123 // has only one use.
9124 SDNode *N = Op.getNode();
9125 SDValue LHS = N->getOperand(0);
9126 SDValue RHS = N->getOperand(1);
9127 unsigned BaseOp = 0;
9129 DebugLoc DL = Op.getDebugLoc();
9130 switch (Op.getOpcode()) {
9131 default: llvm_unreachable("Unknown ovf instruction!");
9133 // A subtract of one will be selected as a INC. Note that INC doesn't
9134 // set CF, so we can't do this for UADDO.
9135 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
9137 BaseOp = X86ISD::INC;
9141 BaseOp = X86ISD::ADD;
9145 BaseOp = X86ISD::ADD;
9149 // A subtract of one will be selected as a DEC. Note that DEC doesn't
9150 // set CF, so we can't do this for USUBO.
9151 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
9153 BaseOp = X86ISD::DEC;
9157 BaseOp = X86ISD::SUB;
9161 BaseOp = X86ISD::SUB;
9165 BaseOp = X86ISD::SMUL;
9168 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
9169 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
9171 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
9174 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9175 DAG.getConstant(X86::COND_O, MVT::i32),
9176 SDValue(Sum.getNode(), 2));
9178 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
9182 // Also sets EFLAGS.
9183 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
9184 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
9187 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
9188 DAG.getConstant(Cond, MVT::i32),
9189 SDValue(Sum.getNode(), 1));
9191 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
9194 SDValue X86TargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const{
9195 DebugLoc dl = Op.getDebugLoc();
9196 SDNode* Node = Op.getNode();
9197 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
9198 EVT VT = Node->getValueType(0);
9200 if (Subtarget->hasSSE2() && VT.isVector()) {
9201 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
9202 ExtraVT.getScalarType().getSizeInBits();
9203 SDValue ShAmt = DAG.getConstant(BitsDiff, MVT::i32);
9205 unsigned SHLIntrinsicsID = 0;
9206 unsigned SRAIntrinsicsID = 0;
9207 switch (VT.getSimpleVT().SimpleTy) {
9211 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_q;
9212 SRAIntrinsicsID = 0;
9216 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_d;
9217 SRAIntrinsicsID = Intrinsic::x86_sse2_psrai_d;
9221 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_w;
9222 SRAIntrinsicsID = Intrinsic::x86_sse2_psrai_w;
9227 SDValue Tmp1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9228 DAG.getConstant(SHLIntrinsicsID, MVT::i32),
9229 Node->getOperand(0), ShAmt);
9231 // In case of 1 bit sext, no need to shr
9232 if (ExtraVT.getScalarType().getSizeInBits() == 1) return Tmp1;
9234 if (SRAIntrinsicsID) {
9235 Tmp1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9236 DAG.getConstant(SRAIntrinsicsID, MVT::i32),
9246 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
9247 DebugLoc dl = Op.getDebugLoc();
9249 // Go ahead and emit the fence on x86-64 even if we asked for no-sse2.
9250 // There isn't any reason to disable it if the target processor supports it.
9251 if (!Subtarget->hasSSE2() && !Subtarget->is64Bit()) {
9252 SDValue Chain = Op.getOperand(0);
9253 SDValue Zero = DAG.getConstant(0, MVT::i32);
9255 DAG.getRegister(X86::ESP, MVT::i32), // Base
9256 DAG.getTargetConstant(1, MVT::i8), // Scale
9257 DAG.getRegister(0, MVT::i32), // Index
9258 DAG.getTargetConstant(0, MVT::i32), // Disp
9259 DAG.getRegister(0, MVT::i32), // Segment.
9264 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
9265 array_lengthof(Ops));
9266 return SDValue(Res, 0);
9269 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
9271 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
9273 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
9274 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
9275 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
9276 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
9278 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
9279 if (!Op1 && !Op2 && !Op3 && Op4)
9280 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
9282 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
9283 if (Op1 && !Op2 && !Op3 && !Op4)
9284 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
9286 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
9288 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
9291 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
9292 EVT T = Op.getValueType();
9293 DebugLoc DL = Op.getDebugLoc();
9296 switch(T.getSimpleVT().SimpleTy) {
9298 assert(false && "Invalid value type!");
9299 case MVT::i8: Reg = X86::AL; size = 1; break;
9300 case MVT::i16: Reg = X86::AX; size = 2; break;
9301 case MVT::i32: Reg = X86::EAX; size = 4; break;
9303 assert(Subtarget->is64Bit() && "Node not type legal!");
9304 Reg = X86::RAX; size = 8;
9307 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
9308 Op.getOperand(2), SDValue());
9309 SDValue Ops[] = { cpIn.getValue(0),
9312 DAG.getTargetConstant(size, MVT::i8),
9314 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9315 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
9316 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
9319 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
9323 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
9324 SelectionDAG &DAG) const {
9325 assert(Subtarget->is64Bit() && "Result not type legalized?");
9326 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9327 SDValue TheChain = Op.getOperand(0);
9328 DebugLoc dl = Op.getDebugLoc();
9329 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
9330 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
9331 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
9333 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
9334 DAG.getConstant(32, MVT::i8));
9336 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
9339 return DAG.getMergeValues(Ops, 2, dl);
9342 SDValue X86TargetLowering::LowerBITCAST(SDValue Op,
9343 SelectionDAG &DAG) const {
9344 EVT SrcVT = Op.getOperand(0).getValueType();
9345 EVT DstVT = Op.getValueType();
9346 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
9347 Subtarget->hasMMX() && "Unexpected custom BITCAST");
9348 assert((DstVT == MVT::i64 ||
9349 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
9350 "Unexpected custom BITCAST");
9351 // i64 <=> MMX conversions are Legal.
9352 if (SrcVT==MVT::i64 && DstVT.isVector())
9354 if (DstVT==MVT::i64 && SrcVT.isVector())
9356 // MMX <=> MMX conversions are Legal.
9357 if (SrcVT.isVector() && DstVT.isVector())
9359 // All other conversions need to be expanded.
9363 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
9364 SDNode *Node = Op.getNode();
9365 DebugLoc dl = Node->getDebugLoc();
9366 EVT T = Node->getValueType(0);
9367 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
9368 DAG.getConstant(0, T), Node->getOperand(2));
9369 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
9370 cast<AtomicSDNode>(Node)->getMemoryVT(),
9371 Node->getOperand(0),
9372 Node->getOperand(1), negOp,
9373 cast<AtomicSDNode>(Node)->getSrcValue(),
9374 cast<AtomicSDNode>(Node)->getAlignment());
9377 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
9378 EVT VT = Op.getNode()->getValueType(0);
9380 // Let legalize expand this if it isn't a legal type yet.
9381 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
9384 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
9387 bool ExtraOp = false;
9388 switch (Op.getOpcode()) {
9389 default: assert(0 && "Invalid code");
9390 case ISD::ADDC: Opc = X86ISD::ADD; break;
9391 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
9392 case ISD::SUBC: Opc = X86ISD::SUB; break;
9393 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
9397 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
9399 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
9400 Op.getOperand(1), Op.getOperand(2));
9403 /// LowerOperation - Provide custom lowering hooks for some operations.
9405 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
9406 switch (Op.getOpcode()) {
9407 default: llvm_unreachable("Should not custom lower this!");
9408 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op,DAG);
9409 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
9410 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
9411 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
9412 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
9413 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
9414 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
9415 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
9416 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
9417 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
9418 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, DAG);
9419 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
9420 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
9421 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
9422 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
9423 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
9424 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
9425 case ISD::SHL_PARTS:
9426 case ISD::SRA_PARTS:
9427 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
9428 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
9429 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
9430 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
9431 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
9432 case ISD::FABS: return LowerFABS(Op, DAG);
9433 case ISD::FNEG: return LowerFNEG(Op, DAG);
9434 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
9435 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
9436 case ISD::SETCC: return LowerSETCC(Op, DAG);
9437 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
9438 case ISD::SELECT: return LowerSELECT(Op, DAG);
9439 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
9440 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
9441 case ISD::VASTART: return LowerVASTART(Op, DAG);
9442 case ISD::VAARG: return LowerVAARG(Op, DAG);
9443 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
9444 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
9445 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
9446 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
9447 case ISD::FRAME_TO_ARGS_OFFSET:
9448 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
9449 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
9450 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
9451 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
9452 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
9453 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
9454 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
9455 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
9458 case ISD::SHL: return LowerShift(Op, DAG);
9464 case ISD::UMULO: return LowerXALUO(Op, DAG);
9465 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
9466 case ISD::BITCAST: return LowerBITCAST(Op, DAG);
9470 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
9474 void X86TargetLowering::
9475 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
9476 SelectionDAG &DAG, unsigned NewOp) const {
9477 EVT T = Node->getValueType(0);
9478 DebugLoc dl = Node->getDebugLoc();
9479 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
9481 SDValue Chain = Node->getOperand(0);
9482 SDValue In1 = Node->getOperand(1);
9483 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
9484 Node->getOperand(2), DAG.getIntPtrConstant(0));
9485 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
9486 Node->getOperand(2), DAG.getIntPtrConstant(1));
9487 SDValue Ops[] = { Chain, In1, In2L, In2H };
9488 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
9490 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
9491 cast<MemSDNode>(Node)->getMemOperand());
9492 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
9493 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
9494 Results.push_back(Result.getValue(2));
9497 /// ReplaceNodeResults - Replace a node with an illegal result type
9498 /// with a new node built out of custom code.
9499 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
9500 SmallVectorImpl<SDValue>&Results,
9501 SelectionDAG &DAG) const {
9502 DebugLoc dl = N->getDebugLoc();
9503 switch (N->getOpcode()) {
9505 assert(false && "Do not know how to custom type legalize this operation!");
9507 case ISD::SIGN_EXTEND_INREG:
9512 // We don't want to expand or promote these.
9514 case ISD::FP_TO_SINT: {
9515 std::pair<SDValue,SDValue> Vals =
9516 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
9517 SDValue FIST = Vals.first, StackSlot = Vals.second;
9518 if (FIST.getNode() != 0) {
9519 EVT VT = N->getValueType(0);
9520 // Return a load from the stack slot.
9521 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
9522 MachinePointerInfo(), false, false, 0));
9526 case ISD::READCYCLECOUNTER: {
9527 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9528 SDValue TheChain = N->getOperand(0);
9529 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
9530 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
9532 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
9534 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
9535 SDValue Ops[] = { eax, edx };
9536 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
9537 Results.push_back(edx.getValue(1));
9540 case ISD::ATOMIC_CMP_SWAP: {
9541 EVT T = N->getValueType(0);
9542 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
9543 SDValue cpInL, cpInH;
9544 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
9545 DAG.getConstant(0, MVT::i32));
9546 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
9547 DAG.getConstant(1, MVT::i32));
9548 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
9549 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
9551 SDValue swapInL, swapInH;
9552 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
9553 DAG.getConstant(0, MVT::i32));
9554 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
9555 DAG.getConstant(1, MVT::i32));
9556 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
9558 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
9559 swapInL.getValue(1));
9560 SDValue Ops[] = { swapInH.getValue(0),
9562 swapInH.getValue(1) };
9563 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9564 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
9565 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG8_DAG, dl, Tys,
9567 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
9568 MVT::i32, Result.getValue(1));
9569 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
9570 MVT::i32, cpOutL.getValue(2));
9571 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
9572 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
9573 Results.push_back(cpOutH.getValue(1));
9576 case ISD::ATOMIC_LOAD_ADD:
9577 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
9579 case ISD::ATOMIC_LOAD_AND:
9580 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
9582 case ISD::ATOMIC_LOAD_NAND:
9583 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
9585 case ISD::ATOMIC_LOAD_OR:
9586 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
9588 case ISD::ATOMIC_LOAD_SUB:
9589 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
9591 case ISD::ATOMIC_LOAD_XOR:
9592 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
9594 case ISD::ATOMIC_SWAP:
9595 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
9600 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
9602 default: return NULL;
9603 case X86ISD::BSF: return "X86ISD::BSF";
9604 case X86ISD::BSR: return "X86ISD::BSR";
9605 case X86ISD::SHLD: return "X86ISD::SHLD";
9606 case X86ISD::SHRD: return "X86ISD::SHRD";
9607 case X86ISD::FAND: return "X86ISD::FAND";
9608 case X86ISD::FOR: return "X86ISD::FOR";
9609 case X86ISD::FXOR: return "X86ISD::FXOR";
9610 case X86ISD::FSRL: return "X86ISD::FSRL";
9611 case X86ISD::FILD: return "X86ISD::FILD";
9612 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
9613 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
9614 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
9615 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
9616 case X86ISD::FLD: return "X86ISD::FLD";
9617 case X86ISD::FST: return "X86ISD::FST";
9618 case X86ISD::CALL: return "X86ISD::CALL";
9619 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
9620 case X86ISD::BT: return "X86ISD::BT";
9621 case X86ISD::CMP: return "X86ISD::CMP";
9622 case X86ISD::COMI: return "X86ISD::COMI";
9623 case X86ISD::UCOMI: return "X86ISD::UCOMI";
9624 case X86ISD::SETCC: return "X86ISD::SETCC";
9625 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
9626 case X86ISD::FSETCCsd: return "X86ISD::FSETCCsd";
9627 case X86ISD::FSETCCss: return "X86ISD::FSETCCss";
9628 case X86ISD::CMOV: return "X86ISD::CMOV";
9629 case X86ISD::BRCOND: return "X86ISD::BRCOND";
9630 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
9631 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
9632 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
9633 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
9634 case X86ISD::Wrapper: return "X86ISD::Wrapper";
9635 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
9636 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
9637 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
9638 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
9639 case X86ISD::PINSRB: return "X86ISD::PINSRB";
9640 case X86ISD::PINSRW: return "X86ISD::PINSRW";
9641 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
9642 case X86ISD::ANDNP: return "X86ISD::ANDNP";
9643 case X86ISD::PSIGNB: return "X86ISD::PSIGNB";
9644 case X86ISD::PSIGNW: return "X86ISD::PSIGNW";
9645 case X86ISD::PSIGND: return "X86ISD::PSIGND";
9646 case X86ISD::PBLENDVB: return "X86ISD::PBLENDVB";
9647 case X86ISD::FMAX: return "X86ISD::FMAX";
9648 case X86ISD::FMIN: return "X86ISD::FMIN";
9649 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
9650 case X86ISD::FRCP: return "X86ISD::FRCP";
9651 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
9652 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
9653 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
9654 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
9655 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
9656 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
9657 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
9658 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
9659 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
9660 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
9661 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
9662 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
9663 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
9664 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
9665 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
9666 case X86ISD::VSHL: return "X86ISD::VSHL";
9667 case X86ISD::VSRL: return "X86ISD::VSRL";
9668 case X86ISD::CMPPD: return "X86ISD::CMPPD";
9669 case X86ISD::CMPPS: return "X86ISD::CMPPS";
9670 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
9671 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
9672 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
9673 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
9674 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
9675 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
9676 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
9677 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
9678 case X86ISD::ADD: return "X86ISD::ADD";
9679 case X86ISD::SUB: return "X86ISD::SUB";
9680 case X86ISD::ADC: return "X86ISD::ADC";
9681 case X86ISD::SBB: return "X86ISD::SBB";
9682 case X86ISD::SMUL: return "X86ISD::SMUL";
9683 case X86ISD::UMUL: return "X86ISD::UMUL";
9684 case X86ISD::INC: return "X86ISD::INC";
9685 case X86ISD::DEC: return "X86ISD::DEC";
9686 case X86ISD::OR: return "X86ISD::OR";
9687 case X86ISD::XOR: return "X86ISD::XOR";
9688 case X86ISD::AND: return "X86ISD::AND";
9689 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
9690 case X86ISD::PTEST: return "X86ISD::PTEST";
9691 case X86ISD::TESTP: return "X86ISD::TESTP";
9692 case X86ISD::PALIGN: return "X86ISD::PALIGN";
9693 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
9694 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
9695 case X86ISD::PSHUFHW_LD: return "X86ISD::PSHUFHW_LD";
9696 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
9697 case X86ISD::PSHUFLW_LD: return "X86ISD::PSHUFLW_LD";
9698 case X86ISD::SHUFPS: return "X86ISD::SHUFPS";
9699 case X86ISD::SHUFPD: return "X86ISD::SHUFPD";
9700 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
9701 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
9702 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
9703 case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
9704 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
9705 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
9706 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
9707 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
9708 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
9709 case X86ISD::MOVSHDUP_LD: return "X86ISD::MOVSHDUP_LD";
9710 case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
9711 case X86ISD::MOVSD: return "X86ISD::MOVSD";
9712 case X86ISD::MOVSS: return "X86ISD::MOVSS";
9713 case X86ISD::UNPCKLPS: return "X86ISD::UNPCKLPS";
9714 case X86ISD::UNPCKLPD: return "X86ISD::UNPCKLPD";
9715 case X86ISD::VUNPCKLPDY: return "X86ISD::VUNPCKLPDY";
9716 case X86ISD::UNPCKHPS: return "X86ISD::UNPCKHPS";
9717 case X86ISD::UNPCKHPD: return "X86ISD::UNPCKHPD";
9718 case X86ISD::PUNPCKLBW: return "X86ISD::PUNPCKLBW";
9719 case X86ISD::PUNPCKLWD: return "X86ISD::PUNPCKLWD";
9720 case X86ISD::PUNPCKLDQ: return "X86ISD::PUNPCKLDQ";
9721 case X86ISD::PUNPCKLQDQ: return "X86ISD::PUNPCKLQDQ";
9722 case X86ISD::PUNPCKHBW: return "X86ISD::PUNPCKHBW";
9723 case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
9724 case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
9725 case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
9726 case X86ISD::VPERMIL: return "X86ISD::VPERMIL";
9727 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
9728 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
9729 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
9733 // isLegalAddressingMode - Return true if the addressing mode represented
9734 // by AM is legal for this target, for a load/store of the specified type.
9735 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
9737 // X86 supports extremely general addressing modes.
9738 CodeModel::Model M = getTargetMachine().getCodeModel();
9739 Reloc::Model R = getTargetMachine().getRelocationModel();
9741 // X86 allows a sign-extended 32-bit immediate field as a displacement.
9742 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
9747 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
9749 // If a reference to this global requires an extra load, we can't fold it.
9750 if (isGlobalStubReference(GVFlags))
9753 // If BaseGV requires a register for the PIC base, we cannot also have a
9754 // BaseReg specified.
9755 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
9758 // If lower 4G is not available, then we must use rip-relative addressing.
9759 if ((M != CodeModel::Small || R != Reloc::Static) &&
9760 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
9770 // These scales always work.
9775 // These scales are formed with basereg+scalereg. Only accept if there is
9780 default: // Other stuff never works.
9788 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
9789 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
9791 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
9792 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
9793 if (NumBits1 <= NumBits2)
9798 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
9799 if (!VT1.isInteger() || !VT2.isInteger())
9801 unsigned NumBits1 = VT1.getSizeInBits();
9802 unsigned NumBits2 = VT2.getSizeInBits();
9803 if (NumBits1 <= NumBits2)
9808 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
9809 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
9810 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
9813 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
9814 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
9815 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
9818 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
9819 // i16 instructions are longer (0x66 prefix) and potentially slower.
9820 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
9823 /// isShuffleMaskLegal - Targets can use this to indicate that they only
9824 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
9825 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
9826 /// are assumed to be legal.
9828 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
9830 // Very little shuffling can be done for 64-bit vectors right now.
9831 if (VT.getSizeInBits() == 64)
9832 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
9834 // FIXME: pshufb, blends, shifts.
9835 return (VT.getVectorNumElements() == 2 ||
9836 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
9837 isMOVLMask(M, VT) ||
9838 isSHUFPMask(M, VT) ||
9839 isPSHUFDMask(M, VT) ||
9840 isPSHUFHWMask(M, VT) ||
9841 isPSHUFLWMask(M, VT) ||
9842 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
9843 isUNPCKLMask(M, VT) ||
9844 isUNPCKHMask(M, VT) ||
9845 isUNPCKL_v_undef_Mask(M, VT) ||
9846 isUNPCKH_v_undef_Mask(M, VT));
9850 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
9852 unsigned NumElts = VT.getVectorNumElements();
9853 // FIXME: This collection of masks seems suspect.
9856 if (NumElts == 4 && VT.getSizeInBits() == 128) {
9857 return (isMOVLMask(Mask, VT) ||
9858 isCommutedMOVLMask(Mask, VT, true) ||
9859 isSHUFPMask(Mask, VT) ||
9860 isCommutedSHUFPMask(Mask, VT));
9865 //===----------------------------------------------------------------------===//
9866 // X86 Scheduler Hooks
9867 //===----------------------------------------------------------------------===//
9869 // private utility function
9871 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
9872 MachineBasicBlock *MBB,
9879 TargetRegisterClass *RC,
9880 bool invSrc) const {
9881 // For the atomic bitwise operator, we generate
9884 // ld t1 = [bitinstr.addr]
9885 // op t2 = t1, [bitinstr.val]
9887 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
9889 // fallthrough -->nextMBB
9890 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9891 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9892 MachineFunction::iterator MBBIter = MBB;
9895 /// First build the CFG
9896 MachineFunction *F = MBB->getParent();
9897 MachineBasicBlock *thisMBB = MBB;
9898 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9899 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9900 F->insert(MBBIter, newMBB);
9901 F->insert(MBBIter, nextMBB);
9903 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9904 nextMBB->splice(nextMBB->begin(), thisMBB,
9905 llvm::next(MachineBasicBlock::iterator(bInstr)),
9907 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9909 // Update thisMBB to fall through to newMBB
9910 thisMBB->addSuccessor(newMBB);
9912 // newMBB jumps to itself and fall through to nextMBB
9913 newMBB->addSuccessor(nextMBB);
9914 newMBB->addSuccessor(newMBB);
9916 // Insert instructions into newMBB based on incoming instruction
9917 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
9918 "unexpected number of operands");
9919 DebugLoc dl = bInstr->getDebugLoc();
9920 MachineOperand& destOper = bInstr->getOperand(0);
9921 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9922 int numArgs = bInstr->getNumOperands() - 1;
9923 for (int i=0; i < numArgs; ++i)
9924 argOpers[i] = &bInstr->getOperand(i+1);
9926 // x86 address has 4 operands: base, index, scale, and displacement
9927 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9928 int valArgIndx = lastAddrIndx + 1;
9930 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
9931 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
9932 for (int i=0; i <= lastAddrIndx; ++i)
9933 (*MIB).addOperand(*argOpers[i]);
9935 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
9937 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
9942 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
9943 assert((argOpers[valArgIndx]->isReg() ||
9944 argOpers[valArgIndx]->isImm()) &&
9946 if (argOpers[valArgIndx]->isReg())
9947 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
9949 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
9951 (*MIB).addOperand(*argOpers[valArgIndx]);
9953 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
9956 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
9957 for (int i=0; i <= lastAddrIndx; ++i)
9958 (*MIB).addOperand(*argOpers[i]);
9960 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9961 (*MIB).setMemRefs(bInstr->memoperands_begin(),
9962 bInstr->memoperands_end());
9964 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
9968 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9970 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
9974 // private utility function: 64 bit atomics on 32 bit host.
9976 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
9977 MachineBasicBlock *MBB,
9982 bool invSrc) const {
9983 // For the atomic bitwise operator, we generate
9984 // thisMBB (instructions are in pairs, except cmpxchg8b)
9985 // ld t1,t2 = [bitinstr.addr]
9987 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
9988 // op t5, t6 <- out1, out2, [bitinstr.val]
9989 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
9990 // mov ECX, EBX <- t5, t6
9991 // mov EAX, EDX <- t1, t2
9992 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
9993 // mov t3, t4 <- EAX, EDX
9995 // result in out1, out2
9996 // fallthrough -->nextMBB
9998 const TargetRegisterClass *RC = X86::GR32RegisterClass;
9999 const unsigned LoadOpc = X86::MOV32rm;
10000 const unsigned NotOpc = X86::NOT32r;
10001 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10002 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10003 MachineFunction::iterator MBBIter = MBB;
10006 /// First build the CFG
10007 MachineFunction *F = MBB->getParent();
10008 MachineBasicBlock *thisMBB = MBB;
10009 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
10010 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
10011 F->insert(MBBIter, newMBB);
10012 F->insert(MBBIter, nextMBB);
10014 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
10015 nextMBB->splice(nextMBB->begin(), thisMBB,
10016 llvm::next(MachineBasicBlock::iterator(bInstr)),
10018 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10020 // Update thisMBB to fall through to newMBB
10021 thisMBB->addSuccessor(newMBB);
10023 // newMBB jumps to itself and fall through to nextMBB
10024 newMBB->addSuccessor(nextMBB);
10025 newMBB->addSuccessor(newMBB);
10027 DebugLoc dl = bInstr->getDebugLoc();
10028 // Insert instructions into newMBB based on incoming instruction
10029 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
10030 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
10031 "unexpected number of operands");
10032 MachineOperand& dest1Oper = bInstr->getOperand(0);
10033 MachineOperand& dest2Oper = bInstr->getOperand(1);
10034 MachineOperand* argOpers[2 + X86::AddrNumOperands];
10035 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
10036 argOpers[i] = &bInstr->getOperand(i+2);
10038 // We use some of the operands multiple times, so conservatively just
10039 // clear any kill flags that might be present.
10040 if (argOpers[i]->isReg() && argOpers[i]->isUse())
10041 argOpers[i]->setIsKill(false);
10044 // x86 address has 5 operands: base, index, scale, displacement, and segment.
10045 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
10047 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
10048 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
10049 for (int i=0; i <= lastAddrIndx; ++i)
10050 (*MIB).addOperand(*argOpers[i]);
10051 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
10052 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
10053 // add 4 to displacement.
10054 for (int i=0; i <= lastAddrIndx-2; ++i)
10055 (*MIB).addOperand(*argOpers[i]);
10056 MachineOperand newOp3 = *(argOpers[3]);
10057 if (newOp3.isImm())
10058 newOp3.setImm(newOp3.getImm()+4);
10060 newOp3.setOffset(newOp3.getOffset()+4);
10061 (*MIB).addOperand(newOp3);
10062 (*MIB).addOperand(*argOpers[lastAddrIndx]);
10064 // t3/4 are defined later, at the bottom of the loop
10065 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
10066 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
10067 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
10068 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
10069 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
10070 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
10072 // The subsequent operations should be using the destination registers of
10073 //the PHI instructions.
10075 t1 = F->getRegInfo().createVirtualRegister(RC);
10076 t2 = F->getRegInfo().createVirtualRegister(RC);
10077 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
10078 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
10080 t1 = dest1Oper.getReg();
10081 t2 = dest2Oper.getReg();
10084 int valArgIndx = lastAddrIndx + 1;
10085 assert((argOpers[valArgIndx]->isReg() ||
10086 argOpers[valArgIndx]->isImm()) &&
10087 "invalid operand");
10088 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
10089 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
10090 if (argOpers[valArgIndx]->isReg())
10091 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
10093 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
10094 if (regOpcL != X86::MOV32rr)
10096 (*MIB).addOperand(*argOpers[valArgIndx]);
10097 assert(argOpers[valArgIndx + 1]->isReg() ==
10098 argOpers[valArgIndx]->isReg());
10099 assert(argOpers[valArgIndx + 1]->isImm() ==
10100 argOpers[valArgIndx]->isImm());
10101 if (argOpers[valArgIndx + 1]->isReg())
10102 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
10104 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
10105 if (regOpcH != X86::MOV32rr)
10107 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
10109 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
10111 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
10114 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
10116 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
10119 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
10120 for (int i=0; i <= lastAddrIndx; ++i)
10121 (*MIB).addOperand(*argOpers[i]);
10123 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
10124 (*MIB).setMemRefs(bInstr->memoperands_begin(),
10125 bInstr->memoperands_end());
10127 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
10128 MIB.addReg(X86::EAX);
10129 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
10130 MIB.addReg(X86::EDX);
10133 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
10135 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
10139 // private utility function
10140 MachineBasicBlock *
10141 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
10142 MachineBasicBlock *MBB,
10143 unsigned cmovOpc) const {
10144 // For the atomic min/max operator, we generate
10147 // ld t1 = [min/max.addr]
10148 // mov t2 = [min/max.val]
10150 // cmov[cond] t2 = t1
10152 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
10154 // fallthrough -->nextMBB
10156 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10157 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10158 MachineFunction::iterator MBBIter = MBB;
10161 /// First build the CFG
10162 MachineFunction *F = MBB->getParent();
10163 MachineBasicBlock *thisMBB = MBB;
10164 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
10165 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
10166 F->insert(MBBIter, newMBB);
10167 F->insert(MBBIter, nextMBB);
10169 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
10170 nextMBB->splice(nextMBB->begin(), thisMBB,
10171 llvm::next(MachineBasicBlock::iterator(mInstr)),
10173 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10175 // Update thisMBB to fall through to newMBB
10176 thisMBB->addSuccessor(newMBB);
10178 // newMBB jumps to newMBB and fall through to nextMBB
10179 newMBB->addSuccessor(nextMBB);
10180 newMBB->addSuccessor(newMBB);
10182 DebugLoc dl = mInstr->getDebugLoc();
10183 // Insert instructions into newMBB based on incoming instruction
10184 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
10185 "unexpected number of operands");
10186 MachineOperand& destOper = mInstr->getOperand(0);
10187 MachineOperand* argOpers[2 + X86::AddrNumOperands];
10188 int numArgs = mInstr->getNumOperands() - 1;
10189 for (int i=0; i < numArgs; ++i)
10190 argOpers[i] = &mInstr->getOperand(i+1);
10192 // x86 address has 4 operands: base, index, scale, and displacement
10193 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
10194 int valArgIndx = lastAddrIndx + 1;
10196 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
10197 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
10198 for (int i=0; i <= lastAddrIndx; ++i)
10199 (*MIB).addOperand(*argOpers[i]);
10201 // We only support register and immediate values
10202 assert((argOpers[valArgIndx]->isReg() ||
10203 argOpers[valArgIndx]->isImm()) &&
10204 "invalid operand");
10206 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
10207 if (argOpers[valArgIndx]->isReg())
10208 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
10210 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
10211 (*MIB).addOperand(*argOpers[valArgIndx]);
10213 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
10216 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
10221 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
10222 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
10226 // Cmp and exchange if none has modified the memory location
10227 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
10228 for (int i=0; i <= lastAddrIndx; ++i)
10229 (*MIB).addOperand(*argOpers[i]);
10231 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
10232 (*MIB).setMemRefs(mInstr->memoperands_begin(),
10233 mInstr->memoperands_end());
10235 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
10236 MIB.addReg(X86::EAX);
10239 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
10241 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
10245 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
10246 // or XMM0_V32I8 in AVX all of this code can be replaced with that
10247 // in the .td file.
10248 MachineBasicBlock *
10249 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
10250 unsigned numArgs, bool memArg) const {
10251 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
10252 "Target must have SSE4.2 or AVX features enabled");
10254 DebugLoc dl = MI->getDebugLoc();
10255 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10257 if (!Subtarget->hasAVX()) {
10259 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
10261 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
10264 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
10266 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
10269 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
10270 for (unsigned i = 0; i < numArgs; ++i) {
10271 MachineOperand &Op = MI->getOperand(i+1);
10272 if (!(Op.isReg() && Op.isImplicit()))
10273 MIB.addOperand(Op);
10275 BuildMI(*BB, MI, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
10276 .addReg(X86::XMM0);
10278 MI->eraseFromParent();
10282 MachineBasicBlock *
10283 X86TargetLowering::EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB) const {
10284 DebugLoc dl = MI->getDebugLoc();
10285 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10287 // Address into RAX/EAX, other two args into ECX, EDX.
10288 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
10289 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
10290 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
10291 for (int i = 0; i < X86::AddrNumOperands; ++i)
10292 MIB.addOperand(MI->getOperand(i));
10294 unsigned ValOps = X86::AddrNumOperands;
10295 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
10296 .addReg(MI->getOperand(ValOps).getReg());
10297 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
10298 .addReg(MI->getOperand(ValOps+1).getReg());
10300 // The instruction doesn't actually take any operands though.
10301 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
10303 MI->eraseFromParent(); // The pseudo is gone now.
10307 MachineBasicBlock *
10308 X86TargetLowering::EmitMwait(MachineInstr *MI, MachineBasicBlock *BB) const {
10309 DebugLoc dl = MI->getDebugLoc();
10310 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10312 // First arg in ECX, the second in EAX.
10313 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
10314 .addReg(MI->getOperand(0).getReg());
10315 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EAX)
10316 .addReg(MI->getOperand(1).getReg());
10318 // The instruction doesn't actually take any operands though.
10319 BuildMI(*BB, MI, dl, TII->get(X86::MWAITrr));
10321 MI->eraseFromParent(); // The pseudo is gone now.
10325 MachineBasicBlock *
10326 X86TargetLowering::EmitVAARG64WithCustomInserter(
10328 MachineBasicBlock *MBB) const {
10329 // Emit va_arg instruction on X86-64.
10331 // Operands to this pseudo-instruction:
10332 // 0 ) Output : destination address (reg)
10333 // 1-5) Input : va_list address (addr, i64mem)
10334 // 6 ) ArgSize : Size (in bytes) of vararg type
10335 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
10336 // 8 ) Align : Alignment of type
10337 // 9 ) EFLAGS (implicit-def)
10339 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
10340 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
10342 unsigned DestReg = MI->getOperand(0).getReg();
10343 MachineOperand &Base = MI->getOperand(1);
10344 MachineOperand &Scale = MI->getOperand(2);
10345 MachineOperand &Index = MI->getOperand(3);
10346 MachineOperand &Disp = MI->getOperand(4);
10347 MachineOperand &Segment = MI->getOperand(5);
10348 unsigned ArgSize = MI->getOperand(6).getImm();
10349 unsigned ArgMode = MI->getOperand(7).getImm();
10350 unsigned Align = MI->getOperand(8).getImm();
10352 // Memory Reference
10353 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
10354 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
10355 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
10357 // Machine Information
10358 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10359 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
10360 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
10361 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
10362 DebugLoc DL = MI->getDebugLoc();
10364 // struct va_list {
10367 // i64 overflow_area (address)
10368 // i64 reg_save_area (address)
10370 // sizeof(va_list) = 24
10371 // alignment(va_list) = 8
10373 unsigned TotalNumIntRegs = 6;
10374 unsigned TotalNumXMMRegs = 8;
10375 bool UseGPOffset = (ArgMode == 1);
10376 bool UseFPOffset = (ArgMode == 2);
10377 unsigned MaxOffset = TotalNumIntRegs * 8 +
10378 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
10380 /* Align ArgSize to a multiple of 8 */
10381 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
10382 bool NeedsAlign = (Align > 8);
10384 MachineBasicBlock *thisMBB = MBB;
10385 MachineBasicBlock *overflowMBB;
10386 MachineBasicBlock *offsetMBB;
10387 MachineBasicBlock *endMBB;
10389 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
10390 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
10391 unsigned OffsetReg = 0;
10393 if (!UseGPOffset && !UseFPOffset) {
10394 // If we only pull from the overflow region, we don't create a branch.
10395 // We don't need to alter control flow.
10396 OffsetDestReg = 0; // unused
10397 OverflowDestReg = DestReg;
10400 overflowMBB = thisMBB;
10403 // First emit code to check if gp_offset (or fp_offset) is below the bound.
10404 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
10405 // If not, pull from overflow_area. (branch to overflowMBB)
10410 // offsetMBB overflowMBB
10415 // Registers for the PHI in endMBB
10416 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
10417 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
10419 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10420 MachineFunction *MF = MBB->getParent();
10421 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10422 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10423 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10425 MachineFunction::iterator MBBIter = MBB;
10428 // Insert the new basic blocks
10429 MF->insert(MBBIter, offsetMBB);
10430 MF->insert(MBBIter, overflowMBB);
10431 MF->insert(MBBIter, endMBB);
10433 // Transfer the remainder of MBB and its successor edges to endMBB.
10434 endMBB->splice(endMBB->begin(), thisMBB,
10435 llvm::next(MachineBasicBlock::iterator(MI)),
10437 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10439 // Make offsetMBB and overflowMBB successors of thisMBB
10440 thisMBB->addSuccessor(offsetMBB);
10441 thisMBB->addSuccessor(overflowMBB);
10443 // endMBB is a successor of both offsetMBB and overflowMBB
10444 offsetMBB->addSuccessor(endMBB);
10445 overflowMBB->addSuccessor(endMBB);
10447 // Load the offset value into a register
10448 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
10449 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
10453 .addDisp(Disp, UseFPOffset ? 4 : 0)
10454 .addOperand(Segment)
10455 .setMemRefs(MMOBegin, MMOEnd);
10457 // Check if there is enough room left to pull this argument.
10458 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
10460 .addImm(MaxOffset + 8 - ArgSizeA8);
10462 // Branch to "overflowMBB" if offset >= max
10463 // Fall through to "offsetMBB" otherwise
10464 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
10465 .addMBB(overflowMBB);
10468 // In offsetMBB, emit code to use the reg_save_area.
10470 assert(OffsetReg != 0);
10472 // Read the reg_save_area address.
10473 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
10474 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
10479 .addOperand(Segment)
10480 .setMemRefs(MMOBegin, MMOEnd);
10482 // Zero-extend the offset
10483 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
10484 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
10487 .addImm(X86::sub_32bit);
10489 // Add the offset to the reg_save_area to get the final address.
10490 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
10491 .addReg(OffsetReg64)
10492 .addReg(RegSaveReg);
10494 // Compute the offset for the next argument
10495 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
10496 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
10498 .addImm(UseFPOffset ? 16 : 8);
10500 // Store it back into the va_list.
10501 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
10505 .addDisp(Disp, UseFPOffset ? 4 : 0)
10506 .addOperand(Segment)
10507 .addReg(NextOffsetReg)
10508 .setMemRefs(MMOBegin, MMOEnd);
10511 BuildMI(offsetMBB, DL, TII->get(X86::JMP_4))
10516 // Emit code to use overflow area
10519 // Load the overflow_area address into a register.
10520 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
10521 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
10526 .addOperand(Segment)
10527 .setMemRefs(MMOBegin, MMOEnd);
10529 // If we need to align it, do so. Otherwise, just copy the address
10530 // to OverflowDestReg.
10532 // Align the overflow address
10533 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
10534 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
10536 // aligned_addr = (addr + (align-1)) & ~(align-1)
10537 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
10538 .addReg(OverflowAddrReg)
10541 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
10543 .addImm(~(uint64_t)(Align-1));
10545 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
10546 .addReg(OverflowAddrReg);
10549 // Compute the next overflow address after this argument.
10550 // (the overflow address should be kept 8-byte aligned)
10551 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
10552 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
10553 .addReg(OverflowDestReg)
10554 .addImm(ArgSizeA8);
10556 // Store the new overflow address.
10557 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
10562 .addOperand(Segment)
10563 .addReg(NextAddrReg)
10564 .setMemRefs(MMOBegin, MMOEnd);
10566 // If we branched, emit the PHI to the front of endMBB.
10568 BuildMI(*endMBB, endMBB->begin(), DL,
10569 TII->get(X86::PHI), DestReg)
10570 .addReg(OffsetDestReg).addMBB(offsetMBB)
10571 .addReg(OverflowDestReg).addMBB(overflowMBB);
10574 // Erase the pseudo instruction
10575 MI->eraseFromParent();
10580 MachineBasicBlock *
10581 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
10583 MachineBasicBlock *MBB) const {
10584 // Emit code to save XMM registers to the stack. The ABI says that the
10585 // number of registers to save is given in %al, so it's theoretically
10586 // possible to do an indirect jump trick to avoid saving all of them,
10587 // however this code takes a simpler approach and just executes all
10588 // of the stores if %al is non-zero. It's less code, and it's probably
10589 // easier on the hardware branch predictor, and stores aren't all that
10590 // expensive anyway.
10592 // Create the new basic blocks. One block contains all the XMM stores,
10593 // and one block is the final destination regardless of whether any
10594 // stores were performed.
10595 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10596 MachineFunction *F = MBB->getParent();
10597 MachineFunction::iterator MBBIter = MBB;
10599 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
10600 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
10601 F->insert(MBBIter, XMMSaveMBB);
10602 F->insert(MBBIter, EndMBB);
10604 // Transfer the remainder of MBB and its successor edges to EndMBB.
10605 EndMBB->splice(EndMBB->begin(), MBB,
10606 llvm::next(MachineBasicBlock::iterator(MI)),
10608 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
10610 // The original block will now fall through to the XMM save block.
10611 MBB->addSuccessor(XMMSaveMBB);
10612 // The XMMSaveMBB will fall through to the end block.
10613 XMMSaveMBB->addSuccessor(EndMBB);
10615 // Now add the instructions.
10616 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10617 DebugLoc DL = MI->getDebugLoc();
10619 unsigned CountReg = MI->getOperand(0).getReg();
10620 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
10621 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
10623 if (!Subtarget->isTargetWin64()) {
10624 // If %al is 0, branch around the XMM save block.
10625 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
10626 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
10627 MBB->addSuccessor(EndMBB);
10630 // In the XMM save block, save all the XMM argument registers.
10631 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
10632 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
10633 MachineMemOperand *MMO =
10634 F->getMachineMemOperand(
10635 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
10636 MachineMemOperand::MOStore,
10637 /*Size=*/16, /*Align=*/16);
10638 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
10639 .addFrameIndex(RegSaveFrameIndex)
10640 .addImm(/*Scale=*/1)
10641 .addReg(/*IndexReg=*/0)
10642 .addImm(/*Disp=*/Offset)
10643 .addReg(/*Segment=*/0)
10644 .addReg(MI->getOperand(i).getReg())
10645 .addMemOperand(MMO);
10648 MI->eraseFromParent(); // The pseudo instruction is gone now.
10653 MachineBasicBlock *
10654 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
10655 MachineBasicBlock *BB) const {
10656 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10657 DebugLoc DL = MI->getDebugLoc();
10659 // To "insert" a SELECT_CC instruction, we actually have to insert the
10660 // diamond control-flow pattern. The incoming instruction knows the
10661 // destination vreg to set, the condition code register to branch on, the
10662 // true/false values to select between, and a branch opcode to use.
10663 const BasicBlock *LLVM_BB = BB->getBasicBlock();
10664 MachineFunction::iterator It = BB;
10670 // cmpTY ccX, r1, r2
10672 // fallthrough --> copy0MBB
10673 MachineBasicBlock *thisMBB = BB;
10674 MachineFunction *F = BB->getParent();
10675 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
10676 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
10677 F->insert(It, copy0MBB);
10678 F->insert(It, sinkMBB);
10680 // If the EFLAGS register isn't dead in the terminator, then claim that it's
10681 // live into the sink and copy blocks.
10682 const MachineFunction *MF = BB->getParent();
10683 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
10684 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
10686 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
10687 const MachineOperand &MO = MI->getOperand(I);
10688 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
10689 unsigned Reg = MO.getReg();
10690 if (Reg != X86::EFLAGS) continue;
10691 copy0MBB->addLiveIn(Reg);
10692 sinkMBB->addLiveIn(Reg);
10695 // Transfer the remainder of BB and its successor edges to sinkMBB.
10696 sinkMBB->splice(sinkMBB->begin(), BB,
10697 llvm::next(MachineBasicBlock::iterator(MI)),
10699 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
10701 // Add the true and fallthrough blocks as its successors.
10702 BB->addSuccessor(copy0MBB);
10703 BB->addSuccessor(sinkMBB);
10705 // Create the conditional branch instruction.
10707 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
10708 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
10711 // %FalseValue = ...
10712 // # fallthrough to sinkMBB
10713 copy0MBB->addSuccessor(sinkMBB);
10716 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
10718 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
10719 TII->get(X86::PHI), MI->getOperand(0).getReg())
10720 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
10721 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
10723 MI->eraseFromParent(); // The pseudo instruction is gone now.
10727 MachineBasicBlock *
10728 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
10729 MachineBasicBlock *BB) const {
10730 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10731 DebugLoc DL = MI->getDebugLoc();
10733 assert(!Subtarget->isTargetEnvMacho());
10735 // The lowering is pretty easy: we're just emitting the call to _alloca. The
10736 // non-trivial part is impdef of ESP.
10738 if (Subtarget->isTargetWin64()) {
10739 if (Subtarget->isTargetCygMing()) {
10740 // ___chkstk(Mingw64):
10741 // Clobbers R10, R11, RAX and EFLAGS.
10743 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
10744 .addExternalSymbol("___chkstk")
10745 .addReg(X86::RAX, RegState::Implicit)
10746 .addReg(X86::RSP, RegState::Implicit)
10747 .addReg(X86::RAX, RegState::Define | RegState::Implicit)
10748 .addReg(X86::RSP, RegState::Define | RegState::Implicit)
10749 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
10751 // __chkstk(MSVCRT): does not update stack pointer.
10752 // Clobbers R10, R11 and EFLAGS.
10753 // FIXME: RAX(allocated size) might be reused and not killed.
10754 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
10755 .addExternalSymbol("__chkstk")
10756 .addReg(X86::RAX, RegState::Implicit)
10757 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
10758 // RAX has the offset to subtracted from RSP.
10759 BuildMI(*BB, MI, DL, TII->get(X86::SUB64rr), X86::RSP)
10764 const char *StackProbeSymbol =
10765 Subtarget->isTargetWindows() ? "_chkstk" : "_alloca";
10767 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
10768 .addExternalSymbol(StackProbeSymbol)
10769 .addReg(X86::EAX, RegState::Implicit)
10770 .addReg(X86::ESP, RegState::Implicit)
10771 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
10772 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
10773 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
10776 MI->eraseFromParent(); // The pseudo instruction is gone now.
10780 MachineBasicBlock *
10781 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
10782 MachineBasicBlock *BB) const {
10783 // This is pretty easy. We're taking the value that we received from
10784 // our load from the relocation, sticking it in either RDI (x86-64)
10785 // or EAX and doing an indirect call. The return value will then
10786 // be in the normal return register.
10787 const X86InstrInfo *TII
10788 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
10789 DebugLoc DL = MI->getDebugLoc();
10790 MachineFunction *F = BB->getParent();
10792 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
10793 assert(MI->getOperand(3).isGlobal() && "This should be a global");
10795 if (Subtarget->is64Bit()) {
10796 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10797 TII->get(X86::MOV64rm), X86::RDI)
10799 .addImm(0).addReg(0)
10800 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10801 MI->getOperand(3).getTargetFlags())
10803 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
10804 addDirectMem(MIB, X86::RDI);
10805 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
10806 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10807 TII->get(X86::MOV32rm), X86::EAX)
10809 .addImm(0).addReg(0)
10810 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10811 MI->getOperand(3).getTargetFlags())
10813 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
10814 addDirectMem(MIB, X86::EAX);
10816 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10817 TII->get(X86::MOV32rm), X86::EAX)
10818 .addReg(TII->getGlobalBaseReg(F))
10819 .addImm(0).addReg(0)
10820 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10821 MI->getOperand(3).getTargetFlags())
10823 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
10824 addDirectMem(MIB, X86::EAX);
10827 MI->eraseFromParent(); // The pseudo instruction is gone now.
10831 MachineBasicBlock *
10832 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
10833 MachineBasicBlock *BB) const {
10834 switch (MI->getOpcode()) {
10835 default: assert(false && "Unexpected instr type to insert");
10836 case X86::TAILJMPd64:
10837 case X86::TAILJMPr64:
10838 case X86::TAILJMPm64:
10839 assert(!"TAILJMP64 would not be touched here.");
10840 case X86::TCRETURNdi64:
10841 case X86::TCRETURNri64:
10842 case X86::TCRETURNmi64:
10843 // Defs of TCRETURNxx64 has Win64's callee-saved registers, as subset.
10844 // On AMD64, additional defs should be added before register allocation.
10845 if (!Subtarget->isTargetWin64()) {
10846 MI->addRegisterDefined(X86::RSI);
10847 MI->addRegisterDefined(X86::RDI);
10848 MI->addRegisterDefined(X86::XMM6);
10849 MI->addRegisterDefined(X86::XMM7);
10850 MI->addRegisterDefined(X86::XMM8);
10851 MI->addRegisterDefined(X86::XMM9);
10852 MI->addRegisterDefined(X86::XMM10);
10853 MI->addRegisterDefined(X86::XMM11);
10854 MI->addRegisterDefined(X86::XMM12);
10855 MI->addRegisterDefined(X86::XMM13);
10856 MI->addRegisterDefined(X86::XMM14);
10857 MI->addRegisterDefined(X86::XMM15);
10860 case X86::WIN_ALLOCA:
10861 return EmitLoweredWinAlloca(MI, BB);
10862 case X86::TLSCall_32:
10863 case X86::TLSCall_64:
10864 return EmitLoweredTLSCall(MI, BB);
10865 case X86::CMOV_GR8:
10866 case X86::CMOV_FR32:
10867 case X86::CMOV_FR64:
10868 case X86::CMOV_V4F32:
10869 case X86::CMOV_V2F64:
10870 case X86::CMOV_V2I64:
10871 case X86::CMOV_GR16:
10872 case X86::CMOV_GR32:
10873 case X86::CMOV_RFP32:
10874 case X86::CMOV_RFP64:
10875 case X86::CMOV_RFP80:
10876 return EmitLoweredSelect(MI, BB);
10878 case X86::FP32_TO_INT16_IN_MEM:
10879 case X86::FP32_TO_INT32_IN_MEM:
10880 case X86::FP32_TO_INT64_IN_MEM:
10881 case X86::FP64_TO_INT16_IN_MEM:
10882 case X86::FP64_TO_INT32_IN_MEM:
10883 case X86::FP64_TO_INT64_IN_MEM:
10884 case X86::FP80_TO_INT16_IN_MEM:
10885 case X86::FP80_TO_INT32_IN_MEM:
10886 case X86::FP80_TO_INT64_IN_MEM: {
10887 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10888 DebugLoc DL = MI->getDebugLoc();
10890 // Change the floating point control register to use "round towards zero"
10891 // mode when truncating to an integer value.
10892 MachineFunction *F = BB->getParent();
10893 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
10894 addFrameReference(BuildMI(*BB, MI, DL,
10895 TII->get(X86::FNSTCW16m)), CWFrameIdx);
10897 // Load the old value of the high byte of the control word...
10899 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
10900 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
10903 // Set the high part to be round to zero...
10904 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
10907 // Reload the modified control word now...
10908 addFrameReference(BuildMI(*BB, MI, DL,
10909 TII->get(X86::FLDCW16m)), CWFrameIdx);
10911 // Restore the memory image of control word to original value
10912 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
10915 // Get the X86 opcode to use.
10917 switch (MI->getOpcode()) {
10918 default: llvm_unreachable("illegal opcode!");
10919 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
10920 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
10921 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
10922 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
10923 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
10924 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
10925 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
10926 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
10927 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
10931 MachineOperand &Op = MI->getOperand(0);
10933 AM.BaseType = X86AddressMode::RegBase;
10934 AM.Base.Reg = Op.getReg();
10936 AM.BaseType = X86AddressMode::FrameIndexBase;
10937 AM.Base.FrameIndex = Op.getIndex();
10939 Op = MI->getOperand(1);
10941 AM.Scale = Op.getImm();
10942 Op = MI->getOperand(2);
10944 AM.IndexReg = Op.getImm();
10945 Op = MI->getOperand(3);
10946 if (Op.isGlobal()) {
10947 AM.GV = Op.getGlobal();
10949 AM.Disp = Op.getImm();
10951 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
10952 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
10954 // Reload the original control word now.
10955 addFrameReference(BuildMI(*BB, MI, DL,
10956 TII->get(X86::FLDCW16m)), CWFrameIdx);
10958 MI->eraseFromParent(); // The pseudo instruction is gone now.
10961 // String/text processing lowering.
10962 case X86::PCMPISTRM128REG:
10963 case X86::VPCMPISTRM128REG:
10964 return EmitPCMP(MI, BB, 3, false /* in-mem */);
10965 case X86::PCMPISTRM128MEM:
10966 case X86::VPCMPISTRM128MEM:
10967 return EmitPCMP(MI, BB, 3, true /* in-mem */);
10968 case X86::PCMPESTRM128REG:
10969 case X86::VPCMPESTRM128REG:
10970 return EmitPCMP(MI, BB, 5, false /* in mem */);
10971 case X86::PCMPESTRM128MEM:
10972 case X86::VPCMPESTRM128MEM:
10973 return EmitPCMP(MI, BB, 5, true /* in mem */);
10975 // Thread synchronization.
10977 return EmitMonitor(MI, BB);
10979 return EmitMwait(MI, BB);
10981 // Atomic Lowering.
10982 case X86::ATOMAND32:
10983 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
10984 X86::AND32ri, X86::MOV32rm,
10986 X86::NOT32r, X86::EAX,
10987 X86::GR32RegisterClass);
10988 case X86::ATOMOR32:
10989 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
10990 X86::OR32ri, X86::MOV32rm,
10992 X86::NOT32r, X86::EAX,
10993 X86::GR32RegisterClass);
10994 case X86::ATOMXOR32:
10995 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
10996 X86::XOR32ri, X86::MOV32rm,
10998 X86::NOT32r, X86::EAX,
10999 X86::GR32RegisterClass);
11000 case X86::ATOMNAND32:
11001 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
11002 X86::AND32ri, X86::MOV32rm,
11004 X86::NOT32r, X86::EAX,
11005 X86::GR32RegisterClass, true);
11006 case X86::ATOMMIN32:
11007 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
11008 case X86::ATOMMAX32:
11009 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
11010 case X86::ATOMUMIN32:
11011 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
11012 case X86::ATOMUMAX32:
11013 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
11015 case X86::ATOMAND16:
11016 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
11017 X86::AND16ri, X86::MOV16rm,
11019 X86::NOT16r, X86::AX,
11020 X86::GR16RegisterClass);
11021 case X86::ATOMOR16:
11022 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
11023 X86::OR16ri, X86::MOV16rm,
11025 X86::NOT16r, X86::AX,
11026 X86::GR16RegisterClass);
11027 case X86::ATOMXOR16:
11028 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
11029 X86::XOR16ri, X86::MOV16rm,
11031 X86::NOT16r, X86::AX,
11032 X86::GR16RegisterClass);
11033 case X86::ATOMNAND16:
11034 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
11035 X86::AND16ri, X86::MOV16rm,
11037 X86::NOT16r, X86::AX,
11038 X86::GR16RegisterClass, true);
11039 case X86::ATOMMIN16:
11040 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
11041 case X86::ATOMMAX16:
11042 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
11043 case X86::ATOMUMIN16:
11044 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
11045 case X86::ATOMUMAX16:
11046 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
11048 case X86::ATOMAND8:
11049 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
11050 X86::AND8ri, X86::MOV8rm,
11052 X86::NOT8r, X86::AL,
11053 X86::GR8RegisterClass);
11055 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
11056 X86::OR8ri, X86::MOV8rm,
11058 X86::NOT8r, X86::AL,
11059 X86::GR8RegisterClass);
11060 case X86::ATOMXOR8:
11061 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
11062 X86::XOR8ri, X86::MOV8rm,
11064 X86::NOT8r, X86::AL,
11065 X86::GR8RegisterClass);
11066 case X86::ATOMNAND8:
11067 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
11068 X86::AND8ri, X86::MOV8rm,
11070 X86::NOT8r, X86::AL,
11071 X86::GR8RegisterClass, true);
11072 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
11073 // This group is for 64-bit host.
11074 case X86::ATOMAND64:
11075 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
11076 X86::AND64ri32, X86::MOV64rm,
11078 X86::NOT64r, X86::RAX,
11079 X86::GR64RegisterClass);
11080 case X86::ATOMOR64:
11081 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
11082 X86::OR64ri32, X86::MOV64rm,
11084 X86::NOT64r, X86::RAX,
11085 X86::GR64RegisterClass);
11086 case X86::ATOMXOR64:
11087 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
11088 X86::XOR64ri32, X86::MOV64rm,
11090 X86::NOT64r, X86::RAX,
11091 X86::GR64RegisterClass);
11092 case X86::ATOMNAND64:
11093 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
11094 X86::AND64ri32, X86::MOV64rm,
11096 X86::NOT64r, X86::RAX,
11097 X86::GR64RegisterClass, true);
11098 case X86::ATOMMIN64:
11099 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
11100 case X86::ATOMMAX64:
11101 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
11102 case X86::ATOMUMIN64:
11103 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
11104 case X86::ATOMUMAX64:
11105 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
11107 // This group does 64-bit operations on a 32-bit host.
11108 case X86::ATOMAND6432:
11109 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11110 X86::AND32rr, X86::AND32rr,
11111 X86::AND32ri, X86::AND32ri,
11113 case X86::ATOMOR6432:
11114 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11115 X86::OR32rr, X86::OR32rr,
11116 X86::OR32ri, X86::OR32ri,
11118 case X86::ATOMXOR6432:
11119 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11120 X86::XOR32rr, X86::XOR32rr,
11121 X86::XOR32ri, X86::XOR32ri,
11123 case X86::ATOMNAND6432:
11124 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11125 X86::AND32rr, X86::AND32rr,
11126 X86::AND32ri, X86::AND32ri,
11128 case X86::ATOMADD6432:
11129 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11130 X86::ADD32rr, X86::ADC32rr,
11131 X86::ADD32ri, X86::ADC32ri,
11133 case X86::ATOMSUB6432:
11134 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11135 X86::SUB32rr, X86::SBB32rr,
11136 X86::SUB32ri, X86::SBB32ri,
11138 case X86::ATOMSWAP6432:
11139 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11140 X86::MOV32rr, X86::MOV32rr,
11141 X86::MOV32ri, X86::MOV32ri,
11143 case X86::VASTART_SAVE_XMM_REGS:
11144 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
11146 case X86::VAARG_64:
11147 return EmitVAARG64WithCustomInserter(MI, BB);
11151 //===----------------------------------------------------------------------===//
11152 // X86 Optimization Hooks
11153 //===----------------------------------------------------------------------===//
11155 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
11159 const SelectionDAG &DAG,
11160 unsigned Depth) const {
11161 unsigned Opc = Op.getOpcode();
11162 assert((Opc >= ISD::BUILTIN_OP_END ||
11163 Opc == ISD::INTRINSIC_WO_CHAIN ||
11164 Opc == ISD::INTRINSIC_W_CHAIN ||
11165 Opc == ISD::INTRINSIC_VOID) &&
11166 "Should use MaskedValueIsZero if you don't know whether Op"
11167 " is a target node!");
11169 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
11183 // These nodes' second result is a boolean.
11184 if (Op.getResNo() == 0)
11187 case X86ISD::SETCC:
11188 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
11189 Mask.getBitWidth() - 1);
11194 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
11195 unsigned Depth) const {
11196 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
11197 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
11198 return Op.getValueType().getScalarType().getSizeInBits();
11204 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
11205 /// node is a GlobalAddress + offset.
11206 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
11207 const GlobalValue* &GA,
11208 int64_t &Offset) const {
11209 if (N->getOpcode() == X86ISD::Wrapper) {
11210 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
11211 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
11212 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
11216 return TargetLowering::isGAPlusOffset(N, GA, Offset);
11219 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
11220 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
11221 TargetLowering::DAGCombinerInfo &DCI) {
11222 DebugLoc dl = N->getDebugLoc();
11223 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
11224 SDValue V1 = SVOp->getOperand(0);
11225 SDValue V2 = SVOp->getOperand(1);
11226 EVT VT = SVOp->getValueType(0);
11228 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
11229 V2.getOpcode() == ISD::CONCAT_VECTORS) {
11233 // V UNDEF BUILD_VECTOR UNDEF
11235 // CONCAT_VECTOR CONCAT_VECTOR
11238 // RESULT: V + zero extended
11240 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
11241 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
11242 V1.getOperand(1).getOpcode() != ISD::UNDEF)
11245 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
11248 // To match the shuffle mask, the first half of the mask should
11249 // be exactly the first vector, and all the rest a splat with the
11250 // first element of the second one.
11251 int NumElems = VT.getVectorNumElements();
11252 for (int i = 0; i < NumElems/2; ++i)
11253 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
11254 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
11257 // Emit a zeroed vector and insert the desired subvector on its
11259 SDValue Zeros = getZeroVector(VT, true /* HasSSE2 */, DAG, dl);
11260 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0),
11261 DAG.getConstant(0, MVT::i32), DAG, dl);
11262 return DCI.CombineTo(N, InsV);
11268 /// PerformShuffleCombine - Performs several different shuffle combines.
11269 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
11270 TargetLowering::DAGCombinerInfo &DCI) {
11271 DebugLoc dl = N->getDebugLoc();
11272 EVT VT = N->getValueType(0);
11274 // Don't create instructions with illegal types after legalize types has run.
11275 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11276 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
11279 // Only handle pure VECTOR_SHUFFLE nodes.
11280 if (VT.getSizeInBits() == 256 && N->getOpcode() == ISD::VECTOR_SHUFFLE)
11281 return PerformShuffleCombine256(N, DAG, DCI);
11283 // Only handle 128 wide vector from here on.
11284 if (VT.getSizeInBits() != 128)
11287 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
11288 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
11289 // consecutive, non-overlapping, and in the right order.
11290 SmallVector<SDValue, 16> Elts;
11291 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
11292 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
11294 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
11297 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
11298 /// generation and convert it from being a bunch of shuffles and extracts
11299 /// to a simple store and scalar loads to extract the elements.
11300 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
11301 const TargetLowering &TLI) {
11302 SDValue InputVector = N->getOperand(0);
11304 // Only operate on vectors of 4 elements, where the alternative shuffling
11305 // gets to be more expensive.
11306 if (InputVector.getValueType() != MVT::v4i32)
11309 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
11310 // single use which is a sign-extend or zero-extend, and all elements are
11312 SmallVector<SDNode *, 4> Uses;
11313 unsigned ExtractedElements = 0;
11314 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
11315 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
11316 if (UI.getUse().getResNo() != InputVector.getResNo())
11319 SDNode *Extract = *UI;
11320 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
11323 if (Extract->getValueType(0) != MVT::i32)
11325 if (!Extract->hasOneUse())
11327 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
11328 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
11330 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
11333 // Record which element was extracted.
11334 ExtractedElements |=
11335 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
11337 Uses.push_back(Extract);
11340 // If not all the elements were used, this may not be worthwhile.
11341 if (ExtractedElements != 15)
11344 // Ok, we've now decided to do the transformation.
11345 DebugLoc dl = InputVector.getDebugLoc();
11347 // Store the value to a temporary stack slot.
11348 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
11349 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
11350 MachinePointerInfo(), false, false, 0);
11352 // Replace each use (extract) with a load of the appropriate element.
11353 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
11354 UE = Uses.end(); UI != UE; ++UI) {
11355 SDNode *Extract = *UI;
11357 // cOMpute the element's address.
11358 SDValue Idx = Extract->getOperand(1);
11360 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
11361 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
11362 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
11364 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
11365 StackPtr, OffsetVal);
11367 // Load the scalar.
11368 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
11369 ScalarAddr, MachinePointerInfo(),
11372 // Replace the exact with the load.
11373 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
11376 // The replacement was made in place; don't return anything.
11380 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
11381 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
11382 const X86Subtarget *Subtarget) {
11383 DebugLoc DL = N->getDebugLoc();
11384 SDValue Cond = N->getOperand(0);
11385 // Get the LHS/RHS of the select.
11386 SDValue LHS = N->getOperand(1);
11387 SDValue RHS = N->getOperand(2);
11389 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
11390 // instructions match the semantics of the common C idiom x<y?x:y but not
11391 // x<=y?x:y, because of how they handle negative zero (which can be
11392 // ignored in unsafe-math mode).
11393 if (Subtarget->hasSSE2() &&
11394 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
11395 Cond.getOpcode() == ISD::SETCC) {
11396 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
11398 unsigned Opcode = 0;
11399 // Check for x CC y ? x : y.
11400 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
11401 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
11405 // Converting this to a min would handle NaNs incorrectly, and swapping
11406 // the operands would cause it to handle comparisons between positive
11407 // and negative zero incorrectly.
11408 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
11409 if (!UnsafeFPMath &&
11410 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
11412 std::swap(LHS, RHS);
11414 Opcode = X86ISD::FMIN;
11417 // Converting this to a min would handle comparisons between positive
11418 // and negative zero incorrectly.
11419 if (!UnsafeFPMath &&
11420 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
11422 Opcode = X86ISD::FMIN;
11425 // Converting this to a min would handle both negative zeros and NaNs
11426 // incorrectly, but we can swap the operands to fix both.
11427 std::swap(LHS, RHS);
11431 Opcode = X86ISD::FMIN;
11435 // Converting this to a max would handle comparisons between positive
11436 // and negative zero incorrectly.
11437 if (!UnsafeFPMath &&
11438 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
11440 Opcode = X86ISD::FMAX;
11443 // Converting this to a max would handle NaNs incorrectly, and swapping
11444 // the operands would cause it to handle comparisons between positive
11445 // and negative zero incorrectly.
11446 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
11447 if (!UnsafeFPMath &&
11448 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
11450 std::swap(LHS, RHS);
11452 Opcode = X86ISD::FMAX;
11455 // Converting this to a max would handle both negative zeros and NaNs
11456 // incorrectly, but we can swap the operands to fix both.
11457 std::swap(LHS, RHS);
11461 Opcode = X86ISD::FMAX;
11464 // Check for x CC y ? y : x -- a min/max with reversed arms.
11465 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
11466 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
11470 // Converting this to a min would handle comparisons between positive
11471 // and negative zero incorrectly, and swapping the operands would
11472 // cause it to handle NaNs incorrectly.
11473 if (!UnsafeFPMath &&
11474 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
11475 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11477 std::swap(LHS, RHS);
11479 Opcode = X86ISD::FMIN;
11482 // Converting this to a min would handle NaNs incorrectly.
11483 if (!UnsafeFPMath &&
11484 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
11486 Opcode = X86ISD::FMIN;
11489 // Converting this to a min would handle both negative zeros and NaNs
11490 // incorrectly, but we can swap the operands to fix both.
11491 std::swap(LHS, RHS);
11495 Opcode = X86ISD::FMIN;
11499 // Converting this to a max would handle NaNs incorrectly.
11500 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11502 Opcode = X86ISD::FMAX;
11505 // Converting this to a max would handle comparisons between positive
11506 // and negative zero incorrectly, and swapping the operands would
11507 // cause it to handle NaNs incorrectly.
11508 if (!UnsafeFPMath &&
11509 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
11510 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11512 std::swap(LHS, RHS);
11514 Opcode = X86ISD::FMAX;
11517 // Converting this to a max would handle both negative zeros and NaNs
11518 // incorrectly, but we can swap the operands to fix both.
11519 std::swap(LHS, RHS);
11523 Opcode = X86ISD::FMAX;
11529 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
11532 // If this is a select between two integer constants, try to do some
11534 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
11535 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
11536 // Don't do this for crazy integer types.
11537 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
11538 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
11539 // so that TrueC (the true value) is larger than FalseC.
11540 bool NeedsCondInvert = false;
11542 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
11543 // Efficiently invertible.
11544 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
11545 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
11546 isa<ConstantSDNode>(Cond.getOperand(1))))) {
11547 NeedsCondInvert = true;
11548 std::swap(TrueC, FalseC);
11551 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
11552 if (FalseC->getAPIntValue() == 0 &&
11553 TrueC->getAPIntValue().isPowerOf2()) {
11554 if (NeedsCondInvert) // Invert the condition if needed.
11555 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11556 DAG.getConstant(1, Cond.getValueType()));
11558 // Zero extend the condition if needed.
11559 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
11561 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
11562 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
11563 DAG.getConstant(ShAmt, MVT::i8));
11566 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
11567 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
11568 if (NeedsCondInvert) // Invert the condition if needed.
11569 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11570 DAG.getConstant(1, Cond.getValueType()));
11572 // Zero extend the condition if needed.
11573 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
11574 FalseC->getValueType(0), Cond);
11575 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11576 SDValue(FalseC, 0));
11579 // Optimize cases that will turn into an LEA instruction. This requires
11580 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
11581 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
11582 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
11583 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
11585 bool isFastMultiplier = false;
11587 switch ((unsigned char)Diff) {
11589 case 1: // result = add base, cond
11590 case 2: // result = lea base( , cond*2)
11591 case 3: // result = lea base(cond, cond*2)
11592 case 4: // result = lea base( , cond*4)
11593 case 5: // result = lea base(cond, cond*4)
11594 case 8: // result = lea base( , cond*8)
11595 case 9: // result = lea base(cond, cond*8)
11596 isFastMultiplier = true;
11601 if (isFastMultiplier) {
11602 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
11603 if (NeedsCondInvert) // Invert the condition if needed.
11604 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11605 DAG.getConstant(1, Cond.getValueType()));
11607 // Zero extend the condition if needed.
11608 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
11610 // Scale the condition by the difference.
11612 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
11613 DAG.getConstant(Diff, Cond.getValueType()));
11615 // Add the base if non-zero.
11616 if (FalseC->getAPIntValue() != 0)
11617 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11618 SDValue(FalseC, 0));
11628 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
11629 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
11630 TargetLowering::DAGCombinerInfo &DCI) {
11631 DebugLoc DL = N->getDebugLoc();
11633 // If the flag operand isn't dead, don't touch this CMOV.
11634 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
11637 SDValue FalseOp = N->getOperand(0);
11638 SDValue TrueOp = N->getOperand(1);
11639 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
11640 SDValue Cond = N->getOperand(3);
11641 if (CC == X86::COND_E || CC == X86::COND_NE) {
11642 switch (Cond.getOpcode()) {
11646 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
11647 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
11648 return (CC == X86::COND_E) ? FalseOp : TrueOp;
11652 // If this is a select between two integer constants, try to do some
11653 // optimizations. Note that the operands are ordered the opposite of SELECT
11655 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
11656 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
11657 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
11658 // larger than FalseC (the false value).
11659 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
11660 CC = X86::GetOppositeBranchCondition(CC);
11661 std::swap(TrueC, FalseC);
11664 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
11665 // This is efficient for any integer data type (including i8/i16) and
11667 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
11668 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11669 DAG.getConstant(CC, MVT::i8), Cond);
11671 // Zero extend the condition if needed.
11672 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
11674 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
11675 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
11676 DAG.getConstant(ShAmt, MVT::i8));
11677 if (N->getNumValues() == 2) // Dead flag value?
11678 return DCI.CombineTo(N, Cond, SDValue());
11682 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
11683 // for any integer data type, including i8/i16.
11684 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
11685 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11686 DAG.getConstant(CC, MVT::i8), Cond);
11688 // Zero extend the condition if needed.
11689 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
11690 FalseC->getValueType(0), Cond);
11691 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11692 SDValue(FalseC, 0));
11694 if (N->getNumValues() == 2) // Dead flag value?
11695 return DCI.CombineTo(N, Cond, SDValue());
11699 // Optimize cases that will turn into an LEA instruction. This requires
11700 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
11701 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
11702 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
11703 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
11705 bool isFastMultiplier = false;
11707 switch ((unsigned char)Diff) {
11709 case 1: // result = add base, cond
11710 case 2: // result = lea base( , cond*2)
11711 case 3: // result = lea base(cond, cond*2)
11712 case 4: // result = lea base( , cond*4)
11713 case 5: // result = lea base(cond, cond*4)
11714 case 8: // result = lea base( , cond*8)
11715 case 9: // result = lea base(cond, cond*8)
11716 isFastMultiplier = true;
11721 if (isFastMultiplier) {
11722 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
11723 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11724 DAG.getConstant(CC, MVT::i8), Cond);
11725 // Zero extend the condition if needed.
11726 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
11728 // Scale the condition by the difference.
11730 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
11731 DAG.getConstant(Diff, Cond.getValueType()));
11733 // Add the base if non-zero.
11734 if (FalseC->getAPIntValue() != 0)
11735 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11736 SDValue(FalseC, 0));
11737 if (N->getNumValues() == 2) // Dead flag value?
11738 return DCI.CombineTo(N, Cond, SDValue());
11748 /// PerformMulCombine - Optimize a single multiply with constant into two
11749 /// in order to implement it with two cheaper instructions, e.g.
11750 /// LEA + SHL, LEA + LEA.
11751 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
11752 TargetLowering::DAGCombinerInfo &DCI) {
11753 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
11756 EVT VT = N->getValueType(0);
11757 if (VT != MVT::i64)
11760 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
11763 uint64_t MulAmt = C->getZExtValue();
11764 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
11767 uint64_t MulAmt1 = 0;
11768 uint64_t MulAmt2 = 0;
11769 if ((MulAmt % 9) == 0) {
11771 MulAmt2 = MulAmt / 9;
11772 } else if ((MulAmt % 5) == 0) {
11774 MulAmt2 = MulAmt / 5;
11775 } else if ((MulAmt % 3) == 0) {
11777 MulAmt2 = MulAmt / 3;
11780 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
11781 DebugLoc DL = N->getDebugLoc();
11783 if (isPowerOf2_64(MulAmt2) &&
11784 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
11785 // If second multiplifer is pow2, issue it first. We want the multiply by
11786 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
11788 std::swap(MulAmt1, MulAmt2);
11791 if (isPowerOf2_64(MulAmt1))
11792 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
11793 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
11795 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
11796 DAG.getConstant(MulAmt1, VT));
11798 if (isPowerOf2_64(MulAmt2))
11799 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
11800 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
11802 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
11803 DAG.getConstant(MulAmt2, VT));
11805 // Do not add new nodes to DAG combiner worklist.
11806 DCI.CombineTo(N, NewMul, false);
11811 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
11812 SDValue N0 = N->getOperand(0);
11813 SDValue N1 = N->getOperand(1);
11814 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
11815 EVT VT = N0.getValueType();
11817 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
11818 // since the result of setcc_c is all zero's or all ones.
11819 if (N1C && N0.getOpcode() == ISD::AND &&
11820 N0.getOperand(1).getOpcode() == ISD::Constant) {
11821 SDValue N00 = N0.getOperand(0);
11822 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
11823 ((N00.getOpcode() == ISD::ANY_EXTEND ||
11824 N00.getOpcode() == ISD::ZERO_EXTEND) &&
11825 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
11826 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
11827 APInt ShAmt = N1C->getAPIntValue();
11828 Mask = Mask.shl(ShAmt);
11830 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
11831 N00, DAG.getConstant(Mask, VT));
11838 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
11840 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
11841 const X86Subtarget *Subtarget) {
11842 EVT VT = N->getValueType(0);
11843 if (!VT.isVector() && VT.isInteger() &&
11844 N->getOpcode() == ISD::SHL)
11845 return PerformSHLCombine(N, DAG);
11847 // On X86 with SSE2 support, we can transform this to a vector shift if
11848 // all elements are shifted by the same amount. We can't do this in legalize
11849 // because the a constant vector is typically transformed to a constant pool
11850 // so we have no knowledge of the shift amount.
11851 if (!Subtarget->hasSSE2())
11854 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
11857 SDValue ShAmtOp = N->getOperand(1);
11858 EVT EltVT = VT.getVectorElementType();
11859 DebugLoc DL = N->getDebugLoc();
11860 SDValue BaseShAmt = SDValue();
11861 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
11862 unsigned NumElts = VT.getVectorNumElements();
11864 for (; i != NumElts; ++i) {
11865 SDValue Arg = ShAmtOp.getOperand(i);
11866 if (Arg.getOpcode() == ISD::UNDEF) continue;
11870 for (; i != NumElts; ++i) {
11871 SDValue Arg = ShAmtOp.getOperand(i);
11872 if (Arg.getOpcode() == ISD::UNDEF) continue;
11873 if (Arg != BaseShAmt) {
11877 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
11878 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
11879 SDValue InVec = ShAmtOp.getOperand(0);
11880 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
11881 unsigned NumElts = InVec.getValueType().getVectorNumElements();
11883 for (; i != NumElts; ++i) {
11884 SDValue Arg = InVec.getOperand(i);
11885 if (Arg.getOpcode() == ISD::UNDEF) continue;
11889 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
11890 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
11891 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
11892 if (C->getZExtValue() == SplatIdx)
11893 BaseShAmt = InVec.getOperand(1);
11896 if (BaseShAmt.getNode() == 0)
11897 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
11898 DAG.getIntPtrConstant(0));
11902 // The shift amount is an i32.
11903 if (EltVT.bitsGT(MVT::i32))
11904 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
11905 else if (EltVT.bitsLT(MVT::i32))
11906 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
11908 // The shift amount is identical so we can do a vector shift.
11909 SDValue ValOp = N->getOperand(0);
11910 switch (N->getOpcode()) {
11912 llvm_unreachable("Unknown shift opcode!");
11915 if (VT == MVT::v2i64)
11916 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11917 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
11919 if (VT == MVT::v4i32)
11920 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11921 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
11923 if (VT == MVT::v8i16)
11924 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11925 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
11929 if (VT == MVT::v4i32)
11930 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11931 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
11933 if (VT == MVT::v8i16)
11934 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11935 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
11939 if (VT == MVT::v2i64)
11940 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11941 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
11943 if (VT == MVT::v4i32)
11944 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11945 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
11947 if (VT == MVT::v8i16)
11948 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11949 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
11957 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
11958 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
11959 // and friends. Likewise for OR -> CMPNEQSS.
11960 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
11961 TargetLowering::DAGCombinerInfo &DCI,
11962 const X86Subtarget *Subtarget) {
11965 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
11966 // we're requiring SSE2 for both.
11967 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
11968 SDValue N0 = N->getOperand(0);
11969 SDValue N1 = N->getOperand(1);
11970 SDValue CMP0 = N0->getOperand(1);
11971 SDValue CMP1 = N1->getOperand(1);
11972 DebugLoc DL = N->getDebugLoc();
11974 // The SETCCs should both refer to the same CMP.
11975 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
11978 SDValue CMP00 = CMP0->getOperand(0);
11979 SDValue CMP01 = CMP0->getOperand(1);
11980 EVT VT = CMP00.getValueType();
11982 if (VT == MVT::f32 || VT == MVT::f64) {
11983 bool ExpectingFlags = false;
11984 // Check for any users that want flags:
11985 for (SDNode::use_iterator UI = N->use_begin(),
11987 !ExpectingFlags && UI != UE; ++UI)
11988 switch (UI->getOpcode()) {
11993 ExpectingFlags = true;
11995 case ISD::CopyToReg:
11996 case ISD::SIGN_EXTEND:
11997 case ISD::ZERO_EXTEND:
11998 case ISD::ANY_EXTEND:
12002 if (!ExpectingFlags) {
12003 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
12004 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
12006 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
12007 X86::CondCode tmp = cc0;
12012 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
12013 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
12014 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
12015 X86ISD::NodeType NTOperator = is64BitFP ?
12016 X86ISD::FSETCCsd : X86ISD::FSETCCss;
12017 // FIXME: need symbolic constants for these magic numbers.
12018 // See X86ATTInstPrinter.cpp:printSSECC().
12019 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
12020 SDValue OnesOrZeroesF = DAG.getNode(NTOperator, DL, MVT::f32, CMP00, CMP01,
12021 DAG.getConstant(x86cc, MVT::i8));
12022 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, MVT::i32,
12024 SDValue ANDed = DAG.getNode(ISD::AND, DL, MVT::i32, OnesOrZeroesI,
12025 DAG.getConstant(1, MVT::i32));
12026 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
12027 return OneBitOfTruth;
12035 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
12036 /// so it can be folded inside ANDNP.
12037 static bool CanFoldXORWithAllOnes(const SDNode *N) {
12038 EVT VT = N->getValueType(0);
12040 // Match direct AllOnes for 128 and 256-bit vectors
12041 if (ISD::isBuildVectorAllOnes(N))
12044 // Look through a bit convert.
12045 if (N->getOpcode() == ISD::BITCAST)
12046 N = N->getOperand(0).getNode();
12048 // Sometimes the operand may come from a insert_subvector building a 256-bit
12050 SDValue V1 = N->getOperand(0);
12051 SDValue V2 = N->getOperand(1);
12053 if (VT.getSizeInBits() == 256 &&
12054 N->getOpcode() == ISD::INSERT_SUBVECTOR &&
12055 V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
12056 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
12057 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
12058 ISD::isBuildVectorAllOnes(V2.getNode()))
12064 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
12065 TargetLowering::DAGCombinerInfo &DCI,
12066 const X86Subtarget *Subtarget) {
12067 if (DCI.isBeforeLegalizeOps())
12070 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
12074 // Want to form ANDNP nodes:
12075 // 1) In the hopes of then easily combining them with OR and AND nodes
12076 // to form PBLEND/PSIGN.
12077 // 2) To match ANDN packed intrinsics
12078 EVT VT = N->getValueType(0);
12079 if (VT != MVT::v2i64 && VT != MVT::v4i64)
12082 SDValue N0 = N->getOperand(0);
12083 SDValue N1 = N->getOperand(1);
12084 DebugLoc DL = N->getDebugLoc();
12086 // Check LHS for vnot
12087 if (N0.getOpcode() == ISD::XOR &&
12088 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
12089 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
12090 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
12092 // Check RHS for vnot
12093 if (N1.getOpcode() == ISD::XOR &&
12094 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
12095 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
12096 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
12101 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
12102 TargetLowering::DAGCombinerInfo &DCI,
12103 const X86Subtarget *Subtarget) {
12104 if (DCI.isBeforeLegalizeOps())
12107 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
12111 EVT VT = N->getValueType(0);
12112 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64 && VT != MVT::v2i64)
12115 SDValue N0 = N->getOperand(0);
12116 SDValue N1 = N->getOperand(1);
12118 // look for psign/blend
12119 if (Subtarget->hasSSSE3()) {
12120 if (VT == MVT::v2i64) {
12121 // Canonicalize pandn to RHS
12122 if (N0.getOpcode() == X86ISD::ANDNP)
12124 // or (and (m, x), (pandn m, y))
12125 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
12126 SDValue Mask = N1.getOperand(0);
12127 SDValue X = N1.getOperand(1);
12129 if (N0.getOperand(0) == Mask)
12130 Y = N0.getOperand(1);
12131 if (N0.getOperand(1) == Mask)
12132 Y = N0.getOperand(0);
12134 // Check to see if the mask appeared in both the AND and ANDNP and
12138 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
12139 if (Mask.getOpcode() != ISD::BITCAST ||
12140 X.getOpcode() != ISD::BITCAST ||
12141 Y.getOpcode() != ISD::BITCAST)
12144 // Look through mask bitcast.
12145 Mask = Mask.getOperand(0);
12146 EVT MaskVT = Mask.getValueType();
12148 // Validate that the Mask operand is a vector sra node. The sra node
12149 // will be an intrinsic.
12150 if (Mask.getOpcode() != ISD::INTRINSIC_WO_CHAIN)
12153 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
12154 // there is no psrai.b
12155 switch (cast<ConstantSDNode>(Mask.getOperand(0))->getZExtValue()) {
12156 case Intrinsic::x86_sse2_psrai_w:
12157 case Intrinsic::x86_sse2_psrai_d:
12159 default: return SDValue();
12162 // Check that the SRA is all signbits.
12163 SDValue SraC = Mask.getOperand(2);
12164 unsigned SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
12165 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
12166 if ((SraAmt + 1) != EltBits)
12169 DebugLoc DL = N->getDebugLoc();
12171 // Now we know we at least have a plendvb with the mask val. See if
12172 // we can form a psignb/w/d.
12173 // psign = x.type == y.type == mask.type && y = sub(0, x);
12174 X = X.getOperand(0);
12175 Y = Y.getOperand(0);
12176 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
12177 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
12178 X.getValueType() == MaskVT && X.getValueType() == Y.getValueType()){
12181 case 8: Opc = X86ISD::PSIGNB; break;
12182 case 16: Opc = X86ISD::PSIGNW; break;
12183 case 32: Opc = X86ISD::PSIGND; break;
12187 SDValue Sign = DAG.getNode(Opc, DL, MaskVT, X, Mask.getOperand(1));
12188 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Sign);
12191 // PBLENDVB only available on SSE 4.1
12192 if (!Subtarget->hasSSE41())
12195 X = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, X);
12196 Y = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Y);
12197 Mask = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Mask);
12198 Mask = DAG.getNode(X86ISD::PBLENDVB, DL, MVT::v16i8, X, Y, Mask);
12199 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Mask);
12204 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
12205 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
12207 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
12209 if (!N0.hasOneUse() || !N1.hasOneUse())
12212 SDValue ShAmt0 = N0.getOperand(1);
12213 if (ShAmt0.getValueType() != MVT::i8)
12215 SDValue ShAmt1 = N1.getOperand(1);
12216 if (ShAmt1.getValueType() != MVT::i8)
12218 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
12219 ShAmt0 = ShAmt0.getOperand(0);
12220 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
12221 ShAmt1 = ShAmt1.getOperand(0);
12223 DebugLoc DL = N->getDebugLoc();
12224 unsigned Opc = X86ISD::SHLD;
12225 SDValue Op0 = N0.getOperand(0);
12226 SDValue Op1 = N1.getOperand(0);
12227 if (ShAmt0.getOpcode() == ISD::SUB) {
12228 Opc = X86ISD::SHRD;
12229 std::swap(Op0, Op1);
12230 std::swap(ShAmt0, ShAmt1);
12233 unsigned Bits = VT.getSizeInBits();
12234 if (ShAmt1.getOpcode() == ISD::SUB) {
12235 SDValue Sum = ShAmt1.getOperand(0);
12236 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
12237 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
12238 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
12239 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
12240 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
12241 return DAG.getNode(Opc, DL, VT,
12243 DAG.getNode(ISD::TRUNCATE, DL,
12246 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
12247 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
12249 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
12250 return DAG.getNode(Opc, DL, VT,
12251 N0.getOperand(0), N1.getOperand(0),
12252 DAG.getNode(ISD::TRUNCATE, DL,
12259 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
12260 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
12261 const X86Subtarget *Subtarget) {
12262 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
12263 // the FP state in cases where an emms may be missing.
12264 // A preferable solution to the general problem is to figure out the right
12265 // places to insert EMMS. This qualifies as a quick hack.
12267 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
12268 StoreSDNode *St = cast<StoreSDNode>(N);
12269 EVT VT = St->getValue().getValueType();
12270 if (VT.getSizeInBits() != 64)
12273 const Function *F = DAG.getMachineFunction().getFunction();
12274 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
12275 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
12276 && Subtarget->hasSSE2();
12277 if ((VT.isVector() ||
12278 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
12279 isa<LoadSDNode>(St->getValue()) &&
12280 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
12281 St->getChain().hasOneUse() && !St->isVolatile()) {
12282 SDNode* LdVal = St->getValue().getNode();
12283 LoadSDNode *Ld = 0;
12284 int TokenFactorIndex = -1;
12285 SmallVector<SDValue, 8> Ops;
12286 SDNode* ChainVal = St->getChain().getNode();
12287 // Must be a store of a load. We currently handle two cases: the load
12288 // is a direct child, and it's under an intervening TokenFactor. It is
12289 // possible to dig deeper under nested TokenFactors.
12290 if (ChainVal == LdVal)
12291 Ld = cast<LoadSDNode>(St->getChain());
12292 else if (St->getValue().hasOneUse() &&
12293 ChainVal->getOpcode() == ISD::TokenFactor) {
12294 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
12295 if (ChainVal->getOperand(i).getNode() == LdVal) {
12296 TokenFactorIndex = i;
12297 Ld = cast<LoadSDNode>(St->getValue());
12299 Ops.push_back(ChainVal->getOperand(i));
12303 if (!Ld || !ISD::isNormalLoad(Ld))
12306 // If this is not the MMX case, i.e. we are just turning i64 load/store
12307 // into f64 load/store, avoid the transformation if there are multiple
12308 // uses of the loaded value.
12309 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
12312 DebugLoc LdDL = Ld->getDebugLoc();
12313 DebugLoc StDL = N->getDebugLoc();
12314 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
12315 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
12317 if (Subtarget->is64Bit() || F64IsLegal) {
12318 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
12319 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
12320 Ld->getPointerInfo(), Ld->isVolatile(),
12321 Ld->isNonTemporal(), Ld->getAlignment());
12322 SDValue NewChain = NewLd.getValue(1);
12323 if (TokenFactorIndex != -1) {
12324 Ops.push_back(NewChain);
12325 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
12328 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
12329 St->getPointerInfo(),
12330 St->isVolatile(), St->isNonTemporal(),
12331 St->getAlignment());
12334 // Otherwise, lower to two pairs of 32-bit loads / stores.
12335 SDValue LoAddr = Ld->getBasePtr();
12336 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
12337 DAG.getConstant(4, MVT::i32));
12339 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
12340 Ld->getPointerInfo(),
12341 Ld->isVolatile(), Ld->isNonTemporal(),
12342 Ld->getAlignment());
12343 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
12344 Ld->getPointerInfo().getWithOffset(4),
12345 Ld->isVolatile(), Ld->isNonTemporal(),
12346 MinAlign(Ld->getAlignment(), 4));
12348 SDValue NewChain = LoLd.getValue(1);
12349 if (TokenFactorIndex != -1) {
12350 Ops.push_back(LoLd);
12351 Ops.push_back(HiLd);
12352 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
12356 LoAddr = St->getBasePtr();
12357 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
12358 DAG.getConstant(4, MVT::i32));
12360 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
12361 St->getPointerInfo(),
12362 St->isVolatile(), St->isNonTemporal(),
12363 St->getAlignment());
12364 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
12365 St->getPointerInfo().getWithOffset(4),
12367 St->isNonTemporal(),
12368 MinAlign(St->getAlignment(), 4));
12369 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
12374 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
12375 /// X86ISD::FXOR nodes.
12376 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
12377 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
12378 // F[X]OR(0.0, x) -> x
12379 // F[X]OR(x, 0.0) -> x
12380 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
12381 if (C->getValueAPF().isPosZero())
12382 return N->getOperand(1);
12383 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
12384 if (C->getValueAPF().isPosZero())
12385 return N->getOperand(0);
12389 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
12390 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
12391 // FAND(0.0, x) -> 0.0
12392 // FAND(x, 0.0) -> 0.0
12393 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
12394 if (C->getValueAPF().isPosZero())
12395 return N->getOperand(0);
12396 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
12397 if (C->getValueAPF().isPosZero())
12398 return N->getOperand(1);
12402 static SDValue PerformBTCombine(SDNode *N,
12404 TargetLowering::DAGCombinerInfo &DCI) {
12405 // BT ignores high bits in the bit index operand.
12406 SDValue Op1 = N->getOperand(1);
12407 if (Op1.hasOneUse()) {
12408 unsigned BitWidth = Op1.getValueSizeInBits();
12409 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
12410 APInt KnownZero, KnownOne;
12411 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
12412 !DCI.isBeforeLegalizeOps());
12413 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12414 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
12415 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
12416 DCI.CommitTargetLoweringOpt(TLO);
12421 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
12422 SDValue Op = N->getOperand(0);
12423 if (Op.getOpcode() == ISD::BITCAST)
12424 Op = Op.getOperand(0);
12425 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
12426 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
12427 VT.getVectorElementType().getSizeInBits() ==
12428 OpVT.getVectorElementType().getSizeInBits()) {
12429 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
12434 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
12435 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
12436 // (and (i32 x86isd::setcc_carry), 1)
12437 // This eliminates the zext. This transformation is necessary because
12438 // ISD::SETCC is always legalized to i8.
12439 DebugLoc dl = N->getDebugLoc();
12440 SDValue N0 = N->getOperand(0);
12441 EVT VT = N->getValueType(0);
12442 if (N0.getOpcode() == ISD::AND &&
12444 N0.getOperand(0).hasOneUse()) {
12445 SDValue N00 = N0.getOperand(0);
12446 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
12448 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
12449 if (!C || C->getZExtValue() != 1)
12451 return DAG.getNode(ISD::AND, dl, VT,
12452 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
12453 N00.getOperand(0), N00.getOperand(1)),
12454 DAG.getConstant(1, VT));
12460 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
12461 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG) {
12462 unsigned X86CC = N->getConstantOperandVal(0);
12463 SDValue EFLAG = N->getOperand(1);
12464 DebugLoc DL = N->getDebugLoc();
12466 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
12467 // a zext and produces an all-ones bit which is more useful than 0/1 in some
12469 if (X86CC == X86::COND_B)
12470 return DAG.getNode(ISD::AND, DL, MVT::i8,
12471 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
12472 DAG.getConstant(X86CC, MVT::i8), EFLAG),
12473 DAG.getConstant(1, MVT::i8));
12478 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
12479 const X86TargetLowering *XTLI) {
12480 SDValue Op0 = N->getOperand(0);
12481 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
12482 // a 32-bit target where SSE doesn't support i64->FP operations.
12483 if (Op0.getOpcode() == ISD::LOAD) {
12484 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
12485 EVT VT = Ld->getValueType(0);
12486 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
12487 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
12488 !XTLI->getSubtarget()->is64Bit() &&
12489 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
12490 SDValue FILDChain = XTLI->BuildFILD(SDValue(N, 0), Ld->getValueType(0),
12491 Ld->getChain(), Op0, DAG);
12492 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
12499 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
12500 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
12501 X86TargetLowering::DAGCombinerInfo &DCI) {
12502 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
12503 // the result is either zero or one (depending on the input carry bit).
12504 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
12505 if (X86::isZeroNode(N->getOperand(0)) &&
12506 X86::isZeroNode(N->getOperand(1)) &&
12507 // We don't have a good way to replace an EFLAGS use, so only do this when
12509 SDValue(N, 1).use_empty()) {
12510 DebugLoc DL = N->getDebugLoc();
12511 EVT VT = N->getValueType(0);
12512 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
12513 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
12514 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
12515 DAG.getConstant(X86::COND_B,MVT::i8),
12517 DAG.getConstant(1, VT));
12518 return DCI.CombineTo(N, Res1, CarryOut);
12524 // fold (add Y, (sete X, 0)) -> adc 0, Y
12525 // (add Y, (setne X, 0)) -> sbb -1, Y
12526 // (sub (sete X, 0), Y) -> sbb 0, Y
12527 // (sub (setne X, 0), Y) -> adc -1, Y
12528 static SDValue OptimizeConditonalInDecrement(SDNode *N, SelectionDAG &DAG) {
12529 DebugLoc DL = N->getDebugLoc();
12531 // Look through ZExts.
12532 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
12533 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
12536 SDValue SetCC = Ext.getOperand(0);
12537 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
12540 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
12541 if (CC != X86::COND_E && CC != X86::COND_NE)
12544 SDValue Cmp = SetCC.getOperand(1);
12545 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
12546 !X86::isZeroNode(Cmp.getOperand(1)) ||
12547 !Cmp.getOperand(0).getValueType().isInteger())
12550 SDValue CmpOp0 = Cmp.getOperand(0);
12551 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
12552 DAG.getConstant(1, CmpOp0.getValueType()));
12554 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
12555 if (CC == X86::COND_NE)
12556 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
12557 DL, OtherVal.getValueType(), OtherVal,
12558 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
12559 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
12560 DL, OtherVal.getValueType(), OtherVal,
12561 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
12564 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
12565 DAGCombinerInfo &DCI) const {
12566 SelectionDAG &DAG = DCI.DAG;
12567 switch (N->getOpcode()) {
12569 case ISD::EXTRACT_VECTOR_ELT:
12570 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
12571 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
12572 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
12574 case ISD::SUB: return OptimizeConditonalInDecrement(N, DAG);
12575 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
12576 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
12579 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
12580 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
12581 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
12582 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
12583 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, this);
12585 case X86ISD::FOR: return PerformFORCombine(N, DAG);
12586 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
12587 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
12588 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
12589 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
12590 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG);
12591 case X86ISD::SHUFPS: // Handle all target specific shuffles
12592 case X86ISD::SHUFPD:
12593 case X86ISD::PALIGN:
12594 case X86ISD::PUNPCKHBW:
12595 case X86ISD::PUNPCKHWD:
12596 case X86ISD::PUNPCKHDQ:
12597 case X86ISD::PUNPCKHQDQ:
12598 case X86ISD::UNPCKHPS:
12599 case X86ISD::UNPCKHPD:
12600 case X86ISD::PUNPCKLBW:
12601 case X86ISD::PUNPCKLWD:
12602 case X86ISD::PUNPCKLDQ:
12603 case X86ISD::PUNPCKLQDQ:
12604 case X86ISD::UNPCKLPS:
12605 case X86ISD::UNPCKLPD:
12606 case X86ISD::VUNPCKLPSY:
12607 case X86ISD::VUNPCKLPDY:
12608 case X86ISD::MOVHLPS:
12609 case X86ISD::MOVLHPS:
12610 case X86ISD::PSHUFD:
12611 case X86ISD::PSHUFHW:
12612 case X86ISD::PSHUFLW:
12613 case X86ISD::MOVSS:
12614 case X86ISD::MOVSD:
12615 case X86ISD::VPERMIL:
12616 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI);
12622 /// isTypeDesirableForOp - Return true if the target has native support for
12623 /// the specified value type and it is 'desirable' to use the type for the
12624 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
12625 /// instruction encodings are longer and some i16 instructions are slow.
12626 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
12627 if (!isTypeLegal(VT))
12629 if (VT != MVT::i16)
12636 case ISD::SIGN_EXTEND:
12637 case ISD::ZERO_EXTEND:
12638 case ISD::ANY_EXTEND:
12651 /// IsDesirableToPromoteOp - This method query the target whether it is
12652 /// beneficial for dag combiner to promote the specified node. If true, it
12653 /// should return the desired promotion type by reference.
12654 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
12655 EVT VT = Op.getValueType();
12656 if (VT != MVT::i16)
12659 bool Promote = false;
12660 bool Commute = false;
12661 switch (Op.getOpcode()) {
12664 LoadSDNode *LD = cast<LoadSDNode>(Op);
12665 // If the non-extending load has a single use and it's not live out, then it
12666 // might be folded.
12667 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
12668 Op.hasOneUse()*/) {
12669 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12670 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
12671 // The only case where we'd want to promote LOAD (rather then it being
12672 // promoted as an operand is when it's only use is liveout.
12673 if (UI->getOpcode() != ISD::CopyToReg)
12680 case ISD::SIGN_EXTEND:
12681 case ISD::ZERO_EXTEND:
12682 case ISD::ANY_EXTEND:
12687 SDValue N0 = Op.getOperand(0);
12688 // Look out for (store (shl (load), x)).
12689 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
12702 SDValue N0 = Op.getOperand(0);
12703 SDValue N1 = Op.getOperand(1);
12704 if (!Commute && MayFoldLoad(N1))
12706 // Avoid disabling potential load folding opportunities.
12707 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
12709 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
12719 //===----------------------------------------------------------------------===//
12720 // X86 Inline Assembly Support
12721 //===----------------------------------------------------------------------===//
12723 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
12724 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
12726 std::string AsmStr = IA->getAsmString();
12728 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
12729 SmallVector<StringRef, 4> AsmPieces;
12730 SplitString(AsmStr, AsmPieces, ";\n");
12732 switch (AsmPieces.size()) {
12733 default: return false;
12735 AsmStr = AsmPieces[0];
12737 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
12739 // FIXME: this should verify that we are targeting a 486 or better. If not,
12740 // we will turn this bswap into something that will be lowered to logical ops
12741 // instead of emitting the bswap asm. For now, we don't support 486 or lower
12742 // so don't worry about this.
12744 if (AsmPieces.size() == 2 &&
12745 (AsmPieces[0] == "bswap" ||
12746 AsmPieces[0] == "bswapq" ||
12747 AsmPieces[0] == "bswapl") &&
12748 (AsmPieces[1] == "$0" ||
12749 AsmPieces[1] == "${0:q}")) {
12750 // No need to check constraints, nothing other than the equivalent of
12751 // "=r,0" would be valid here.
12752 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12753 if (!Ty || Ty->getBitWidth() % 16 != 0)
12755 return IntrinsicLowering::LowerToByteSwap(CI);
12757 // rorw $$8, ${0:w} --> llvm.bswap.i16
12758 if (CI->getType()->isIntegerTy(16) &&
12759 AsmPieces.size() == 3 &&
12760 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
12761 AsmPieces[1] == "$$8," &&
12762 AsmPieces[2] == "${0:w}" &&
12763 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
12765 const std::string &ConstraintsStr = IA->getConstraintString();
12766 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
12767 std::sort(AsmPieces.begin(), AsmPieces.end());
12768 if (AsmPieces.size() == 4 &&
12769 AsmPieces[0] == "~{cc}" &&
12770 AsmPieces[1] == "~{dirflag}" &&
12771 AsmPieces[2] == "~{flags}" &&
12772 AsmPieces[3] == "~{fpsr}") {
12773 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12774 if (!Ty || Ty->getBitWidth() % 16 != 0)
12776 return IntrinsicLowering::LowerToByteSwap(CI);
12781 if (CI->getType()->isIntegerTy(32) &&
12782 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
12783 SmallVector<StringRef, 4> Words;
12784 SplitString(AsmPieces[0], Words, " \t,");
12785 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
12786 Words[2] == "${0:w}") {
12788 SplitString(AsmPieces[1], Words, " \t,");
12789 if (Words.size() == 3 && Words[0] == "rorl" && Words[1] == "$$16" &&
12790 Words[2] == "$0") {
12792 SplitString(AsmPieces[2], Words, " \t,");
12793 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
12794 Words[2] == "${0:w}") {
12796 const std::string &ConstraintsStr = IA->getConstraintString();
12797 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
12798 std::sort(AsmPieces.begin(), AsmPieces.end());
12799 if (AsmPieces.size() == 4 &&
12800 AsmPieces[0] == "~{cc}" &&
12801 AsmPieces[1] == "~{dirflag}" &&
12802 AsmPieces[2] == "~{flags}" &&
12803 AsmPieces[3] == "~{fpsr}") {
12804 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12805 if (!Ty || Ty->getBitWidth() % 16 != 0)
12807 return IntrinsicLowering::LowerToByteSwap(CI);
12814 if (CI->getType()->isIntegerTy(64)) {
12815 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
12816 if (Constraints.size() >= 2 &&
12817 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
12818 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
12819 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
12820 SmallVector<StringRef, 4> Words;
12821 SplitString(AsmPieces[0], Words, " \t");
12822 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
12824 SplitString(AsmPieces[1], Words, " \t");
12825 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
12827 SplitString(AsmPieces[2], Words, " \t,");
12828 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
12829 Words[2] == "%edx") {
12830 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12831 if (!Ty || Ty->getBitWidth() % 16 != 0)
12833 return IntrinsicLowering::LowerToByteSwap(CI);
12846 /// getConstraintType - Given a constraint letter, return the type of
12847 /// constraint it is for this target.
12848 X86TargetLowering::ConstraintType
12849 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
12850 if (Constraint.size() == 1) {
12851 switch (Constraint[0]) {
12862 return C_RegisterClass;
12886 return TargetLowering::getConstraintType(Constraint);
12889 /// Examine constraint type and operand type and determine a weight value.
12890 /// This object must already have been set up with the operand type
12891 /// and the current alternative constraint selected.
12892 TargetLowering::ConstraintWeight
12893 X86TargetLowering::getSingleConstraintMatchWeight(
12894 AsmOperandInfo &info, const char *constraint) const {
12895 ConstraintWeight weight = CW_Invalid;
12896 Value *CallOperandVal = info.CallOperandVal;
12897 // If we don't have a value, we can't do a match,
12898 // but allow it at the lowest weight.
12899 if (CallOperandVal == NULL)
12901 Type *type = CallOperandVal->getType();
12902 // Look at the constraint type.
12903 switch (*constraint) {
12905 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
12916 if (CallOperandVal->getType()->isIntegerTy())
12917 weight = CW_SpecificReg;
12922 if (type->isFloatingPointTy())
12923 weight = CW_SpecificReg;
12926 if (type->isX86_MMXTy() && Subtarget->hasMMX())
12927 weight = CW_SpecificReg;
12931 if ((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasXMM())
12932 weight = CW_Register;
12935 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
12936 if (C->getZExtValue() <= 31)
12937 weight = CW_Constant;
12941 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12942 if (C->getZExtValue() <= 63)
12943 weight = CW_Constant;
12947 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12948 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
12949 weight = CW_Constant;
12953 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12954 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
12955 weight = CW_Constant;
12959 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12960 if (C->getZExtValue() <= 3)
12961 weight = CW_Constant;
12965 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12966 if (C->getZExtValue() <= 0xff)
12967 weight = CW_Constant;
12972 if (dyn_cast<ConstantFP>(CallOperandVal)) {
12973 weight = CW_Constant;
12977 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12978 if ((C->getSExtValue() >= -0x80000000LL) &&
12979 (C->getSExtValue() <= 0x7fffffffLL))
12980 weight = CW_Constant;
12984 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12985 if (C->getZExtValue() <= 0xffffffff)
12986 weight = CW_Constant;
12993 /// LowerXConstraint - try to replace an X constraint, which matches anything,
12994 /// with another that has more specific requirements based on the type of the
12995 /// corresponding operand.
12996 const char *X86TargetLowering::
12997 LowerXConstraint(EVT ConstraintVT) const {
12998 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
12999 // 'f' like normal targets.
13000 if (ConstraintVT.isFloatingPoint()) {
13001 if (Subtarget->hasXMMInt())
13003 if (Subtarget->hasXMM())
13007 return TargetLowering::LowerXConstraint(ConstraintVT);
13010 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
13011 /// vector. If it is invalid, don't add anything to Ops.
13012 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
13013 std::string &Constraint,
13014 std::vector<SDValue>&Ops,
13015 SelectionDAG &DAG) const {
13016 SDValue Result(0, 0);
13018 // Only support length 1 constraints for now.
13019 if (Constraint.length() > 1) return;
13021 char ConstraintLetter = Constraint[0];
13022 switch (ConstraintLetter) {
13025 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13026 if (C->getZExtValue() <= 31) {
13027 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
13033 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13034 if (C->getZExtValue() <= 63) {
13035 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
13041 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13042 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
13043 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
13049 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13050 if (C->getZExtValue() <= 255) {
13051 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
13057 // 32-bit signed value
13058 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13059 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
13060 C->getSExtValue())) {
13061 // Widen to 64 bits here to get it sign extended.
13062 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
13065 // FIXME gcc accepts some relocatable values here too, but only in certain
13066 // memory models; it's complicated.
13071 // 32-bit unsigned value
13072 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13073 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
13074 C->getZExtValue())) {
13075 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
13079 // FIXME gcc accepts some relocatable values here too, but only in certain
13080 // memory models; it's complicated.
13084 // Literal immediates are always ok.
13085 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
13086 // Widen to 64 bits here to get it sign extended.
13087 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
13091 // In any sort of PIC mode addresses need to be computed at runtime by
13092 // adding in a register or some sort of table lookup. These can't
13093 // be used as immediates.
13094 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
13097 // If we are in non-pic codegen mode, we allow the address of a global (with
13098 // an optional displacement) to be used with 'i'.
13099 GlobalAddressSDNode *GA = 0;
13100 int64_t Offset = 0;
13102 // Match either (GA), (GA+C), (GA+C1+C2), etc.
13104 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
13105 Offset += GA->getOffset();
13107 } else if (Op.getOpcode() == ISD::ADD) {
13108 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
13109 Offset += C->getZExtValue();
13110 Op = Op.getOperand(0);
13113 } else if (Op.getOpcode() == ISD::SUB) {
13114 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
13115 Offset += -C->getZExtValue();
13116 Op = Op.getOperand(0);
13121 // Otherwise, this isn't something we can handle, reject it.
13125 const GlobalValue *GV = GA->getGlobal();
13126 // If we require an extra load to get this address, as in PIC mode, we
13127 // can't accept it.
13128 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
13129 getTargetMachine())))
13132 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
13133 GA->getValueType(0), Offset);
13138 if (Result.getNode()) {
13139 Ops.push_back(Result);
13142 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
13145 std::pair<unsigned, const TargetRegisterClass*>
13146 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
13148 // First, see if this is a constraint that directly corresponds to an LLVM
13150 if (Constraint.size() == 1) {
13151 // GCC Constraint Letters
13152 switch (Constraint[0]) {
13154 // TODO: Slight differences here in allocation order and leaving
13155 // RIP in the class. Do they matter any more here than they do
13156 // in the normal allocation?
13157 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
13158 if (Subtarget->is64Bit()) {
13159 if (VT == MVT::i32 || VT == MVT::f32)
13160 return std::make_pair(0U, X86::GR32RegisterClass);
13161 else if (VT == MVT::i16)
13162 return std::make_pair(0U, X86::GR16RegisterClass);
13163 else if (VT == MVT::i8 || VT == MVT::i1)
13164 return std::make_pair(0U, X86::GR8RegisterClass);
13165 else if (VT == MVT::i64 || VT == MVT::f64)
13166 return std::make_pair(0U, X86::GR64RegisterClass);
13169 // 32-bit fallthrough
13170 case 'Q': // Q_REGS
13171 if (VT == MVT::i32 || VT == MVT::f32)
13172 return std::make_pair(0U, X86::GR32_ABCDRegisterClass);
13173 else if (VT == MVT::i16)
13174 return std::make_pair(0U, X86::GR16_ABCDRegisterClass);
13175 else if (VT == MVT::i8 || VT == MVT::i1)
13176 return std::make_pair(0U, X86::GR8_ABCD_LRegisterClass);
13177 else if (VT == MVT::i64)
13178 return std::make_pair(0U, X86::GR64_ABCDRegisterClass);
13180 case 'r': // GENERAL_REGS
13181 case 'l': // INDEX_REGS
13182 if (VT == MVT::i8 || VT == MVT::i1)
13183 return std::make_pair(0U, X86::GR8RegisterClass);
13184 if (VT == MVT::i16)
13185 return std::make_pair(0U, X86::GR16RegisterClass);
13186 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
13187 return std::make_pair(0U, X86::GR32RegisterClass);
13188 return std::make_pair(0U, X86::GR64RegisterClass);
13189 case 'R': // LEGACY_REGS
13190 if (VT == MVT::i8 || VT == MVT::i1)
13191 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
13192 if (VT == MVT::i16)
13193 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
13194 if (VT == MVT::i32 || !Subtarget->is64Bit())
13195 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
13196 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
13197 case 'f': // FP Stack registers.
13198 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
13199 // value to the correct fpstack register class.
13200 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
13201 return std::make_pair(0U, X86::RFP32RegisterClass);
13202 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
13203 return std::make_pair(0U, X86::RFP64RegisterClass);
13204 return std::make_pair(0U, X86::RFP80RegisterClass);
13205 case 'y': // MMX_REGS if MMX allowed.
13206 if (!Subtarget->hasMMX()) break;
13207 return std::make_pair(0U, X86::VR64RegisterClass);
13208 case 'Y': // SSE_REGS if SSE2 allowed
13209 if (!Subtarget->hasXMMInt()) break;
13211 case 'x': // SSE_REGS if SSE1 allowed
13212 if (!Subtarget->hasXMM()) break;
13214 switch (VT.getSimpleVT().SimpleTy) {
13216 // Scalar SSE types.
13219 return std::make_pair(0U, X86::FR32RegisterClass);
13222 return std::make_pair(0U, X86::FR64RegisterClass);
13230 return std::make_pair(0U, X86::VR128RegisterClass);
13236 // Use the default implementation in TargetLowering to convert the register
13237 // constraint into a member of a register class.
13238 std::pair<unsigned, const TargetRegisterClass*> Res;
13239 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
13241 // Not found as a standard register?
13242 if (Res.second == 0) {
13243 // Map st(0) -> st(7) -> ST0
13244 if (Constraint.size() == 7 && Constraint[0] == '{' &&
13245 tolower(Constraint[1]) == 's' &&
13246 tolower(Constraint[2]) == 't' &&
13247 Constraint[3] == '(' &&
13248 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
13249 Constraint[5] == ')' &&
13250 Constraint[6] == '}') {
13252 Res.first = X86::ST0+Constraint[4]-'0';
13253 Res.second = X86::RFP80RegisterClass;
13257 // GCC allows "st(0)" to be called just plain "st".
13258 if (StringRef("{st}").equals_lower(Constraint)) {
13259 Res.first = X86::ST0;
13260 Res.second = X86::RFP80RegisterClass;
13265 if (StringRef("{flags}").equals_lower(Constraint)) {
13266 Res.first = X86::EFLAGS;
13267 Res.second = X86::CCRRegisterClass;
13271 // 'A' means EAX + EDX.
13272 if (Constraint == "A") {
13273 Res.first = X86::EAX;
13274 Res.second = X86::GR32_ADRegisterClass;
13280 // Otherwise, check to see if this is a register class of the wrong value
13281 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
13282 // turn into {ax},{dx}.
13283 if (Res.second->hasType(VT))
13284 return Res; // Correct type already, nothing to do.
13286 // All of the single-register GCC register classes map their values onto
13287 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
13288 // really want an 8-bit or 32-bit register, map to the appropriate register
13289 // class and return the appropriate register.
13290 if (Res.second == X86::GR16RegisterClass) {
13291 if (VT == MVT::i8) {
13292 unsigned DestReg = 0;
13293 switch (Res.first) {
13295 case X86::AX: DestReg = X86::AL; break;
13296 case X86::DX: DestReg = X86::DL; break;
13297 case X86::CX: DestReg = X86::CL; break;
13298 case X86::BX: DestReg = X86::BL; break;
13301 Res.first = DestReg;
13302 Res.second = X86::GR8RegisterClass;
13304 } else if (VT == MVT::i32) {
13305 unsigned DestReg = 0;
13306 switch (Res.first) {
13308 case X86::AX: DestReg = X86::EAX; break;
13309 case X86::DX: DestReg = X86::EDX; break;
13310 case X86::CX: DestReg = X86::ECX; break;
13311 case X86::BX: DestReg = X86::EBX; break;
13312 case X86::SI: DestReg = X86::ESI; break;
13313 case X86::DI: DestReg = X86::EDI; break;
13314 case X86::BP: DestReg = X86::EBP; break;
13315 case X86::SP: DestReg = X86::ESP; break;
13318 Res.first = DestReg;
13319 Res.second = X86::GR32RegisterClass;
13321 } else if (VT == MVT::i64) {
13322 unsigned DestReg = 0;
13323 switch (Res.first) {
13325 case X86::AX: DestReg = X86::RAX; break;
13326 case X86::DX: DestReg = X86::RDX; break;
13327 case X86::CX: DestReg = X86::RCX; break;
13328 case X86::BX: DestReg = X86::RBX; break;
13329 case X86::SI: DestReg = X86::RSI; break;
13330 case X86::DI: DestReg = X86::RDI; break;
13331 case X86::BP: DestReg = X86::RBP; break;
13332 case X86::SP: DestReg = X86::RSP; break;
13335 Res.first = DestReg;
13336 Res.second = X86::GR64RegisterClass;
13339 } else if (Res.second == X86::FR32RegisterClass ||
13340 Res.second == X86::FR64RegisterClass ||
13341 Res.second == X86::VR128RegisterClass) {
13342 // Handle references to XMM physical registers that got mapped into the
13343 // wrong class. This can happen with constraints like {xmm0} where the
13344 // target independent register mapper will just pick the first match it can
13345 // find, ignoring the required type.
13346 if (VT == MVT::f32)
13347 Res.second = X86::FR32RegisterClass;
13348 else if (VT == MVT::f64)
13349 Res.second = X86::FR64RegisterClass;
13350 else if (X86::VR128RegisterClass->hasType(VT))
13351 Res.second = X86::VR128RegisterClass;