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::INSERT_SUBVECTOR, SVT, Custom);
989 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
990 for (unsigned i = (unsigned)MVT::v32i8; i != (unsigned)MVT::v4i64; ++i) {
991 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
994 // Do not attempt to promote non-256-bit vectors
995 if (!VT.is256BitVector())
998 setOperationAction(ISD::AND, SVT, Promote);
999 AddPromotedToType (ISD::AND, SVT, MVT::v4i64);
1000 setOperationAction(ISD::OR, SVT, Promote);
1001 AddPromotedToType (ISD::OR, SVT, MVT::v4i64);
1002 setOperationAction(ISD::XOR, SVT, Promote);
1003 AddPromotedToType (ISD::XOR, SVT, MVT::v4i64);
1004 setOperationAction(ISD::LOAD, SVT, Promote);
1005 AddPromotedToType (ISD::LOAD, SVT, MVT::v4i64);
1006 setOperationAction(ISD::SELECT, SVT, Promote);
1007 AddPromotedToType (ISD::SELECT, SVT, MVT::v4i64);
1011 // SIGN_EXTEND_INREGs are evaluated by the extend type. Handle the expansion
1012 // of this type with custom code.
1013 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1014 VT != (unsigned)MVT::LAST_VECTOR_VALUETYPE; VT++) {
1015 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT, Custom);
1018 // We want to custom lower some of our intrinsics.
1019 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1022 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1023 // handle type legalization for these operations here.
1025 // FIXME: We really should do custom legalization for addition and
1026 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1027 // than generic legalization for 64-bit multiplication-with-overflow, though.
1028 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1029 // Add/Sub/Mul with overflow operations are custom lowered.
1031 setOperationAction(ISD::SADDO, VT, Custom);
1032 setOperationAction(ISD::UADDO, VT, Custom);
1033 setOperationAction(ISD::SSUBO, VT, Custom);
1034 setOperationAction(ISD::USUBO, VT, Custom);
1035 setOperationAction(ISD::SMULO, VT, Custom);
1036 setOperationAction(ISD::UMULO, VT, Custom);
1039 // There are no 8-bit 3-address imul/mul instructions
1040 setOperationAction(ISD::SMULO, MVT::i8, Expand);
1041 setOperationAction(ISD::UMULO, MVT::i8, Expand);
1043 if (!Subtarget->is64Bit()) {
1044 // These libcalls are not available in 32-bit.
1045 setLibcallName(RTLIB::SHL_I128, 0);
1046 setLibcallName(RTLIB::SRL_I128, 0);
1047 setLibcallName(RTLIB::SRA_I128, 0);
1050 // We have target-specific dag combine patterns for the following nodes:
1051 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1052 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1053 setTargetDAGCombine(ISD::BUILD_VECTOR);
1054 setTargetDAGCombine(ISD::SELECT);
1055 setTargetDAGCombine(ISD::SHL);
1056 setTargetDAGCombine(ISD::SRA);
1057 setTargetDAGCombine(ISD::SRL);
1058 setTargetDAGCombine(ISD::OR);
1059 setTargetDAGCombine(ISD::AND);
1060 setTargetDAGCombine(ISD::ADD);
1061 setTargetDAGCombine(ISD::SUB);
1062 setTargetDAGCombine(ISD::STORE);
1063 setTargetDAGCombine(ISD::ZERO_EXTEND);
1064 setTargetDAGCombine(ISD::SINT_TO_FP);
1065 if (Subtarget->is64Bit())
1066 setTargetDAGCombine(ISD::MUL);
1068 computeRegisterProperties();
1070 // On Darwin, -Os means optimize for size without hurting performance,
1071 // do not reduce the limit.
1072 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1073 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1074 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1075 maxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1076 maxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1077 maxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1078 setPrefLoopAlignment(16);
1079 benefitFromCodePlacementOpt = true;
1081 setPrefFunctionAlignment(4);
1085 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1090 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1091 /// the desired ByVal argument alignment.
1092 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1095 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1096 if (VTy->getBitWidth() == 128)
1098 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1099 unsigned EltAlign = 0;
1100 getMaxByValAlign(ATy->getElementType(), EltAlign);
1101 if (EltAlign > MaxAlign)
1102 MaxAlign = EltAlign;
1103 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1104 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1105 unsigned EltAlign = 0;
1106 getMaxByValAlign(STy->getElementType(i), EltAlign);
1107 if (EltAlign > MaxAlign)
1108 MaxAlign = EltAlign;
1116 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1117 /// function arguments in the caller parameter area. For X86, aggregates
1118 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1119 /// are at 4-byte boundaries.
1120 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1121 if (Subtarget->is64Bit()) {
1122 // Max of 8 and alignment of type.
1123 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1130 if (Subtarget->hasXMM())
1131 getMaxByValAlign(Ty, Align);
1135 /// getOptimalMemOpType - Returns the target specific optimal type for load
1136 /// and store operations as a result of memset, memcpy, and memmove
1137 /// lowering. If DstAlign is zero that means it's safe to destination
1138 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1139 /// means there isn't a need to check it against alignment requirement,
1140 /// probably because the source does not need to be loaded. If
1141 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1142 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1143 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1144 /// constant so it does not need to be loaded.
1145 /// It returns EVT::Other if the type should be determined using generic
1146 /// target-independent logic.
1148 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1149 unsigned DstAlign, unsigned SrcAlign,
1150 bool NonScalarIntSafe,
1152 MachineFunction &MF) const {
1153 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1154 // linux. This is because the stack realignment code can't handle certain
1155 // cases like PR2962. This should be removed when PR2962 is fixed.
1156 const Function *F = MF.getFunction();
1157 if (NonScalarIntSafe &&
1158 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1160 (Subtarget->isUnalignedMemAccessFast() ||
1161 ((DstAlign == 0 || DstAlign >= 16) &&
1162 (SrcAlign == 0 || SrcAlign >= 16))) &&
1163 Subtarget->getStackAlignment() >= 16) {
1164 if (Subtarget->hasSSE2())
1166 if (Subtarget->hasSSE1())
1168 } else if (!MemcpyStrSrc && Size >= 8 &&
1169 !Subtarget->is64Bit() &&
1170 Subtarget->getStackAlignment() >= 8 &&
1171 Subtarget->hasXMMInt()) {
1172 // Do not use f64 to lower memcpy if source is string constant. It's
1173 // better to use i32 to avoid the loads.
1177 if (Subtarget->is64Bit() && Size >= 8)
1182 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1183 /// current function. The returned value is a member of the
1184 /// MachineJumpTableInfo::JTEntryKind enum.
1185 unsigned X86TargetLowering::getJumpTableEncoding() const {
1186 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1188 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1189 Subtarget->isPICStyleGOT())
1190 return MachineJumpTableInfo::EK_Custom32;
1192 // Otherwise, use the normal jump table encoding heuristics.
1193 return TargetLowering::getJumpTableEncoding();
1197 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1198 const MachineBasicBlock *MBB,
1199 unsigned uid,MCContext &Ctx) const{
1200 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1201 Subtarget->isPICStyleGOT());
1202 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1204 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1205 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1208 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1210 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1211 SelectionDAG &DAG) const {
1212 if (!Subtarget->is64Bit())
1213 // This doesn't have DebugLoc associated with it, but is not really the
1214 // same as a Register.
1215 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1219 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1220 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1222 const MCExpr *X86TargetLowering::
1223 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1224 MCContext &Ctx) const {
1225 // X86-64 uses RIP relative addressing based on the jump table label.
1226 if (Subtarget->isPICStyleRIPRel())
1227 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1229 // Otherwise, the reference is relative to the PIC base.
1230 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1233 // FIXME: Why this routine is here? Move to RegInfo!
1234 std::pair<const TargetRegisterClass*, uint8_t>
1235 X86TargetLowering::findRepresentativeClass(EVT VT) const{
1236 const TargetRegisterClass *RRC = 0;
1238 switch (VT.getSimpleVT().SimpleTy) {
1240 return TargetLowering::findRepresentativeClass(VT);
1241 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1242 RRC = (Subtarget->is64Bit()
1243 ? X86::GR64RegisterClass : X86::GR32RegisterClass);
1246 RRC = X86::VR64RegisterClass;
1248 case MVT::f32: case MVT::f64:
1249 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1250 case MVT::v4f32: case MVT::v2f64:
1251 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1253 RRC = X86::VR128RegisterClass;
1256 return std::make_pair(RRC, Cost);
1259 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1260 unsigned &Offset) const {
1261 if (!Subtarget->isTargetLinux())
1264 if (Subtarget->is64Bit()) {
1265 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1267 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1280 //===----------------------------------------------------------------------===//
1281 // Return Value Calling Convention Implementation
1282 //===----------------------------------------------------------------------===//
1284 #include "X86GenCallingConv.inc"
1287 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1288 MachineFunction &MF, bool isVarArg,
1289 const SmallVectorImpl<ISD::OutputArg> &Outs,
1290 LLVMContext &Context) const {
1291 SmallVector<CCValAssign, 16> RVLocs;
1292 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1294 return CCInfo.CheckReturn(Outs, RetCC_X86);
1298 X86TargetLowering::LowerReturn(SDValue Chain,
1299 CallingConv::ID CallConv, bool isVarArg,
1300 const SmallVectorImpl<ISD::OutputArg> &Outs,
1301 const SmallVectorImpl<SDValue> &OutVals,
1302 DebugLoc dl, SelectionDAG &DAG) const {
1303 MachineFunction &MF = DAG.getMachineFunction();
1304 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1306 SmallVector<CCValAssign, 16> RVLocs;
1307 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1308 RVLocs, *DAG.getContext());
1309 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1311 // Add the regs to the liveout set for the function.
1312 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1313 for (unsigned i = 0; i != RVLocs.size(); ++i)
1314 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1315 MRI.addLiveOut(RVLocs[i].getLocReg());
1319 SmallVector<SDValue, 6> RetOps;
1320 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1321 // Operand #1 = Bytes To Pop
1322 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1325 // Copy the result values into the output registers.
1326 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1327 CCValAssign &VA = RVLocs[i];
1328 assert(VA.isRegLoc() && "Can only return in registers!");
1329 SDValue ValToCopy = OutVals[i];
1330 EVT ValVT = ValToCopy.getValueType();
1332 // If this is x86-64, and we disabled SSE, we can't return FP values,
1333 // or SSE or MMX vectors.
1334 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1335 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1336 (Subtarget->is64Bit() && !Subtarget->hasXMM())) {
1337 report_fatal_error("SSE register return with SSE disabled");
1339 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1340 // llvm-gcc has never done it right and no one has noticed, so this
1341 // should be OK for now.
1342 if (ValVT == MVT::f64 &&
1343 (Subtarget->is64Bit() && !Subtarget->hasXMMInt()))
1344 report_fatal_error("SSE2 register return with SSE2 disabled");
1346 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1347 // the RET instruction and handled by the FP Stackifier.
1348 if (VA.getLocReg() == X86::ST0 ||
1349 VA.getLocReg() == X86::ST1) {
1350 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1351 // change the value to the FP stack register class.
1352 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1353 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1354 RetOps.push_back(ValToCopy);
1355 // Don't emit a copytoreg.
1359 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1360 // which is returned in RAX / RDX.
1361 if (Subtarget->is64Bit()) {
1362 if (ValVT == MVT::x86mmx) {
1363 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1364 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
1365 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1367 // If we don't have SSE2 available, convert to v4f32 so the generated
1368 // register is legal.
1369 if (!Subtarget->hasSSE2())
1370 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
1375 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1376 Flag = Chain.getValue(1);
1379 // The x86-64 ABI for returning structs by value requires that we copy
1380 // the sret argument into %rax for the return. We saved the argument into
1381 // a virtual register in the entry block, so now we copy the value out
1383 if (Subtarget->is64Bit() &&
1384 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1385 MachineFunction &MF = DAG.getMachineFunction();
1386 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1387 unsigned Reg = FuncInfo->getSRetReturnReg();
1389 "SRetReturnReg should have been set in LowerFormalArguments().");
1390 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1392 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1393 Flag = Chain.getValue(1);
1395 // RAX now acts like a return value.
1396 MRI.addLiveOut(X86::RAX);
1399 RetOps[0] = Chain; // Update chain.
1401 // Add the flag if we have it.
1403 RetOps.push_back(Flag);
1405 return DAG.getNode(X86ISD::RET_FLAG, dl,
1406 MVT::Other, &RetOps[0], RetOps.size());
1409 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N) const {
1410 if (N->getNumValues() != 1)
1412 if (!N->hasNUsesOfValue(1, 0))
1415 SDNode *Copy = *N->use_begin();
1416 if (Copy->getOpcode() != ISD::CopyToReg &&
1417 Copy->getOpcode() != ISD::FP_EXTEND)
1420 bool HasRet = false;
1421 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1423 if (UI->getOpcode() != X86ISD::RET_FLAG)
1432 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
1433 ISD::NodeType ExtendKind) const {
1435 // TODO: Is this also valid on 32-bit?
1436 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
1437 ReturnMVT = MVT::i8;
1439 ReturnMVT = MVT::i32;
1441 EVT MinVT = getRegisterType(Context, ReturnMVT);
1442 return VT.bitsLT(MinVT) ? MinVT : VT;
1445 /// LowerCallResult - Lower the result values of a call into the
1446 /// appropriate copies out of appropriate physical registers.
1449 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1450 CallingConv::ID CallConv, bool isVarArg,
1451 const SmallVectorImpl<ISD::InputArg> &Ins,
1452 DebugLoc dl, SelectionDAG &DAG,
1453 SmallVectorImpl<SDValue> &InVals) const {
1455 // Assign locations to each value returned by this call.
1456 SmallVector<CCValAssign, 16> RVLocs;
1457 bool Is64Bit = Subtarget->is64Bit();
1458 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1459 getTargetMachine(), RVLocs, *DAG.getContext());
1460 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1462 // Copy all of the result registers out of their specified physreg.
1463 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1464 CCValAssign &VA = RVLocs[i];
1465 EVT CopyVT = VA.getValVT();
1467 // If this is x86-64, and we disabled SSE, we can't return FP values
1468 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1469 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasXMM())) {
1470 report_fatal_error("SSE register return with SSE disabled");
1475 // If this is a call to a function that returns an fp value on the floating
1476 // point stack, we must guarantee the the value is popped from the stack, so
1477 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1478 // if the return value is not used. We use the FpPOP_RETVAL instruction
1480 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1481 // If we prefer to use the value in xmm registers, copy it out as f80 and
1482 // use a truncate to move it from fp stack reg to xmm reg.
1483 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1484 SDValue Ops[] = { Chain, InFlag };
1485 Chain = SDValue(DAG.getMachineNode(X86::FpPOP_RETVAL, dl, CopyVT,
1486 MVT::Other, MVT::Glue, Ops, 2), 1);
1487 Val = Chain.getValue(0);
1489 // Round the f80 to the right size, which also moves it to the appropriate
1491 if (CopyVT != VA.getValVT())
1492 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1493 // This truncation won't change the value.
1494 DAG.getIntPtrConstant(1));
1496 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1497 CopyVT, InFlag).getValue(1);
1498 Val = Chain.getValue(0);
1500 InFlag = Chain.getValue(2);
1501 InVals.push_back(Val);
1508 //===----------------------------------------------------------------------===//
1509 // C & StdCall & Fast Calling Convention implementation
1510 //===----------------------------------------------------------------------===//
1511 // StdCall calling convention seems to be standard for many Windows' API
1512 // routines and around. It differs from C calling convention just a little:
1513 // callee should clean up the stack, not caller. Symbols should be also
1514 // decorated in some fancy way :) It doesn't support any vector arguments.
1515 // For info on fast calling convention see Fast Calling Convention (tail call)
1516 // implementation LowerX86_32FastCCCallTo.
1518 /// CallIsStructReturn - Determines whether a call uses struct return
1520 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1524 return Outs[0].Flags.isSRet();
1527 /// ArgsAreStructReturn - Determines whether a function uses struct
1528 /// return semantics.
1530 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1534 return Ins[0].Flags.isSRet();
1537 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1538 /// by "Src" to address "Dst" with size and alignment information specified by
1539 /// the specific parameter attribute. The copy will be passed as a byval
1540 /// function parameter.
1542 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1543 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1545 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1547 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1548 /*isVolatile*/false, /*AlwaysInline=*/true,
1549 MachinePointerInfo(), MachinePointerInfo());
1552 /// IsTailCallConvention - Return true if the calling convention is one that
1553 /// supports tail call optimization.
1554 static bool IsTailCallConvention(CallingConv::ID CC) {
1555 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1558 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1559 if (!CI->isTailCall())
1563 CallingConv::ID CalleeCC = CS.getCallingConv();
1564 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
1570 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1571 /// a tailcall target by changing its ABI.
1572 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1573 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1577 X86TargetLowering::LowerMemArgument(SDValue Chain,
1578 CallingConv::ID CallConv,
1579 const SmallVectorImpl<ISD::InputArg> &Ins,
1580 DebugLoc dl, SelectionDAG &DAG,
1581 const CCValAssign &VA,
1582 MachineFrameInfo *MFI,
1584 // Create the nodes corresponding to a load from this parameter slot.
1585 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1586 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1587 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1590 // If value is passed by pointer we have address passed instead of the value
1592 if (VA.getLocInfo() == CCValAssign::Indirect)
1593 ValVT = VA.getLocVT();
1595 ValVT = VA.getValVT();
1597 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1598 // changed with more analysis.
1599 // In case of tail call optimization mark all arguments mutable. Since they
1600 // could be overwritten by lowering of arguments in case of a tail call.
1601 if (Flags.isByVal()) {
1602 unsigned Bytes = Flags.getByValSize();
1603 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
1604 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
1605 return DAG.getFrameIndex(FI, getPointerTy());
1607 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1608 VA.getLocMemOffset(), isImmutable);
1609 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1610 return DAG.getLoad(ValVT, dl, Chain, FIN,
1611 MachinePointerInfo::getFixedStack(FI),
1617 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1618 CallingConv::ID CallConv,
1620 const SmallVectorImpl<ISD::InputArg> &Ins,
1623 SmallVectorImpl<SDValue> &InVals)
1625 MachineFunction &MF = DAG.getMachineFunction();
1626 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1628 const Function* Fn = MF.getFunction();
1629 if (Fn->hasExternalLinkage() &&
1630 Subtarget->isTargetCygMing() &&
1631 Fn->getName() == "main")
1632 FuncInfo->setForceFramePointer(true);
1634 MachineFrameInfo *MFI = MF.getFrameInfo();
1635 bool Is64Bit = Subtarget->is64Bit();
1636 bool IsWin64 = Subtarget->isTargetWin64();
1638 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1639 "Var args not supported with calling convention fastcc or ghc");
1641 // Assign locations to all of the incoming arguments.
1642 SmallVector<CCValAssign, 16> ArgLocs;
1643 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1644 ArgLocs, *DAG.getContext());
1646 // Allocate shadow area for Win64
1648 CCInfo.AllocateStack(32, 8);
1651 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
1653 unsigned LastVal = ~0U;
1655 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1656 CCValAssign &VA = ArgLocs[i];
1657 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1659 assert(VA.getValNo() != LastVal &&
1660 "Don't support value assigned to multiple locs yet");
1661 LastVal = VA.getValNo();
1663 if (VA.isRegLoc()) {
1664 EVT RegVT = VA.getLocVT();
1665 TargetRegisterClass *RC = NULL;
1666 if (RegVT == MVT::i32)
1667 RC = X86::GR32RegisterClass;
1668 else if (Is64Bit && RegVT == MVT::i64)
1669 RC = X86::GR64RegisterClass;
1670 else if (RegVT == MVT::f32)
1671 RC = X86::FR32RegisterClass;
1672 else if (RegVT == MVT::f64)
1673 RC = X86::FR64RegisterClass;
1674 else if (RegVT.isVector() && RegVT.getSizeInBits() == 256)
1675 RC = X86::VR256RegisterClass;
1676 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1677 RC = X86::VR128RegisterClass;
1678 else if (RegVT == MVT::x86mmx)
1679 RC = X86::VR64RegisterClass;
1681 llvm_unreachable("Unknown argument type!");
1683 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1684 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1686 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1687 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1689 if (VA.getLocInfo() == CCValAssign::SExt)
1690 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1691 DAG.getValueType(VA.getValVT()));
1692 else if (VA.getLocInfo() == CCValAssign::ZExt)
1693 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1694 DAG.getValueType(VA.getValVT()));
1695 else if (VA.getLocInfo() == CCValAssign::BCvt)
1696 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
1698 if (VA.isExtInLoc()) {
1699 // Handle MMX values passed in XMM regs.
1700 if (RegVT.isVector()) {
1701 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(),
1704 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1707 assert(VA.isMemLoc());
1708 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1711 // If value is passed via pointer - do a load.
1712 if (VA.getLocInfo() == CCValAssign::Indirect)
1713 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
1714 MachinePointerInfo(), false, false, 0);
1716 InVals.push_back(ArgValue);
1719 // The x86-64 ABI for returning structs by value requires that we copy
1720 // the sret argument into %rax for the return. Save the argument into
1721 // a virtual register so that we can access it from the return points.
1722 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1723 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1724 unsigned Reg = FuncInfo->getSRetReturnReg();
1726 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1727 FuncInfo->setSRetReturnReg(Reg);
1729 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1730 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1733 unsigned StackSize = CCInfo.getNextStackOffset();
1734 // Align stack specially for tail calls.
1735 if (FuncIsMadeTailCallSafe(CallConv))
1736 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1738 // If the function takes variable number of arguments, make a frame index for
1739 // the start of the first vararg value... for expansion of llvm.va_start.
1741 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1742 CallConv != CallingConv::X86_ThisCall)) {
1743 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1746 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1748 // FIXME: We should really autogenerate these arrays
1749 static const unsigned GPR64ArgRegsWin64[] = {
1750 X86::RCX, X86::RDX, X86::R8, X86::R9
1752 static const unsigned GPR64ArgRegs64Bit[] = {
1753 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1755 static const unsigned XMMArgRegs64Bit[] = {
1756 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1757 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1759 const unsigned *GPR64ArgRegs;
1760 unsigned NumXMMRegs = 0;
1763 // The XMM registers which might contain var arg parameters are shadowed
1764 // in their paired GPR. So we only need to save the GPR to their home
1766 TotalNumIntRegs = 4;
1767 GPR64ArgRegs = GPR64ArgRegsWin64;
1769 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1770 GPR64ArgRegs = GPR64ArgRegs64Bit;
1772 NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs64Bit, TotalNumXMMRegs);
1774 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1777 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1778 assert(!(NumXMMRegs && !Subtarget->hasXMM()) &&
1779 "SSE register cannot be used when SSE is disabled!");
1780 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1781 "SSE register cannot be used when SSE is disabled!");
1782 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasXMM())
1783 // Kernel mode asks for SSE to be disabled, so don't push them
1785 TotalNumXMMRegs = 0;
1788 const TargetFrameLowering &TFI = *getTargetMachine().getFrameLowering();
1789 // Get to the caller-allocated home save location. Add 8 to account
1790 // for the return address.
1791 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
1792 FuncInfo->setRegSaveFrameIndex(
1793 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
1794 // Fixup to set vararg frame on shadow area (4 x i64).
1796 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
1798 // For X86-64, if there are vararg parameters that are passed via
1799 // registers, then we must store them to their spots on the stack so they
1800 // may be loaded by deferencing the result of va_next.
1801 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1802 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1803 FuncInfo->setRegSaveFrameIndex(
1804 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1808 // Store the integer parameter registers.
1809 SmallVector<SDValue, 8> MemOps;
1810 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1812 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1813 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1814 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1815 DAG.getIntPtrConstant(Offset));
1816 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1817 X86::GR64RegisterClass);
1818 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1820 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1821 MachinePointerInfo::getFixedStack(
1822 FuncInfo->getRegSaveFrameIndex(), Offset),
1824 MemOps.push_back(Store);
1828 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1829 // Now store the XMM (fp + vector) parameter registers.
1830 SmallVector<SDValue, 11> SaveXMMOps;
1831 SaveXMMOps.push_back(Chain);
1833 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1834 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1835 SaveXMMOps.push_back(ALVal);
1837 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1838 FuncInfo->getRegSaveFrameIndex()));
1839 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1840 FuncInfo->getVarArgsFPOffset()));
1842 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1843 unsigned VReg = MF.addLiveIn(XMMArgRegs64Bit[NumXMMRegs],
1844 X86::VR128RegisterClass);
1845 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1846 SaveXMMOps.push_back(Val);
1848 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1850 &SaveXMMOps[0], SaveXMMOps.size()));
1853 if (!MemOps.empty())
1854 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1855 &MemOps[0], MemOps.size());
1859 // Some CCs need callee pop.
1860 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, GuaranteedTailCallOpt)) {
1861 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1863 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1864 // If this is an sret function, the return should pop the hidden pointer.
1865 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1866 FuncInfo->setBytesToPopOnReturn(4);
1870 // RegSaveFrameIndex is X86-64 only.
1871 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1872 if (CallConv == CallingConv::X86_FastCall ||
1873 CallConv == CallingConv::X86_ThisCall)
1874 // fastcc functions can't have varargs.
1875 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1882 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1883 SDValue StackPtr, SDValue Arg,
1884 DebugLoc dl, SelectionDAG &DAG,
1885 const CCValAssign &VA,
1886 ISD::ArgFlagsTy Flags) const {
1887 unsigned LocMemOffset = VA.getLocMemOffset();
1888 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1889 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1890 if (Flags.isByVal())
1891 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1893 return DAG.getStore(Chain, dl, Arg, PtrOff,
1894 MachinePointerInfo::getStack(LocMemOffset),
1898 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1899 /// optimization is performed and it is required.
1901 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1902 SDValue &OutRetAddr, SDValue Chain,
1903 bool IsTailCall, bool Is64Bit,
1904 int FPDiff, DebugLoc dl) const {
1905 // Adjust the Return address stack slot.
1906 EVT VT = getPointerTy();
1907 OutRetAddr = getReturnAddressFrameIndex(DAG);
1909 // Load the "old" Return address.
1910 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
1912 return SDValue(OutRetAddr.getNode(), 1);
1915 /// EmitTailCallStoreRetAddr - Emit a store of the return address if tail call
1916 /// optimization is performed and it is required (FPDiff!=0).
1918 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1919 SDValue Chain, SDValue RetAddrFrIdx,
1920 bool Is64Bit, int FPDiff, DebugLoc dl) {
1921 // Store the return address to the appropriate stack slot.
1922 if (!FPDiff) return Chain;
1923 // Calculate the new stack slot for the return address.
1924 int SlotSize = Is64Bit ? 8 : 4;
1925 int NewReturnAddrFI =
1926 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1927 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1928 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1929 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1930 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
1936 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1937 CallingConv::ID CallConv, bool isVarArg,
1939 const SmallVectorImpl<ISD::OutputArg> &Outs,
1940 const SmallVectorImpl<SDValue> &OutVals,
1941 const SmallVectorImpl<ISD::InputArg> &Ins,
1942 DebugLoc dl, SelectionDAG &DAG,
1943 SmallVectorImpl<SDValue> &InVals) const {
1944 MachineFunction &MF = DAG.getMachineFunction();
1945 bool Is64Bit = Subtarget->is64Bit();
1946 bool IsWin64 = Subtarget->isTargetWin64();
1947 bool IsStructRet = CallIsStructReturn(Outs);
1948 bool IsSibcall = false;
1951 // Check if it's really possible to do a tail call.
1952 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1953 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1954 Outs, OutVals, Ins, DAG);
1956 // Sibcalls are automatically detected tailcalls which do not require
1958 if (!GuaranteedTailCallOpt && isTailCall)
1965 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1966 "Var args not supported with calling convention fastcc or ghc");
1968 // Analyze operands of the call, assigning locations to each operand.
1969 SmallVector<CCValAssign, 16> ArgLocs;
1970 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1971 ArgLocs, *DAG.getContext());
1973 // Allocate shadow area for Win64
1975 CCInfo.AllocateStack(32, 8);
1978 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
1980 // Get a count of how many bytes are to be pushed on the stack.
1981 unsigned NumBytes = CCInfo.getNextStackOffset();
1983 // This is a sibcall. The memory operands are available in caller's
1984 // own caller's stack.
1986 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1987 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1990 if (isTailCall && !IsSibcall) {
1991 // Lower arguments at fp - stackoffset + fpdiff.
1992 unsigned NumBytesCallerPushed =
1993 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1994 FPDiff = NumBytesCallerPushed - NumBytes;
1996 // Set the delta of movement of the returnaddr stackslot.
1997 // But only set if delta is greater than previous delta.
1998 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1999 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
2003 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2005 SDValue RetAddrFrIdx;
2006 // Load return address for tail calls.
2007 if (isTailCall && FPDiff)
2008 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2009 Is64Bit, FPDiff, dl);
2011 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2012 SmallVector<SDValue, 8> MemOpChains;
2015 // Walk the register/memloc assignments, inserting copies/loads. In the case
2016 // of tail call optimization arguments are handle later.
2017 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2018 CCValAssign &VA = ArgLocs[i];
2019 EVT RegVT = VA.getLocVT();
2020 SDValue Arg = OutVals[i];
2021 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2022 bool isByVal = Flags.isByVal();
2024 // Promote the value if needed.
2025 switch (VA.getLocInfo()) {
2026 default: llvm_unreachable("Unknown loc info!");
2027 case CCValAssign::Full: break;
2028 case CCValAssign::SExt:
2029 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2031 case CCValAssign::ZExt:
2032 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2034 case CCValAssign::AExt:
2035 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
2036 // Special case: passing MMX values in XMM registers.
2037 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2038 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2039 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2041 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2043 case CCValAssign::BCvt:
2044 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2046 case CCValAssign::Indirect: {
2047 // Store the argument.
2048 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2049 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2050 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2051 MachinePointerInfo::getFixedStack(FI),
2058 if (VA.isRegLoc()) {
2059 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2060 if (isVarArg && IsWin64) {
2061 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2062 // shadow reg if callee is a varargs function.
2063 unsigned ShadowReg = 0;
2064 switch (VA.getLocReg()) {
2065 case X86::XMM0: ShadowReg = X86::RCX; break;
2066 case X86::XMM1: ShadowReg = X86::RDX; break;
2067 case X86::XMM2: ShadowReg = X86::R8; break;
2068 case X86::XMM3: ShadowReg = X86::R9; break;
2071 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2073 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2074 assert(VA.isMemLoc());
2075 if (StackPtr.getNode() == 0)
2076 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
2077 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2078 dl, DAG, VA, Flags));
2082 if (!MemOpChains.empty())
2083 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2084 &MemOpChains[0], MemOpChains.size());
2086 // Build a sequence of copy-to-reg nodes chained together with token chain
2087 // and flag operands which copy the outgoing args into registers.
2089 // Tail call byval lowering might overwrite argument registers so in case of
2090 // tail call optimization the copies to registers are lowered later.
2092 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2093 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2094 RegsToPass[i].second, InFlag);
2095 InFlag = Chain.getValue(1);
2098 if (Subtarget->isPICStyleGOT()) {
2099 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2102 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
2103 DAG.getNode(X86ISD::GlobalBaseReg,
2104 DebugLoc(), getPointerTy()),
2106 InFlag = Chain.getValue(1);
2108 // If we are tail calling and generating PIC/GOT style code load the
2109 // address of the callee into ECX. The value in ecx is used as target of
2110 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2111 // for tail calls on PIC/GOT architectures. Normally we would just put the
2112 // address of GOT into ebx and then call target@PLT. But for tail calls
2113 // ebx would be restored (since ebx is callee saved) before jumping to the
2116 // Note: The actual moving to ECX is done further down.
2117 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2118 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2119 !G->getGlobal()->hasProtectedVisibility())
2120 Callee = LowerGlobalAddress(Callee, DAG);
2121 else if (isa<ExternalSymbolSDNode>(Callee))
2122 Callee = LowerExternalSymbol(Callee, DAG);
2126 if (Is64Bit && isVarArg && !IsWin64) {
2127 // From AMD64 ABI document:
2128 // For calls that may call functions that use varargs or stdargs
2129 // (prototype-less calls or calls to functions containing ellipsis (...) in
2130 // the declaration) %al is used as hidden argument to specify the number
2131 // of SSE registers used. The contents of %al do not need to match exactly
2132 // the number of registers, but must be an ubound on the number of SSE
2133 // registers used and is in the range 0 - 8 inclusive.
2135 // Count the number of XMM registers allocated.
2136 static const unsigned XMMArgRegs[] = {
2137 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2138 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2140 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2141 assert((Subtarget->hasXMM() || !NumXMMRegs)
2142 && "SSE registers cannot be used when SSE is disabled");
2144 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2145 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2146 InFlag = Chain.getValue(1);
2150 // For tail calls lower the arguments to the 'real' stack slot.
2152 // Force all the incoming stack arguments to be loaded from the stack
2153 // before any new outgoing arguments are stored to the stack, because the
2154 // outgoing stack slots may alias the incoming argument stack slots, and
2155 // the alias isn't otherwise explicit. This is slightly more conservative
2156 // than necessary, because it means that each store effectively depends
2157 // on every argument instead of just those arguments it would clobber.
2158 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2160 SmallVector<SDValue, 8> MemOpChains2;
2163 // Do not flag preceding copytoreg stuff together with the following stuff.
2165 if (GuaranteedTailCallOpt) {
2166 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2167 CCValAssign &VA = ArgLocs[i];
2170 assert(VA.isMemLoc());
2171 SDValue Arg = OutVals[i];
2172 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2173 // Create frame index.
2174 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2175 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2176 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2177 FIN = DAG.getFrameIndex(FI, getPointerTy());
2179 if (Flags.isByVal()) {
2180 // Copy relative to framepointer.
2181 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2182 if (StackPtr.getNode() == 0)
2183 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2185 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2187 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2191 // Store relative to framepointer.
2192 MemOpChains2.push_back(
2193 DAG.getStore(ArgChain, dl, Arg, FIN,
2194 MachinePointerInfo::getFixedStack(FI),
2200 if (!MemOpChains2.empty())
2201 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2202 &MemOpChains2[0], MemOpChains2.size());
2204 // Copy arguments to their registers.
2205 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2206 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2207 RegsToPass[i].second, InFlag);
2208 InFlag = Chain.getValue(1);
2212 // Store the return address to the appropriate stack slot.
2213 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2217 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2218 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2219 // In the 64-bit large code model, we have to make all calls
2220 // through a register, since the call instruction's 32-bit
2221 // pc-relative offset may not be large enough to hold the whole
2223 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2224 // If the callee is a GlobalAddress node (quite common, every direct call
2225 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2228 // We should use extra load for direct calls to dllimported functions in
2230 const GlobalValue *GV = G->getGlobal();
2231 if (!GV->hasDLLImportLinkage()) {
2232 unsigned char OpFlags = 0;
2233 bool ExtraLoad = false;
2234 unsigned WrapperKind = ISD::DELETED_NODE;
2236 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2237 // external symbols most go through the PLT in PIC mode. If the symbol
2238 // has hidden or protected visibility, or if it is static or local, then
2239 // we don't need to use the PLT - we can directly call it.
2240 if (Subtarget->isTargetELF() &&
2241 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2242 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2243 OpFlags = X86II::MO_PLT;
2244 } else if (Subtarget->isPICStyleStubAny() &&
2245 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2246 (!Subtarget->getTargetTriple().isMacOSX() ||
2247 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2248 // PC-relative references to external symbols should go through $stub,
2249 // unless we're building with the leopard linker or later, which
2250 // automatically synthesizes these stubs.
2251 OpFlags = X86II::MO_DARWIN_STUB;
2252 } else if (Subtarget->isPICStyleRIPRel() &&
2253 isa<Function>(GV) &&
2254 cast<Function>(GV)->hasFnAttr(Attribute::NonLazyBind)) {
2255 // If the function is marked as non-lazy, generate an indirect call
2256 // which loads from the GOT directly. This avoids runtime overhead
2257 // at the cost of eager binding (and one extra byte of encoding).
2258 OpFlags = X86II::MO_GOTPCREL;
2259 WrapperKind = X86ISD::WrapperRIP;
2263 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2264 G->getOffset(), OpFlags);
2266 // Add a wrapper if needed.
2267 if (WrapperKind != ISD::DELETED_NODE)
2268 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
2269 // Add extra indirection if needed.
2271 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
2272 MachinePointerInfo::getGOT(),
2275 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2276 unsigned char OpFlags = 0;
2278 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
2279 // external symbols should go through the PLT.
2280 if (Subtarget->isTargetELF() &&
2281 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2282 OpFlags = X86II::MO_PLT;
2283 } else if (Subtarget->isPICStyleStubAny() &&
2284 (!Subtarget->getTargetTriple().isMacOSX() ||
2285 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2286 // PC-relative references to external symbols should go through $stub,
2287 // unless we're building with the leopard linker or later, which
2288 // automatically synthesizes these stubs.
2289 OpFlags = X86II::MO_DARWIN_STUB;
2292 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2296 // Returns a chain & a flag for retval copy to use.
2297 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2298 SmallVector<SDValue, 8> Ops;
2300 if (!IsSibcall && isTailCall) {
2301 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2302 DAG.getIntPtrConstant(0, true), InFlag);
2303 InFlag = Chain.getValue(1);
2306 Ops.push_back(Chain);
2307 Ops.push_back(Callee);
2310 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2312 // Add argument registers to the end of the list so that they are known live
2314 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2315 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2316 RegsToPass[i].second.getValueType()));
2318 // Add an implicit use GOT pointer in EBX.
2319 if (!isTailCall && Subtarget->isPICStyleGOT())
2320 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2322 // Add an implicit use of AL for non-Windows x86 64-bit vararg functions.
2323 if (Is64Bit && isVarArg && !IsWin64)
2324 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2326 if (InFlag.getNode())
2327 Ops.push_back(InFlag);
2331 //// If this is the first return lowered for this function, add the regs
2332 //// to the liveout set for the function.
2333 // This isn't right, although it's probably harmless on x86; liveouts
2334 // should be computed from returns not tail calls. Consider a void
2335 // function making a tail call to a function returning int.
2336 return DAG.getNode(X86ISD::TC_RETURN, dl,
2337 NodeTys, &Ops[0], Ops.size());
2340 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2341 InFlag = Chain.getValue(1);
2343 // Create the CALLSEQ_END node.
2344 unsigned NumBytesForCalleeToPush;
2345 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, GuaranteedTailCallOpt))
2346 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2347 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2348 // If this is a call to a struct-return function, the callee
2349 // pops the hidden struct pointer, so we have to push it back.
2350 // This is common for Darwin/X86, Linux & Mingw32 targets.
2351 NumBytesForCalleeToPush = 4;
2353 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2355 // Returns a flag for retval copy to use.
2357 Chain = DAG.getCALLSEQ_END(Chain,
2358 DAG.getIntPtrConstant(NumBytes, true),
2359 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2362 InFlag = Chain.getValue(1);
2365 // Handle result values, copying them out of physregs into vregs that we
2367 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2368 Ins, dl, DAG, InVals);
2372 //===----------------------------------------------------------------------===//
2373 // Fast Calling Convention (tail call) implementation
2374 //===----------------------------------------------------------------------===//
2376 // Like std call, callee cleans arguments, convention except that ECX is
2377 // reserved for storing the tail called function address. Only 2 registers are
2378 // free for argument passing (inreg). Tail call optimization is performed
2380 // * tailcallopt is enabled
2381 // * caller/callee are fastcc
2382 // On X86_64 architecture with GOT-style position independent code only local
2383 // (within module) calls are supported at the moment.
2384 // To keep the stack aligned according to platform abi the function
2385 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2386 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2387 // If a tail called function callee has more arguments than the caller the
2388 // caller needs to make sure that there is room to move the RETADDR to. This is
2389 // achieved by reserving an area the size of the argument delta right after the
2390 // original REtADDR, but before the saved framepointer or the spilled registers
2391 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2403 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2404 /// for a 16 byte align requirement.
2406 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2407 SelectionDAG& DAG) const {
2408 MachineFunction &MF = DAG.getMachineFunction();
2409 const TargetMachine &TM = MF.getTarget();
2410 const TargetFrameLowering &TFI = *TM.getFrameLowering();
2411 unsigned StackAlignment = TFI.getStackAlignment();
2412 uint64_t AlignMask = StackAlignment - 1;
2413 int64_t Offset = StackSize;
2414 uint64_t SlotSize = TD->getPointerSize();
2415 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2416 // Number smaller than 12 so just add the difference.
2417 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2419 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2420 Offset = ((~AlignMask) & Offset) + StackAlignment +
2421 (StackAlignment-SlotSize);
2426 /// MatchingStackOffset - Return true if the given stack call argument is
2427 /// already available in the same position (relatively) of the caller's
2428 /// incoming argument stack.
2430 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2431 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2432 const X86InstrInfo *TII) {
2433 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2435 if (Arg.getOpcode() == ISD::CopyFromReg) {
2436 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2437 if (!TargetRegisterInfo::isVirtualRegister(VR))
2439 MachineInstr *Def = MRI->getVRegDef(VR);
2442 if (!Flags.isByVal()) {
2443 if (!TII->isLoadFromStackSlot(Def, FI))
2446 unsigned Opcode = Def->getOpcode();
2447 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2448 Def->getOperand(1).isFI()) {
2449 FI = Def->getOperand(1).getIndex();
2450 Bytes = Flags.getByValSize();
2454 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2455 if (Flags.isByVal())
2456 // ByVal argument is passed in as a pointer but it's now being
2457 // dereferenced. e.g.
2458 // define @foo(%struct.X* %A) {
2459 // tail call @bar(%struct.X* byval %A)
2462 SDValue Ptr = Ld->getBasePtr();
2463 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2466 FI = FINode->getIndex();
2467 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
2468 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
2469 FI = FINode->getIndex();
2470 Bytes = Flags.getByValSize();
2474 assert(FI != INT_MAX);
2475 if (!MFI->isFixedObjectIndex(FI))
2477 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2480 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2481 /// for tail call optimization. Targets which want to do tail call
2482 /// optimization should implement this function.
2484 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2485 CallingConv::ID CalleeCC,
2487 bool isCalleeStructRet,
2488 bool isCallerStructRet,
2489 const SmallVectorImpl<ISD::OutputArg> &Outs,
2490 const SmallVectorImpl<SDValue> &OutVals,
2491 const SmallVectorImpl<ISD::InputArg> &Ins,
2492 SelectionDAG& DAG) const {
2493 if (!IsTailCallConvention(CalleeCC) &&
2494 CalleeCC != CallingConv::C)
2497 // If -tailcallopt is specified, make fastcc functions tail-callable.
2498 const MachineFunction &MF = DAG.getMachineFunction();
2499 const Function *CallerF = DAG.getMachineFunction().getFunction();
2500 CallingConv::ID CallerCC = CallerF->getCallingConv();
2501 bool CCMatch = CallerCC == CalleeCC;
2503 if (GuaranteedTailCallOpt) {
2504 if (IsTailCallConvention(CalleeCC) && CCMatch)
2509 // Look for obvious safe cases to perform tail call optimization that do not
2510 // require ABI changes. This is what gcc calls sibcall.
2512 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2513 // emit a special epilogue.
2514 if (RegInfo->needsStackRealignment(MF))
2517 // Also avoid sibcall optimization if either caller or callee uses struct
2518 // return semantics.
2519 if (isCalleeStructRet || isCallerStructRet)
2522 // An stdcall caller is expected to clean up its arguments; the callee
2523 // isn't going to do that.
2524 if (!CCMatch && CallerCC==CallingConv::X86_StdCall)
2527 // Do not sibcall optimize vararg calls unless all arguments are passed via
2529 if (isVarArg && !Outs.empty()) {
2531 // Optimizing for varargs on Win64 is unlikely to be safe without
2532 // additional testing.
2533 if (Subtarget->isTargetWin64())
2536 SmallVector<CCValAssign, 16> ArgLocs;
2537 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2538 getTargetMachine(), ArgLocs, *DAG.getContext());
2540 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2541 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
2542 if (!ArgLocs[i].isRegLoc())
2546 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2547 // Therefore if it's not used by the call it is not safe to optimize this into
2549 bool Unused = false;
2550 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2557 SmallVector<CCValAssign, 16> RVLocs;
2558 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(),
2559 getTargetMachine(), RVLocs, *DAG.getContext());
2560 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2561 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2562 CCValAssign &VA = RVLocs[i];
2563 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2568 // If the calling conventions do not match, then we'd better make sure the
2569 // results are returned in the same way as what the caller expects.
2571 SmallVector<CCValAssign, 16> RVLocs1;
2572 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
2573 getTargetMachine(), RVLocs1, *DAG.getContext());
2574 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2576 SmallVector<CCValAssign, 16> RVLocs2;
2577 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
2578 getTargetMachine(), RVLocs2, *DAG.getContext());
2579 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2581 if (RVLocs1.size() != RVLocs2.size())
2583 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2584 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2586 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2588 if (RVLocs1[i].isRegLoc()) {
2589 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2592 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2598 // If the callee takes no arguments then go on to check the results of the
2600 if (!Outs.empty()) {
2601 // Check if stack adjustment is needed. For now, do not do this if any
2602 // argument is passed on the stack.
2603 SmallVector<CCValAssign, 16> ArgLocs;
2604 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2605 getTargetMachine(), ArgLocs, *DAG.getContext());
2607 // Allocate shadow area for Win64
2608 if (Subtarget->isTargetWin64()) {
2609 CCInfo.AllocateStack(32, 8);
2612 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2613 if (CCInfo.getNextStackOffset()) {
2614 MachineFunction &MF = DAG.getMachineFunction();
2615 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2618 // Check if the arguments are already laid out in the right way as
2619 // the caller's fixed stack objects.
2620 MachineFrameInfo *MFI = MF.getFrameInfo();
2621 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2622 const X86InstrInfo *TII =
2623 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2624 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2625 CCValAssign &VA = ArgLocs[i];
2626 SDValue Arg = OutVals[i];
2627 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2628 if (VA.getLocInfo() == CCValAssign::Indirect)
2630 if (!VA.isRegLoc()) {
2631 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2638 // If the tailcall address may be in a register, then make sure it's
2639 // possible to register allocate for it. In 32-bit, the call address can
2640 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2641 // callee-saved registers are restored. These happen to be the same
2642 // registers used to pass 'inreg' arguments so watch out for those.
2643 if (!Subtarget->is64Bit() &&
2644 !isa<GlobalAddressSDNode>(Callee) &&
2645 !isa<ExternalSymbolSDNode>(Callee)) {
2646 unsigned NumInRegs = 0;
2647 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2648 CCValAssign &VA = ArgLocs[i];
2651 unsigned Reg = VA.getLocReg();
2654 case X86::EAX: case X86::EDX: case X86::ECX:
2655 if (++NumInRegs == 3)
2667 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2668 return X86::createFastISel(funcInfo);
2672 //===----------------------------------------------------------------------===//
2673 // Other Lowering Hooks
2674 //===----------------------------------------------------------------------===//
2676 static bool MayFoldLoad(SDValue Op) {
2677 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
2680 static bool MayFoldIntoStore(SDValue Op) {
2681 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
2684 static bool isTargetShuffle(unsigned Opcode) {
2686 default: return false;
2687 case X86ISD::PSHUFD:
2688 case X86ISD::PSHUFHW:
2689 case X86ISD::PSHUFLW:
2690 case X86ISD::SHUFPD:
2691 case X86ISD::PALIGN:
2692 case X86ISD::SHUFPS:
2693 case X86ISD::MOVLHPS:
2694 case X86ISD::MOVLHPD:
2695 case X86ISD::MOVHLPS:
2696 case X86ISD::MOVLPS:
2697 case X86ISD::MOVLPD:
2698 case X86ISD::MOVSHDUP:
2699 case X86ISD::MOVSLDUP:
2700 case X86ISD::MOVDDUP:
2703 case X86ISD::UNPCKLPS:
2704 case X86ISD::UNPCKLPD:
2705 case X86ISD::VUNPCKLPSY:
2706 case X86ISD::VUNPCKLPDY:
2707 case X86ISD::PUNPCKLWD:
2708 case X86ISD::PUNPCKLBW:
2709 case X86ISD::PUNPCKLDQ:
2710 case X86ISD::PUNPCKLQDQ:
2711 case X86ISD::UNPCKHPS:
2712 case X86ISD::UNPCKHPD:
2713 case X86ISD::PUNPCKHWD:
2714 case X86ISD::PUNPCKHBW:
2715 case X86ISD::PUNPCKHDQ:
2716 case X86ISD::PUNPCKHQDQ:
2717 case X86ISD::VPERMIL:
2723 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2724 SDValue V1, SelectionDAG &DAG) {
2726 default: llvm_unreachable("Unknown x86 shuffle node");
2727 case X86ISD::MOVSHDUP:
2728 case X86ISD::MOVSLDUP:
2729 case X86ISD::MOVDDUP:
2730 return DAG.getNode(Opc, dl, VT, V1);
2736 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2737 SDValue V1, unsigned TargetMask, SelectionDAG &DAG) {
2739 default: llvm_unreachable("Unknown x86 shuffle node");
2740 case X86ISD::PSHUFD:
2741 case X86ISD::PSHUFHW:
2742 case X86ISD::PSHUFLW:
2743 case X86ISD::VPERMIL:
2744 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
2750 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2751 SDValue V1, SDValue V2, unsigned TargetMask, SelectionDAG &DAG) {
2753 default: llvm_unreachable("Unknown x86 shuffle node");
2754 case X86ISD::PALIGN:
2755 case X86ISD::SHUFPD:
2756 case X86ISD::SHUFPS:
2757 return DAG.getNode(Opc, dl, VT, V1, V2,
2758 DAG.getConstant(TargetMask, MVT::i8));
2763 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2764 SDValue V1, SDValue V2, SelectionDAG &DAG) {
2766 default: llvm_unreachable("Unknown x86 shuffle node");
2767 case X86ISD::MOVLHPS:
2768 case X86ISD::MOVLHPD:
2769 case X86ISD::MOVHLPS:
2770 case X86ISD::MOVLPS:
2771 case X86ISD::MOVLPD:
2774 case X86ISD::UNPCKLPS:
2775 case X86ISD::UNPCKLPD:
2776 case X86ISD::VUNPCKLPSY:
2777 case X86ISD::VUNPCKLPDY:
2778 case X86ISD::PUNPCKLWD:
2779 case X86ISD::PUNPCKLBW:
2780 case X86ISD::PUNPCKLDQ:
2781 case X86ISD::PUNPCKLQDQ:
2782 case X86ISD::UNPCKHPS:
2783 case X86ISD::UNPCKHPD:
2784 case X86ISD::PUNPCKHWD:
2785 case X86ISD::PUNPCKHBW:
2786 case X86ISD::PUNPCKHDQ:
2787 case X86ISD::PUNPCKHQDQ:
2788 return DAG.getNode(Opc, dl, VT, V1, V2);
2793 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2794 MachineFunction &MF = DAG.getMachineFunction();
2795 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2796 int ReturnAddrIndex = FuncInfo->getRAIndex();
2798 if (ReturnAddrIndex == 0) {
2799 // Set up a frame object for the return address.
2800 uint64_t SlotSize = TD->getPointerSize();
2801 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2803 FuncInfo->setRAIndex(ReturnAddrIndex);
2806 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2810 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2811 bool hasSymbolicDisplacement) {
2812 // Offset should fit into 32 bit immediate field.
2813 if (!isInt<32>(Offset))
2816 // If we don't have a symbolic displacement - we don't have any extra
2818 if (!hasSymbolicDisplacement)
2821 // FIXME: Some tweaks might be needed for medium code model.
2822 if (M != CodeModel::Small && M != CodeModel::Kernel)
2825 // For small code model we assume that latest object is 16MB before end of 31
2826 // bits boundary. We may also accept pretty large negative constants knowing
2827 // that all objects are in the positive half of address space.
2828 if (M == CodeModel::Small && Offset < 16*1024*1024)
2831 // For kernel code model we know that all object resist in the negative half
2832 // of 32bits address space. We may not accept negative offsets, since they may
2833 // be just off and we may accept pretty large positive ones.
2834 if (M == CodeModel::Kernel && Offset > 0)
2840 /// isCalleePop - Determines whether the callee is required to pop its
2841 /// own arguments. Callee pop is necessary to support tail calls.
2842 bool X86::isCalleePop(CallingConv::ID CallingConv,
2843 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
2847 switch (CallingConv) {
2850 case CallingConv::X86_StdCall:
2852 case CallingConv::X86_FastCall:
2854 case CallingConv::X86_ThisCall:
2856 case CallingConv::Fast:
2858 case CallingConv::GHC:
2863 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2864 /// specific condition code, returning the condition code and the LHS/RHS of the
2865 /// comparison to make.
2866 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2867 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2869 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2870 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2871 // X > -1 -> X == 0, jump !sign.
2872 RHS = DAG.getConstant(0, RHS.getValueType());
2873 return X86::COND_NS;
2874 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2875 // X < 0 -> X == 0, jump on sign.
2877 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2879 RHS = DAG.getConstant(0, RHS.getValueType());
2880 return X86::COND_LE;
2884 switch (SetCCOpcode) {
2885 default: llvm_unreachable("Invalid integer condition!");
2886 case ISD::SETEQ: return X86::COND_E;
2887 case ISD::SETGT: return X86::COND_G;
2888 case ISD::SETGE: return X86::COND_GE;
2889 case ISD::SETLT: return X86::COND_L;
2890 case ISD::SETLE: return X86::COND_LE;
2891 case ISD::SETNE: return X86::COND_NE;
2892 case ISD::SETULT: return X86::COND_B;
2893 case ISD::SETUGT: return X86::COND_A;
2894 case ISD::SETULE: return X86::COND_BE;
2895 case ISD::SETUGE: return X86::COND_AE;
2899 // First determine if it is required or is profitable to flip the operands.
2901 // If LHS is a foldable load, but RHS is not, flip the condition.
2902 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
2903 !ISD::isNON_EXTLoad(RHS.getNode())) {
2904 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2905 std::swap(LHS, RHS);
2908 switch (SetCCOpcode) {
2914 std::swap(LHS, RHS);
2918 // On a floating point condition, the flags are set as follows:
2920 // 0 | 0 | 0 | X > Y
2921 // 0 | 0 | 1 | X < Y
2922 // 1 | 0 | 0 | X == Y
2923 // 1 | 1 | 1 | unordered
2924 switch (SetCCOpcode) {
2925 default: llvm_unreachable("Condcode should be pre-legalized away");
2927 case ISD::SETEQ: return X86::COND_E;
2928 case ISD::SETOLT: // flipped
2930 case ISD::SETGT: return X86::COND_A;
2931 case ISD::SETOLE: // flipped
2933 case ISD::SETGE: return X86::COND_AE;
2934 case ISD::SETUGT: // flipped
2936 case ISD::SETLT: return X86::COND_B;
2937 case ISD::SETUGE: // flipped
2939 case ISD::SETLE: return X86::COND_BE;
2941 case ISD::SETNE: return X86::COND_NE;
2942 case ISD::SETUO: return X86::COND_P;
2943 case ISD::SETO: return X86::COND_NP;
2945 case ISD::SETUNE: return X86::COND_INVALID;
2949 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2950 /// code. Current x86 isa includes the following FP cmov instructions:
2951 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2952 static bool hasFPCMov(unsigned X86CC) {
2968 /// isFPImmLegal - Returns true if the target can instruction select the
2969 /// specified FP immediate natively. If false, the legalizer will
2970 /// materialize the FP immediate as a load from a constant pool.
2971 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2972 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2973 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2979 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2980 /// the specified range (L, H].
2981 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2982 return (Val < 0) || (Val >= Low && Val < Hi);
2985 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2986 /// specified value.
2987 static bool isUndefOrEqual(int Val, int CmpVal) {
2988 if (Val < 0 || Val == CmpVal)
2993 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2994 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2995 /// the second operand.
2996 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2997 if (VT == MVT::v4f32 || VT == MVT::v4i32 )
2998 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2999 if (VT == MVT::v2f64 || VT == MVT::v2i64)
3000 return (Mask[0] < 2 && Mask[1] < 2);
3004 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
3005 SmallVector<int, 8> M;
3007 return ::isPSHUFDMask(M, N->getValueType(0));
3010 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
3011 /// is suitable for input to PSHUFHW.
3012 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3013 if (VT != MVT::v8i16)
3016 // Lower quadword copied in order or undef.
3017 for (int i = 0; i != 4; ++i)
3018 if (Mask[i] >= 0 && Mask[i] != i)
3021 // Upper quadword shuffled.
3022 for (int i = 4; i != 8; ++i)
3023 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
3029 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
3030 SmallVector<int, 8> M;
3032 return ::isPSHUFHWMask(M, N->getValueType(0));
3035 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
3036 /// is suitable for input to PSHUFLW.
3037 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3038 if (VT != MVT::v8i16)
3041 // Upper quadword copied in order.
3042 for (int i = 4; i != 8; ++i)
3043 if (Mask[i] >= 0 && Mask[i] != i)
3046 // Lower quadword shuffled.
3047 for (int i = 0; i != 4; ++i)
3054 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
3055 SmallVector<int, 8> M;
3057 return ::isPSHUFLWMask(M, N->getValueType(0));
3060 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
3061 /// is suitable for input to PALIGNR.
3062 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
3064 int i, e = VT.getVectorNumElements();
3066 // Do not handle v2i64 / v2f64 shuffles with palignr.
3067 if (e < 4 || !hasSSSE3)
3070 for (i = 0; i != e; ++i)
3074 // All undef, not a palignr.
3078 // Determine if it's ok to perform a palignr with only the LHS, since we
3079 // don't have access to the actual shuffle elements to see if RHS is undef.
3080 bool Unary = Mask[i] < (int)e;
3081 bool NeedsUnary = false;
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) {
3091 Unary = Unary && (m < (int)e);
3092 NeedsUnary = NeedsUnary || (m < s);
3094 if (NeedsUnary && !Unary)
3096 if (Unary && m != ((s+i) & (e-1)))
3098 if (!Unary && m != (s+i))
3104 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
3105 SmallVector<int, 8> M;
3107 return ::isPALIGNRMask(M, N->getValueType(0), true);
3110 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
3111 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
3112 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3113 int NumElems = VT.getVectorNumElements();
3114 if (NumElems != 2 && NumElems != 4)
3117 int Half = NumElems / 2;
3118 for (int i = 0; i < Half; ++i)
3119 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3121 for (int i = Half; i < NumElems; ++i)
3122 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3128 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
3129 SmallVector<int, 8> M;
3131 return ::isSHUFPMask(M, N->getValueType(0));
3134 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
3135 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
3136 /// half elements to come from vector 1 (which would equal the dest.) and
3137 /// the upper half to come from vector 2.
3138 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3139 int NumElems = VT.getVectorNumElements();
3141 if (NumElems != 2 && NumElems != 4)
3144 int Half = NumElems / 2;
3145 for (int i = 0; i < Half; ++i)
3146 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3148 for (int i = Half; i < NumElems; ++i)
3149 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3154 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
3155 SmallVector<int, 8> M;
3157 return isCommutedSHUFPMask(M, N->getValueType(0));
3160 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3161 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3162 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
3163 if (N->getValueType(0).getVectorNumElements() != 4)
3166 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3167 return isUndefOrEqual(N->getMaskElt(0), 6) &&
3168 isUndefOrEqual(N->getMaskElt(1), 7) &&
3169 isUndefOrEqual(N->getMaskElt(2), 2) &&
3170 isUndefOrEqual(N->getMaskElt(3), 3);
3173 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3174 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3176 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
3177 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3182 return isUndefOrEqual(N->getMaskElt(0), 2) &&
3183 isUndefOrEqual(N->getMaskElt(1), 3) &&
3184 isUndefOrEqual(N->getMaskElt(2), 2) &&
3185 isUndefOrEqual(N->getMaskElt(3), 3);
3188 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3189 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3190 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
3191 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3193 if (NumElems != 2 && NumElems != 4)
3196 for (unsigned i = 0; i < NumElems/2; ++i)
3197 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
3200 for (unsigned i = NumElems/2; i < NumElems; ++i)
3201 if (!isUndefOrEqual(N->getMaskElt(i), i))
3207 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3208 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3209 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
3210 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3212 if ((NumElems != 2 && NumElems != 4)
3213 || N->getValueType(0).getSizeInBits() > 128)
3216 for (unsigned i = 0; i < NumElems/2; ++i)
3217 if (!isUndefOrEqual(N->getMaskElt(i), i))
3220 for (unsigned i = 0; i < NumElems/2; ++i)
3221 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
3227 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3228 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3229 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3230 bool V2IsSplat = false) {
3231 int NumElts = VT.getVectorNumElements();
3232 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3235 // Handle vector lengths > 128 bits. Define a "section" as a set of
3236 // 128 bits. AVX defines UNPCK* to operate independently on 128-bit
3238 unsigned NumSections = VT.getSizeInBits() / 128;
3239 if (NumSections == 0 ) NumSections = 1; // Handle MMX
3240 unsigned NumSectionElts = NumElts / NumSections;
3243 unsigned End = NumSectionElts;
3244 for (unsigned s = 0; s < NumSections; ++s) {
3245 for (unsigned i = Start, j = s * NumSectionElts;
3249 int BitI1 = Mask[i+1];
3250 if (!isUndefOrEqual(BitI, j))
3253 if (!isUndefOrEqual(BitI1, NumElts))
3256 if (!isUndefOrEqual(BitI1, j + NumElts))
3260 // Process the next 128 bits.
3261 Start += NumSectionElts;
3262 End += NumSectionElts;
3268 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3269 SmallVector<int, 8> M;
3271 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
3274 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3275 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3276 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3277 bool V2IsSplat = false) {
3278 int NumElts = VT.getVectorNumElements();
3279 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3282 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3284 int BitI1 = Mask[i+1];
3285 if (!isUndefOrEqual(BitI, j + NumElts/2))
3288 if (isUndefOrEqual(BitI1, NumElts))
3291 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
3298 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3299 SmallVector<int, 8> M;
3301 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3304 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3305 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3307 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3308 int NumElems = VT.getVectorNumElements();
3309 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3312 // Handle vector lengths > 128 bits. Define a "section" as a set of
3313 // 128 bits. AVX defines UNPCK* to operate independently on 128-bit
3315 unsigned NumSections = VT.getSizeInBits() / 128;
3316 if (NumSections == 0 ) NumSections = 1; // Handle MMX
3317 unsigned NumSectionElts = NumElems / NumSections;
3319 for (unsigned s = 0; s < NumSections; ++s) {
3320 for (unsigned i = s * NumSectionElts, j = s * NumSectionElts;
3321 i != NumSectionElts * (s + 1);
3324 int BitI1 = Mask[i+1];
3326 if (!isUndefOrEqual(BitI, j))
3328 if (!isUndefOrEqual(BitI1, j))
3336 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3337 SmallVector<int, 8> M;
3339 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3342 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3343 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3345 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3346 int NumElems = VT.getVectorNumElements();
3347 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3350 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3352 int BitI1 = Mask[i+1];
3353 if (!isUndefOrEqual(BitI, j))
3355 if (!isUndefOrEqual(BitI1, j))
3361 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3362 SmallVector<int, 8> M;
3364 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3367 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3368 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3369 /// MOVSD, and MOVD, i.e. setting the lowest element.
3370 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3371 if (VT.getVectorElementType().getSizeInBits() < 32)
3374 int NumElts = VT.getVectorNumElements();
3376 if (!isUndefOrEqual(Mask[0], NumElts))
3379 for (int i = 1; i < NumElts; ++i)
3380 if (!isUndefOrEqual(Mask[i], i))
3386 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3387 SmallVector<int, 8> M;
3389 return ::isMOVLMask(M, N->getValueType(0));
3392 /// isVPERMILMask - Return true if the specified VECTOR_SHUFFLE operand
3393 /// specifies a shuffle of elements that is suitable for input to VPERMIL*.
3394 static bool isVPERMILMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3395 unsigned NumElts = VT.getVectorNumElements();
3396 unsigned NumLanes = VT.getSizeInBits()/128;
3398 // Match any permutation of 128-bit vector with 32/64-bit types
3399 if (NumLanes == 1) {
3400 if (NumElts == 4 || NumElts == 2)
3405 // Only match 256-bit with 32/64-bit types
3406 if (NumElts != 8 && NumElts != 4)
3409 // The mask on the high lane should be the same as the low. Actually,
3410 // they can differ if any of the corresponding index in a lane is undef.
3411 int LaneSize = NumElts/NumLanes;
3412 for (int i = 0; i < LaneSize; ++i) {
3413 int HighElt = i+LaneSize;
3414 if (Mask[i] < 0 || Mask[HighElt] < 0)
3417 if (Mask[HighElt]-Mask[i] != LaneSize)
3424 /// getShuffleVPERMILImmediateediate - Return the appropriate immediate to shuffle
3425 /// the specified VECTOR_MASK mask with VPERMIL* instructions.
3426 static unsigned getShuffleVPERMILImmediate(SDNode *N) {
3427 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3428 EVT VT = SVOp->getValueType(0);
3430 int NumElts = VT.getVectorNumElements();
3431 int NumLanes = VT.getSizeInBits()/128;
3434 for (int i = 0; i < NumElts/NumLanes /* lane size */; ++i)
3435 Mask |= SVOp->getMaskElt(i) << (i*2);
3440 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3441 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3442 /// element of vector 2 and the other elements to come from vector 1 in order.
3443 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3444 bool V2IsSplat = false, bool V2IsUndef = false) {
3445 int NumOps = VT.getVectorNumElements();
3446 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3449 if (!isUndefOrEqual(Mask[0], 0))
3452 for (int i = 1; i < NumOps; ++i)
3453 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3454 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3455 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3461 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3462 bool V2IsUndef = false) {
3463 SmallVector<int, 8> M;
3465 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3468 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3469 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3470 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3471 if (N->getValueType(0).getVectorNumElements() != 4)
3474 // Expect 1, 1, 3, 3
3475 for (unsigned i = 0; i < 2; ++i) {
3476 int Elt = N->getMaskElt(i);
3477 if (Elt >= 0 && Elt != 1)
3482 for (unsigned i = 2; i < 4; ++i) {
3483 int Elt = N->getMaskElt(i);
3484 if (Elt >= 0 && Elt != 3)
3489 // Don't use movshdup if it can be done with a shufps.
3490 // FIXME: verify that matching u, u, 3, 3 is what we want.
3494 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3495 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3496 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3497 if (N->getValueType(0).getVectorNumElements() != 4)
3500 // Expect 0, 0, 2, 2
3501 for (unsigned i = 0; i < 2; ++i)
3502 if (N->getMaskElt(i) > 0)
3506 for (unsigned i = 2; i < 4; ++i) {
3507 int Elt = N->getMaskElt(i);
3508 if (Elt >= 0 && Elt != 2)
3513 // Don't use movsldup if it can be done with a shufps.
3517 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3518 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3519 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3520 int e = N->getValueType(0).getVectorNumElements() / 2;
3522 for (int i = 0; i < e; ++i)
3523 if (!isUndefOrEqual(N->getMaskElt(i), i))
3525 for (int i = 0; i < e; ++i)
3526 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3531 /// isVEXTRACTF128Index - Return true if the specified
3532 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
3533 /// suitable for input to VEXTRACTF128.
3534 bool X86::isVEXTRACTF128Index(SDNode *N) {
3535 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3538 // The index should be aligned on a 128-bit boundary.
3540 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3542 unsigned VL = N->getValueType(0).getVectorNumElements();
3543 unsigned VBits = N->getValueType(0).getSizeInBits();
3544 unsigned ElSize = VBits / VL;
3545 bool Result = (Index * ElSize) % 128 == 0;
3550 /// isVINSERTF128Index - Return true if the specified INSERT_SUBVECTOR
3551 /// operand specifies a subvector insert that is suitable for input to
3553 bool X86::isVINSERTF128Index(SDNode *N) {
3554 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3557 // The index should be aligned on a 128-bit boundary.
3559 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3561 unsigned VL = N->getValueType(0).getVectorNumElements();
3562 unsigned VBits = N->getValueType(0).getSizeInBits();
3563 unsigned ElSize = VBits / VL;
3564 bool Result = (Index * ElSize) % 128 == 0;
3569 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3570 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3571 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3572 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3573 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3575 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3577 for (int i = 0; i < NumOperands; ++i) {
3578 int Val = SVOp->getMaskElt(NumOperands-i-1);
3579 if (Val < 0) Val = 0;
3580 if (Val >= NumOperands) Val -= NumOperands;
3582 if (i != NumOperands - 1)
3588 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3589 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3590 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3591 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3593 // 8 nodes, but we only care about the last 4.
3594 for (unsigned i = 7; i >= 4; --i) {
3595 int Val = SVOp->getMaskElt(i);
3604 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3605 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3606 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3607 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3609 // 8 nodes, but we only care about the first 4.
3610 for (int i = 3; i >= 0; --i) {
3611 int Val = SVOp->getMaskElt(i);
3620 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3621 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3622 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3623 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3624 EVT VVT = N->getValueType(0);
3625 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3629 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3630 Val = SVOp->getMaskElt(i);
3634 return (Val - i) * EltSize;
3637 /// getExtractVEXTRACTF128Immediate - Return the appropriate immediate
3638 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
3640 unsigned X86::getExtractVEXTRACTF128Immediate(SDNode *N) {
3641 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3642 llvm_unreachable("Illegal extract subvector for VEXTRACTF128");
3645 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3647 EVT VecVT = N->getOperand(0).getValueType();
3648 EVT ElVT = VecVT.getVectorElementType();
3650 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
3651 return Index / NumElemsPerChunk;
3654 /// getInsertVINSERTF128Immediate - Return the appropriate immediate
3655 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
3657 unsigned X86::getInsertVINSERTF128Immediate(SDNode *N) {
3658 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3659 llvm_unreachable("Illegal insert subvector for VINSERTF128");
3662 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3664 EVT VecVT = N->getValueType(0);
3665 EVT ElVT = VecVT.getVectorElementType();
3667 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
3668 return Index / NumElemsPerChunk;
3671 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3673 bool X86::isZeroNode(SDValue Elt) {
3674 return ((isa<ConstantSDNode>(Elt) &&
3675 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3676 (isa<ConstantFPSDNode>(Elt) &&
3677 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3680 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3681 /// their permute mask.
3682 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3683 SelectionDAG &DAG) {
3684 EVT VT = SVOp->getValueType(0);
3685 unsigned NumElems = VT.getVectorNumElements();
3686 SmallVector<int, 8> MaskVec;
3688 for (unsigned i = 0; i != NumElems; ++i) {
3689 int idx = SVOp->getMaskElt(i);
3691 MaskVec.push_back(idx);
3692 else if (idx < (int)NumElems)
3693 MaskVec.push_back(idx + NumElems);
3695 MaskVec.push_back(idx - NumElems);
3697 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3698 SVOp->getOperand(0), &MaskVec[0]);
3701 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3702 /// the two vector operands have swapped position.
3703 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3704 unsigned NumElems = VT.getVectorNumElements();
3705 for (unsigned i = 0; i != NumElems; ++i) {
3709 else if (idx < (int)NumElems)
3710 Mask[i] = idx + NumElems;
3712 Mask[i] = idx - NumElems;
3716 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3717 /// match movhlps. The lower half elements should come from upper half of
3718 /// V1 (and in order), and the upper half elements should come from the upper
3719 /// half of V2 (and in order).
3720 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3721 if (Op->getValueType(0).getVectorNumElements() != 4)
3723 for (unsigned i = 0, e = 2; i != e; ++i)
3724 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3726 for (unsigned i = 2; i != 4; ++i)
3727 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3732 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3733 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3735 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3736 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3738 N = N->getOperand(0).getNode();
3739 if (!ISD::isNON_EXTLoad(N))
3742 *LD = cast<LoadSDNode>(N);
3746 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3747 /// match movlp{s|d}. The lower half elements should come from lower half of
3748 /// V1 (and in order), and the upper half elements should come from the upper
3749 /// half of V2 (and in order). And since V1 will become the source of the
3750 /// MOVLP, it must be either a vector load or a scalar load to vector.
3751 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3752 ShuffleVectorSDNode *Op) {
3753 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3755 // Is V2 is a vector load, don't do this transformation. We will try to use
3756 // load folding shufps op.
3757 if (ISD::isNON_EXTLoad(V2))
3760 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3762 if (NumElems != 2 && NumElems != 4)
3764 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3765 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3767 for (unsigned i = NumElems/2; i != NumElems; ++i)
3768 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3773 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3775 static bool isSplatVector(SDNode *N) {
3776 if (N->getOpcode() != ISD::BUILD_VECTOR)
3779 SDValue SplatValue = N->getOperand(0);
3780 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3781 if (N->getOperand(i) != SplatValue)
3786 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3787 /// to an zero vector.
3788 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3789 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3790 SDValue V1 = N->getOperand(0);
3791 SDValue V2 = N->getOperand(1);
3792 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3793 for (unsigned i = 0; i != NumElems; ++i) {
3794 int Idx = N->getMaskElt(i);
3795 if (Idx >= (int)NumElems) {
3796 unsigned Opc = V2.getOpcode();
3797 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3799 if (Opc != ISD::BUILD_VECTOR ||
3800 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3802 } else if (Idx >= 0) {
3803 unsigned Opc = V1.getOpcode();
3804 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3806 if (Opc != ISD::BUILD_VECTOR ||
3807 !X86::isZeroNode(V1.getOperand(Idx)))
3814 /// getZeroVector - Returns a vector of specified type with all zero elements.
3816 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3818 assert(VT.isVector() && "Expected a vector type");
3820 // Always build SSE zero vectors as <4 x i32> bitcasted
3821 // to their dest type. This ensures they get CSE'd.
3823 if (VT.getSizeInBits() == 128) { // SSE
3824 if (HasSSE2) { // SSE2
3825 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3826 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3828 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3829 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3831 } else if (VT.getSizeInBits() == 256) { // AVX
3832 // 256-bit logic and arithmetic instructions in AVX are
3833 // all floating-point, no support for integer ops. Default
3834 // to emitting fp zeroed vectors then.
3835 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3836 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3837 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
3839 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
3842 /// getOnesVector - Returns a vector of specified type with all bits set.
3843 /// Always build ones vectors as <4 x i32> or <8 x i32> bitcasted to
3844 /// their original type, ensuring they get CSE'd.
3845 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3846 assert(VT.isVector() && "Expected a vector type");
3847 assert((VT.is128BitVector() || VT.is256BitVector())
3848 && "Expected a 128-bit or 256-bit vector type");
3850 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3853 if (VT.is256BitVector()) {
3854 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3855 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops, 8);
3857 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3858 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
3861 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3862 /// that point to V2 points to its first element.
3863 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3864 EVT VT = SVOp->getValueType(0);
3865 unsigned NumElems = VT.getVectorNumElements();
3867 bool Changed = false;
3868 SmallVector<int, 8> MaskVec;
3869 SVOp->getMask(MaskVec);
3871 for (unsigned i = 0; i != NumElems; ++i) {
3872 if (MaskVec[i] > (int)NumElems) {
3873 MaskVec[i] = NumElems;
3878 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3879 SVOp->getOperand(1), &MaskVec[0]);
3880 return SDValue(SVOp, 0);
3883 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3884 /// operation of specified width.
3885 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3887 unsigned NumElems = VT.getVectorNumElements();
3888 SmallVector<int, 8> Mask;
3889 Mask.push_back(NumElems);
3890 for (unsigned i = 1; i != NumElems; ++i)
3892 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3895 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3896 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3898 unsigned NumElems = VT.getVectorNumElements();
3899 SmallVector<int, 8> Mask;
3900 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3902 Mask.push_back(i + NumElems);
3904 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3907 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
3908 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3910 unsigned NumElems = VT.getVectorNumElements();
3911 unsigned Half = NumElems/2;
3912 SmallVector<int, 8> Mask;
3913 for (unsigned i = 0; i != Half; ++i) {
3914 Mask.push_back(i + Half);
3915 Mask.push_back(i + NumElems + Half);
3917 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3920 // PromoteSplatv8v16 - All i16 and i8 vector types can't be used directly by
3921 // a generic shuffle instruction because the target has no such instructions.
3922 // Generate shuffles which repeat i16 and i8 several times until they can be
3923 // represented by v4f32 and then be manipulated by target suported shuffles.
3924 static SDValue PromoteSplatv8v16(SDValue V, SelectionDAG &DAG, int &EltNo) {
3925 EVT VT = V.getValueType();
3926 int NumElems = VT.getVectorNumElements();
3927 DebugLoc dl = V.getDebugLoc();
3929 while (NumElems > 4) {
3930 if (EltNo < NumElems/2) {
3931 V = getUnpackl(DAG, dl, VT, V, V);
3933 V = getUnpackh(DAG, dl, VT, V, V);
3934 EltNo -= NumElems/2;
3941 /// getLegalSplat - Generate a legal splat with supported x86 shuffles
3942 static SDValue getLegalSplat(SelectionDAG &DAG, SDValue V, int EltNo) {
3943 EVT VT = V.getValueType();
3944 DebugLoc dl = V.getDebugLoc();
3945 assert((VT.getSizeInBits() == 128 || VT.getSizeInBits() == 256)
3946 && "Vector size not supported");
3948 bool Is128 = VT.getSizeInBits() == 128;
3949 EVT NVT = Is128 ? MVT::v4f32 : MVT::v8f32;
3950 V = DAG.getNode(ISD::BITCAST, dl, NVT, V);
3953 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3954 V = DAG.getVectorShuffle(NVT, dl, V, DAG.getUNDEF(NVT), &SplatMask[0]);
3956 // The second half of indicies refer to the higher part, which is a
3957 // duplication of the lower one. This makes this shuffle a perfect match
3958 // for the VPERM instruction.
3959 int SplatMask[8] = { EltNo, EltNo, EltNo, EltNo,
3960 EltNo+4, EltNo+4, EltNo+4, EltNo+4 };
3961 V = DAG.getVectorShuffle(NVT, dl, V, DAG.getUNDEF(NVT), &SplatMask[0]);
3964 return DAG.getNode(ISD::BITCAST, dl, VT, V);
3967 /// PromoteSplat - Promote a splat of v4i32, v8i16 or v16i8 to v4f32 and
3968 /// v8i32, v16i16 or v32i8 to v8f32.
3969 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
3970 EVT SrcVT = SV->getValueType(0);
3971 SDValue V1 = SV->getOperand(0);
3972 DebugLoc dl = SV->getDebugLoc();
3974 int EltNo = SV->getSplatIndex();
3975 int NumElems = SrcVT.getVectorNumElements();
3976 unsigned Size = SrcVT.getSizeInBits();
3978 // Extract the 128-bit part containing the splat element and update
3979 // the splat element index when it refers to the higher register.
3981 unsigned Idx = (EltNo > NumElems/2) ? NumElems/2 : 0;
3982 V1 = Extract128BitVector(V1, DAG.getConstant(Idx, MVT::i32), DAG, dl);
3984 EltNo -= NumElems/2;
3987 // Make this 128-bit vector duplicate i8 and i16 elements
3989 V1 = PromoteSplatv8v16(V1, DAG, EltNo);
3991 // Recreate the 256-bit vector and place the same 128-bit vector
3992 // into the low and high part. This is necessary because we want
3993 // to use VPERM to shuffle the v8f32 vector, and VPERM only shuffles
3994 // inside each separate v4f32 lane.
3996 SDValue InsV = Insert128BitVector(DAG.getUNDEF(SrcVT), V1,
3997 DAG.getConstant(0, MVT::i32), DAG, dl);
3998 V1 = Insert128BitVector(InsV, V1,
3999 DAG.getConstant(NumElems/2, MVT::i32), DAG, dl);
4002 return getLegalSplat(DAG, V1, EltNo);
4005 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
4006 /// vector of zero or undef vector. This produces a shuffle where the low
4007 /// element of V2 is swizzled into the zero/undef vector, landing at element
4008 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4009 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4010 bool isZero, bool HasSSE2,
4011 SelectionDAG &DAG) {
4012 EVT VT = V2.getValueType();
4014 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
4015 unsigned NumElems = VT.getVectorNumElements();
4016 SmallVector<int, 16> MaskVec;
4017 for (unsigned i = 0; i != NumElems; ++i)
4018 // If this is the insertion idx, put the low elt of V2 here.
4019 MaskVec.push_back(i == Idx ? NumElems : i);
4020 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
4023 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4024 /// element of the result of the vector shuffle.
4025 static SDValue getShuffleScalarElt(SDNode *N, int Index, SelectionDAG &DAG,
4028 return SDValue(); // Limit search depth.
4030 SDValue V = SDValue(N, 0);
4031 EVT VT = V.getValueType();
4032 unsigned Opcode = V.getOpcode();
4034 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4035 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4036 Index = SV->getMaskElt(Index);
4039 return DAG.getUNDEF(VT.getVectorElementType());
4041 int NumElems = VT.getVectorNumElements();
4042 SDValue NewV = (Index < NumElems) ? SV->getOperand(0) : SV->getOperand(1);
4043 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG, Depth+1);
4046 // Recurse into target specific vector shuffles to find scalars.
4047 if (isTargetShuffle(Opcode)) {
4048 int NumElems = VT.getVectorNumElements();
4049 SmallVector<unsigned, 16> ShuffleMask;
4053 case X86ISD::SHUFPS:
4054 case X86ISD::SHUFPD:
4055 ImmN = N->getOperand(N->getNumOperands()-1);
4056 DecodeSHUFPSMask(NumElems,
4057 cast<ConstantSDNode>(ImmN)->getZExtValue(),
4060 case X86ISD::PUNPCKHBW:
4061 case X86ISD::PUNPCKHWD:
4062 case X86ISD::PUNPCKHDQ:
4063 case X86ISD::PUNPCKHQDQ:
4064 DecodePUNPCKHMask(NumElems, ShuffleMask);
4066 case X86ISD::UNPCKHPS:
4067 case X86ISD::UNPCKHPD:
4068 DecodeUNPCKHPMask(NumElems, ShuffleMask);
4070 case X86ISD::PUNPCKLBW:
4071 case X86ISD::PUNPCKLWD:
4072 case X86ISD::PUNPCKLDQ:
4073 case X86ISD::PUNPCKLQDQ:
4074 DecodePUNPCKLMask(VT, ShuffleMask);
4076 case X86ISD::UNPCKLPS:
4077 case X86ISD::UNPCKLPD:
4078 case X86ISD::VUNPCKLPSY:
4079 case X86ISD::VUNPCKLPDY:
4080 DecodeUNPCKLPMask(VT, ShuffleMask);
4082 case X86ISD::MOVHLPS:
4083 DecodeMOVHLPSMask(NumElems, ShuffleMask);
4085 case X86ISD::MOVLHPS:
4086 DecodeMOVLHPSMask(NumElems, ShuffleMask);
4088 case X86ISD::PSHUFD:
4089 ImmN = N->getOperand(N->getNumOperands()-1);
4090 DecodePSHUFMask(NumElems,
4091 cast<ConstantSDNode>(ImmN)->getZExtValue(),
4094 case X86ISD::PSHUFHW:
4095 ImmN = N->getOperand(N->getNumOperands()-1);
4096 DecodePSHUFHWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
4099 case X86ISD::PSHUFLW:
4100 ImmN = N->getOperand(N->getNumOperands()-1);
4101 DecodePSHUFLWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
4105 case X86ISD::MOVSD: {
4106 // The index 0 always comes from the first element of the second source,
4107 // this is why MOVSS and MOVSD are used in the first place. The other
4108 // elements come from the other positions of the first source vector.
4109 unsigned OpNum = (Index == 0) ? 1 : 0;
4110 return getShuffleScalarElt(V.getOperand(OpNum).getNode(), Index, DAG,
4113 case X86ISD::VPERMIL:
4114 ImmN = N->getOperand(N->getNumOperands()-1);
4115 DecodeVPERMILMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4118 assert("not implemented for target shuffle node");
4122 Index = ShuffleMask[Index];
4124 return DAG.getUNDEF(VT.getVectorElementType());
4126 SDValue NewV = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
4127 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG,
4131 // Actual nodes that may contain scalar elements
4132 if (Opcode == ISD::BITCAST) {
4133 V = V.getOperand(0);
4134 EVT SrcVT = V.getValueType();
4135 unsigned NumElems = VT.getVectorNumElements();
4137 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4141 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4142 return (Index == 0) ? V.getOperand(0)
4143 : DAG.getUNDEF(VT.getVectorElementType());
4145 if (V.getOpcode() == ISD::BUILD_VECTOR)
4146 return V.getOperand(Index);
4151 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
4152 /// shuffle operation which come from a consecutively from a zero. The
4153 /// search can start in two different directions, from left or right.
4155 unsigned getNumOfConsecutiveZeros(SDNode *N, int NumElems,
4156 bool ZerosFromLeft, SelectionDAG &DAG) {
4159 while (i < NumElems) {
4160 unsigned Index = ZerosFromLeft ? i : NumElems-i-1;
4161 SDValue Elt = getShuffleScalarElt(N, Index, DAG, 0);
4162 if (!(Elt.getNode() &&
4163 (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt))))
4171 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies from MaskI to
4172 /// MaskE correspond consecutively to elements from one of the vector operands,
4173 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
4175 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp, int MaskI, int MaskE,
4176 int OpIdx, int NumElems, unsigned &OpNum) {
4177 bool SeenV1 = false;
4178 bool SeenV2 = false;
4180 for (int i = MaskI; i <= MaskE; ++i, ++OpIdx) {
4181 int Idx = SVOp->getMaskElt(i);
4182 // Ignore undef indicies
4191 // Only accept consecutive elements from the same vector
4192 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
4196 OpNum = SeenV1 ? 0 : 1;
4200 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
4201 /// logical left shift of a vector.
4202 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4203 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4204 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4205 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4206 false /* check zeros from right */, DAG);
4212 // Considering the elements in the mask that are not consecutive zeros,
4213 // check if they consecutively come from only one of the source vectors.
4215 // V1 = {X, A, B, C} 0
4217 // vector_shuffle V1, V2 <1, 2, 3, X>
4219 if (!isShuffleMaskConsecutive(SVOp,
4220 0, // Mask Start Index
4221 NumElems-NumZeros-1, // Mask End Index
4222 NumZeros, // Where to start looking in the src vector
4223 NumElems, // Number of elements in vector
4224 OpSrc)) // Which source operand ?
4229 ShVal = SVOp->getOperand(OpSrc);
4233 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
4234 /// logical left shift of a vector.
4235 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4236 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4237 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4238 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4239 true /* check zeros from left */, DAG);
4245 // Considering the elements in the mask that are not consecutive zeros,
4246 // check if they consecutively come from only one of the source vectors.
4248 // 0 { A, B, X, X } = V2
4250 // vector_shuffle V1, V2 <X, X, 4, 5>
4252 if (!isShuffleMaskConsecutive(SVOp,
4253 NumZeros, // Mask Start Index
4254 NumElems-1, // Mask End Index
4255 0, // Where to start looking in the src vector
4256 NumElems, // Number of elements in vector
4257 OpSrc)) // Which source operand ?
4262 ShVal = SVOp->getOperand(OpSrc);
4266 /// isVectorShift - Returns true if the shuffle can be implemented as a
4267 /// logical left or right shift of a vector.
4268 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4269 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4270 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
4271 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
4277 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
4279 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4280 unsigned NumNonZero, unsigned NumZero,
4282 const TargetLowering &TLI) {
4286 DebugLoc dl = Op.getDebugLoc();
4289 for (unsigned i = 0; i < 16; ++i) {
4290 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4291 if (ThisIsNonZero && First) {
4293 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4295 V = DAG.getUNDEF(MVT::v8i16);
4300 SDValue ThisElt(0, 0), LastElt(0, 0);
4301 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4302 if (LastIsNonZero) {
4303 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4304 MVT::i16, Op.getOperand(i-1));
4306 if (ThisIsNonZero) {
4307 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4308 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4309 ThisElt, DAG.getConstant(8, MVT::i8));
4311 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4315 if (ThisElt.getNode())
4316 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4317 DAG.getIntPtrConstant(i/2));
4321 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
4324 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
4326 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4327 unsigned NumNonZero, unsigned NumZero,
4329 const TargetLowering &TLI) {
4333 DebugLoc dl = Op.getDebugLoc();
4336 for (unsigned i = 0; i < 8; ++i) {
4337 bool isNonZero = (NonZeros & (1 << i)) != 0;
4341 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4343 V = DAG.getUNDEF(MVT::v8i16);
4346 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4347 MVT::v8i16, V, Op.getOperand(i),
4348 DAG.getIntPtrConstant(i));
4355 /// getVShift - Return a vector logical shift node.
4357 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4358 unsigned NumBits, SelectionDAG &DAG,
4359 const TargetLowering &TLI, DebugLoc dl) {
4360 EVT ShVT = MVT::v2i64;
4361 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
4362 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
4363 return DAG.getNode(ISD::BITCAST, dl, VT,
4364 DAG.getNode(Opc, dl, ShVT, SrcOp,
4365 DAG.getConstant(NumBits,
4366 TLI.getShiftAmountTy(SrcOp.getValueType()))));
4370 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
4371 SelectionDAG &DAG) const {
4373 // Check if the scalar load can be widened into a vector load. And if
4374 // the address is "base + cst" see if the cst can be "absorbed" into
4375 // the shuffle mask.
4376 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4377 SDValue Ptr = LD->getBasePtr();
4378 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4380 EVT PVT = LD->getValueType(0);
4381 if (PVT != MVT::i32 && PVT != MVT::f32)
4386 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4387 FI = FINode->getIndex();
4389 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4390 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4391 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4392 Offset = Ptr.getConstantOperandVal(1);
4393 Ptr = Ptr.getOperand(0);
4398 SDValue Chain = LD->getChain();
4399 // Make sure the stack object alignment is at least 16.
4400 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4401 if (DAG.InferPtrAlignment(Ptr) < 16) {
4402 if (MFI->isFixedObjectIndex(FI)) {
4403 // Can't change the alignment. FIXME: It's possible to compute
4404 // the exact stack offset and reference FI + adjust offset instead.
4405 // If someone *really* cares about this. That's the way to implement it.
4408 MFI->setObjectAlignment(FI, 16);
4412 // (Offset % 16) must be multiple of 4. Then address is then
4413 // Ptr + (Offset & ~15).
4416 if ((Offset % 16) & 3)
4418 int64_t StartOffset = Offset & ~15;
4420 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
4421 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4423 int EltNo = (Offset - StartOffset) >> 2;
4424 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
4425 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
4426 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,
4427 LD->getPointerInfo().getWithOffset(StartOffset),
4429 // Canonicalize it to a v4i32 shuffle.
4430 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1);
4431 return DAG.getNode(ISD::BITCAST, dl, VT,
4432 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
4433 DAG.getUNDEF(MVT::v4i32),&Mask[0]));
4439 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
4440 /// vector of type 'VT', see if the elements can be replaced by a single large
4441 /// load which has the same value as a build_vector whose operands are 'elts'.
4443 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4445 /// FIXME: we'd also like to handle the case where the last elements are zero
4446 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4447 /// There's even a handy isZeroNode for that purpose.
4448 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
4449 DebugLoc &DL, SelectionDAG &DAG) {
4450 EVT EltVT = VT.getVectorElementType();
4451 unsigned NumElems = Elts.size();
4453 LoadSDNode *LDBase = NULL;
4454 unsigned LastLoadedElt = -1U;
4456 // For each element in the initializer, see if we've found a load or an undef.
4457 // If we don't find an initial load element, or later load elements are
4458 // non-consecutive, bail out.
4459 for (unsigned i = 0; i < NumElems; ++i) {
4460 SDValue Elt = Elts[i];
4462 if (!Elt.getNode() ||
4463 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4466 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4468 LDBase = cast<LoadSDNode>(Elt.getNode());
4472 if (Elt.getOpcode() == ISD::UNDEF)
4475 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4476 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
4481 // If we have found an entire vector of loads and undefs, then return a large
4482 // load of the entire vector width starting at the base pointer. If we found
4483 // consecutive loads for the low half, generate a vzext_load node.
4484 if (LastLoadedElt == NumElems - 1) {
4485 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
4486 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4487 LDBase->getPointerInfo(),
4488 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
4489 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4490 LDBase->getPointerInfo(),
4491 LDBase->isVolatile(), LDBase->isNonTemporal(),
4492 LDBase->getAlignment());
4493 } else if (NumElems == 4 && LastLoadedElt == 1) {
4494 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4495 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4496 SDValue ResNode = DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys,
4498 LDBase->getMemOperand());
4499 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
4505 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
4506 DebugLoc dl = Op.getDebugLoc();
4508 EVT VT = Op.getValueType();
4509 EVT ExtVT = VT.getVectorElementType();
4510 unsigned NumElems = Op.getNumOperands();
4513 // - pxor (SSE2), xorps (SSE1), vpxor (128 AVX), xorp[s|d] (256 AVX)
4515 // - pcmpeqd (SSE2 and 128 AVX), fallback to constant pools (256 AVX)
4516 if (ISD::isBuildVectorAllZeros(Op.getNode()) ||
4517 ISD::isBuildVectorAllOnes(Op.getNode())) {
4518 // Canonicalize this to <4 x i32> or <8 x 32> (SSE) to
4519 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
4520 // eliminated on x86-32 hosts.
4521 if (Op.getValueType() == MVT::v4i32 ||
4522 Op.getValueType() == MVT::v8i32)
4525 if (ISD::isBuildVectorAllOnes(Op.getNode()))
4526 return getOnesVector(Op.getValueType(), DAG, dl);
4527 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
4530 unsigned EVTBits = ExtVT.getSizeInBits();
4532 unsigned NumZero = 0;
4533 unsigned NumNonZero = 0;
4534 unsigned NonZeros = 0;
4535 bool IsAllConstants = true;
4536 SmallSet<SDValue, 8> Values;
4537 for (unsigned i = 0; i < NumElems; ++i) {
4538 SDValue Elt = Op.getOperand(i);
4539 if (Elt.getOpcode() == ISD::UNDEF)
4542 if (Elt.getOpcode() != ISD::Constant &&
4543 Elt.getOpcode() != ISD::ConstantFP)
4544 IsAllConstants = false;
4545 if (X86::isZeroNode(Elt))
4548 NonZeros |= (1 << i);
4553 // All undef vector. Return an UNDEF. All zero vectors were handled above.
4554 if (NumNonZero == 0)
4555 return DAG.getUNDEF(VT);
4557 // Special case for single non-zero, non-undef, element.
4558 if (NumNonZero == 1) {
4559 unsigned Idx = CountTrailingZeros_32(NonZeros);
4560 SDValue Item = Op.getOperand(Idx);
4562 // If this is an insertion of an i64 value on x86-32, and if the top bits of
4563 // the value are obviously zero, truncate the value to i32 and do the
4564 // insertion that way. Only do this if the value is non-constant or if the
4565 // value is a constant being inserted into element 0. It is cheaper to do
4566 // a constant pool load than it is to do a movd + shuffle.
4567 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
4568 (!IsAllConstants || Idx == 0)) {
4569 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
4571 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
4572 EVT VecVT = MVT::v4i32;
4573 unsigned VecElts = 4;
4575 // Truncate the value (which may itself be a constant) to i32, and
4576 // convert it to a vector with movd (S2V+shuffle to zero extend).
4577 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
4578 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
4579 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4580 Subtarget->hasSSE2(), DAG);
4582 // Now we have our 32-bit value zero extended in the low element of
4583 // a vector. If Idx != 0, swizzle it into place.
4585 SmallVector<int, 4> Mask;
4586 Mask.push_back(Idx);
4587 for (unsigned i = 1; i != VecElts; ++i)
4589 Item = DAG.getVectorShuffle(VecVT, dl, Item,
4590 DAG.getUNDEF(Item.getValueType()),
4593 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Item);
4597 // If we have a constant or non-constant insertion into the low element of
4598 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
4599 // the rest of the elements. This will be matched as movd/movq/movss/movsd
4600 // depending on what the source datatype is.
4603 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4604 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
4605 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
4606 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4607 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
4608 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
4610 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
4611 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
4612 assert(VT.getSizeInBits() == 128 && "Expected an SSE value type!");
4613 EVT MiddleVT = MVT::v4i32;
4614 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
4615 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4616 Subtarget->hasSSE2(), DAG);
4617 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
4621 // Is it a vector logical left shift?
4622 if (NumElems == 2 && Idx == 1 &&
4623 X86::isZeroNode(Op.getOperand(0)) &&
4624 !X86::isZeroNode(Op.getOperand(1))) {
4625 unsigned NumBits = VT.getSizeInBits();
4626 return getVShift(true, VT,
4627 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4628 VT, Op.getOperand(1)),
4629 NumBits/2, DAG, *this, dl);
4632 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
4635 // Otherwise, if this is a vector with i32 or f32 elements, and the element
4636 // is a non-constant being inserted into an element other than the low one,
4637 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
4638 // movd/movss) to move this into the low element, then shuffle it into
4640 if (EVTBits == 32) {
4641 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4643 // Turn it into a shuffle of zero and zero-extended scalar to vector.
4644 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
4645 Subtarget->hasSSE2(), DAG);
4646 SmallVector<int, 8> MaskVec;
4647 for (unsigned i = 0; i < NumElems; i++)
4648 MaskVec.push_back(i == Idx ? 0 : 1);
4649 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
4653 // Splat is obviously ok. Let legalizer expand it to a shuffle.
4654 if (Values.size() == 1) {
4655 if (EVTBits == 32) {
4656 // Instead of a shuffle like this:
4657 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
4658 // Check if it's possible to issue this instead.
4659 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
4660 unsigned Idx = CountTrailingZeros_32(NonZeros);
4661 SDValue Item = Op.getOperand(Idx);
4662 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
4663 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
4668 // A vector full of immediates; various special cases are already
4669 // handled, so this is best done with a single constant-pool load.
4673 // For AVX-length vectors, build the individual 128-bit pieces and use
4674 // shuffles to put them in place.
4675 if (VT.getSizeInBits() == 256 && !ISD::isBuildVectorAllZeros(Op.getNode())) {
4676 SmallVector<SDValue, 32> V;
4677 for (unsigned i = 0; i < NumElems; ++i)
4678 V.push_back(Op.getOperand(i));
4680 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
4682 // Build both the lower and upper subvector.
4683 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[0], NumElems/2);
4684 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[NumElems / 2],
4687 // Recreate the wider vector with the lower and upper part.
4688 SDValue Vec = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT), Upper,
4689 DAG.getConstant(NumElems/2, MVT::i32), DAG, dl);
4690 return Insert128BitVector(Vec, Lower, DAG.getConstant(0, MVT::i32),
4694 // Let legalizer expand 2-wide build_vectors.
4695 if (EVTBits == 64) {
4696 if (NumNonZero == 1) {
4697 // One half is zero or undef.
4698 unsigned Idx = CountTrailingZeros_32(NonZeros);
4699 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
4700 Op.getOperand(Idx));
4701 return getShuffleVectorZeroOrUndef(V2, Idx, true,
4702 Subtarget->hasSSE2(), DAG);
4707 // If element VT is < 32 bits, convert it to inserts into a zero vector.
4708 if (EVTBits == 8 && NumElems == 16) {
4709 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
4711 if (V.getNode()) return V;
4714 if (EVTBits == 16 && NumElems == 8) {
4715 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
4717 if (V.getNode()) return V;
4720 // If element VT is == 32 bits, turn it into a number of shuffles.
4721 SmallVector<SDValue, 8> V;
4723 if (NumElems == 4 && NumZero > 0) {
4724 for (unsigned i = 0; i < 4; ++i) {
4725 bool isZero = !(NonZeros & (1 << i));
4727 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4729 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4732 for (unsigned i = 0; i < 2; ++i) {
4733 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
4736 V[i] = V[i*2]; // Must be a zero vector.
4739 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
4742 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
4745 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
4750 SmallVector<int, 8> MaskVec;
4751 bool Reverse = (NonZeros & 0x3) == 2;
4752 for (unsigned i = 0; i < 2; ++i)
4753 MaskVec.push_back(Reverse ? 1-i : i);
4754 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4755 for (unsigned i = 0; i < 2; ++i)
4756 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4757 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4760 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4761 // Check for a build vector of consecutive loads.
4762 for (unsigned i = 0; i < NumElems; ++i)
4763 V[i] = Op.getOperand(i);
4765 // Check for elements which are consecutive loads.
4766 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4770 // For SSE 4.1, use insertps to put the high elements into the low element.
4771 if (getSubtarget()->hasSSE41()) {
4773 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
4774 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
4776 Result = DAG.getUNDEF(VT);
4778 for (unsigned i = 1; i < NumElems; ++i) {
4779 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
4780 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
4781 Op.getOperand(i), DAG.getIntPtrConstant(i));
4786 // Otherwise, expand into a number of unpckl*, start by extending each of
4787 // our (non-undef) elements to the full vector width with the element in the
4788 // bottom slot of the vector (which generates no code for SSE).
4789 for (unsigned i = 0; i < NumElems; ++i) {
4790 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4791 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4793 V[i] = DAG.getUNDEF(VT);
4796 // Next, we iteratively mix elements, e.g. for v4f32:
4797 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4798 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4799 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4800 unsigned EltStride = NumElems >> 1;
4801 while (EltStride != 0) {
4802 for (unsigned i = 0; i < EltStride; ++i) {
4803 // If V[i+EltStride] is undef and this is the first round of mixing,
4804 // then it is safe to just drop this shuffle: V[i] is already in the
4805 // right place, the one element (since it's the first round) being
4806 // inserted as undef can be dropped. This isn't safe for successive
4807 // rounds because they will permute elements within both vectors.
4808 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
4809 EltStride == NumElems/2)
4812 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
4822 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4823 // We support concatenate two MMX registers and place them in a MMX
4824 // register. This is better than doing a stack convert.
4825 DebugLoc dl = Op.getDebugLoc();
4826 EVT ResVT = Op.getValueType();
4827 assert(Op.getNumOperands() == 2);
4828 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4829 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4831 SDValue InVec = DAG.getNode(ISD::BITCAST,dl, MVT::v1i64, Op.getOperand(0));
4832 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4833 InVec = Op.getOperand(1);
4834 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4835 unsigned NumElts = ResVT.getVectorNumElements();
4836 VecOp = DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
4837 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4838 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4840 InVec = DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, InVec);
4841 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4842 Mask[0] = 0; Mask[1] = 2;
4843 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4845 return DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
4848 // v8i16 shuffles - Prefer shuffles in the following order:
4849 // 1. [all] pshuflw, pshufhw, optional move
4850 // 2. [ssse3] 1 x pshufb
4851 // 3. [ssse3] 2 x pshufb + 1 x por
4852 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4854 X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
4855 SelectionDAG &DAG) const {
4856 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4857 SDValue V1 = SVOp->getOperand(0);
4858 SDValue V2 = SVOp->getOperand(1);
4859 DebugLoc dl = SVOp->getDebugLoc();
4860 SmallVector<int, 8> MaskVals;
4862 // Determine if more than 1 of the words in each of the low and high quadwords
4863 // of the result come from the same quadword of one of the two inputs. Undef
4864 // mask values count as coming from any quadword, for better codegen.
4865 SmallVector<unsigned, 4> LoQuad(4);
4866 SmallVector<unsigned, 4> HiQuad(4);
4867 BitVector InputQuads(4);
4868 for (unsigned i = 0; i < 8; ++i) {
4869 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4870 int EltIdx = SVOp->getMaskElt(i);
4871 MaskVals.push_back(EltIdx);
4880 InputQuads.set(EltIdx / 4);
4883 int BestLoQuad = -1;
4884 unsigned MaxQuad = 1;
4885 for (unsigned i = 0; i < 4; ++i) {
4886 if (LoQuad[i] > MaxQuad) {
4888 MaxQuad = LoQuad[i];
4892 int BestHiQuad = -1;
4894 for (unsigned i = 0; i < 4; ++i) {
4895 if (HiQuad[i] > MaxQuad) {
4897 MaxQuad = HiQuad[i];
4901 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4902 // of the two input vectors, shuffle them into one input vector so only a
4903 // single pshufb instruction is necessary. If There are more than 2 input
4904 // quads, disable the next transformation since it does not help SSSE3.
4905 bool V1Used = InputQuads[0] || InputQuads[1];
4906 bool V2Used = InputQuads[2] || InputQuads[3];
4907 if (Subtarget->hasSSSE3()) {
4908 if (InputQuads.count() == 2 && V1Used && V2Used) {
4909 BestLoQuad = InputQuads.find_first();
4910 BestHiQuad = InputQuads.find_next(BestLoQuad);
4912 if (InputQuads.count() > 2) {
4918 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4919 // the shuffle mask. If a quad is scored as -1, that means that it contains
4920 // words from all 4 input quadwords.
4922 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4923 SmallVector<int, 8> MaskV;
4924 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4925 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4926 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4927 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
4928 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
4929 NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
4931 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4932 // source words for the shuffle, to aid later transformations.
4933 bool AllWordsInNewV = true;
4934 bool InOrder[2] = { true, true };
4935 for (unsigned i = 0; i != 8; ++i) {
4936 int idx = MaskVals[i];
4938 InOrder[i/4] = false;
4939 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4941 AllWordsInNewV = false;
4945 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4946 if (AllWordsInNewV) {
4947 for (int i = 0; i != 8; ++i) {
4948 int idx = MaskVals[i];
4951 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4952 if ((idx != i) && idx < 4)
4954 if ((idx != i) && idx > 3)
4963 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4964 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4965 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4966 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
4967 unsigned TargetMask = 0;
4968 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4969 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4970 TargetMask = pshufhw ? X86::getShufflePSHUFHWImmediate(NewV.getNode()):
4971 X86::getShufflePSHUFLWImmediate(NewV.getNode());
4972 V1 = NewV.getOperand(0);
4973 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
4977 // If we have SSSE3, and all words of the result are from 1 input vector,
4978 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4979 // is present, fall back to case 4.
4980 if (Subtarget->hasSSSE3()) {
4981 SmallVector<SDValue,16> pshufbMask;
4983 // If we have elements from both input vectors, set the high bit of the
4984 // shuffle mask element to zero out elements that come from V2 in the V1
4985 // mask, and elements that come from V1 in the V2 mask, so that the two
4986 // results can be OR'd together.
4987 bool TwoInputs = V1Used && V2Used;
4988 for (unsigned i = 0; i != 8; ++i) {
4989 int EltIdx = MaskVals[i] * 2;
4990 if (TwoInputs && (EltIdx >= 16)) {
4991 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4992 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4995 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4996 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4998 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V1);
4999 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
5000 DAG.getNode(ISD::BUILD_VECTOR, dl,
5001 MVT::v16i8, &pshufbMask[0], 16));
5003 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5005 // Calculate the shuffle mask for the second input, shuffle it, and
5006 // OR it with the first shuffled input.
5008 for (unsigned i = 0; i != 8; ++i) {
5009 int EltIdx = MaskVals[i] * 2;
5011 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5012 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5015 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
5016 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
5018 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V2);
5019 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
5020 DAG.getNode(ISD::BUILD_VECTOR, dl,
5021 MVT::v16i8, &pshufbMask[0], 16));
5022 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
5023 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5026 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
5027 // and update MaskVals with new element order.
5028 BitVector InOrder(8);
5029 if (BestLoQuad >= 0) {
5030 SmallVector<int, 8> MaskV;
5031 for (int i = 0; i != 4; ++i) {
5032 int idx = MaskVals[i];
5034 MaskV.push_back(-1);
5036 } else if ((idx / 4) == BestLoQuad) {
5037 MaskV.push_back(idx & 3);
5040 MaskV.push_back(-1);
5043 for (unsigned i = 4; i != 8; ++i)
5045 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
5048 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
5049 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
5051 X86::getShufflePSHUFLWImmediate(NewV.getNode()),
5055 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
5056 // and update MaskVals with the new element order.
5057 if (BestHiQuad >= 0) {
5058 SmallVector<int, 8> MaskV;
5059 for (unsigned i = 0; i != 4; ++i)
5061 for (unsigned i = 4; i != 8; ++i) {
5062 int idx = MaskVals[i];
5064 MaskV.push_back(-1);
5066 } else if ((idx / 4) == BestHiQuad) {
5067 MaskV.push_back((idx & 3) + 4);
5070 MaskV.push_back(-1);
5073 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
5076 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
5077 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
5079 X86::getShufflePSHUFHWImmediate(NewV.getNode()),
5083 // In case BestHi & BestLo were both -1, which means each quadword has a word
5084 // from each of the four input quadwords, calculate the InOrder bitvector now
5085 // before falling through to the insert/extract cleanup.
5086 if (BestLoQuad == -1 && BestHiQuad == -1) {
5088 for (int i = 0; i != 8; ++i)
5089 if (MaskVals[i] < 0 || MaskVals[i] == i)
5093 // The other elements are put in the right place using pextrw and pinsrw.
5094 for (unsigned i = 0; i != 8; ++i) {
5097 int EltIdx = MaskVals[i];
5100 SDValue ExtOp = (EltIdx < 8)
5101 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
5102 DAG.getIntPtrConstant(EltIdx))
5103 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
5104 DAG.getIntPtrConstant(EltIdx - 8));
5105 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
5106 DAG.getIntPtrConstant(i));
5111 // v16i8 shuffles - Prefer shuffles in the following order:
5112 // 1. [ssse3] 1 x pshufb
5113 // 2. [ssse3] 2 x pshufb + 1 x por
5114 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
5116 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
5118 const X86TargetLowering &TLI) {
5119 SDValue V1 = SVOp->getOperand(0);
5120 SDValue V2 = SVOp->getOperand(1);
5121 DebugLoc dl = SVOp->getDebugLoc();
5122 SmallVector<int, 16> MaskVals;
5123 SVOp->getMask(MaskVals);
5125 // If we have SSSE3, case 1 is generated when all result bytes come from
5126 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
5127 // present, fall back to case 3.
5128 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
5131 for (unsigned i = 0; i < 16; ++i) {
5132 int EltIdx = MaskVals[i];
5141 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
5142 if (TLI.getSubtarget()->hasSSSE3()) {
5143 SmallVector<SDValue,16> pshufbMask;
5145 // If all result elements are from one input vector, then only translate
5146 // undef mask values to 0x80 (zero out result) in the pshufb mask.
5148 // Otherwise, we have elements from both input vectors, and must zero out
5149 // elements that come from V2 in the first mask, and V1 in the second mask
5150 // so that we can OR them together.
5151 bool TwoInputs = !(V1Only || V2Only);
5152 for (unsigned i = 0; i != 16; ++i) {
5153 int EltIdx = MaskVals[i];
5154 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
5155 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5158 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
5160 // If all the elements are from V2, assign it to V1 and return after
5161 // building the first pshufb.
5164 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
5165 DAG.getNode(ISD::BUILD_VECTOR, dl,
5166 MVT::v16i8, &pshufbMask[0], 16));
5170 // Calculate the shuffle mask for the second input, shuffle it, and
5171 // OR it with the first shuffled input.
5173 for (unsigned i = 0; i != 16; ++i) {
5174 int EltIdx = MaskVals[i];
5176 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5179 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
5181 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
5182 DAG.getNode(ISD::BUILD_VECTOR, dl,
5183 MVT::v16i8, &pshufbMask[0], 16));
5184 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
5187 // No SSSE3 - Calculate in place words and then fix all out of place words
5188 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
5189 // the 16 different words that comprise the two doublequadword input vectors.
5190 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5191 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
5192 SDValue NewV = V2Only ? V2 : V1;
5193 for (int i = 0; i != 8; ++i) {
5194 int Elt0 = MaskVals[i*2];
5195 int Elt1 = MaskVals[i*2+1];
5197 // This word of the result is all undef, skip it.
5198 if (Elt0 < 0 && Elt1 < 0)
5201 // This word of the result is already in the correct place, skip it.
5202 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
5204 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
5207 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
5208 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
5211 // If Elt0 and Elt1 are defined, are consecutive, and can be load
5212 // using a single extract together, load it and store it.
5213 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
5214 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
5215 DAG.getIntPtrConstant(Elt1 / 2));
5216 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5217 DAG.getIntPtrConstant(i));
5221 // If Elt1 is defined, extract it from the appropriate source. If the
5222 // source byte is not also odd, shift the extracted word left 8 bits
5223 // otherwise clear the bottom 8 bits if we need to do an or.
5225 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
5226 DAG.getIntPtrConstant(Elt1 / 2));
5227 if ((Elt1 & 1) == 0)
5228 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
5230 TLI.getShiftAmountTy(InsElt.getValueType())));
5232 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
5233 DAG.getConstant(0xFF00, MVT::i16));
5235 // If Elt0 is defined, extract it from the appropriate source. If the
5236 // source byte is not also even, shift the extracted word right 8 bits. If
5237 // Elt1 was also defined, OR the extracted values together before
5238 // inserting them in the result.
5240 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
5241 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
5242 if ((Elt0 & 1) != 0)
5243 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
5245 TLI.getShiftAmountTy(InsElt0.getValueType())));
5247 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
5248 DAG.getConstant(0x00FF, MVT::i16));
5249 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
5252 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5253 DAG.getIntPtrConstant(i));
5255 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
5258 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
5259 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
5260 /// done when every pair / quad of shuffle mask elements point to elements in
5261 /// the right sequence. e.g.
5262 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
5264 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
5265 SelectionDAG &DAG, DebugLoc dl) {
5266 EVT VT = SVOp->getValueType(0);
5267 SDValue V1 = SVOp->getOperand(0);
5268 SDValue V2 = SVOp->getOperand(1);
5269 unsigned NumElems = VT.getVectorNumElements();
5270 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
5272 switch (VT.getSimpleVT().SimpleTy) {
5273 default: assert(false && "Unexpected!");
5274 case MVT::v4f32: NewVT = MVT::v2f64; break;
5275 case MVT::v4i32: NewVT = MVT::v2i64; break;
5276 case MVT::v8i16: NewVT = MVT::v4i32; break;
5277 case MVT::v16i8: NewVT = MVT::v4i32; break;
5280 int Scale = NumElems / NewWidth;
5281 SmallVector<int, 8> MaskVec;
5282 for (unsigned i = 0; i < NumElems; i += Scale) {
5284 for (int j = 0; j < Scale; ++j) {
5285 int EltIdx = SVOp->getMaskElt(i+j);
5289 StartIdx = EltIdx - (EltIdx % Scale);
5290 if (EltIdx != StartIdx + j)
5294 MaskVec.push_back(-1);
5296 MaskVec.push_back(StartIdx / Scale);
5299 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
5300 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
5301 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
5304 /// getVZextMovL - Return a zero-extending vector move low node.
5306 static SDValue getVZextMovL(EVT VT, EVT OpVT,
5307 SDValue SrcOp, SelectionDAG &DAG,
5308 const X86Subtarget *Subtarget, DebugLoc dl) {
5309 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
5310 LoadSDNode *LD = NULL;
5311 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
5312 LD = dyn_cast<LoadSDNode>(SrcOp);
5314 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
5316 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
5317 if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
5318 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
5319 SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
5320 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
5322 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
5323 return DAG.getNode(ISD::BITCAST, dl, VT,
5324 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5325 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5333 return DAG.getNode(ISD::BITCAST, dl, VT,
5334 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5335 DAG.getNode(ISD::BITCAST, dl,
5339 /// LowerVECTOR_SHUFFLE_256 - Handle all 256-bit wide vectors shuffles
5340 /// which could not be matched by any known target speficic shuffle
5342 LowerVECTOR_SHUFFLE_256(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
5346 /// LowerVECTOR_SHUFFLE_128v4 - Handle all 128-bit wide vectors with
5347 /// 4 elements, and match them with several different shuffle types.
5349 LowerVECTOR_SHUFFLE_128v4(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
5350 SDValue V1 = SVOp->getOperand(0);
5351 SDValue V2 = SVOp->getOperand(1);
5352 DebugLoc dl = SVOp->getDebugLoc();
5353 EVT VT = SVOp->getValueType(0);
5355 assert(VT.getSizeInBits() == 128 && "Unsupported vector size");
5357 SmallVector<std::pair<int, int>, 8> Locs;
5359 SmallVector<int, 8> Mask1(4U, -1);
5360 SmallVector<int, 8> PermMask;
5361 SVOp->getMask(PermMask);
5365 for (unsigned i = 0; i != 4; ++i) {
5366 int Idx = PermMask[i];
5368 Locs[i] = std::make_pair(-1, -1);
5370 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
5372 Locs[i] = std::make_pair(0, NumLo);
5376 Locs[i] = std::make_pair(1, NumHi);
5378 Mask1[2+NumHi] = Idx;
5384 if (NumLo <= 2 && NumHi <= 2) {
5385 // If no more than two elements come from either vector. This can be
5386 // implemented with two shuffles. First shuffle gather the elements.
5387 // The second shuffle, which takes the first shuffle as both of its
5388 // vector operands, put the elements into the right order.
5389 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5391 SmallVector<int, 8> Mask2(4U, -1);
5393 for (unsigned i = 0; i != 4; ++i) {
5394 if (Locs[i].first == -1)
5397 unsigned Idx = (i < 2) ? 0 : 4;
5398 Idx += Locs[i].first * 2 + Locs[i].second;
5403 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
5404 } else if (NumLo == 3 || NumHi == 3) {
5405 // Otherwise, we must have three elements from one vector, call it X, and
5406 // one element from the other, call it Y. First, use a shufps to build an
5407 // intermediate vector with the one element from Y and the element from X
5408 // that will be in the same half in the final destination (the indexes don't
5409 // matter). Then, use a shufps to build the final vector, taking the half
5410 // containing the element from Y from the intermediate, and the other half
5413 // Normalize it so the 3 elements come from V1.
5414 CommuteVectorShuffleMask(PermMask, VT);
5418 // Find the element from V2.
5420 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
5421 int Val = PermMask[HiIndex];
5428 Mask1[0] = PermMask[HiIndex];
5430 Mask1[2] = PermMask[HiIndex^1];
5432 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5435 Mask1[0] = PermMask[0];
5436 Mask1[1] = PermMask[1];
5437 Mask1[2] = HiIndex & 1 ? 6 : 4;
5438 Mask1[3] = HiIndex & 1 ? 4 : 6;
5439 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5441 Mask1[0] = HiIndex & 1 ? 2 : 0;
5442 Mask1[1] = HiIndex & 1 ? 0 : 2;
5443 Mask1[2] = PermMask[2];
5444 Mask1[3] = PermMask[3];
5449 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
5453 // Break it into (shuffle shuffle_hi, shuffle_lo).
5456 SmallVector<int,8> LoMask(4U, -1);
5457 SmallVector<int,8> HiMask(4U, -1);
5459 SmallVector<int,8> *MaskPtr = &LoMask;
5460 unsigned MaskIdx = 0;
5463 for (unsigned i = 0; i != 4; ++i) {
5470 int Idx = PermMask[i];
5472 Locs[i] = std::make_pair(-1, -1);
5473 } else if (Idx < 4) {
5474 Locs[i] = std::make_pair(MaskIdx, LoIdx);
5475 (*MaskPtr)[LoIdx] = Idx;
5478 Locs[i] = std::make_pair(MaskIdx, HiIdx);
5479 (*MaskPtr)[HiIdx] = Idx;
5484 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
5485 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
5486 SmallVector<int, 8> MaskOps;
5487 for (unsigned i = 0; i != 4; ++i) {
5488 if (Locs[i].first == -1) {
5489 MaskOps.push_back(-1);
5491 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
5492 MaskOps.push_back(Idx);
5495 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
5498 static bool MayFoldVectorLoad(SDValue V) {
5499 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
5500 V = V.getOperand(0);
5501 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5502 V = V.getOperand(0);
5508 // FIXME: the version above should always be used. Since there's
5509 // a bug where several vector shuffles can't be folded because the
5510 // DAG is not updated during lowering and a node claims to have two
5511 // uses while it only has one, use this version, and let isel match
5512 // another instruction if the load really happens to have more than
5513 // one use. Remove this version after this bug get fixed.
5514 // rdar://8434668, PR8156
5515 static bool RelaxedMayFoldVectorLoad(SDValue V) {
5516 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
5517 V = V.getOperand(0);
5518 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5519 V = V.getOperand(0);
5520 if (ISD::isNormalLoad(V.getNode()))
5525 /// CanFoldShuffleIntoVExtract - Check if the current shuffle is used by
5526 /// a vector extract, and if both can be later optimized into a single load.
5527 /// This is done in visitEXTRACT_VECTOR_ELT and the conditions are checked
5528 /// here because otherwise a target specific shuffle node is going to be
5529 /// emitted for this shuffle, and the optimization not done.
5530 /// FIXME: This is probably not the best approach, but fix the problem
5531 /// until the right path is decided.
5533 bool CanXFormVExtractWithShuffleIntoLoad(SDValue V, SelectionDAG &DAG,
5534 const TargetLowering &TLI) {
5535 EVT VT = V.getValueType();
5536 ShuffleVectorSDNode *SVOp = dyn_cast<ShuffleVectorSDNode>(V);
5538 // Be sure that the vector shuffle is present in a pattern like this:
5539 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), c) -> (f32 load $addr)
5543 SDNode *N = *V.getNode()->use_begin();
5544 if (N->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5547 SDValue EltNo = N->getOperand(1);
5548 if (!isa<ConstantSDNode>(EltNo))
5551 // If the bit convert changed the number of elements, it is unsafe
5552 // to examine the mask.
5553 bool HasShuffleIntoBitcast = false;
5554 if (V.getOpcode() == ISD::BITCAST) {
5555 EVT SrcVT = V.getOperand(0).getValueType();
5556 if (SrcVT.getVectorNumElements() != VT.getVectorNumElements())
5558 V = V.getOperand(0);
5559 HasShuffleIntoBitcast = true;
5562 // Select the input vector, guarding against out of range extract vector.
5563 unsigned NumElems = VT.getVectorNumElements();
5564 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5565 int Idx = (Elt > NumElems) ? -1 : SVOp->getMaskElt(Elt);
5566 V = (Idx < (int)NumElems) ? V.getOperand(0) : V.getOperand(1);
5568 // Skip one more bit_convert if necessary
5569 if (V.getOpcode() == ISD::BITCAST)
5570 V = V.getOperand(0);
5572 if (ISD::isNormalLoad(V.getNode())) {
5573 // Is the original load suitable?
5574 LoadSDNode *LN0 = cast<LoadSDNode>(V);
5576 // FIXME: avoid the multi-use bug that is preventing lots of
5577 // of foldings to be detected, this is still wrong of course, but
5578 // give the temporary desired behavior, and if it happens that
5579 // the load has real more uses, during isel it will not fold, and
5580 // will generate poor code.
5581 if (!LN0 || LN0->isVolatile()) // || !LN0->hasOneUse()
5584 if (!HasShuffleIntoBitcast)
5587 // If there's a bitcast before the shuffle, check if the load type and
5588 // alignment is valid.
5589 unsigned Align = LN0->getAlignment();
5591 TLI.getTargetData()->getABITypeAlignment(
5592 VT.getTypeForEVT(*DAG.getContext()));
5594 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
5602 SDValue getMOVDDup(SDValue &Op, DebugLoc &dl, SDValue V1, SelectionDAG &DAG) {
5603 EVT VT = Op.getValueType();
5605 // Canonizalize to v2f64.
5606 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
5607 return DAG.getNode(ISD::BITCAST, dl, VT,
5608 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
5613 SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
5615 SDValue V1 = Op.getOperand(0);
5616 SDValue V2 = Op.getOperand(1);
5617 EVT VT = Op.getValueType();
5619 assert(VT != MVT::v2i64 && "unsupported shuffle type");
5621 if (HasSSE2 && VT == MVT::v2f64)
5622 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
5625 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V2, DAG);
5629 SDValue getMOVHighToLow(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG) {
5630 SDValue V1 = Op.getOperand(0);
5631 SDValue V2 = Op.getOperand(1);
5632 EVT VT = Op.getValueType();
5634 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
5635 "unsupported shuffle type");
5637 if (V2.getOpcode() == ISD::UNDEF)
5641 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
5645 SDValue getMOVLP(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
5646 SDValue V1 = Op.getOperand(0);
5647 SDValue V2 = Op.getOperand(1);
5648 EVT VT = Op.getValueType();
5649 unsigned NumElems = VT.getVectorNumElements();
5651 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
5652 // operand of these instructions is only memory, so check if there's a
5653 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
5655 bool CanFoldLoad = false;
5657 // Trivial case, when V2 comes from a load.
5658 if (MayFoldVectorLoad(V2))
5661 // When V1 is a load, it can be folded later into a store in isel, example:
5662 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
5664 // (MOVLPSmr addr:$src1, VR128:$src2)
5665 // So, recognize this potential and also use MOVLPS or MOVLPD
5666 if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
5669 // Both of them can't be memory operations though.
5670 if (MayFoldVectorLoad(V1) && MayFoldVectorLoad(V2))
5671 CanFoldLoad = false;
5674 if (HasSSE2 && NumElems == 2)
5675 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
5678 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
5681 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5682 // movl and movlp will both match v2i64, but v2i64 is never matched by
5683 // movl earlier because we make it strict to avoid messing with the movlp load
5684 // folding logic (see the code above getMOVLP call). Match it here then,
5685 // this is horrible, but will stay like this until we move all shuffle
5686 // matching to x86 specific nodes. Note that for the 1st condition all
5687 // types are matched with movsd.
5688 if ((HasSSE2 && NumElems == 2) || !X86::isMOVLMask(SVOp))
5689 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5691 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5694 assert(VT != MVT::v4i32 && "unsupported shuffle type");
5696 // Invert the operand order and use SHUFPS to match it.
5697 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V2, V1,
5698 X86::getShuffleSHUFImmediate(SVOp), DAG);
5701 static inline unsigned getUNPCKLOpcode(EVT VT, const X86Subtarget *Subtarget) {
5702 switch(VT.getSimpleVT().SimpleTy) {
5703 case MVT::v4i32: return X86ISD::PUNPCKLDQ;
5704 case MVT::v2i64: return X86ISD::PUNPCKLQDQ;
5705 case MVT::v4f32: return X86ISD::UNPCKLPS;
5706 case MVT::v2f64: return X86ISD::UNPCKLPD;
5707 case MVT::v8f32: return X86ISD::VUNPCKLPSY;
5708 case MVT::v4f64: return X86ISD::VUNPCKLPDY;
5709 case MVT::v16i8: return X86ISD::PUNPCKLBW;
5710 case MVT::v8i16: return X86ISD::PUNPCKLWD;
5712 llvm_unreachable("Unknown type for unpckl");
5717 static inline unsigned getUNPCKHOpcode(EVT VT) {
5718 switch(VT.getSimpleVT().SimpleTy) {
5719 case MVT::v4i32: return X86ISD::PUNPCKHDQ;
5720 case MVT::v2i64: return X86ISD::PUNPCKHQDQ;
5721 case MVT::v4f32: return X86ISD::UNPCKHPS;
5722 case MVT::v2f64: return X86ISD::UNPCKHPD;
5723 case MVT::v16i8: return X86ISD::PUNPCKHBW;
5724 case MVT::v8i16: return X86ISD::PUNPCKHWD;
5726 llvm_unreachable("Unknown type for unpckh");
5732 SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG,
5733 const TargetLowering &TLI,
5734 const X86Subtarget *Subtarget) {
5735 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5736 EVT VT = Op.getValueType();
5737 DebugLoc dl = Op.getDebugLoc();
5738 SDValue V1 = Op.getOperand(0);
5739 SDValue V2 = Op.getOperand(1);
5741 if (isZeroShuffle(SVOp))
5742 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
5744 // Handle splat operations
5745 if (SVOp->isSplat()) {
5746 unsigned NumElem = VT.getVectorNumElements();
5747 // Special case, this is the only place now where it's allowed to return
5748 // a vector_shuffle operation without using a target specific node, because
5749 // *hopefully* it will be optimized away by the dag combiner. FIXME: should
5750 // this be moved to DAGCombine instead?
5751 if (NumElem <= 4 && CanXFormVExtractWithShuffleIntoLoad(Op, DAG, TLI))
5754 // Handle splats by matching through known masks
5755 if ((VT.is128BitVector() && NumElem <= 4) ||
5756 (VT.is256BitVector() && NumElem <= 8))
5759 // All i16 and i8 vector types can't be used directly by a generic shuffle
5760 // instruction because the target has no such instruction. Generate shuffles
5761 // which repeat i16 and i8 several times until they fit in i32, and then can
5762 // be manipulated by target suported shuffles. After the insertion of the
5763 // necessary shuffles, the result is bitcasted back to v4f32 or v8f32.
5764 return PromoteSplat(SVOp, DAG);
5767 // If the shuffle can be profitably rewritten as a narrower shuffle, then
5769 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
5770 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5771 if (NewOp.getNode())
5772 return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
5773 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
5774 // FIXME: Figure out a cleaner way to do this.
5775 // Try to make use of movq to zero out the top part.
5776 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
5777 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5778 if (NewOp.getNode()) {
5779 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
5780 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
5781 DAG, Subtarget, dl);
5783 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
5784 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5785 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
5786 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
5787 DAG, Subtarget, dl);
5794 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
5795 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5796 SDValue V1 = Op.getOperand(0);
5797 SDValue V2 = Op.getOperand(1);
5798 EVT VT = Op.getValueType();
5799 DebugLoc dl = Op.getDebugLoc();
5800 unsigned NumElems = VT.getVectorNumElements();
5801 bool isMMX = VT.getSizeInBits() == 64;
5802 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
5803 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
5804 bool V1IsSplat = false;
5805 bool V2IsSplat = false;
5806 bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
5807 bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
5808 bool HasSSSE3 = Subtarget->hasSSSE3() || Subtarget->hasAVX();
5809 MachineFunction &MF = DAG.getMachineFunction();
5810 bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
5812 // Shuffle operations on MMX not supported.
5816 // Vector shuffle lowering takes 3 steps:
5818 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
5819 // narrowing and commutation of operands should be handled.
5820 // 2) Matching of shuffles with known shuffle masks to x86 target specific
5822 // 3) Rewriting of unmatched masks into new generic shuffle operations,
5823 // so the shuffle can be broken into other shuffles and the legalizer can
5824 // try the lowering again.
5826 // The general ideia is that no vector_shuffle operation should be left to
5827 // be matched during isel, all of them must be converted to a target specific
5830 // Normalize the input vectors. Here splats, zeroed vectors, profitable
5831 // narrowing and commutation of operands should be handled. The actual code
5832 // doesn't include all of those, work in progress...
5833 SDValue NewOp = NormalizeVectorShuffle(Op, DAG, *this, Subtarget);
5834 if (NewOp.getNode())
5837 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
5838 // unpckh_undef). Only use pshufd if speed is more important than size.
5839 if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp))
5840 return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()), dl, VT, V1, V1, DAG);
5841 if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp))
5842 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5844 if (X86::isMOVDDUPMask(SVOp) && HasSSE3 && V2IsUndef &&
5845 RelaxedMayFoldVectorLoad(V1))
5846 return getMOVDDup(Op, dl, V1, DAG);
5848 if (X86::isMOVHLPS_v_undef_Mask(SVOp))
5849 return getMOVHighToLow(Op, dl, DAG);
5851 // Use to match splats
5852 if (HasSSE2 && X86::isUNPCKHMask(SVOp) && V2IsUndef &&
5853 (VT == MVT::v2f64 || VT == MVT::v2i64))
5854 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5856 if (X86::isPSHUFDMask(SVOp)) {
5857 // The actual implementation will match the mask in the if above and then
5858 // during isel it can match several different instructions, not only pshufd
5859 // as its name says, sad but true, emulate the behavior for now...
5860 if (X86::isMOVDDUPMask(SVOp) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
5861 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
5863 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5865 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
5866 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
5868 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5869 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V1,
5872 if (VT == MVT::v4f32)
5873 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V1,
5877 // Check if this can be converted into a logical shift.
5878 bool isLeft = false;
5881 bool isShift = getSubtarget()->hasSSE2() &&
5882 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
5883 if (isShift && ShVal.hasOneUse()) {
5884 // If the shifted value has multiple uses, it may be cheaper to use
5885 // v_set0 + movlhps or movhlps, etc.
5886 EVT EltVT = VT.getVectorElementType();
5887 ShAmt *= EltVT.getSizeInBits();
5888 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5891 if (X86::isMOVLMask(SVOp)) {
5894 if (ISD::isBuildVectorAllZeros(V1.getNode()))
5895 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
5896 if (!X86::isMOVLPMask(SVOp)) {
5897 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5898 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5900 if (VT == MVT::v4i32 || VT == MVT::v4f32)
5901 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5905 // FIXME: fold these into legal mask.
5906 if (X86::isMOVLHPSMask(SVOp) && !X86::isUNPCKLMask(SVOp))
5907 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
5909 if (X86::isMOVHLPSMask(SVOp))
5910 return getMOVHighToLow(Op, dl, DAG);
5912 if (X86::isMOVSHDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5913 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
5915 if (X86::isMOVSLDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5916 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
5918 if (X86::isMOVLPMask(SVOp))
5919 return getMOVLP(Op, dl, DAG, HasSSE2);
5921 if (ShouldXformToMOVHLPS(SVOp) ||
5922 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
5923 return CommuteVectorShuffle(SVOp, DAG);
5926 // No better options. Use a vshl / vsrl.
5927 EVT EltVT = VT.getVectorElementType();
5928 ShAmt *= EltVT.getSizeInBits();
5929 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5932 bool Commuted = false;
5933 // FIXME: This should also accept a bitcast of a splat? Be careful, not
5934 // 1,1,1,1 -> v8i16 though.
5935 V1IsSplat = isSplatVector(V1.getNode());
5936 V2IsSplat = isSplatVector(V2.getNode());
5938 // Canonicalize the splat or undef, if present, to be on the RHS.
5939 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
5940 Op = CommuteVectorShuffle(SVOp, DAG);
5941 SVOp = cast<ShuffleVectorSDNode>(Op);
5942 V1 = SVOp->getOperand(0);
5943 V2 = SVOp->getOperand(1);
5944 std::swap(V1IsSplat, V2IsSplat);
5945 std::swap(V1IsUndef, V2IsUndef);
5949 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
5950 // Shuffling low element of v1 into undef, just return v1.
5953 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
5954 // the instruction selector will not match, so get a canonical MOVL with
5955 // swapped operands to undo the commute.
5956 return getMOVL(DAG, dl, VT, V2, V1);
5959 if (X86::isUNPCKLMask(SVOp))
5960 return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()),
5961 dl, VT, V1, V2, DAG);
5963 if (X86::isUNPCKHMask(SVOp))
5964 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V2, DAG);
5967 // Normalize mask so all entries that point to V2 points to its first
5968 // element then try to match unpck{h|l} again. If match, return a
5969 // new vector_shuffle with the corrected mask.
5970 SDValue NewMask = NormalizeMask(SVOp, DAG);
5971 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
5972 if (NSVOp != SVOp) {
5973 if (X86::isUNPCKLMask(NSVOp, true)) {
5975 } else if (X86::isUNPCKHMask(NSVOp, true)) {
5982 // Commute is back and try unpck* again.
5983 // FIXME: this seems wrong.
5984 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
5985 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
5987 if (X86::isUNPCKLMask(NewSVOp))
5988 return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()),
5989 dl, VT, V2, V1, DAG);
5991 if (X86::isUNPCKHMask(NewSVOp))
5992 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V2, V1, DAG);
5995 // Normalize the node to match x86 shuffle ops if needed
5996 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
5997 return CommuteVectorShuffle(SVOp, DAG);
5999 // The checks below are all present in isShuffleMaskLegal, but they are
6000 // inlined here right now to enable us to directly emit target specific
6001 // nodes, and remove one by one until they don't return Op anymore.
6002 SmallVector<int, 16> M;
6005 if (isPALIGNRMask(M, VT, HasSSSE3))
6006 return getTargetShuffleNode(X86ISD::PALIGN, dl, VT, V1, V2,
6007 X86::getShufflePALIGNRImmediate(SVOp),
6010 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
6011 SVOp->getSplatIndex() == 0 && V2IsUndef) {
6012 if (VT == MVT::v2f64)
6013 return getTargetShuffleNode(X86ISD::UNPCKLPD, dl, VT, V1, V1, DAG);
6014 if (VT == MVT::v2i64)
6015 return getTargetShuffleNode(X86ISD::PUNPCKLQDQ, dl, VT, V1, V1, DAG);
6018 if (isPSHUFHWMask(M, VT))
6019 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
6020 X86::getShufflePSHUFHWImmediate(SVOp),
6023 if (isPSHUFLWMask(M, VT))
6024 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
6025 X86::getShufflePSHUFLWImmediate(SVOp),
6028 if (isSHUFPMask(M, VT)) {
6029 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
6030 if (VT == MVT::v4f32 || VT == MVT::v4i32)
6031 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V2,
6033 if (VT == MVT::v2f64 || VT == MVT::v2i64)
6034 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V2,
6038 if (X86::isUNPCKL_v_undef_Mask(SVOp))
6039 return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()),
6040 dl, VT, V1, V1, DAG);
6041 if (X86::isUNPCKH_v_undef_Mask(SVOp))
6042 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
6044 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
6045 if (VT == MVT::v8i16) {
6046 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
6047 if (NewOp.getNode())
6051 if (VT == MVT::v16i8) {
6052 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
6053 if (NewOp.getNode())
6057 // Handle all 128-bit wide vectors with 4 elements, and match them with
6058 // several different shuffle types.
6059 if (NumElems == 4 && VT.getSizeInBits() == 128)
6060 return LowerVECTOR_SHUFFLE_128v4(SVOp, DAG);
6062 //===--------------------------------------------------------------------===//
6063 // Custom lower or generate target specific nodes for 256-bit shuffles.
6065 // Handle VPERMIL permutations
6066 if (isVPERMILMask(M, VT)) {
6067 unsigned TargetMask = getShuffleVPERMILImmediate(SVOp);
6068 if (VT == MVT::v8f32)
6069 return getTargetShuffleNode(X86ISD::VPERMIL, dl, VT, V1, TargetMask, DAG);
6072 // Handle general 256-bit shuffles
6073 if (VT.is256BitVector())
6074 return LowerVECTOR_SHUFFLE_256(SVOp, DAG);
6080 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
6081 SelectionDAG &DAG) const {
6082 EVT VT = Op.getValueType();
6083 DebugLoc dl = Op.getDebugLoc();
6084 if (VT.getSizeInBits() == 8) {
6085 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
6086 Op.getOperand(0), Op.getOperand(1));
6087 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
6088 DAG.getValueType(VT));
6089 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6090 } else if (VT.getSizeInBits() == 16) {
6091 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6092 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
6094 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
6095 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6096 DAG.getNode(ISD::BITCAST, dl,
6100 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
6101 Op.getOperand(0), Op.getOperand(1));
6102 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
6103 DAG.getValueType(VT));
6104 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6105 } else if (VT == MVT::f32) {
6106 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
6107 // the result back to FR32 register. It's only worth matching if the
6108 // result has a single use which is a store or a bitcast to i32. And in
6109 // the case of a store, it's not worth it if the index is a constant 0,
6110 // because a MOVSSmr can be used instead, which is smaller and faster.
6111 if (!Op.hasOneUse())
6113 SDNode *User = *Op.getNode()->use_begin();
6114 if ((User->getOpcode() != ISD::STORE ||
6115 (isa<ConstantSDNode>(Op.getOperand(1)) &&
6116 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
6117 (User->getOpcode() != ISD::BITCAST ||
6118 User->getValueType(0) != MVT::i32))
6120 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6121 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
6124 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
6125 } else if (VT == MVT::i32) {
6126 // ExtractPS works with constant index.
6127 if (isa<ConstantSDNode>(Op.getOperand(1)))
6135 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
6136 SelectionDAG &DAG) const {
6137 if (!isa<ConstantSDNode>(Op.getOperand(1)))
6140 SDValue Vec = Op.getOperand(0);
6141 EVT VecVT = Vec.getValueType();
6143 // If this is a 256-bit vector result, first extract the 128-bit
6144 // vector and then extract from the 128-bit vector.
6145 if (VecVT.getSizeInBits() > 128) {
6146 DebugLoc dl = Op.getNode()->getDebugLoc();
6147 unsigned NumElems = VecVT.getVectorNumElements();
6148 SDValue Idx = Op.getOperand(1);
6150 if (!isa<ConstantSDNode>(Idx))
6153 unsigned ExtractNumElems = NumElems / (VecVT.getSizeInBits() / 128);
6154 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
6156 // Get the 128-bit vector.
6157 bool Upper = IdxVal >= ExtractNumElems;
6158 Vec = Extract128BitVector(Vec, Idx, DAG, dl);
6161 SDValue ScaledIdx = Idx;
6163 ScaledIdx = DAG.getNode(ISD::SUB, dl, Idx.getValueType(), Idx,
6164 DAG.getConstant(ExtractNumElems,
6165 Idx.getValueType()));
6166 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
6170 assert(Vec.getValueSizeInBits() <= 128 && "Unexpected vector length");
6172 if (Subtarget->hasSSE41()) {
6173 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
6178 EVT VT = Op.getValueType();
6179 DebugLoc dl = Op.getDebugLoc();
6180 // TODO: handle v16i8.
6181 if (VT.getSizeInBits() == 16) {
6182 SDValue Vec = Op.getOperand(0);
6183 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6185 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
6186 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6187 DAG.getNode(ISD::BITCAST, dl,
6190 // Transform it so it match pextrw which produces a 32-bit result.
6191 EVT EltVT = MVT::i32;
6192 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
6193 Op.getOperand(0), Op.getOperand(1));
6194 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
6195 DAG.getValueType(VT));
6196 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6197 } else if (VT.getSizeInBits() == 32) {
6198 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6202 // SHUFPS the element to the lowest double word, then movss.
6203 int Mask[4] = { Idx, -1, -1, -1 };
6204 EVT VVT = Op.getOperand(0).getValueType();
6205 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
6206 DAG.getUNDEF(VVT), Mask);
6207 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
6208 DAG.getIntPtrConstant(0));
6209 } else if (VT.getSizeInBits() == 64) {
6210 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
6211 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
6212 // to match extract_elt for f64.
6213 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6217 // UNPCKHPD the element to the lowest double word, then movsd.
6218 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
6219 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
6220 int Mask[2] = { 1, -1 };
6221 EVT VVT = Op.getOperand(0).getValueType();
6222 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
6223 DAG.getUNDEF(VVT), Mask);
6224 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
6225 DAG.getIntPtrConstant(0));
6232 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
6233 SelectionDAG &DAG) const {
6234 EVT VT = Op.getValueType();
6235 EVT EltVT = VT.getVectorElementType();
6236 DebugLoc dl = Op.getDebugLoc();
6238 SDValue N0 = Op.getOperand(0);
6239 SDValue N1 = Op.getOperand(1);
6240 SDValue N2 = Op.getOperand(2);
6242 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
6243 isa<ConstantSDNode>(N2)) {
6245 if (VT == MVT::v8i16)
6246 Opc = X86ISD::PINSRW;
6247 else if (VT == MVT::v16i8)
6248 Opc = X86ISD::PINSRB;
6250 Opc = X86ISD::PINSRB;
6252 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
6254 if (N1.getValueType() != MVT::i32)
6255 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
6256 if (N2.getValueType() != MVT::i32)
6257 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
6258 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
6259 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
6260 // Bits [7:6] of the constant are the source select. This will always be
6261 // zero here. The DAG Combiner may combine an extract_elt index into these
6262 // bits. For example (insert (extract, 3), 2) could be matched by putting
6263 // the '3' into bits [7:6] of X86ISD::INSERTPS.
6264 // Bits [5:4] of the constant are the destination select. This is the
6265 // value of the incoming immediate.
6266 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
6267 // combine either bitwise AND or insert of float 0.0 to set these bits.
6268 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
6269 // Create this as a scalar to vector..
6270 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
6271 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
6272 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
6273 // PINSR* works with constant index.
6280 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
6281 EVT VT = Op.getValueType();
6282 EVT EltVT = VT.getVectorElementType();
6284 DebugLoc dl = Op.getDebugLoc();
6285 SDValue N0 = Op.getOperand(0);
6286 SDValue N1 = Op.getOperand(1);
6287 SDValue N2 = Op.getOperand(2);
6289 // If this is a 256-bit vector result, first insert into a 128-bit
6290 // vector and then insert into the 256-bit vector.
6291 if (VT.getSizeInBits() > 128) {
6292 if (!isa<ConstantSDNode>(N2))
6295 // Get the 128-bit vector.
6296 unsigned NumElems = VT.getVectorNumElements();
6297 unsigned IdxVal = cast<ConstantSDNode>(N2)->getZExtValue();
6298 bool Upper = IdxVal >= NumElems / 2;
6300 SDValue SubN0 = Extract128BitVector(N0, N2, DAG, dl);
6303 SDValue ScaledN2 = N2;
6305 ScaledN2 = DAG.getNode(ISD::SUB, dl, N2.getValueType(), N2,
6306 DAG.getConstant(NumElems /
6307 (VT.getSizeInBits() / 128),
6308 N2.getValueType()));
6309 Op = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, SubN0.getValueType(), SubN0,
6312 // Insert the 128-bit vector
6313 // FIXME: Why UNDEF?
6314 return Insert128BitVector(N0, Op, N2, DAG, dl);
6317 if (Subtarget->hasSSE41())
6318 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
6320 if (EltVT == MVT::i8)
6323 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
6324 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
6325 // as its second argument.
6326 if (N1.getValueType() != MVT::i32)
6327 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
6328 if (N2.getValueType() != MVT::i32)
6329 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
6330 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
6336 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6337 DebugLoc dl = Op.getDebugLoc();
6338 EVT OpVT = Op.getValueType();
6340 if (Op.getValueType() == MVT::v1i64 &&
6341 Op.getOperand(0).getValueType() == MVT::i64)
6342 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
6344 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
6345 assert(Op.getValueType().getSimpleVT().getSizeInBits() == 128 &&
6346 "Expected an SSE type!");
6347 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(),
6348 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
6351 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
6352 // a simple subregister reference or explicit instructions to grab
6353 // upper bits of a vector.
6355 X86TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
6356 if (Subtarget->hasAVX()) {
6357 DebugLoc dl = Op.getNode()->getDebugLoc();
6358 SDValue Vec = Op.getNode()->getOperand(0);
6359 SDValue Idx = Op.getNode()->getOperand(1);
6361 if (Op.getNode()->getValueType(0).getSizeInBits() == 128
6362 && Vec.getNode()->getValueType(0).getSizeInBits() == 256) {
6363 return Extract128BitVector(Vec, Idx, DAG, dl);
6369 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
6370 // simple superregister reference or explicit instructions to insert
6371 // the upper bits of a vector.
6373 X86TargetLowering::LowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
6374 if (Subtarget->hasAVX()) {
6375 DebugLoc dl = Op.getNode()->getDebugLoc();
6376 SDValue Vec = Op.getNode()->getOperand(0);
6377 SDValue SubVec = Op.getNode()->getOperand(1);
6378 SDValue Idx = Op.getNode()->getOperand(2);
6380 if (Op.getNode()->getValueType(0).getSizeInBits() == 256
6381 && SubVec.getNode()->getValueType(0).getSizeInBits() == 128) {
6382 return Insert128BitVector(Vec, SubVec, Idx, DAG, dl);
6388 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
6389 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
6390 // one of the above mentioned nodes. It has to be wrapped because otherwise
6391 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
6392 // be used to form addressing mode. These wrapped nodes will be selected
6395 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
6396 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
6398 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6400 unsigned char OpFlag = 0;
6401 unsigned WrapperKind = X86ISD::Wrapper;
6402 CodeModel::Model M = getTargetMachine().getCodeModel();
6404 if (Subtarget->isPICStyleRIPRel() &&
6405 (M == CodeModel::Small || M == CodeModel::Kernel))
6406 WrapperKind = X86ISD::WrapperRIP;
6407 else if (Subtarget->isPICStyleGOT())
6408 OpFlag = X86II::MO_GOTOFF;
6409 else if (Subtarget->isPICStyleStubPIC())
6410 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6412 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
6414 CP->getOffset(), OpFlag);
6415 DebugLoc DL = CP->getDebugLoc();
6416 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6417 // With PIC, the address is actually $g + Offset.
6419 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6420 DAG.getNode(X86ISD::GlobalBaseReg,
6421 DebugLoc(), getPointerTy()),
6428 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
6429 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
6431 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6433 unsigned char OpFlag = 0;
6434 unsigned WrapperKind = X86ISD::Wrapper;
6435 CodeModel::Model M = getTargetMachine().getCodeModel();
6437 if (Subtarget->isPICStyleRIPRel() &&
6438 (M == CodeModel::Small || M == CodeModel::Kernel))
6439 WrapperKind = X86ISD::WrapperRIP;
6440 else if (Subtarget->isPICStyleGOT())
6441 OpFlag = X86II::MO_GOTOFF;
6442 else if (Subtarget->isPICStyleStubPIC())
6443 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6445 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
6447 DebugLoc DL = JT->getDebugLoc();
6448 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6450 // With PIC, the address is actually $g + Offset.
6452 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6453 DAG.getNode(X86ISD::GlobalBaseReg,
6454 DebugLoc(), getPointerTy()),
6461 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
6462 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
6464 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6466 unsigned char OpFlag = 0;
6467 unsigned WrapperKind = X86ISD::Wrapper;
6468 CodeModel::Model M = getTargetMachine().getCodeModel();
6470 if (Subtarget->isPICStyleRIPRel() &&
6471 (M == CodeModel::Small || M == CodeModel::Kernel))
6472 WrapperKind = X86ISD::WrapperRIP;
6473 else if (Subtarget->isPICStyleGOT())
6474 OpFlag = X86II::MO_GOTOFF;
6475 else if (Subtarget->isPICStyleStubPIC())
6476 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6478 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
6480 DebugLoc DL = Op.getDebugLoc();
6481 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6484 // With PIC, the address is actually $g + Offset.
6485 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
6486 !Subtarget->is64Bit()) {
6487 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6488 DAG.getNode(X86ISD::GlobalBaseReg,
6489 DebugLoc(), getPointerTy()),
6497 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
6498 // Create the TargetBlockAddressAddress node.
6499 unsigned char OpFlags =
6500 Subtarget->ClassifyBlockAddressReference();
6501 CodeModel::Model M = getTargetMachine().getCodeModel();
6502 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
6503 DebugLoc dl = Op.getDebugLoc();
6504 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
6505 /*isTarget=*/true, OpFlags);
6507 if (Subtarget->isPICStyleRIPRel() &&
6508 (M == CodeModel::Small || M == CodeModel::Kernel))
6509 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6511 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6513 // With PIC, the address is actually $g + Offset.
6514 if (isGlobalRelativeToPICBase(OpFlags)) {
6515 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6516 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6524 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
6526 SelectionDAG &DAG) const {
6527 // Create the TargetGlobalAddress node, folding in the constant
6528 // offset if it is legal.
6529 unsigned char OpFlags =
6530 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
6531 CodeModel::Model M = getTargetMachine().getCodeModel();
6533 if (OpFlags == X86II::MO_NO_FLAG &&
6534 X86::isOffsetSuitableForCodeModel(Offset, M)) {
6535 // A direct static reference to a global.
6536 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
6539 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
6542 if (Subtarget->isPICStyleRIPRel() &&
6543 (M == CodeModel::Small || M == CodeModel::Kernel))
6544 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6546 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6548 // With PIC, the address is actually $g + Offset.
6549 if (isGlobalRelativeToPICBase(OpFlags)) {
6550 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6551 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6555 // For globals that require a load from a stub to get the address, emit the
6557 if (isGlobalStubReference(OpFlags))
6558 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
6559 MachinePointerInfo::getGOT(), false, false, 0);
6561 // If there was a non-zero offset that we didn't fold, create an explicit
6564 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
6565 DAG.getConstant(Offset, getPointerTy()));
6571 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
6572 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
6573 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
6574 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
6578 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
6579 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
6580 unsigned char OperandFlags) {
6581 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6582 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6583 DebugLoc dl = GA->getDebugLoc();
6584 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6585 GA->getValueType(0),
6589 SDValue Ops[] = { Chain, TGA, *InFlag };
6590 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
6592 SDValue Ops[] = { Chain, TGA };
6593 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
6596 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
6597 MFI->setAdjustsStack(true);
6599 SDValue Flag = Chain.getValue(1);
6600 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
6603 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
6605 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6608 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
6609 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
6610 DAG.getNode(X86ISD::GlobalBaseReg,
6611 DebugLoc(), PtrVT), InFlag);
6612 InFlag = Chain.getValue(1);
6614 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
6617 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
6619 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6621 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
6622 X86::RAX, X86II::MO_TLSGD);
6625 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
6626 // "local exec" model.
6627 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6628 const EVT PtrVT, TLSModel::Model model,
6630 DebugLoc dl = GA->getDebugLoc();
6632 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
6633 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
6634 is64Bit ? 257 : 256));
6636 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
6637 DAG.getIntPtrConstant(0),
6638 MachinePointerInfo(Ptr), false, false, 0);
6640 unsigned char OperandFlags = 0;
6641 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
6643 unsigned WrapperKind = X86ISD::Wrapper;
6644 if (model == TLSModel::LocalExec) {
6645 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
6646 } else if (is64Bit) {
6647 assert(model == TLSModel::InitialExec);
6648 OperandFlags = X86II::MO_GOTTPOFF;
6649 WrapperKind = X86ISD::WrapperRIP;
6651 assert(model == TLSModel::InitialExec);
6652 OperandFlags = X86II::MO_INDNTPOFF;
6655 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
6657 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6658 GA->getValueType(0),
6659 GA->getOffset(), OperandFlags);
6660 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
6662 if (model == TLSModel::InitialExec)
6663 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
6664 MachinePointerInfo::getGOT(), false, false, 0);
6666 // The address of the thread local variable is the add of the thread
6667 // pointer with the offset of the variable.
6668 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
6672 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
6674 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
6675 const GlobalValue *GV = GA->getGlobal();
6677 if (Subtarget->isTargetELF()) {
6678 // TODO: implement the "local dynamic" model
6679 // TODO: implement the "initial exec"model for pic executables
6681 // If GV is an alias then use the aliasee for determining
6682 // thread-localness.
6683 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
6684 GV = GA->resolveAliasedGlobal(false);
6686 TLSModel::Model model
6687 = getTLSModel(GV, getTargetMachine().getRelocationModel());
6690 case TLSModel::GeneralDynamic:
6691 case TLSModel::LocalDynamic: // not implemented
6692 if (Subtarget->is64Bit())
6693 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
6694 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
6696 case TLSModel::InitialExec:
6697 case TLSModel::LocalExec:
6698 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
6699 Subtarget->is64Bit());
6701 } else if (Subtarget->isTargetDarwin()) {
6702 // Darwin only has one model of TLS. Lower to that.
6703 unsigned char OpFlag = 0;
6704 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
6705 X86ISD::WrapperRIP : X86ISD::Wrapper;
6707 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6709 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
6710 !Subtarget->is64Bit();
6712 OpFlag = X86II::MO_TLVP_PIC_BASE;
6714 OpFlag = X86II::MO_TLVP;
6715 DebugLoc DL = Op.getDebugLoc();
6716 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
6717 GA->getValueType(0),
6718 GA->getOffset(), OpFlag);
6719 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6721 // With PIC32, the address is actually $g + Offset.
6723 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6724 DAG.getNode(X86ISD::GlobalBaseReg,
6725 DebugLoc(), getPointerTy()),
6728 // Lowering the machine isd will make sure everything is in the right
6730 SDValue Chain = DAG.getEntryNode();
6731 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6732 SDValue Args[] = { Chain, Offset };
6733 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args, 2);
6735 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
6736 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6737 MFI->setAdjustsStack(true);
6739 // And our return value (tls address) is in the standard call return value
6741 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
6742 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
6746 "TLS not implemented for this target.");
6748 llvm_unreachable("Unreachable");
6753 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values and
6754 /// take a 2 x i32 value to shift plus a shift amount.
6755 SDValue X86TargetLowering::LowerShiftParts(SDValue Op, SelectionDAG &DAG) const {
6756 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
6757 EVT VT = Op.getValueType();
6758 unsigned VTBits = VT.getSizeInBits();
6759 DebugLoc dl = Op.getDebugLoc();
6760 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
6761 SDValue ShOpLo = Op.getOperand(0);
6762 SDValue ShOpHi = Op.getOperand(1);
6763 SDValue ShAmt = Op.getOperand(2);
6764 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
6765 DAG.getConstant(VTBits - 1, MVT::i8))
6766 : DAG.getConstant(0, VT);
6769 if (Op.getOpcode() == ISD::SHL_PARTS) {
6770 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
6771 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
6773 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
6774 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
6777 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
6778 DAG.getConstant(VTBits, MVT::i8));
6779 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6780 AndNode, DAG.getConstant(0, MVT::i8));
6783 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6784 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
6785 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
6787 if (Op.getOpcode() == ISD::SHL_PARTS) {
6788 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6789 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6791 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6792 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6795 SDValue Ops[2] = { Lo, Hi };
6796 return DAG.getMergeValues(Ops, 2, dl);
6799 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
6800 SelectionDAG &DAG) const {
6801 EVT SrcVT = Op.getOperand(0).getValueType();
6803 if (SrcVT.isVector())
6806 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
6807 "Unknown SINT_TO_FP to lower!");
6809 // These are really Legal; return the operand so the caller accepts it as
6811 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
6813 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
6814 Subtarget->is64Bit()) {
6818 DebugLoc dl = Op.getDebugLoc();
6819 unsigned Size = SrcVT.getSizeInBits()/8;
6820 MachineFunction &MF = DAG.getMachineFunction();
6821 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
6822 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6823 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6825 MachinePointerInfo::getFixedStack(SSFI),
6827 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
6830 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
6832 SelectionDAG &DAG) const {
6834 DebugLoc DL = Op.getDebugLoc();
6836 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
6838 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
6840 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
6842 unsigned ByteSize = SrcVT.getSizeInBits()/8;
6844 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
6845 MachineMemOperand *MMO;
6847 int SSFI = FI->getIndex();
6849 DAG.getMachineFunction()
6850 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6851 MachineMemOperand::MOLoad, ByteSize, ByteSize);
6853 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
6854 StackSlot = StackSlot.getOperand(1);
6856 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
6857 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
6859 Tys, Ops, array_lengthof(Ops),
6863 Chain = Result.getValue(1);
6864 SDValue InFlag = Result.getValue(2);
6866 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
6867 // shouldn't be necessary except that RFP cannot be live across
6868 // multiple blocks. When stackifier is fixed, they can be uncoupled.
6869 MachineFunction &MF = DAG.getMachineFunction();
6870 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
6871 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
6872 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6873 Tys = DAG.getVTList(MVT::Other);
6875 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
6877 MachineMemOperand *MMO =
6878 DAG.getMachineFunction()
6879 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6880 MachineMemOperand::MOStore, SSFISize, SSFISize);
6882 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
6883 Ops, array_lengthof(Ops),
6884 Op.getValueType(), MMO);
6885 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
6886 MachinePointerInfo::getFixedStack(SSFI),
6893 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
6894 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
6895 SelectionDAG &DAG) const {
6896 // This algorithm is not obvious. Here it is in C code, more or less:
6898 double uint64_to_double( uint32_t hi, uint32_t lo ) {
6899 static const __m128i exp = { 0x4330000045300000ULL, 0 };
6900 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
6902 // Copy ints to xmm registers.
6903 __m128i xh = _mm_cvtsi32_si128( hi );
6904 __m128i xl = _mm_cvtsi32_si128( lo );
6906 // Combine into low half of a single xmm register.
6907 __m128i x = _mm_unpacklo_epi32( xh, xl );
6911 // Merge in appropriate exponents to give the integer bits the right
6913 x = _mm_unpacklo_epi32( x, exp );
6915 // Subtract away the biases to deal with the IEEE-754 double precision
6917 d = _mm_sub_pd( (__m128d) x, bias );
6919 // All conversions up to here are exact. The correctly rounded result is
6920 // calculated using the current rounding mode using the following
6922 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
6923 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
6924 // store doesn't really need to be here (except
6925 // maybe to zero the other double)
6930 DebugLoc dl = Op.getDebugLoc();
6931 LLVMContext *Context = DAG.getContext();
6933 // Build some magic constants.
6934 std::vector<Constant*> CV0;
6935 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
6936 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
6937 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6938 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6939 Constant *C0 = ConstantVector::get(CV0);
6940 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
6942 std::vector<Constant*> CV1;
6944 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
6946 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
6947 Constant *C1 = ConstantVector::get(CV1);
6948 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
6950 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6951 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6953 DAG.getIntPtrConstant(1)));
6954 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6955 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6957 DAG.getIntPtrConstant(0)));
6958 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
6959 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
6960 MachinePointerInfo::getConstantPool(),
6962 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
6963 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck2);
6964 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
6965 MachinePointerInfo::getConstantPool(),
6967 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
6969 // Add the halves; easiest way is to swap them into another reg first.
6970 int ShufMask[2] = { 1, -1 };
6971 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
6972 DAG.getUNDEF(MVT::v2f64), ShufMask);
6973 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
6974 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
6975 DAG.getIntPtrConstant(0));
6978 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
6979 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
6980 SelectionDAG &DAG) const {
6981 DebugLoc dl = Op.getDebugLoc();
6982 // FP constant to bias correct the final result.
6983 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
6986 // Load the 32-bit value into an XMM register.
6987 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6988 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6990 DAG.getIntPtrConstant(0)));
6992 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6993 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
6994 DAG.getIntPtrConstant(0));
6996 // Or the load with the bias.
6997 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
6998 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
6999 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
7001 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7002 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
7003 MVT::v2f64, Bias)));
7004 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7005 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
7006 DAG.getIntPtrConstant(0));
7008 // Subtract the bias.
7009 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
7011 // Handle final rounding.
7012 EVT DestVT = Op.getValueType();
7014 if (DestVT.bitsLT(MVT::f64)) {
7015 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
7016 DAG.getIntPtrConstant(0));
7017 } else if (DestVT.bitsGT(MVT::f64)) {
7018 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
7021 // Handle final rounding.
7025 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
7026 SelectionDAG &DAG) const {
7027 SDValue N0 = Op.getOperand(0);
7028 DebugLoc dl = Op.getDebugLoc();
7030 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
7031 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
7032 // the optimization here.
7033 if (DAG.SignBitIsZero(N0))
7034 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
7036 EVT SrcVT = N0.getValueType();
7037 EVT DstVT = Op.getValueType();
7038 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
7039 return LowerUINT_TO_FP_i64(Op, DAG);
7040 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
7041 return LowerUINT_TO_FP_i32(Op, DAG);
7043 // Make a 64-bit buffer, and use it to build an FILD.
7044 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
7045 if (SrcVT == MVT::i32) {
7046 SDValue WordOff = DAG.getConstant(4, getPointerTy());
7047 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
7048 getPointerTy(), StackSlot, WordOff);
7049 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7050 StackSlot, MachinePointerInfo(),
7052 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
7053 OffsetSlot, MachinePointerInfo(),
7055 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
7059 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
7060 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7061 StackSlot, MachinePointerInfo(),
7063 // For i64 source, we need to add the appropriate power of 2 if the input
7064 // was negative. This is the same as the optimization in
7065 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
7066 // we must be careful to do the computation in x87 extended precision, not
7067 // in SSE. (The generic code can't know it's OK to do this, or how to.)
7068 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
7069 MachineMemOperand *MMO =
7070 DAG.getMachineFunction()
7071 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7072 MachineMemOperand::MOLoad, 8, 8);
7074 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
7075 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
7076 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops, 3,
7079 APInt FF(32, 0x5F800000ULL);
7081 // Check whether the sign bit is set.
7082 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
7083 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
7086 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
7087 SDValue FudgePtr = DAG.getConstantPool(
7088 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
7091 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
7092 SDValue Zero = DAG.getIntPtrConstant(0);
7093 SDValue Four = DAG.getIntPtrConstant(4);
7094 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
7096 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
7098 // Load the value out, extending it from f32 to f80.
7099 // FIXME: Avoid the extend by constructing the right constant pool?
7100 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
7101 FudgePtr, MachinePointerInfo::getConstantPool(),
7102 MVT::f32, false, false, 4);
7103 // Extend everything to 80 bits to force it to be done on x87.
7104 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
7105 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
7108 std::pair<SDValue,SDValue> X86TargetLowering::
7109 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
7110 DebugLoc DL = Op.getDebugLoc();
7112 EVT DstTy = Op.getValueType();
7115 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
7119 assert(DstTy.getSimpleVT() <= MVT::i64 &&
7120 DstTy.getSimpleVT() >= MVT::i16 &&
7121 "Unknown FP_TO_SINT to lower!");
7123 // These are really Legal.
7124 if (DstTy == MVT::i32 &&
7125 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
7126 return std::make_pair(SDValue(), SDValue());
7127 if (Subtarget->is64Bit() &&
7128 DstTy == MVT::i64 &&
7129 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
7130 return std::make_pair(SDValue(), SDValue());
7132 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
7134 MachineFunction &MF = DAG.getMachineFunction();
7135 unsigned MemSize = DstTy.getSizeInBits()/8;
7136 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
7137 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7142 switch (DstTy.getSimpleVT().SimpleTy) {
7143 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
7144 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
7145 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
7146 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
7149 SDValue Chain = DAG.getEntryNode();
7150 SDValue Value = Op.getOperand(0);
7151 EVT TheVT = Op.getOperand(0).getValueType();
7152 if (isScalarFPTypeInSSEReg(TheVT)) {
7153 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
7154 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
7155 MachinePointerInfo::getFixedStack(SSFI),
7157 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
7159 Chain, StackSlot, DAG.getValueType(TheVT)
7162 MachineMemOperand *MMO =
7163 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7164 MachineMemOperand::MOLoad, MemSize, MemSize);
7165 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, 3,
7167 Chain = Value.getValue(1);
7168 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
7169 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7172 MachineMemOperand *MMO =
7173 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7174 MachineMemOperand::MOStore, MemSize, MemSize);
7176 // Build the FP_TO_INT*_IN_MEM
7177 SDValue Ops[] = { Chain, Value, StackSlot };
7178 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
7179 Ops, 3, DstTy, MMO);
7181 return std::make_pair(FIST, StackSlot);
7184 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
7185 SelectionDAG &DAG) const {
7186 if (Op.getValueType().isVector())
7189 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
7190 SDValue FIST = Vals.first, StackSlot = Vals.second;
7191 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
7192 if (FIST.getNode() == 0) return Op;
7195 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
7196 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
7199 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
7200 SelectionDAG &DAG) const {
7201 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
7202 SDValue FIST = Vals.first, StackSlot = Vals.second;
7203 assert(FIST.getNode() && "Unexpected failure");
7206 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
7207 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
7210 SDValue X86TargetLowering::LowerFABS(SDValue Op,
7211 SelectionDAG &DAG) const {
7212 LLVMContext *Context = DAG.getContext();
7213 DebugLoc dl = Op.getDebugLoc();
7214 EVT VT = Op.getValueType();
7217 EltVT = VT.getVectorElementType();
7218 std::vector<Constant*> CV;
7219 if (EltVT == MVT::f64) {
7220 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
7224 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
7230 Constant *C = ConstantVector::get(CV);
7231 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7232 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7233 MachinePointerInfo::getConstantPool(),
7235 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
7238 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
7239 LLVMContext *Context = DAG.getContext();
7240 DebugLoc dl = Op.getDebugLoc();
7241 EVT VT = Op.getValueType();
7244 EltVT = VT.getVectorElementType();
7245 std::vector<Constant*> CV;
7246 if (EltVT == MVT::f64) {
7247 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
7251 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
7257 Constant *C = ConstantVector::get(CV);
7258 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7259 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7260 MachinePointerInfo::getConstantPool(),
7262 if (VT.isVector()) {
7263 return DAG.getNode(ISD::BITCAST, dl, VT,
7264 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
7265 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7267 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, Mask)));
7269 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
7273 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
7274 LLVMContext *Context = DAG.getContext();
7275 SDValue Op0 = Op.getOperand(0);
7276 SDValue Op1 = Op.getOperand(1);
7277 DebugLoc dl = Op.getDebugLoc();
7278 EVT VT = Op.getValueType();
7279 EVT SrcVT = Op1.getValueType();
7281 // If second operand is smaller, extend it first.
7282 if (SrcVT.bitsLT(VT)) {
7283 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
7286 // And if it is bigger, shrink it first.
7287 if (SrcVT.bitsGT(VT)) {
7288 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
7292 // At this point the operands and the result should have the same
7293 // type, and that won't be f80 since that is not custom lowered.
7295 // First get the sign bit of second operand.
7296 std::vector<Constant*> CV;
7297 if (SrcVT == MVT::f64) {
7298 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
7299 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
7301 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
7302 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7303 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7304 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7306 Constant *C = ConstantVector::get(CV);
7307 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7308 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
7309 MachinePointerInfo::getConstantPool(),
7311 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
7313 // Shift sign bit right or left if the two operands have different types.
7314 if (SrcVT.bitsGT(VT)) {
7315 // Op0 is MVT::f32, Op1 is MVT::f64.
7316 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
7317 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
7318 DAG.getConstant(32, MVT::i32));
7319 SignBit = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, SignBit);
7320 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
7321 DAG.getIntPtrConstant(0));
7324 // Clear first operand sign bit.
7326 if (VT == MVT::f64) {
7327 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
7328 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
7330 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
7331 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7332 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7333 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7335 C = ConstantVector::get(CV);
7336 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7337 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7338 MachinePointerInfo::getConstantPool(),
7340 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
7342 // Or the value with the sign bit.
7343 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
7346 SDValue X86TargetLowering::LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) const {
7347 SDValue N0 = Op.getOperand(0);
7348 DebugLoc dl = Op.getDebugLoc();
7349 EVT VT = Op.getValueType();
7351 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
7352 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
7353 DAG.getConstant(1, VT));
7354 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT));
7357 /// Emit nodes that will be selected as "test Op0,Op0", or something
7359 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
7360 SelectionDAG &DAG) const {
7361 DebugLoc dl = Op.getDebugLoc();
7363 // CF and OF aren't always set the way we want. Determine which
7364 // of these we need.
7365 bool NeedCF = false;
7366 bool NeedOF = false;
7369 case X86::COND_A: case X86::COND_AE:
7370 case X86::COND_B: case X86::COND_BE:
7373 case X86::COND_G: case X86::COND_GE:
7374 case X86::COND_L: case X86::COND_LE:
7375 case X86::COND_O: case X86::COND_NO:
7380 // See if we can use the EFLAGS value from the operand instead of
7381 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
7382 // we prove that the arithmetic won't overflow, we can't use OF or CF.
7383 if (Op.getResNo() != 0 || NeedOF || NeedCF)
7384 // Emit a CMP with 0, which is the TEST pattern.
7385 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
7386 DAG.getConstant(0, Op.getValueType()));
7388 unsigned Opcode = 0;
7389 unsigned NumOperands = 0;
7390 switch (Op.getNode()->getOpcode()) {
7392 // Due to an isel shortcoming, be conservative if this add is likely to be
7393 // selected as part of a load-modify-store instruction. When the root node
7394 // in a match is a store, isel doesn't know how to remap non-chain non-flag
7395 // uses of other nodes in the match, such as the ADD in this case. This
7396 // leads to the ADD being left around and reselected, with the result being
7397 // two adds in the output. Alas, even if none our users are stores, that
7398 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
7399 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
7400 // climbing the DAG back to the root, and it doesn't seem to be worth the
7402 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7403 UE = Op.getNode()->use_end(); UI != UE; ++UI)
7404 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
7407 if (ConstantSDNode *C =
7408 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
7409 // An add of one will be selected as an INC.
7410 if (C->getAPIntValue() == 1) {
7411 Opcode = X86ISD::INC;
7416 // An add of negative one (subtract of one) will be selected as a DEC.
7417 if (C->getAPIntValue().isAllOnesValue()) {
7418 Opcode = X86ISD::DEC;
7424 // Otherwise use a regular EFLAGS-setting add.
7425 Opcode = X86ISD::ADD;
7429 // If the primary and result isn't used, don't bother using X86ISD::AND,
7430 // because a TEST instruction will be better.
7431 bool NonFlagUse = false;
7432 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7433 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
7435 unsigned UOpNo = UI.getOperandNo();
7436 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
7437 // Look pass truncate.
7438 UOpNo = User->use_begin().getOperandNo();
7439 User = *User->use_begin();
7442 if (User->getOpcode() != ISD::BRCOND &&
7443 User->getOpcode() != ISD::SETCC &&
7444 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
7457 // Due to the ISEL shortcoming noted above, be conservative if this op is
7458 // likely to be selected as part of a load-modify-store instruction.
7459 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7460 UE = Op.getNode()->use_end(); UI != UE; ++UI)
7461 if (UI->getOpcode() == ISD::STORE)
7464 // Otherwise use a regular EFLAGS-setting instruction.
7465 switch (Op.getNode()->getOpcode()) {
7466 default: llvm_unreachable("unexpected operator!");
7467 case ISD::SUB: Opcode = X86ISD::SUB; break;
7468 case ISD::OR: Opcode = X86ISD::OR; break;
7469 case ISD::XOR: Opcode = X86ISD::XOR; break;
7470 case ISD::AND: Opcode = X86ISD::AND; break;
7482 return SDValue(Op.getNode(), 1);
7489 // Emit a CMP with 0, which is the TEST pattern.
7490 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
7491 DAG.getConstant(0, Op.getValueType()));
7493 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
7494 SmallVector<SDValue, 4> Ops;
7495 for (unsigned i = 0; i != NumOperands; ++i)
7496 Ops.push_back(Op.getOperand(i));
7498 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
7499 DAG.ReplaceAllUsesWith(Op, New);
7500 return SDValue(New.getNode(), 1);
7503 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
7505 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
7506 SelectionDAG &DAG) const {
7507 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
7508 if (C->getAPIntValue() == 0)
7509 return EmitTest(Op0, X86CC, DAG);
7511 DebugLoc dl = Op0.getDebugLoc();
7512 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
7515 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
7516 /// if it's possible.
7517 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
7518 DebugLoc dl, SelectionDAG &DAG) const {
7519 SDValue Op0 = And.getOperand(0);
7520 SDValue Op1 = And.getOperand(1);
7521 if (Op0.getOpcode() == ISD::TRUNCATE)
7522 Op0 = Op0.getOperand(0);
7523 if (Op1.getOpcode() == ISD::TRUNCATE)
7524 Op1 = Op1.getOperand(0);
7527 if (Op1.getOpcode() == ISD::SHL)
7528 std::swap(Op0, Op1);
7529 if (Op0.getOpcode() == ISD::SHL) {
7530 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
7531 if (And00C->getZExtValue() == 1) {
7532 // If we looked past a truncate, check that it's only truncating away
7534 unsigned BitWidth = Op0.getValueSizeInBits();
7535 unsigned AndBitWidth = And.getValueSizeInBits();
7536 if (BitWidth > AndBitWidth) {
7537 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
7538 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
7539 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
7543 RHS = Op0.getOperand(1);
7545 } else if (Op1.getOpcode() == ISD::Constant) {
7546 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
7547 SDValue AndLHS = Op0;
7548 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
7549 LHS = AndLHS.getOperand(0);
7550 RHS = AndLHS.getOperand(1);
7554 if (LHS.getNode()) {
7555 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
7556 // instruction. Since the shift amount is in-range-or-undefined, we know
7557 // that doing a bittest on the i32 value is ok. We extend to i32 because
7558 // the encoding for the i16 version is larger than the i32 version.
7559 // Also promote i16 to i32 for performance / code size reason.
7560 if (LHS.getValueType() == MVT::i8 ||
7561 LHS.getValueType() == MVT::i16)
7562 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
7564 // If the operand types disagree, extend the shift amount to match. Since
7565 // BT ignores high bits (like shifts) we can use anyextend.
7566 if (LHS.getValueType() != RHS.getValueType())
7567 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
7569 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
7570 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
7571 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7572 DAG.getConstant(Cond, MVT::i8), BT);
7578 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
7579 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
7580 SDValue Op0 = Op.getOperand(0);
7581 SDValue Op1 = Op.getOperand(1);
7582 DebugLoc dl = Op.getDebugLoc();
7583 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
7585 // Optimize to BT if possible.
7586 // Lower (X & (1 << N)) == 0 to BT(X, N).
7587 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
7588 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
7589 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
7590 Op1.getOpcode() == ISD::Constant &&
7591 cast<ConstantSDNode>(Op1)->isNullValue() &&
7592 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7593 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
7594 if (NewSetCC.getNode())
7598 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
7600 if (Op1.getOpcode() == ISD::Constant &&
7601 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
7602 cast<ConstantSDNode>(Op1)->isNullValue()) &&
7603 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7605 // If the input is a setcc, then reuse the input setcc or use a new one with
7606 // the inverted condition.
7607 if (Op0.getOpcode() == X86ISD::SETCC) {
7608 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
7609 bool Invert = (CC == ISD::SETNE) ^
7610 cast<ConstantSDNode>(Op1)->isNullValue();
7611 if (!Invert) return Op0;
7613 CCode = X86::GetOppositeBranchCondition(CCode);
7614 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7615 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
7619 bool isFP = Op1.getValueType().isFloatingPoint();
7620 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
7621 if (X86CC == X86::COND_INVALID)
7624 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
7625 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7626 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
7629 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
7631 SDValue Op0 = Op.getOperand(0);
7632 SDValue Op1 = Op.getOperand(1);
7633 SDValue CC = Op.getOperand(2);
7634 EVT VT = Op.getValueType();
7635 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
7636 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
7637 DebugLoc dl = Op.getDebugLoc();
7641 EVT VT0 = Op0.getValueType();
7642 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
7643 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
7646 switch (SetCCOpcode) {
7649 case ISD::SETEQ: SSECC = 0; break;
7651 case ISD::SETGT: Swap = true; // Fallthrough
7653 case ISD::SETOLT: SSECC = 1; break;
7655 case ISD::SETGE: Swap = true; // Fallthrough
7657 case ISD::SETOLE: SSECC = 2; break;
7658 case ISD::SETUO: SSECC = 3; break;
7660 case ISD::SETNE: SSECC = 4; break;
7661 case ISD::SETULE: Swap = true;
7662 case ISD::SETUGE: SSECC = 5; break;
7663 case ISD::SETULT: Swap = true;
7664 case ISD::SETUGT: SSECC = 6; break;
7665 case ISD::SETO: SSECC = 7; break;
7668 std::swap(Op0, Op1);
7670 // In the two special cases we can't handle, emit two comparisons.
7672 if (SetCCOpcode == ISD::SETUEQ) {
7674 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
7675 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
7676 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
7678 else if (SetCCOpcode == ISD::SETONE) {
7680 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
7681 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
7682 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
7684 llvm_unreachable("Illegal FP comparison");
7686 // Handle all other FP comparisons here.
7687 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
7690 // We are handling one of the integer comparisons here. Since SSE only has
7691 // GT and EQ comparisons for integer, swapping operands and multiple
7692 // operations may be required for some comparisons.
7693 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
7694 bool Swap = false, Invert = false, FlipSigns = false;
7696 switch (VT.getSimpleVT().SimpleTy) {
7698 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
7699 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
7700 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
7701 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
7704 switch (SetCCOpcode) {
7706 case ISD::SETNE: Invert = true;
7707 case ISD::SETEQ: Opc = EQOpc; break;
7708 case ISD::SETLT: Swap = true;
7709 case ISD::SETGT: Opc = GTOpc; break;
7710 case ISD::SETGE: Swap = true;
7711 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
7712 case ISD::SETULT: Swap = true;
7713 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
7714 case ISD::SETUGE: Swap = true;
7715 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
7718 std::swap(Op0, Op1);
7720 // Since SSE has no unsigned integer comparisons, we need to flip the sign
7721 // bits of the inputs before performing those operations.
7723 EVT EltVT = VT.getVectorElementType();
7724 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
7726 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
7727 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
7729 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
7730 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
7733 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
7735 // If the logical-not of the result is required, perform that now.
7737 Result = DAG.getNOT(dl, Result, VT);
7742 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
7743 static bool isX86LogicalCmp(SDValue Op) {
7744 unsigned Opc = Op.getNode()->getOpcode();
7745 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
7747 if (Op.getResNo() == 1 &&
7748 (Opc == X86ISD::ADD ||
7749 Opc == X86ISD::SUB ||
7750 Opc == X86ISD::ADC ||
7751 Opc == X86ISD::SBB ||
7752 Opc == X86ISD::SMUL ||
7753 Opc == X86ISD::UMUL ||
7754 Opc == X86ISD::INC ||
7755 Opc == X86ISD::DEC ||
7756 Opc == X86ISD::OR ||
7757 Opc == X86ISD::XOR ||
7758 Opc == X86ISD::AND))
7761 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
7767 static bool isZero(SDValue V) {
7768 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
7769 return C && C->isNullValue();
7772 static bool isAllOnes(SDValue V) {
7773 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
7774 return C && C->isAllOnesValue();
7777 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
7778 bool addTest = true;
7779 SDValue Cond = Op.getOperand(0);
7780 SDValue Op1 = Op.getOperand(1);
7781 SDValue Op2 = Op.getOperand(2);
7782 DebugLoc DL = Op.getDebugLoc();
7785 if (Cond.getOpcode() == ISD::SETCC) {
7786 SDValue NewCond = LowerSETCC(Cond, DAG);
7787 if (NewCond.getNode())
7791 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
7792 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
7793 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
7794 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
7795 if (Cond.getOpcode() == X86ISD::SETCC &&
7796 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
7797 isZero(Cond.getOperand(1).getOperand(1))) {
7798 SDValue Cmp = Cond.getOperand(1);
7800 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
7802 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
7803 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
7804 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
7806 SDValue CmpOp0 = Cmp.getOperand(0);
7807 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
7808 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
7810 SDValue Res = // Res = 0 or -1.
7811 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
7812 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
7814 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
7815 Res = DAG.getNOT(DL, Res, Res.getValueType());
7817 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
7818 if (N2C == 0 || !N2C->isNullValue())
7819 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
7824 // Look past (and (setcc_carry (cmp ...)), 1).
7825 if (Cond.getOpcode() == ISD::AND &&
7826 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7827 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7828 if (C && C->getAPIntValue() == 1)
7829 Cond = Cond.getOperand(0);
7832 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7833 // setting operand in place of the X86ISD::SETCC.
7834 if (Cond.getOpcode() == X86ISD::SETCC ||
7835 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7836 CC = Cond.getOperand(0);
7838 SDValue Cmp = Cond.getOperand(1);
7839 unsigned Opc = Cmp.getOpcode();
7840 EVT VT = Op.getValueType();
7842 bool IllegalFPCMov = false;
7843 if (VT.isFloatingPoint() && !VT.isVector() &&
7844 !isScalarFPTypeInSSEReg(VT)) // FPStack?
7845 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
7847 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
7848 Opc == X86ISD::BT) { // FIXME
7855 // Look pass the truncate.
7856 if (Cond.getOpcode() == ISD::TRUNCATE)
7857 Cond = Cond.getOperand(0);
7859 // We know the result of AND is compared against zero. Try to match
7861 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7862 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
7863 if (NewSetCC.getNode()) {
7864 CC = NewSetCC.getOperand(0);
7865 Cond = NewSetCC.getOperand(1);
7872 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7873 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7876 // a < b ? -1 : 0 -> RES = ~setcc_carry
7877 // a < b ? 0 : -1 -> RES = setcc_carry
7878 // a >= b ? -1 : 0 -> RES = setcc_carry
7879 // a >= b ? 0 : -1 -> RES = ~setcc_carry
7880 if (Cond.getOpcode() == X86ISD::CMP) {
7881 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
7883 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
7884 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
7885 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
7886 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
7887 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
7888 return DAG.getNOT(DL, Res, Res.getValueType());
7893 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
7894 // condition is true.
7895 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
7896 SDValue Ops[] = { Op2, Op1, CC, Cond };
7897 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops, array_lengthof(Ops));
7900 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
7901 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
7902 // from the AND / OR.
7903 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
7904 Opc = Op.getOpcode();
7905 if (Opc != ISD::OR && Opc != ISD::AND)
7907 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7908 Op.getOperand(0).hasOneUse() &&
7909 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
7910 Op.getOperand(1).hasOneUse());
7913 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
7914 // 1 and that the SETCC node has a single use.
7915 static bool isXor1OfSetCC(SDValue Op) {
7916 if (Op.getOpcode() != ISD::XOR)
7918 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7919 if (N1C && N1C->getAPIntValue() == 1) {
7920 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7921 Op.getOperand(0).hasOneUse();
7926 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
7927 bool addTest = true;
7928 SDValue Chain = Op.getOperand(0);
7929 SDValue Cond = Op.getOperand(1);
7930 SDValue Dest = Op.getOperand(2);
7931 DebugLoc dl = Op.getDebugLoc();
7934 if (Cond.getOpcode() == ISD::SETCC) {
7935 SDValue NewCond = LowerSETCC(Cond, DAG);
7936 if (NewCond.getNode())
7940 // FIXME: LowerXALUO doesn't handle these!!
7941 else if (Cond.getOpcode() == X86ISD::ADD ||
7942 Cond.getOpcode() == X86ISD::SUB ||
7943 Cond.getOpcode() == X86ISD::SMUL ||
7944 Cond.getOpcode() == X86ISD::UMUL)
7945 Cond = LowerXALUO(Cond, DAG);
7948 // Look pass (and (setcc_carry (cmp ...)), 1).
7949 if (Cond.getOpcode() == ISD::AND &&
7950 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7951 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7952 if (C && C->getAPIntValue() == 1)
7953 Cond = Cond.getOperand(0);
7956 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7957 // setting operand in place of the X86ISD::SETCC.
7958 if (Cond.getOpcode() == X86ISD::SETCC ||
7959 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7960 CC = Cond.getOperand(0);
7962 SDValue Cmp = Cond.getOperand(1);
7963 unsigned Opc = Cmp.getOpcode();
7964 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
7965 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
7969 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
7973 // These can only come from an arithmetic instruction with overflow,
7974 // e.g. SADDO, UADDO.
7975 Cond = Cond.getNode()->getOperand(1);
7982 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
7983 SDValue Cmp = Cond.getOperand(0).getOperand(1);
7984 if (CondOpc == ISD::OR) {
7985 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
7986 // two branches instead of an explicit OR instruction with a
7988 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7989 isX86LogicalCmp(Cmp)) {
7990 CC = Cond.getOperand(0).getOperand(0);
7991 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7992 Chain, Dest, CC, Cmp);
7993 CC = Cond.getOperand(1).getOperand(0);
7997 } else { // ISD::AND
7998 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
7999 // two branches instead of an explicit AND instruction with a
8000 // separate test. However, we only do this if this block doesn't
8001 // have a fall-through edge, because this requires an explicit
8002 // jmp when the condition is false.
8003 if (Cmp == Cond.getOperand(1).getOperand(1) &&
8004 isX86LogicalCmp(Cmp) &&
8005 Op.getNode()->hasOneUse()) {
8006 X86::CondCode CCode =
8007 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
8008 CCode = X86::GetOppositeBranchCondition(CCode);
8009 CC = DAG.getConstant(CCode, MVT::i8);
8010 SDNode *User = *Op.getNode()->use_begin();
8011 // Look for an unconditional branch following this conditional branch.
8012 // We need this because we need to reverse the successors in order
8013 // to implement FCMP_OEQ.
8014 if (User->getOpcode() == ISD::BR) {
8015 SDValue FalseBB = User->getOperand(1);
8017 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
8018 assert(NewBR == User);
8022 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8023 Chain, Dest, CC, Cmp);
8024 X86::CondCode CCode =
8025 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
8026 CCode = X86::GetOppositeBranchCondition(CCode);
8027 CC = DAG.getConstant(CCode, MVT::i8);
8033 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
8034 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
8035 // It should be transformed during dag combiner except when the condition
8036 // is set by a arithmetics with overflow node.
8037 X86::CondCode CCode =
8038 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
8039 CCode = X86::GetOppositeBranchCondition(CCode);
8040 CC = DAG.getConstant(CCode, MVT::i8);
8041 Cond = Cond.getOperand(0).getOperand(1);
8047 // Look pass the truncate.
8048 if (Cond.getOpcode() == ISD::TRUNCATE)
8049 Cond = Cond.getOperand(0);
8051 // We know the result of AND is compared against zero. Try to match
8053 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
8054 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
8055 if (NewSetCC.getNode()) {
8056 CC = NewSetCC.getOperand(0);
8057 Cond = NewSetCC.getOperand(1);
8064 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
8065 Cond = EmitTest(Cond, X86::COND_NE, DAG);
8067 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8068 Chain, Dest, CC, Cond);
8072 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
8073 // Calls to _alloca is needed to probe the stack when allocating more than 4k
8074 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
8075 // that the guard pages used by the OS virtual memory manager are allocated in
8076 // correct sequence.
8078 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
8079 SelectionDAG &DAG) const {
8080 assert((Subtarget->isTargetCygMing() || Subtarget->isTargetWindows()) &&
8081 "This should be used only on Windows targets");
8082 assert(!Subtarget->isTargetEnvMacho());
8083 DebugLoc dl = Op.getDebugLoc();
8086 SDValue Chain = Op.getOperand(0);
8087 SDValue Size = Op.getOperand(1);
8088 // FIXME: Ensure alignment here
8092 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
8093 unsigned Reg = (Subtarget->is64Bit() ? X86::RAX : X86::EAX);
8095 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
8096 Flag = Chain.getValue(1);
8098 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
8100 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
8101 Flag = Chain.getValue(1);
8103 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
8105 SDValue Ops1[2] = { Chain.getValue(0), Chain };
8106 return DAG.getMergeValues(Ops1, 2, dl);
8109 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
8110 MachineFunction &MF = DAG.getMachineFunction();
8111 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
8113 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
8114 DebugLoc DL = Op.getDebugLoc();
8116 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
8117 // vastart just stores the address of the VarArgsFrameIndex slot into the
8118 // memory location argument.
8119 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
8121 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
8122 MachinePointerInfo(SV), false, false, 0);
8126 // gp_offset (0 - 6 * 8)
8127 // fp_offset (48 - 48 + 8 * 16)
8128 // overflow_arg_area (point to parameters coming in memory).
8130 SmallVector<SDValue, 8> MemOps;
8131 SDValue FIN = Op.getOperand(1);
8133 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
8134 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
8136 FIN, MachinePointerInfo(SV), false, false, 0);
8137 MemOps.push_back(Store);
8140 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8141 FIN, DAG.getIntPtrConstant(4));
8142 Store = DAG.getStore(Op.getOperand(0), DL,
8143 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
8145 FIN, MachinePointerInfo(SV, 4), false, false, 0);
8146 MemOps.push_back(Store);
8148 // Store ptr to overflow_arg_area
8149 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8150 FIN, DAG.getIntPtrConstant(4));
8151 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
8153 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
8154 MachinePointerInfo(SV, 8),
8156 MemOps.push_back(Store);
8158 // Store ptr to reg_save_area.
8159 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8160 FIN, DAG.getIntPtrConstant(8));
8161 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
8163 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
8164 MachinePointerInfo(SV, 16), false, false, 0);
8165 MemOps.push_back(Store);
8166 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
8167 &MemOps[0], MemOps.size());
8170 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
8171 assert(Subtarget->is64Bit() &&
8172 "LowerVAARG only handles 64-bit va_arg!");
8173 assert((Subtarget->isTargetLinux() ||
8174 Subtarget->isTargetDarwin()) &&
8175 "Unhandled target in LowerVAARG");
8176 assert(Op.getNode()->getNumOperands() == 4);
8177 SDValue Chain = Op.getOperand(0);
8178 SDValue SrcPtr = Op.getOperand(1);
8179 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
8180 unsigned Align = Op.getConstantOperandVal(3);
8181 DebugLoc dl = Op.getDebugLoc();
8183 EVT ArgVT = Op.getNode()->getValueType(0);
8184 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
8185 uint32_t ArgSize = getTargetData()->getTypeAllocSize(ArgTy);
8188 // Decide which area this value should be read from.
8189 // TODO: Implement the AMD64 ABI in its entirety. This simple
8190 // selection mechanism works only for the basic types.
8191 if (ArgVT == MVT::f80) {
8192 llvm_unreachable("va_arg for f80 not yet implemented");
8193 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
8194 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
8195 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
8196 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
8198 llvm_unreachable("Unhandled argument type in LowerVAARG");
8202 // Sanity Check: Make sure using fp_offset makes sense.
8203 assert(!UseSoftFloat &&
8204 !(DAG.getMachineFunction()
8205 .getFunction()->hasFnAttr(Attribute::NoImplicitFloat)) &&
8206 Subtarget->hasXMM());
8209 // Insert VAARG_64 node into the DAG
8210 // VAARG_64 returns two values: Variable Argument Address, Chain
8211 SmallVector<SDValue, 11> InstOps;
8212 InstOps.push_back(Chain);
8213 InstOps.push_back(SrcPtr);
8214 InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
8215 InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
8216 InstOps.push_back(DAG.getConstant(Align, MVT::i32));
8217 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
8218 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
8219 VTs, &InstOps[0], InstOps.size(),
8221 MachinePointerInfo(SV),
8226 Chain = VAARG.getValue(1);
8228 // Load the next argument and return it
8229 return DAG.getLoad(ArgVT, dl,
8232 MachinePointerInfo(),
8236 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
8237 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
8238 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
8239 SDValue Chain = Op.getOperand(0);
8240 SDValue DstPtr = Op.getOperand(1);
8241 SDValue SrcPtr = Op.getOperand(2);
8242 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
8243 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
8244 DebugLoc DL = Op.getDebugLoc();
8246 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
8247 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
8249 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
8253 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
8254 DebugLoc dl = Op.getDebugLoc();
8255 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8257 default: return SDValue(); // Don't custom lower most intrinsics.
8258 // Comparison intrinsics.
8259 case Intrinsic::x86_sse_comieq_ss:
8260 case Intrinsic::x86_sse_comilt_ss:
8261 case Intrinsic::x86_sse_comile_ss:
8262 case Intrinsic::x86_sse_comigt_ss:
8263 case Intrinsic::x86_sse_comige_ss:
8264 case Intrinsic::x86_sse_comineq_ss:
8265 case Intrinsic::x86_sse_ucomieq_ss:
8266 case Intrinsic::x86_sse_ucomilt_ss:
8267 case Intrinsic::x86_sse_ucomile_ss:
8268 case Intrinsic::x86_sse_ucomigt_ss:
8269 case Intrinsic::x86_sse_ucomige_ss:
8270 case Intrinsic::x86_sse_ucomineq_ss:
8271 case Intrinsic::x86_sse2_comieq_sd:
8272 case Intrinsic::x86_sse2_comilt_sd:
8273 case Intrinsic::x86_sse2_comile_sd:
8274 case Intrinsic::x86_sse2_comigt_sd:
8275 case Intrinsic::x86_sse2_comige_sd:
8276 case Intrinsic::x86_sse2_comineq_sd:
8277 case Intrinsic::x86_sse2_ucomieq_sd:
8278 case Intrinsic::x86_sse2_ucomilt_sd:
8279 case Intrinsic::x86_sse2_ucomile_sd:
8280 case Intrinsic::x86_sse2_ucomigt_sd:
8281 case Intrinsic::x86_sse2_ucomige_sd:
8282 case Intrinsic::x86_sse2_ucomineq_sd: {
8284 ISD::CondCode CC = ISD::SETCC_INVALID;
8287 case Intrinsic::x86_sse_comieq_ss:
8288 case Intrinsic::x86_sse2_comieq_sd:
8292 case Intrinsic::x86_sse_comilt_ss:
8293 case Intrinsic::x86_sse2_comilt_sd:
8297 case Intrinsic::x86_sse_comile_ss:
8298 case Intrinsic::x86_sse2_comile_sd:
8302 case Intrinsic::x86_sse_comigt_ss:
8303 case Intrinsic::x86_sse2_comigt_sd:
8307 case Intrinsic::x86_sse_comige_ss:
8308 case Intrinsic::x86_sse2_comige_sd:
8312 case Intrinsic::x86_sse_comineq_ss:
8313 case Intrinsic::x86_sse2_comineq_sd:
8317 case Intrinsic::x86_sse_ucomieq_ss:
8318 case Intrinsic::x86_sse2_ucomieq_sd:
8319 Opc = X86ISD::UCOMI;
8322 case Intrinsic::x86_sse_ucomilt_ss:
8323 case Intrinsic::x86_sse2_ucomilt_sd:
8324 Opc = X86ISD::UCOMI;
8327 case Intrinsic::x86_sse_ucomile_ss:
8328 case Intrinsic::x86_sse2_ucomile_sd:
8329 Opc = X86ISD::UCOMI;
8332 case Intrinsic::x86_sse_ucomigt_ss:
8333 case Intrinsic::x86_sse2_ucomigt_sd:
8334 Opc = X86ISD::UCOMI;
8337 case Intrinsic::x86_sse_ucomige_ss:
8338 case Intrinsic::x86_sse2_ucomige_sd:
8339 Opc = X86ISD::UCOMI;
8342 case Intrinsic::x86_sse_ucomineq_ss:
8343 case Intrinsic::x86_sse2_ucomineq_sd:
8344 Opc = X86ISD::UCOMI;
8349 SDValue LHS = Op.getOperand(1);
8350 SDValue RHS = Op.getOperand(2);
8351 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
8352 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
8353 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
8354 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
8355 DAG.getConstant(X86CC, MVT::i8), Cond);
8356 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
8358 // ptest and testp intrinsics. The intrinsic these come from are designed to
8359 // return an integer value, not just an instruction so lower it to the ptest
8360 // or testp pattern and a setcc for the result.
8361 case Intrinsic::x86_sse41_ptestz:
8362 case Intrinsic::x86_sse41_ptestc:
8363 case Intrinsic::x86_sse41_ptestnzc:
8364 case Intrinsic::x86_avx_ptestz_256:
8365 case Intrinsic::x86_avx_ptestc_256:
8366 case Intrinsic::x86_avx_ptestnzc_256:
8367 case Intrinsic::x86_avx_vtestz_ps:
8368 case Intrinsic::x86_avx_vtestc_ps:
8369 case Intrinsic::x86_avx_vtestnzc_ps:
8370 case Intrinsic::x86_avx_vtestz_pd:
8371 case Intrinsic::x86_avx_vtestc_pd:
8372 case Intrinsic::x86_avx_vtestnzc_pd:
8373 case Intrinsic::x86_avx_vtestz_ps_256:
8374 case Intrinsic::x86_avx_vtestc_ps_256:
8375 case Intrinsic::x86_avx_vtestnzc_ps_256:
8376 case Intrinsic::x86_avx_vtestz_pd_256:
8377 case Intrinsic::x86_avx_vtestc_pd_256:
8378 case Intrinsic::x86_avx_vtestnzc_pd_256: {
8379 bool IsTestPacked = false;
8382 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
8383 case Intrinsic::x86_avx_vtestz_ps:
8384 case Intrinsic::x86_avx_vtestz_pd:
8385 case Intrinsic::x86_avx_vtestz_ps_256:
8386 case Intrinsic::x86_avx_vtestz_pd_256:
8387 IsTestPacked = true; // Fallthrough
8388 case Intrinsic::x86_sse41_ptestz:
8389 case Intrinsic::x86_avx_ptestz_256:
8391 X86CC = X86::COND_E;
8393 case Intrinsic::x86_avx_vtestc_ps:
8394 case Intrinsic::x86_avx_vtestc_pd:
8395 case Intrinsic::x86_avx_vtestc_ps_256:
8396 case Intrinsic::x86_avx_vtestc_pd_256:
8397 IsTestPacked = true; // Fallthrough
8398 case Intrinsic::x86_sse41_ptestc:
8399 case Intrinsic::x86_avx_ptestc_256:
8401 X86CC = X86::COND_B;
8403 case Intrinsic::x86_avx_vtestnzc_ps:
8404 case Intrinsic::x86_avx_vtestnzc_pd:
8405 case Intrinsic::x86_avx_vtestnzc_ps_256:
8406 case Intrinsic::x86_avx_vtestnzc_pd_256:
8407 IsTestPacked = true; // Fallthrough
8408 case Intrinsic::x86_sse41_ptestnzc:
8409 case Intrinsic::x86_avx_ptestnzc_256:
8411 X86CC = X86::COND_A;
8415 SDValue LHS = Op.getOperand(1);
8416 SDValue RHS = Op.getOperand(2);
8417 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
8418 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
8419 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
8420 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
8421 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
8424 // Fix vector shift instructions where the last operand is a non-immediate
8426 case Intrinsic::x86_sse2_pslli_w:
8427 case Intrinsic::x86_sse2_pslli_d:
8428 case Intrinsic::x86_sse2_pslli_q:
8429 case Intrinsic::x86_sse2_psrli_w:
8430 case Intrinsic::x86_sse2_psrli_d:
8431 case Intrinsic::x86_sse2_psrli_q:
8432 case Intrinsic::x86_sse2_psrai_w:
8433 case Intrinsic::x86_sse2_psrai_d:
8434 case Intrinsic::x86_mmx_pslli_w:
8435 case Intrinsic::x86_mmx_pslli_d:
8436 case Intrinsic::x86_mmx_pslli_q:
8437 case Intrinsic::x86_mmx_psrli_w:
8438 case Intrinsic::x86_mmx_psrli_d:
8439 case Intrinsic::x86_mmx_psrli_q:
8440 case Intrinsic::x86_mmx_psrai_w:
8441 case Intrinsic::x86_mmx_psrai_d: {
8442 SDValue ShAmt = Op.getOperand(2);
8443 if (isa<ConstantSDNode>(ShAmt))
8446 unsigned NewIntNo = 0;
8447 EVT ShAmtVT = MVT::v4i32;
8449 case Intrinsic::x86_sse2_pslli_w:
8450 NewIntNo = Intrinsic::x86_sse2_psll_w;
8452 case Intrinsic::x86_sse2_pslli_d:
8453 NewIntNo = Intrinsic::x86_sse2_psll_d;
8455 case Intrinsic::x86_sse2_pslli_q:
8456 NewIntNo = Intrinsic::x86_sse2_psll_q;
8458 case Intrinsic::x86_sse2_psrli_w:
8459 NewIntNo = Intrinsic::x86_sse2_psrl_w;
8461 case Intrinsic::x86_sse2_psrli_d:
8462 NewIntNo = Intrinsic::x86_sse2_psrl_d;
8464 case Intrinsic::x86_sse2_psrli_q:
8465 NewIntNo = Intrinsic::x86_sse2_psrl_q;
8467 case Intrinsic::x86_sse2_psrai_w:
8468 NewIntNo = Intrinsic::x86_sse2_psra_w;
8470 case Intrinsic::x86_sse2_psrai_d:
8471 NewIntNo = Intrinsic::x86_sse2_psra_d;
8474 ShAmtVT = MVT::v2i32;
8476 case Intrinsic::x86_mmx_pslli_w:
8477 NewIntNo = Intrinsic::x86_mmx_psll_w;
8479 case Intrinsic::x86_mmx_pslli_d:
8480 NewIntNo = Intrinsic::x86_mmx_psll_d;
8482 case Intrinsic::x86_mmx_pslli_q:
8483 NewIntNo = Intrinsic::x86_mmx_psll_q;
8485 case Intrinsic::x86_mmx_psrli_w:
8486 NewIntNo = Intrinsic::x86_mmx_psrl_w;
8488 case Intrinsic::x86_mmx_psrli_d:
8489 NewIntNo = Intrinsic::x86_mmx_psrl_d;
8491 case Intrinsic::x86_mmx_psrli_q:
8492 NewIntNo = Intrinsic::x86_mmx_psrl_q;
8494 case Intrinsic::x86_mmx_psrai_w:
8495 NewIntNo = Intrinsic::x86_mmx_psra_w;
8497 case Intrinsic::x86_mmx_psrai_d:
8498 NewIntNo = Intrinsic::x86_mmx_psra_d;
8500 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
8506 // The vector shift intrinsics with scalars uses 32b shift amounts but
8507 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
8511 ShOps[1] = DAG.getConstant(0, MVT::i32);
8512 if (ShAmtVT == MVT::v4i32) {
8513 ShOps[2] = DAG.getUNDEF(MVT::i32);
8514 ShOps[3] = DAG.getUNDEF(MVT::i32);
8515 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
8517 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
8518 // FIXME this must be lowered to get rid of the invalid type.
8521 EVT VT = Op.getValueType();
8522 ShAmt = DAG.getNode(ISD::BITCAST, dl, VT, ShAmt);
8523 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8524 DAG.getConstant(NewIntNo, MVT::i32),
8525 Op.getOperand(1), ShAmt);
8530 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
8531 SelectionDAG &DAG) const {
8532 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8533 MFI->setReturnAddressIsTaken(true);
8535 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8536 DebugLoc dl = Op.getDebugLoc();
8539 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
8541 DAG.getConstant(TD->getPointerSize(),
8542 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
8543 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
8544 DAG.getNode(ISD::ADD, dl, getPointerTy(),
8546 MachinePointerInfo(), false, false, 0);
8549 // Just load the return address.
8550 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
8551 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
8552 RetAddrFI, MachinePointerInfo(), false, false, 0);
8555 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
8556 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8557 MFI->setFrameAddressIsTaken(true);
8559 EVT VT = Op.getValueType();
8560 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
8561 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8562 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
8563 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
8565 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
8566 MachinePointerInfo(),
8571 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
8572 SelectionDAG &DAG) const {
8573 return DAG.getIntPtrConstant(2*TD->getPointerSize());
8576 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
8577 MachineFunction &MF = DAG.getMachineFunction();
8578 SDValue Chain = Op.getOperand(0);
8579 SDValue Offset = Op.getOperand(1);
8580 SDValue Handler = Op.getOperand(2);
8581 DebugLoc dl = Op.getDebugLoc();
8583 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
8584 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
8586 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
8588 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
8589 DAG.getIntPtrConstant(TD->getPointerSize()));
8590 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
8591 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
8593 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
8594 MF.getRegInfo().addLiveOut(StoreAddrReg);
8596 return DAG.getNode(X86ISD::EH_RETURN, dl,
8598 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
8601 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
8602 SelectionDAG &DAG) const {
8603 SDValue Root = Op.getOperand(0);
8604 SDValue Trmp = Op.getOperand(1); // trampoline
8605 SDValue FPtr = Op.getOperand(2); // nested function
8606 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
8607 DebugLoc dl = Op.getDebugLoc();
8609 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
8611 if (Subtarget->is64Bit()) {
8612 SDValue OutChains[6];
8614 // Large code-model.
8615 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
8616 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
8618 const unsigned char N86R10 = X86_MC::getX86RegNum(X86::R10);
8619 const unsigned char N86R11 = X86_MC::getX86RegNum(X86::R11);
8621 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
8623 // Load the pointer to the nested function into R11.
8624 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
8625 SDValue Addr = Trmp;
8626 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8627 Addr, MachinePointerInfo(TrmpAddr),
8630 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8631 DAG.getConstant(2, MVT::i64));
8632 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
8633 MachinePointerInfo(TrmpAddr, 2),
8636 // Load the 'nest' parameter value into R10.
8637 // R10 is specified in X86CallingConv.td
8638 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
8639 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8640 DAG.getConstant(10, MVT::i64));
8641 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8642 Addr, MachinePointerInfo(TrmpAddr, 10),
8645 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8646 DAG.getConstant(12, MVT::i64));
8647 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
8648 MachinePointerInfo(TrmpAddr, 12),
8651 // Jump to the nested function.
8652 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
8653 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8654 DAG.getConstant(20, MVT::i64));
8655 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8656 Addr, MachinePointerInfo(TrmpAddr, 20),
8659 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
8660 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8661 DAG.getConstant(22, MVT::i64));
8662 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
8663 MachinePointerInfo(TrmpAddr, 22),
8667 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
8668 return DAG.getMergeValues(Ops, 2, dl);
8670 const Function *Func =
8671 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
8672 CallingConv::ID CC = Func->getCallingConv();
8677 llvm_unreachable("Unsupported calling convention");
8678 case CallingConv::C:
8679 case CallingConv::X86_StdCall: {
8680 // Pass 'nest' parameter in ECX.
8681 // Must be kept in sync with X86CallingConv.td
8684 // Check that ECX wasn't needed by an 'inreg' parameter.
8685 FunctionType *FTy = Func->getFunctionType();
8686 const AttrListPtr &Attrs = Func->getAttributes();
8688 if (!Attrs.isEmpty() && !Func->isVarArg()) {
8689 unsigned InRegCount = 0;
8692 for (FunctionType::param_iterator I = FTy->param_begin(),
8693 E = FTy->param_end(); I != E; ++I, ++Idx)
8694 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
8695 // FIXME: should only count parameters that are lowered to integers.
8696 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
8698 if (InRegCount > 2) {
8699 report_fatal_error("Nest register in use - reduce number of inreg"
8705 case CallingConv::X86_FastCall:
8706 case CallingConv::X86_ThisCall:
8707 case CallingConv::Fast:
8708 // Pass 'nest' parameter in EAX.
8709 // Must be kept in sync with X86CallingConv.td
8714 SDValue OutChains[4];
8717 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8718 DAG.getConstant(10, MVT::i32));
8719 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
8721 // This is storing the opcode for MOV32ri.
8722 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
8723 const unsigned char N86Reg = X86_MC::getX86RegNum(NestReg);
8724 OutChains[0] = DAG.getStore(Root, dl,
8725 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
8726 Trmp, MachinePointerInfo(TrmpAddr),
8729 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8730 DAG.getConstant(1, MVT::i32));
8731 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
8732 MachinePointerInfo(TrmpAddr, 1),
8735 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
8736 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8737 DAG.getConstant(5, MVT::i32));
8738 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
8739 MachinePointerInfo(TrmpAddr, 5),
8742 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8743 DAG.getConstant(6, MVT::i32));
8744 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
8745 MachinePointerInfo(TrmpAddr, 6),
8749 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
8750 return DAG.getMergeValues(Ops, 2, dl);
8754 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
8755 SelectionDAG &DAG) const {
8757 The rounding mode is in bits 11:10 of FPSR, and has the following
8764 FLT_ROUNDS, on the other hand, expects the following:
8771 To perform the conversion, we do:
8772 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
8775 MachineFunction &MF = DAG.getMachineFunction();
8776 const TargetMachine &TM = MF.getTarget();
8777 const TargetFrameLowering &TFI = *TM.getFrameLowering();
8778 unsigned StackAlignment = TFI.getStackAlignment();
8779 EVT VT = Op.getValueType();
8780 DebugLoc DL = Op.getDebugLoc();
8782 // Save FP Control Word to stack slot
8783 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
8784 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8787 MachineMemOperand *MMO =
8788 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8789 MachineMemOperand::MOStore, 2, 2);
8791 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
8792 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
8793 DAG.getVTList(MVT::Other),
8794 Ops, 2, MVT::i16, MMO);
8796 // Load FP Control Word from stack slot
8797 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
8798 MachinePointerInfo(), false, false, 0);
8800 // Transform as necessary
8802 DAG.getNode(ISD::SRL, DL, MVT::i16,
8803 DAG.getNode(ISD::AND, DL, MVT::i16,
8804 CWD, DAG.getConstant(0x800, MVT::i16)),
8805 DAG.getConstant(11, MVT::i8));
8807 DAG.getNode(ISD::SRL, DL, MVT::i16,
8808 DAG.getNode(ISD::AND, DL, MVT::i16,
8809 CWD, DAG.getConstant(0x400, MVT::i16)),
8810 DAG.getConstant(9, MVT::i8));
8813 DAG.getNode(ISD::AND, DL, MVT::i16,
8814 DAG.getNode(ISD::ADD, DL, MVT::i16,
8815 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
8816 DAG.getConstant(1, MVT::i16)),
8817 DAG.getConstant(3, MVT::i16));
8820 return DAG.getNode((VT.getSizeInBits() < 16 ?
8821 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
8824 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
8825 EVT VT = Op.getValueType();
8827 unsigned NumBits = VT.getSizeInBits();
8828 DebugLoc dl = Op.getDebugLoc();
8830 Op = Op.getOperand(0);
8831 if (VT == MVT::i8) {
8832 // Zero extend to i32 since there is not an i8 bsr.
8834 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8837 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
8838 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8839 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
8841 // If src is zero (i.e. bsr sets ZF), returns NumBits.
8844 DAG.getConstant(NumBits+NumBits-1, OpVT),
8845 DAG.getConstant(X86::COND_E, MVT::i8),
8848 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8850 // Finally xor with NumBits-1.
8851 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
8854 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8858 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
8859 EVT VT = Op.getValueType();
8861 unsigned NumBits = VT.getSizeInBits();
8862 DebugLoc dl = Op.getDebugLoc();
8864 Op = Op.getOperand(0);
8865 if (VT == MVT::i8) {
8867 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8870 // Issue a bsf (scan bits forward) which also sets EFLAGS.
8871 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8872 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
8874 // If src is zero (i.e. bsf sets ZF), returns NumBits.
8877 DAG.getConstant(NumBits, OpVT),
8878 DAG.getConstant(X86::COND_E, MVT::i8),
8881 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8884 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8888 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
8889 EVT VT = Op.getValueType();
8890 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
8891 DebugLoc dl = Op.getDebugLoc();
8893 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
8894 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
8895 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
8896 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
8897 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
8899 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
8900 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
8901 // return AloBlo + AloBhi + AhiBlo;
8903 SDValue A = Op.getOperand(0);
8904 SDValue B = Op.getOperand(1);
8906 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8907 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8908 A, DAG.getConstant(32, MVT::i32));
8909 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8910 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8911 B, DAG.getConstant(32, MVT::i32));
8912 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8913 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8915 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8916 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8918 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8919 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8921 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8922 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8923 AloBhi, DAG.getConstant(32, MVT::i32));
8924 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8925 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8926 AhiBlo, DAG.getConstant(32, MVT::i32));
8927 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
8928 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
8932 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
8934 EVT VT = Op.getValueType();
8935 DebugLoc dl = Op.getDebugLoc();
8936 SDValue R = Op.getOperand(0);
8937 SDValue Amt = Op.getOperand(1);
8939 LLVMContext *Context = DAG.getContext();
8942 if (!Subtarget->hasSSE2()) return SDValue();
8944 // Optimize shl/srl/sra with constant shift amount.
8945 if (isSplatVector(Amt.getNode())) {
8946 SDValue SclrAmt = Amt->getOperand(0);
8947 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(SclrAmt)) {
8948 uint64_t ShiftAmt = C->getZExtValue();
8950 if (VT == MVT::v2i64 && Op.getOpcode() == ISD::SHL)
8951 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8952 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8953 R, DAG.getConstant(ShiftAmt, MVT::i32));
8955 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SHL)
8956 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8957 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8958 R, DAG.getConstant(ShiftAmt, MVT::i32));
8960 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SHL)
8961 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8962 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8963 R, DAG.getConstant(ShiftAmt, MVT::i32));
8965 if (VT == MVT::v2i64 && Op.getOpcode() == ISD::SRL)
8966 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8967 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8968 R, DAG.getConstant(ShiftAmt, MVT::i32));
8970 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SRL)
8971 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8972 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
8973 R, DAG.getConstant(ShiftAmt, MVT::i32));
8975 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SRL)
8976 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8977 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
8978 R, DAG.getConstant(ShiftAmt, MVT::i32));
8980 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SRA)
8981 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8982 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
8983 R, DAG.getConstant(ShiftAmt, MVT::i32));
8985 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SRA)
8986 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8987 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
8988 R, DAG.getConstant(ShiftAmt, MVT::i32));
8992 // Lower SHL with variable shift amount.
8993 // Cannot lower SHL without SSE2 or later.
8994 if (!Subtarget->hasSSE2()) return SDValue();
8996 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
8997 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8998 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8999 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
9001 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
9003 std::vector<Constant*> CV(4, CI);
9004 Constant *C = ConstantVector::get(CV);
9005 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9006 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9007 MachinePointerInfo::getConstantPool(),
9010 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
9011 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
9012 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
9013 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
9015 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
9017 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9018 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9019 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
9021 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
9022 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
9024 std::vector<Constant*> CVM1(16, CM1);
9025 std::vector<Constant*> CVM2(16, CM2);
9026 Constant *C = ConstantVector::get(CVM1);
9027 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9028 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9029 MachinePointerInfo::getConstantPool(),
9032 // r = pblendv(r, psllw(r & (char16)15, 4), a);
9033 M = DAG.getNode(ISD::AND, dl, VT, R, M);
9034 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9035 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
9036 DAG.getConstant(4, MVT::i32));
9037 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
9039 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
9041 C = ConstantVector::get(CVM2);
9042 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9043 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9044 MachinePointerInfo::getConstantPool(),
9047 // r = pblendv(r, psllw(r & (char16)63, 2), a);
9048 M = DAG.getNode(ISD::AND, dl, VT, R, M);
9049 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9050 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
9051 DAG.getConstant(2, MVT::i32));
9052 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
9054 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
9056 // return pblendv(r, r+r, a);
9057 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT,
9058 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
9064 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
9065 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
9066 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
9067 // looks for this combo and may remove the "setcc" instruction if the "setcc"
9068 // has only one use.
9069 SDNode *N = Op.getNode();
9070 SDValue LHS = N->getOperand(0);
9071 SDValue RHS = N->getOperand(1);
9072 unsigned BaseOp = 0;
9074 DebugLoc DL = Op.getDebugLoc();
9075 switch (Op.getOpcode()) {
9076 default: llvm_unreachable("Unknown ovf instruction!");
9078 // A subtract of one will be selected as a INC. Note that INC doesn't
9079 // set CF, so we can't do this for UADDO.
9080 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
9082 BaseOp = X86ISD::INC;
9086 BaseOp = X86ISD::ADD;
9090 BaseOp = X86ISD::ADD;
9094 // A subtract of one will be selected as a DEC. Note that DEC doesn't
9095 // set CF, so we can't do this for USUBO.
9096 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
9098 BaseOp = X86ISD::DEC;
9102 BaseOp = X86ISD::SUB;
9106 BaseOp = X86ISD::SUB;
9110 BaseOp = X86ISD::SMUL;
9113 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
9114 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
9116 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
9119 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9120 DAG.getConstant(X86::COND_O, MVT::i32),
9121 SDValue(Sum.getNode(), 2));
9123 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
9127 // Also sets EFLAGS.
9128 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
9129 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
9132 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
9133 DAG.getConstant(Cond, MVT::i32),
9134 SDValue(Sum.getNode(), 1));
9136 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
9139 SDValue X86TargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const{
9140 DebugLoc dl = Op.getDebugLoc();
9141 SDNode* Node = Op.getNode();
9142 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
9143 EVT VT = Node->getValueType(0);
9145 if (Subtarget->hasSSE2() && VT.isVector()) {
9146 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
9147 ExtraVT.getScalarType().getSizeInBits();
9148 SDValue ShAmt = DAG.getConstant(BitsDiff, MVT::i32);
9150 unsigned SHLIntrinsicsID = 0;
9151 unsigned SRAIntrinsicsID = 0;
9152 switch (VT.getSimpleVT().SimpleTy) {
9156 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_q;
9157 SRAIntrinsicsID = 0;
9161 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_d;
9162 SRAIntrinsicsID = Intrinsic::x86_sse2_psrai_d;
9166 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_w;
9167 SRAIntrinsicsID = Intrinsic::x86_sse2_psrai_w;
9172 SDValue Tmp1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9173 DAG.getConstant(SHLIntrinsicsID, MVT::i32),
9174 Node->getOperand(0), ShAmt);
9176 // In case of 1 bit sext, no need to shr
9177 if (ExtraVT.getScalarType().getSizeInBits() == 1) return Tmp1;
9179 if (SRAIntrinsicsID) {
9180 Tmp1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9181 DAG.getConstant(SRAIntrinsicsID, MVT::i32),
9191 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
9192 DebugLoc dl = Op.getDebugLoc();
9194 // Go ahead and emit the fence on x86-64 even if we asked for no-sse2.
9195 // There isn't any reason to disable it if the target processor supports it.
9196 if (!Subtarget->hasSSE2() && !Subtarget->is64Bit()) {
9197 SDValue Chain = Op.getOperand(0);
9198 SDValue Zero = DAG.getConstant(0, MVT::i32);
9200 DAG.getRegister(X86::ESP, MVT::i32), // Base
9201 DAG.getTargetConstant(1, MVT::i8), // Scale
9202 DAG.getRegister(0, MVT::i32), // Index
9203 DAG.getTargetConstant(0, MVT::i32), // Disp
9204 DAG.getRegister(0, MVT::i32), // Segment.
9209 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
9210 array_lengthof(Ops));
9211 return SDValue(Res, 0);
9214 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
9216 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
9218 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
9219 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
9220 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
9221 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
9223 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
9224 if (!Op1 && !Op2 && !Op3 && Op4)
9225 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
9227 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
9228 if (Op1 && !Op2 && !Op3 && !Op4)
9229 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
9231 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
9233 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
9236 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
9237 EVT T = Op.getValueType();
9238 DebugLoc DL = Op.getDebugLoc();
9241 switch(T.getSimpleVT().SimpleTy) {
9243 assert(false && "Invalid value type!");
9244 case MVT::i8: Reg = X86::AL; size = 1; break;
9245 case MVT::i16: Reg = X86::AX; size = 2; break;
9246 case MVT::i32: Reg = X86::EAX; size = 4; break;
9248 assert(Subtarget->is64Bit() && "Node not type legal!");
9249 Reg = X86::RAX; size = 8;
9252 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
9253 Op.getOperand(2), SDValue());
9254 SDValue Ops[] = { cpIn.getValue(0),
9257 DAG.getTargetConstant(size, MVT::i8),
9259 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9260 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
9261 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
9264 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
9268 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
9269 SelectionDAG &DAG) const {
9270 assert(Subtarget->is64Bit() && "Result not type legalized?");
9271 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9272 SDValue TheChain = Op.getOperand(0);
9273 DebugLoc dl = Op.getDebugLoc();
9274 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
9275 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
9276 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
9278 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
9279 DAG.getConstant(32, MVT::i8));
9281 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
9284 return DAG.getMergeValues(Ops, 2, dl);
9287 SDValue X86TargetLowering::LowerBITCAST(SDValue Op,
9288 SelectionDAG &DAG) const {
9289 EVT SrcVT = Op.getOperand(0).getValueType();
9290 EVT DstVT = Op.getValueType();
9291 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
9292 Subtarget->hasMMX() && "Unexpected custom BITCAST");
9293 assert((DstVT == MVT::i64 ||
9294 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
9295 "Unexpected custom BITCAST");
9296 // i64 <=> MMX conversions are Legal.
9297 if (SrcVT==MVT::i64 && DstVT.isVector())
9299 if (DstVT==MVT::i64 && SrcVT.isVector())
9301 // MMX <=> MMX conversions are Legal.
9302 if (SrcVT.isVector() && DstVT.isVector())
9304 // All other conversions need to be expanded.
9308 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
9309 SDNode *Node = Op.getNode();
9310 DebugLoc dl = Node->getDebugLoc();
9311 EVT T = Node->getValueType(0);
9312 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
9313 DAG.getConstant(0, T), Node->getOperand(2));
9314 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
9315 cast<AtomicSDNode>(Node)->getMemoryVT(),
9316 Node->getOperand(0),
9317 Node->getOperand(1), negOp,
9318 cast<AtomicSDNode>(Node)->getSrcValue(),
9319 cast<AtomicSDNode>(Node)->getAlignment());
9322 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
9323 EVT VT = Op.getNode()->getValueType(0);
9325 // Let legalize expand this if it isn't a legal type yet.
9326 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
9329 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
9332 bool ExtraOp = false;
9333 switch (Op.getOpcode()) {
9334 default: assert(0 && "Invalid code");
9335 case ISD::ADDC: Opc = X86ISD::ADD; break;
9336 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
9337 case ISD::SUBC: Opc = X86ISD::SUB; break;
9338 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
9342 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
9344 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
9345 Op.getOperand(1), Op.getOperand(2));
9348 /// LowerOperation - Provide custom lowering hooks for some operations.
9350 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
9351 switch (Op.getOpcode()) {
9352 default: llvm_unreachable("Should not custom lower this!");
9353 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op,DAG);
9354 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
9355 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
9356 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
9357 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
9358 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
9359 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
9360 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
9361 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
9362 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
9363 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, DAG);
9364 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
9365 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
9366 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
9367 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
9368 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
9369 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
9370 case ISD::SHL_PARTS:
9371 case ISD::SRA_PARTS:
9372 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
9373 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
9374 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
9375 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
9376 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
9377 case ISD::FABS: return LowerFABS(Op, DAG);
9378 case ISD::FNEG: return LowerFNEG(Op, DAG);
9379 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
9380 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
9381 case ISD::SETCC: return LowerSETCC(Op, DAG);
9382 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
9383 case ISD::SELECT: return LowerSELECT(Op, DAG);
9384 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
9385 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
9386 case ISD::VASTART: return LowerVASTART(Op, DAG);
9387 case ISD::VAARG: return LowerVAARG(Op, DAG);
9388 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
9389 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
9390 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
9391 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
9392 case ISD::FRAME_TO_ARGS_OFFSET:
9393 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
9394 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
9395 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
9396 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
9397 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
9398 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
9399 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
9400 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
9403 case ISD::SHL: return LowerShift(Op, DAG);
9409 case ISD::UMULO: return LowerXALUO(Op, DAG);
9410 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
9411 case ISD::BITCAST: return LowerBITCAST(Op, DAG);
9415 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
9419 void X86TargetLowering::
9420 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
9421 SelectionDAG &DAG, unsigned NewOp) const {
9422 EVT T = Node->getValueType(0);
9423 DebugLoc dl = Node->getDebugLoc();
9424 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
9426 SDValue Chain = Node->getOperand(0);
9427 SDValue In1 = Node->getOperand(1);
9428 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
9429 Node->getOperand(2), DAG.getIntPtrConstant(0));
9430 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
9431 Node->getOperand(2), DAG.getIntPtrConstant(1));
9432 SDValue Ops[] = { Chain, In1, In2L, In2H };
9433 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
9435 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
9436 cast<MemSDNode>(Node)->getMemOperand());
9437 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
9438 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
9439 Results.push_back(Result.getValue(2));
9442 /// ReplaceNodeResults - Replace a node with an illegal result type
9443 /// with a new node built out of custom code.
9444 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
9445 SmallVectorImpl<SDValue>&Results,
9446 SelectionDAG &DAG) const {
9447 DebugLoc dl = N->getDebugLoc();
9448 switch (N->getOpcode()) {
9450 assert(false && "Do not know how to custom type legalize this operation!");
9452 case ISD::SIGN_EXTEND_INREG:
9457 // We don't want to expand or promote these.
9459 case ISD::FP_TO_SINT: {
9460 std::pair<SDValue,SDValue> Vals =
9461 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
9462 SDValue FIST = Vals.first, StackSlot = Vals.second;
9463 if (FIST.getNode() != 0) {
9464 EVT VT = N->getValueType(0);
9465 // Return a load from the stack slot.
9466 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
9467 MachinePointerInfo(), false, false, 0));
9471 case ISD::READCYCLECOUNTER: {
9472 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9473 SDValue TheChain = N->getOperand(0);
9474 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
9475 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
9477 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
9479 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
9480 SDValue Ops[] = { eax, edx };
9481 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
9482 Results.push_back(edx.getValue(1));
9485 case ISD::ATOMIC_CMP_SWAP: {
9486 EVT T = N->getValueType(0);
9487 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
9488 SDValue cpInL, cpInH;
9489 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
9490 DAG.getConstant(0, MVT::i32));
9491 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
9492 DAG.getConstant(1, MVT::i32));
9493 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
9494 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
9496 SDValue swapInL, swapInH;
9497 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
9498 DAG.getConstant(0, MVT::i32));
9499 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
9500 DAG.getConstant(1, MVT::i32));
9501 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
9503 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
9504 swapInL.getValue(1));
9505 SDValue Ops[] = { swapInH.getValue(0),
9507 swapInH.getValue(1) };
9508 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9509 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
9510 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG8_DAG, dl, Tys,
9512 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
9513 MVT::i32, Result.getValue(1));
9514 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
9515 MVT::i32, cpOutL.getValue(2));
9516 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
9517 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
9518 Results.push_back(cpOutH.getValue(1));
9521 case ISD::ATOMIC_LOAD_ADD:
9522 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
9524 case ISD::ATOMIC_LOAD_AND:
9525 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
9527 case ISD::ATOMIC_LOAD_NAND:
9528 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
9530 case ISD::ATOMIC_LOAD_OR:
9531 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
9533 case ISD::ATOMIC_LOAD_SUB:
9534 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
9536 case ISD::ATOMIC_LOAD_XOR:
9537 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
9539 case ISD::ATOMIC_SWAP:
9540 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
9545 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
9547 default: return NULL;
9548 case X86ISD::BSF: return "X86ISD::BSF";
9549 case X86ISD::BSR: return "X86ISD::BSR";
9550 case X86ISD::SHLD: return "X86ISD::SHLD";
9551 case X86ISD::SHRD: return "X86ISD::SHRD";
9552 case X86ISD::FAND: return "X86ISD::FAND";
9553 case X86ISD::FOR: return "X86ISD::FOR";
9554 case X86ISD::FXOR: return "X86ISD::FXOR";
9555 case X86ISD::FSRL: return "X86ISD::FSRL";
9556 case X86ISD::FILD: return "X86ISD::FILD";
9557 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
9558 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
9559 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
9560 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
9561 case X86ISD::FLD: return "X86ISD::FLD";
9562 case X86ISD::FST: return "X86ISD::FST";
9563 case X86ISD::CALL: return "X86ISD::CALL";
9564 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
9565 case X86ISD::BT: return "X86ISD::BT";
9566 case X86ISD::CMP: return "X86ISD::CMP";
9567 case X86ISD::COMI: return "X86ISD::COMI";
9568 case X86ISD::UCOMI: return "X86ISD::UCOMI";
9569 case X86ISD::SETCC: return "X86ISD::SETCC";
9570 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
9571 case X86ISD::FSETCCsd: return "X86ISD::FSETCCsd";
9572 case X86ISD::FSETCCss: return "X86ISD::FSETCCss";
9573 case X86ISD::CMOV: return "X86ISD::CMOV";
9574 case X86ISD::BRCOND: return "X86ISD::BRCOND";
9575 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
9576 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
9577 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
9578 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
9579 case X86ISD::Wrapper: return "X86ISD::Wrapper";
9580 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
9581 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
9582 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
9583 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
9584 case X86ISD::PINSRB: return "X86ISD::PINSRB";
9585 case X86ISD::PINSRW: return "X86ISD::PINSRW";
9586 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
9587 case X86ISD::ANDNP: return "X86ISD::ANDNP";
9588 case X86ISD::PSIGNB: return "X86ISD::PSIGNB";
9589 case X86ISD::PSIGNW: return "X86ISD::PSIGNW";
9590 case X86ISD::PSIGND: return "X86ISD::PSIGND";
9591 case X86ISD::PBLENDVB: return "X86ISD::PBLENDVB";
9592 case X86ISD::FMAX: return "X86ISD::FMAX";
9593 case X86ISD::FMIN: return "X86ISD::FMIN";
9594 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
9595 case X86ISD::FRCP: return "X86ISD::FRCP";
9596 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
9597 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
9598 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
9599 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
9600 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
9601 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
9602 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
9603 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
9604 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
9605 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
9606 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
9607 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
9608 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
9609 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
9610 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
9611 case X86ISD::VSHL: return "X86ISD::VSHL";
9612 case X86ISD::VSRL: return "X86ISD::VSRL";
9613 case X86ISD::CMPPD: return "X86ISD::CMPPD";
9614 case X86ISD::CMPPS: return "X86ISD::CMPPS";
9615 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
9616 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
9617 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
9618 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
9619 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
9620 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
9621 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
9622 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
9623 case X86ISD::ADD: return "X86ISD::ADD";
9624 case X86ISD::SUB: return "X86ISD::SUB";
9625 case X86ISD::ADC: return "X86ISD::ADC";
9626 case X86ISD::SBB: return "X86ISD::SBB";
9627 case X86ISD::SMUL: return "X86ISD::SMUL";
9628 case X86ISD::UMUL: return "X86ISD::UMUL";
9629 case X86ISD::INC: return "X86ISD::INC";
9630 case X86ISD::DEC: return "X86ISD::DEC";
9631 case X86ISD::OR: return "X86ISD::OR";
9632 case X86ISD::XOR: return "X86ISD::XOR";
9633 case X86ISD::AND: return "X86ISD::AND";
9634 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
9635 case X86ISD::PTEST: return "X86ISD::PTEST";
9636 case X86ISD::TESTP: return "X86ISD::TESTP";
9637 case X86ISD::PALIGN: return "X86ISD::PALIGN";
9638 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
9639 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
9640 case X86ISD::PSHUFHW_LD: return "X86ISD::PSHUFHW_LD";
9641 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
9642 case X86ISD::PSHUFLW_LD: return "X86ISD::PSHUFLW_LD";
9643 case X86ISD::SHUFPS: return "X86ISD::SHUFPS";
9644 case X86ISD::SHUFPD: return "X86ISD::SHUFPD";
9645 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
9646 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
9647 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
9648 case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
9649 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
9650 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
9651 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
9652 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
9653 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
9654 case X86ISD::MOVSHDUP_LD: return "X86ISD::MOVSHDUP_LD";
9655 case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
9656 case X86ISD::MOVSD: return "X86ISD::MOVSD";
9657 case X86ISD::MOVSS: return "X86ISD::MOVSS";
9658 case X86ISD::UNPCKLPS: return "X86ISD::UNPCKLPS";
9659 case X86ISD::UNPCKLPD: return "X86ISD::UNPCKLPD";
9660 case X86ISD::VUNPCKLPDY: return "X86ISD::VUNPCKLPDY";
9661 case X86ISD::UNPCKHPS: return "X86ISD::UNPCKHPS";
9662 case X86ISD::UNPCKHPD: return "X86ISD::UNPCKHPD";
9663 case X86ISD::PUNPCKLBW: return "X86ISD::PUNPCKLBW";
9664 case X86ISD::PUNPCKLWD: return "X86ISD::PUNPCKLWD";
9665 case X86ISD::PUNPCKLDQ: return "X86ISD::PUNPCKLDQ";
9666 case X86ISD::PUNPCKLQDQ: return "X86ISD::PUNPCKLQDQ";
9667 case X86ISD::PUNPCKHBW: return "X86ISD::PUNPCKHBW";
9668 case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
9669 case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
9670 case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
9671 case X86ISD::VPERMIL: return "X86ISD::VPERMIL";
9672 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
9673 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
9674 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
9678 // isLegalAddressingMode - Return true if the addressing mode represented
9679 // by AM is legal for this target, for a load/store of the specified type.
9680 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
9682 // X86 supports extremely general addressing modes.
9683 CodeModel::Model M = getTargetMachine().getCodeModel();
9684 Reloc::Model R = getTargetMachine().getRelocationModel();
9686 // X86 allows a sign-extended 32-bit immediate field as a displacement.
9687 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
9692 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
9694 // If a reference to this global requires an extra load, we can't fold it.
9695 if (isGlobalStubReference(GVFlags))
9698 // If BaseGV requires a register for the PIC base, we cannot also have a
9699 // BaseReg specified.
9700 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
9703 // If lower 4G is not available, then we must use rip-relative addressing.
9704 if ((M != CodeModel::Small || R != Reloc::Static) &&
9705 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
9715 // These scales always work.
9720 // These scales are formed with basereg+scalereg. Only accept if there is
9725 default: // Other stuff never works.
9733 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
9734 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
9736 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
9737 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
9738 if (NumBits1 <= NumBits2)
9743 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
9744 if (!VT1.isInteger() || !VT2.isInteger())
9746 unsigned NumBits1 = VT1.getSizeInBits();
9747 unsigned NumBits2 = VT2.getSizeInBits();
9748 if (NumBits1 <= NumBits2)
9753 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
9754 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
9755 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
9758 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
9759 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
9760 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
9763 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
9764 // i16 instructions are longer (0x66 prefix) and potentially slower.
9765 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
9768 /// isShuffleMaskLegal - Targets can use this to indicate that they only
9769 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
9770 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
9771 /// are assumed to be legal.
9773 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
9775 // Very little shuffling can be done for 64-bit vectors right now.
9776 if (VT.getSizeInBits() == 64)
9777 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
9779 // FIXME: pshufb, blends, shifts.
9780 return (VT.getVectorNumElements() == 2 ||
9781 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
9782 isMOVLMask(M, VT) ||
9783 isSHUFPMask(M, VT) ||
9784 isPSHUFDMask(M, VT) ||
9785 isPSHUFHWMask(M, VT) ||
9786 isPSHUFLWMask(M, VT) ||
9787 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
9788 isUNPCKLMask(M, VT) ||
9789 isUNPCKHMask(M, VT) ||
9790 isUNPCKL_v_undef_Mask(M, VT) ||
9791 isUNPCKH_v_undef_Mask(M, VT));
9795 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
9797 unsigned NumElts = VT.getVectorNumElements();
9798 // FIXME: This collection of masks seems suspect.
9801 if (NumElts == 4 && VT.getSizeInBits() == 128) {
9802 return (isMOVLMask(Mask, VT) ||
9803 isCommutedMOVLMask(Mask, VT, true) ||
9804 isSHUFPMask(Mask, VT) ||
9805 isCommutedSHUFPMask(Mask, VT));
9810 //===----------------------------------------------------------------------===//
9811 // X86 Scheduler Hooks
9812 //===----------------------------------------------------------------------===//
9814 // private utility function
9816 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
9817 MachineBasicBlock *MBB,
9824 TargetRegisterClass *RC,
9825 bool invSrc) const {
9826 // For the atomic bitwise operator, we generate
9829 // ld t1 = [bitinstr.addr]
9830 // op t2 = t1, [bitinstr.val]
9832 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
9834 // fallthrough -->nextMBB
9835 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9836 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9837 MachineFunction::iterator MBBIter = MBB;
9840 /// First build the CFG
9841 MachineFunction *F = MBB->getParent();
9842 MachineBasicBlock *thisMBB = MBB;
9843 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9844 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9845 F->insert(MBBIter, newMBB);
9846 F->insert(MBBIter, nextMBB);
9848 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9849 nextMBB->splice(nextMBB->begin(), thisMBB,
9850 llvm::next(MachineBasicBlock::iterator(bInstr)),
9852 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9854 // Update thisMBB to fall through to newMBB
9855 thisMBB->addSuccessor(newMBB);
9857 // newMBB jumps to itself and fall through to nextMBB
9858 newMBB->addSuccessor(nextMBB);
9859 newMBB->addSuccessor(newMBB);
9861 // Insert instructions into newMBB based on incoming instruction
9862 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
9863 "unexpected number of operands");
9864 DebugLoc dl = bInstr->getDebugLoc();
9865 MachineOperand& destOper = bInstr->getOperand(0);
9866 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9867 int numArgs = bInstr->getNumOperands() - 1;
9868 for (int i=0; i < numArgs; ++i)
9869 argOpers[i] = &bInstr->getOperand(i+1);
9871 // x86 address has 4 operands: base, index, scale, and displacement
9872 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9873 int valArgIndx = lastAddrIndx + 1;
9875 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
9876 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
9877 for (int i=0; i <= lastAddrIndx; ++i)
9878 (*MIB).addOperand(*argOpers[i]);
9880 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
9882 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
9887 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
9888 assert((argOpers[valArgIndx]->isReg() ||
9889 argOpers[valArgIndx]->isImm()) &&
9891 if (argOpers[valArgIndx]->isReg())
9892 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
9894 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
9896 (*MIB).addOperand(*argOpers[valArgIndx]);
9898 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
9901 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
9902 for (int i=0; i <= lastAddrIndx; ++i)
9903 (*MIB).addOperand(*argOpers[i]);
9905 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9906 (*MIB).setMemRefs(bInstr->memoperands_begin(),
9907 bInstr->memoperands_end());
9909 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
9913 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9915 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
9919 // private utility function: 64 bit atomics on 32 bit host.
9921 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
9922 MachineBasicBlock *MBB,
9927 bool invSrc) const {
9928 // For the atomic bitwise operator, we generate
9929 // thisMBB (instructions are in pairs, except cmpxchg8b)
9930 // ld t1,t2 = [bitinstr.addr]
9932 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
9933 // op t5, t6 <- out1, out2, [bitinstr.val]
9934 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
9935 // mov ECX, EBX <- t5, t6
9936 // mov EAX, EDX <- t1, t2
9937 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
9938 // mov t3, t4 <- EAX, EDX
9940 // result in out1, out2
9941 // fallthrough -->nextMBB
9943 const TargetRegisterClass *RC = X86::GR32RegisterClass;
9944 const unsigned LoadOpc = X86::MOV32rm;
9945 const unsigned NotOpc = X86::NOT32r;
9946 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9947 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9948 MachineFunction::iterator MBBIter = MBB;
9951 /// First build the CFG
9952 MachineFunction *F = MBB->getParent();
9953 MachineBasicBlock *thisMBB = MBB;
9954 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9955 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9956 F->insert(MBBIter, newMBB);
9957 F->insert(MBBIter, nextMBB);
9959 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9960 nextMBB->splice(nextMBB->begin(), thisMBB,
9961 llvm::next(MachineBasicBlock::iterator(bInstr)),
9963 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9965 // Update thisMBB to fall through to newMBB
9966 thisMBB->addSuccessor(newMBB);
9968 // newMBB jumps to itself and fall through to nextMBB
9969 newMBB->addSuccessor(nextMBB);
9970 newMBB->addSuccessor(newMBB);
9972 DebugLoc dl = bInstr->getDebugLoc();
9973 // Insert instructions into newMBB based on incoming instruction
9974 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
9975 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
9976 "unexpected number of operands");
9977 MachineOperand& dest1Oper = bInstr->getOperand(0);
9978 MachineOperand& dest2Oper = bInstr->getOperand(1);
9979 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9980 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
9981 argOpers[i] = &bInstr->getOperand(i+2);
9983 // We use some of the operands multiple times, so conservatively just
9984 // clear any kill flags that might be present.
9985 if (argOpers[i]->isReg() && argOpers[i]->isUse())
9986 argOpers[i]->setIsKill(false);
9989 // x86 address has 5 operands: base, index, scale, displacement, and segment.
9990 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9992 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
9993 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
9994 for (int i=0; i <= lastAddrIndx; ++i)
9995 (*MIB).addOperand(*argOpers[i]);
9996 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
9997 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
9998 // add 4 to displacement.
9999 for (int i=0; i <= lastAddrIndx-2; ++i)
10000 (*MIB).addOperand(*argOpers[i]);
10001 MachineOperand newOp3 = *(argOpers[3]);
10002 if (newOp3.isImm())
10003 newOp3.setImm(newOp3.getImm()+4);
10005 newOp3.setOffset(newOp3.getOffset()+4);
10006 (*MIB).addOperand(newOp3);
10007 (*MIB).addOperand(*argOpers[lastAddrIndx]);
10009 // t3/4 are defined later, at the bottom of the loop
10010 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
10011 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
10012 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
10013 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
10014 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
10015 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
10017 // The subsequent operations should be using the destination registers of
10018 //the PHI instructions.
10020 t1 = F->getRegInfo().createVirtualRegister(RC);
10021 t2 = F->getRegInfo().createVirtualRegister(RC);
10022 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
10023 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
10025 t1 = dest1Oper.getReg();
10026 t2 = dest2Oper.getReg();
10029 int valArgIndx = lastAddrIndx + 1;
10030 assert((argOpers[valArgIndx]->isReg() ||
10031 argOpers[valArgIndx]->isImm()) &&
10032 "invalid operand");
10033 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
10034 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
10035 if (argOpers[valArgIndx]->isReg())
10036 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
10038 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
10039 if (regOpcL != X86::MOV32rr)
10041 (*MIB).addOperand(*argOpers[valArgIndx]);
10042 assert(argOpers[valArgIndx + 1]->isReg() ==
10043 argOpers[valArgIndx]->isReg());
10044 assert(argOpers[valArgIndx + 1]->isImm() ==
10045 argOpers[valArgIndx]->isImm());
10046 if (argOpers[valArgIndx + 1]->isReg())
10047 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
10049 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
10050 if (regOpcH != X86::MOV32rr)
10052 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
10054 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
10056 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
10059 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
10061 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
10064 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
10065 for (int i=0; i <= lastAddrIndx; ++i)
10066 (*MIB).addOperand(*argOpers[i]);
10068 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
10069 (*MIB).setMemRefs(bInstr->memoperands_begin(),
10070 bInstr->memoperands_end());
10072 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
10073 MIB.addReg(X86::EAX);
10074 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
10075 MIB.addReg(X86::EDX);
10078 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
10080 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
10084 // private utility function
10085 MachineBasicBlock *
10086 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
10087 MachineBasicBlock *MBB,
10088 unsigned cmovOpc) const {
10089 // For the atomic min/max operator, we generate
10092 // ld t1 = [min/max.addr]
10093 // mov t2 = [min/max.val]
10095 // cmov[cond] t2 = t1
10097 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
10099 // fallthrough -->nextMBB
10101 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10102 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10103 MachineFunction::iterator MBBIter = MBB;
10106 /// First build the CFG
10107 MachineFunction *F = MBB->getParent();
10108 MachineBasicBlock *thisMBB = MBB;
10109 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
10110 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
10111 F->insert(MBBIter, newMBB);
10112 F->insert(MBBIter, nextMBB);
10114 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
10115 nextMBB->splice(nextMBB->begin(), thisMBB,
10116 llvm::next(MachineBasicBlock::iterator(mInstr)),
10118 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10120 // Update thisMBB to fall through to newMBB
10121 thisMBB->addSuccessor(newMBB);
10123 // newMBB jumps to newMBB and fall through to nextMBB
10124 newMBB->addSuccessor(nextMBB);
10125 newMBB->addSuccessor(newMBB);
10127 DebugLoc dl = mInstr->getDebugLoc();
10128 // Insert instructions into newMBB based on incoming instruction
10129 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
10130 "unexpected number of operands");
10131 MachineOperand& destOper = mInstr->getOperand(0);
10132 MachineOperand* argOpers[2 + X86::AddrNumOperands];
10133 int numArgs = mInstr->getNumOperands() - 1;
10134 for (int i=0; i < numArgs; ++i)
10135 argOpers[i] = &mInstr->getOperand(i+1);
10137 // x86 address has 4 operands: base, index, scale, and displacement
10138 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
10139 int valArgIndx = lastAddrIndx + 1;
10141 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
10142 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
10143 for (int i=0; i <= lastAddrIndx; ++i)
10144 (*MIB).addOperand(*argOpers[i]);
10146 // We only support register and immediate values
10147 assert((argOpers[valArgIndx]->isReg() ||
10148 argOpers[valArgIndx]->isImm()) &&
10149 "invalid operand");
10151 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
10152 if (argOpers[valArgIndx]->isReg())
10153 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
10155 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
10156 (*MIB).addOperand(*argOpers[valArgIndx]);
10158 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
10161 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
10166 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
10167 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
10171 // Cmp and exchange if none has modified the memory location
10172 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
10173 for (int i=0; i <= lastAddrIndx; ++i)
10174 (*MIB).addOperand(*argOpers[i]);
10176 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
10177 (*MIB).setMemRefs(mInstr->memoperands_begin(),
10178 mInstr->memoperands_end());
10180 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
10181 MIB.addReg(X86::EAX);
10184 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
10186 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
10190 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
10191 // or XMM0_V32I8 in AVX all of this code can be replaced with that
10192 // in the .td file.
10193 MachineBasicBlock *
10194 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
10195 unsigned numArgs, bool memArg) const {
10196 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
10197 "Target must have SSE4.2 or AVX features enabled");
10199 DebugLoc dl = MI->getDebugLoc();
10200 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10202 if (!Subtarget->hasAVX()) {
10204 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
10206 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
10209 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
10211 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
10214 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
10215 for (unsigned i = 0; i < numArgs; ++i) {
10216 MachineOperand &Op = MI->getOperand(i+1);
10217 if (!(Op.isReg() && Op.isImplicit()))
10218 MIB.addOperand(Op);
10220 BuildMI(*BB, MI, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
10221 .addReg(X86::XMM0);
10223 MI->eraseFromParent();
10227 MachineBasicBlock *
10228 X86TargetLowering::EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB) const {
10229 DebugLoc dl = MI->getDebugLoc();
10230 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10232 // Address into RAX/EAX, other two args into ECX, EDX.
10233 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
10234 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
10235 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
10236 for (int i = 0; i < X86::AddrNumOperands; ++i)
10237 MIB.addOperand(MI->getOperand(i));
10239 unsigned ValOps = X86::AddrNumOperands;
10240 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
10241 .addReg(MI->getOperand(ValOps).getReg());
10242 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
10243 .addReg(MI->getOperand(ValOps+1).getReg());
10245 // The instruction doesn't actually take any operands though.
10246 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
10248 MI->eraseFromParent(); // The pseudo is gone now.
10252 MachineBasicBlock *
10253 X86TargetLowering::EmitMwait(MachineInstr *MI, MachineBasicBlock *BB) const {
10254 DebugLoc dl = MI->getDebugLoc();
10255 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10257 // First arg in ECX, the second in EAX.
10258 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
10259 .addReg(MI->getOperand(0).getReg());
10260 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EAX)
10261 .addReg(MI->getOperand(1).getReg());
10263 // The instruction doesn't actually take any operands though.
10264 BuildMI(*BB, MI, dl, TII->get(X86::MWAITrr));
10266 MI->eraseFromParent(); // The pseudo is gone now.
10270 MachineBasicBlock *
10271 X86TargetLowering::EmitVAARG64WithCustomInserter(
10273 MachineBasicBlock *MBB) const {
10274 // Emit va_arg instruction on X86-64.
10276 // Operands to this pseudo-instruction:
10277 // 0 ) Output : destination address (reg)
10278 // 1-5) Input : va_list address (addr, i64mem)
10279 // 6 ) ArgSize : Size (in bytes) of vararg type
10280 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
10281 // 8 ) Align : Alignment of type
10282 // 9 ) EFLAGS (implicit-def)
10284 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
10285 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
10287 unsigned DestReg = MI->getOperand(0).getReg();
10288 MachineOperand &Base = MI->getOperand(1);
10289 MachineOperand &Scale = MI->getOperand(2);
10290 MachineOperand &Index = MI->getOperand(3);
10291 MachineOperand &Disp = MI->getOperand(4);
10292 MachineOperand &Segment = MI->getOperand(5);
10293 unsigned ArgSize = MI->getOperand(6).getImm();
10294 unsigned ArgMode = MI->getOperand(7).getImm();
10295 unsigned Align = MI->getOperand(8).getImm();
10297 // Memory Reference
10298 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
10299 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
10300 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
10302 // Machine Information
10303 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10304 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
10305 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
10306 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
10307 DebugLoc DL = MI->getDebugLoc();
10309 // struct va_list {
10312 // i64 overflow_area (address)
10313 // i64 reg_save_area (address)
10315 // sizeof(va_list) = 24
10316 // alignment(va_list) = 8
10318 unsigned TotalNumIntRegs = 6;
10319 unsigned TotalNumXMMRegs = 8;
10320 bool UseGPOffset = (ArgMode == 1);
10321 bool UseFPOffset = (ArgMode == 2);
10322 unsigned MaxOffset = TotalNumIntRegs * 8 +
10323 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
10325 /* Align ArgSize to a multiple of 8 */
10326 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
10327 bool NeedsAlign = (Align > 8);
10329 MachineBasicBlock *thisMBB = MBB;
10330 MachineBasicBlock *overflowMBB;
10331 MachineBasicBlock *offsetMBB;
10332 MachineBasicBlock *endMBB;
10334 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
10335 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
10336 unsigned OffsetReg = 0;
10338 if (!UseGPOffset && !UseFPOffset) {
10339 // If we only pull from the overflow region, we don't create a branch.
10340 // We don't need to alter control flow.
10341 OffsetDestReg = 0; // unused
10342 OverflowDestReg = DestReg;
10345 overflowMBB = thisMBB;
10348 // First emit code to check if gp_offset (or fp_offset) is below the bound.
10349 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
10350 // If not, pull from overflow_area. (branch to overflowMBB)
10355 // offsetMBB overflowMBB
10360 // Registers for the PHI in endMBB
10361 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
10362 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
10364 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10365 MachineFunction *MF = MBB->getParent();
10366 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10367 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10368 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10370 MachineFunction::iterator MBBIter = MBB;
10373 // Insert the new basic blocks
10374 MF->insert(MBBIter, offsetMBB);
10375 MF->insert(MBBIter, overflowMBB);
10376 MF->insert(MBBIter, endMBB);
10378 // Transfer the remainder of MBB and its successor edges to endMBB.
10379 endMBB->splice(endMBB->begin(), thisMBB,
10380 llvm::next(MachineBasicBlock::iterator(MI)),
10382 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10384 // Make offsetMBB and overflowMBB successors of thisMBB
10385 thisMBB->addSuccessor(offsetMBB);
10386 thisMBB->addSuccessor(overflowMBB);
10388 // endMBB is a successor of both offsetMBB and overflowMBB
10389 offsetMBB->addSuccessor(endMBB);
10390 overflowMBB->addSuccessor(endMBB);
10392 // Load the offset value into a register
10393 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
10394 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
10398 .addDisp(Disp, UseFPOffset ? 4 : 0)
10399 .addOperand(Segment)
10400 .setMemRefs(MMOBegin, MMOEnd);
10402 // Check if there is enough room left to pull this argument.
10403 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
10405 .addImm(MaxOffset + 8 - ArgSizeA8);
10407 // Branch to "overflowMBB" if offset >= max
10408 // Fall through to "offsetMBB" otherwise
10409 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
10410 .addMBB(overflowMBB);
10413 // In offsetMBB, emit code to use the reg_save_area.
10415 assert(OffsetReg != 0);
10417 // Read the reg_save_area address.
10418 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
10419 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
10424 .addOperand(Segment)
10425 .setMemRefs(MMOBegin, MMOEnd);
10427 // Zero-extend the offset
10428 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
10429 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
10432 .addImm(X86::sub_32bit);
10434 // Add the offset to the reg_save_area to get the final address.
10435 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
10436 .addReg(OffsetReg64)
10437 .addReg(RegSaveReg);
10439 // Compute the offset for the next argument
10440 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
10441 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
10443 .addImm(UseFPOffset ? 16 : 8);
10445 // Store it back into the va_list.
10446 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
10450 .addDisp(Disp, UseFPOffset ? 4 : 0)
10451 .addOperand(Segment)
10452 .addReg(NextOffsetReg)
10453 .setMemRefs(MMOBegin, MMOEnd);
10456 BuildMI(offsetMBB, DL, TII->get(X86::JMP_4))
10461 // Emit code to use overflow area
10464 // Load the overflow_area address into a register.
10465 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
10466 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
10471 .addOperand(Segment)
10472 .setMemRefs(MMOBegin, MMOEnd);
10474 // If we need to align it, do so. Otherwise, just copy the address
10475 // to OverflowDestReg.
10477 // Align the overflow address
10478 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
10479 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
10481 // aligned_addr = (addr + (align-1)) & ~(align-1)
10482 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
10483 .addReg(OverflowAddrReg)
10486 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
10488 .addImm(~(uint64_t)(Align-1));
10490 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
10491 .addReg(OverflowAddrReg);
10494 // Compute the next overflow address after this argument.
10495 // (the overflow address should be kept 8-byte aligned)
10496 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
10497 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
10498 .addReg(OverflowDestReg)
10499 .addImm(ArgSizeA8);
10501 // Store the new overflow address.
10502 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
10507 .addOperand(Segment)
10508 .addReg(NextAddrReg)
10509 .setMemRefs(MMOBegin, MMOEnd);
10511 // If we branched, emit the PHI to the front of endMBB.
10513 BuildMI(*endMBB, endMBB->begin(), DL,
10514 TII->get(X86::PHI), DestReg)
10515 .addReg(OffsetDestReg).addMBB(offsetMBB)
10516 .addReg(OverflowDestReg).addMBB(overflowMBB);
10519 // Erase the pseudo instruction
10520 MI->eraseFromParent();
10525 MachineBasicBlock *
10526 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
10528 MachineBasicBlock *MBB) const {
10529 // Emit code to save XMM registers to the stack. The ABI says that the
10530 // number of registers to save is given in %al, so it's theoretically
10531 // possible to do an indirect jump trick to avoid saving all of them,
10532 // however this code takes a simpler approach and just executes all
10533 // of the stores if %al is non-zero. It's less code, and it's probably
10534 // easier on the hardware branch predictor, and stores aren't all that
10535 // expensive anyway.
10537 // Create the new basic blocks. One block contains all the XMM stores,
10538 // and one block is the final destination regardless of whether any
10539 // stores were performed.
10540 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10541 MachineFunction *F = MBB->getParent();
10542 MachineFunction::iterator MBBIter = MBB;
10544 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
10545 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
10546 F->insert(MBBIter, XMMSaveMBB);
10547 F->insert(MBBIter, EndMBB);
10549 // Transfer the remainder of MBB and its successor edges to EndMBB.
10550 EndMBB->splice(EndMBB->begin(), MBB,
10551 llvm::next(MachineBasicBlock::iterator(MI)),
10553 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
10555 // The original block will now fall through to the XMM save block.
10556 MBB->addSuccessor(XMMSaveMBB);
10557 // The XMMSaveMBB will fall through to the end block.
10558 XMMSaveMBB->addSuccessor(EndMBB);
10560 // Now add the instructions.
10561 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10562 DebugLoc DL = MI->getDebugLoc();
10564 unsigned CountReg = MI->getOperand(0).getReg();
10565 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
10566 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
10568 if (!Subtarget->isTargetWin64()) {
10569 // If %al is 0, branch around the XMM save block.
10570 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
10571 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
10572 MBB->addSuccessor(EndMBB);
10575 // In the XMM save block, save all the XMM argument registers.
10576 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
10577 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
10578 MachineMemOperand *MMO =
10579 F->getMachineMemOperand(
10580 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
10581 MachineMemOperand::MOStore,
10582 /*Size=*/16, /*Align=*/16);
10583 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
10584 .addFrameIndex(RegSaveFrameIndex)
10585 .addImm(/*Scale=*/1)
10586 .addReg(/*IndexReg=*/0)
10587 .addImm(/*Disp=*/Offset)
10588 .addReg(/*Segment=*/0)
10589 .addReg(MI->getOperand(i).getReg())
10590 .addMemOperand(MMO);
10593 MI->eraseFromParent(); // The pseudo instruction is gone now.
10598 MachineBasicBlock *
10599 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
10600 MachineBasicBlock *BB) const {
10601 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10602 DebugLoc DL = MI->getDebugLoc();
10604 // To "insert" a SELECT_CC instruction, we actually have to insert the
10605 // diamond control-flow pattern. The incoming instruction knows the
10606 // destination vreg to set, the condition code register to branch on, the
10607 // true/false values to select between, and a branch opcode to use.
10608 const BasicBlock *LLVM_BB = BB->getBasicBlock();
10609 MachineFunction::iterator It = BB;
10615 // cmpTY ccX, r1, r2
10617 // fallthrough --> copy0MBB
10618 MachineBasicBlock *thisMBB = BB;
10619 MachineFunction *F = BB->getParent();
10620 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
10621 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
10622 F->insert(It, copy0MBB);
10623 F->insert(It, sinkMBB);
10625 // If the EFLAGS register isn't dead in the terminator, then claim that it's
10626 // live into the sink and copy blocks.
10627 const MachineFunction *MF = BB->getParent();
10628 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
10629 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
10631 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
10632 const MachineOperand &MO = MI->getOperand(I);
10633 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
10634 unsigned Reg = MO.getReg();
10635 if (Reg != X86::EFLAGS) continue;
10636 copy0MBB->addLiveIn(Reg);
10637 sinkMBB->addLiveIn(Reg);
10640 // Transfer the remainder of BB and its successor edges to sinkMBB.
10641 sinkMBB->splice(sinkMBB->begin(), BB,
10642 llvm::next(MachineBasicBlock::iterator(MI)),
10644 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
10646 // Add the true and fallthrough blocks as its successors.
10647 BB->addSuccessor(copy0MBB);
10648 BB->addSuccessor(sinkMBB);
10650 // Create the conditional branch instruction.
10652 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
10653 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
10656 // %FalseValue = ...
10657 // # fallthrough to sinkMBB
10658 copy0MBB->addSuccessor(sinkMBB);
10661 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
10663 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
10664 TII->get(X86::PHI), MI->getOperand(0).getReg())
10665 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
10666 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
10668 MI->eraseFromParent(); // The pseudo instruction is gone now.
10672 MachineBasicBlock *
10673 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
10674 MachineBasicBlock *BB) const {
10675 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10676 DebugLoc DL = MI->getDebugLoc();
10678 assert(!Subtarget->isTargetEnvMacho());
10680 // The lowering is pretty easy: we're just emitting the call to _alloca. The
10681 // non-trivial part is impdef of ESP.
10683 if (Subtarget->isTargetWin64()) {
10684 if (Subtarget->isTargetCygMing()) {
10685 // ___chkstk(Mingw64):
10686 // Clobbers R10, R11, RAX and EFLAGS.
10688 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
10689 .addExternalSymbol("___chkstk")
10690 .addReg(X86::RAX, RegState::Implicit)
10691 .addReg(X86::RSP, RegState::Implicit)
10692 .addReg(X86::RAX, RegState::Define | RegState::Implicit)
10693 .addReg(X86::RSP, RegState::Define | RegState::Implicit)
10694 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
10696 // __chkstk(MSVCRT): does not update stack pointer.
10697 // Clobbers R10, R11 and EFLAGS.
10698 // FIXME: RAX(allocated size) might be reused and not killed.
10699 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
10700 .addExternalSymbol("__chkstk")
10701 .addReg(X86::RAX, RegState::Implicit)
10702 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
10703 // RAX has the offset to subtracted from RSP.
10704 BuildMI(*BB, MI, DL, TII->get(X86::SUB64rr), X86::RSP)
10709 const char *StackProbeSymbol =
10710 Subtarget->isTargetWindows() ? "_chkstk" : "_alloca";
10712 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
10713 .addExternalSymbol(StackProbeSymbol)
10714 .addReg(X86::EAX, RegState::Implicit)
10715 .addReg(X86::ESP, RegState::Implicit)
10716 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
10717 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
10718 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
10721 MI->eraseFromParent(); // The pseudo instruction is gone now.
10725 MachineBasicBlock *
10726 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
10727 MachineBasicBlock *BB) const {
10728 // This is pretty easy. We're taking the value that we received from
10729 // our load from the relocation, sticking it in either RDI (x86-64)
10730 // or EAX and doing an indirect call. The return value will then
10731 // be in the normal return register.
10732 const X86InstrInfo *TII
10733 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
10734 DebugLoc DL = MI->getDebugLoc();
10735 MachineFunction *F = BB->getParent();
10737 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
10738 assert(MI->getOperand(3).isGlobal() && "This should be a global");
10740 if (Subtarget->is64Bit()) {
10741 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10742 TII->get(X86::MOV64rm), X86::RDI)
10744 .addImm(0).addReg(0)
10745 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10746 MI->getOperand(3).getTargetFlags())
10748 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
10749 addDirectMem(MIB, X86::RDI);
10750 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
10751 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10752 TII->get(X86::MOV32rm), X86::EAX)
10754 .addImm(0).addReg(0)
10755 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10756 MI->getOperand(3).getTargetFlags())
10758 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
10759 addDirectMem(MIB, X86::EAX);
10761 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10762 TII->get(X86::MOV32rm), X86::EAX)
10763 .addReg(TII->getGlobalBaseReg(F))
10764 .addImm(0).addReg(0)
10765 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10766 MI->getOperand(3).getTargetFlags())
10768 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
10769 addDirectMem(MIB, X86::EAX);
10772 MI->eraseFromParent(); // The pseudo instruction is gone now.
10776 MachineBasicBlock *
10777 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
10778 MachineBasicBlock *BB) const {
10779 switch (MI->getOpcode()) {
10780 default: assert(false && "Unexpected instr type to insert");
10781 case X86::TAILJMPd64:
10782 case X86::TAILJMPr64:
10783 case X86::TAILJMPm64:
10784 assert(!"TAILJMP64 would not be touched here.");
10785 case X86::TCRETURNdi64:
10786 case X86::TCRETURNri64:
10787 case X86::TCRETURNmi64:
10788 // Defs of TCRETURNxx64 has Win64's callee-saved registers, as subset.
10789 // On AMD64, additional defs should be added before register allocation.
10790 if (!Subtarget->isTargetWin64()) {
10791 MI->addRegisterDefined(X86::RSI);
10792 MI->addRegisterDefined(X86::RDI);
10793 MI->addRegisterDefined(X86::XMM6);
10794 MI->addRegisterDefined(X86::XMM7);
10795 MI->addRegisterDefined(X86::XMM8);
10796 MI->addRegisterDefined(X86::XMM9);
10797 MI->addRegisterDefined(X86::XMM10);
10798 MI->addRegisterDefined(X86::XMM11);
10799 MI->addRegisterDefined(X86::XMM12);
10800 MI->addRegisterDefined(X86::XMM13);
10801 MI->addRegisterDefined(X86::XMM14);
10802 MI->addRegisterDefined(X86::XMM15);
10805 case X86::WIN_ALLOCA:
10806 return EmitLoweredWinAlloca(MI, BB);
10807 case X86::TLSCall_32:
10808 case X86::TLSCall_64:
10809 return EmitLoweredTLSCall(MI, BB);
10810 case X86::CMOV_GR8:
10811 case X86::CMOV_FR32:
10812 case X86::CMOV_FR64:
10813 case X86::CMOV_V4F32:
10814 case X86::CMOV_V2F64:
10815 case X86::CMOV_V2I64:
10816 case X86::CMOV_GR16:
10817 case X86::CMOV_GR32:
10818 case X86::CMOV_RFP32:
10819 case X86::CMOV_RFP64:
10820 case X86::CMOV_RFP80:
10821 return EmitLoweredSelect(MI, BB);
10823 case X86::FP32_TO_INT16_IN_MEM:
10824 case X86::FP32_TO_INT32_IN_MEM:
10825 case X86::FP32_TO_INT64_IN_MEM:
10826 case X86::FP64_TO_INT16_IN_MEM:
10827 case X86::FP64_TO_INT32_IN_MEM:
10828 case X86::FP64_TO_INT64_IN_MEM:
10829 case X86::FP80_TO_INT16_IN_MEM:
10830 case X86::FP80_TO_INT32_IN_MEM:
10831 case X86::FP80_TO_INT64_IN_MEM: {
10832 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10833 DebugLoc DL = MI->getDebugLoc();
10835 // Change the floating point control register to use "round towards zero"
10836 // mode when truncating to an integer value.
10837 MachineFunction *F = BB->getParent();
10838 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
10839 addFrameReference(BuildMI(*BB, MI, DL,
10840 TII->get(X86::FNSTCW16m)), CWFrameIdx);
10842 // Load the old value of the high byte of the control word...
10844 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
10845 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
10848 // Set the high part to be round to zero...
10849 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
10852 // Reload the modified control word now...
10853 addFrameReference(BuildMI(*BB, MI, DL,
10854 TII->get(X86::FLDCW16m)), CWFrameIdx);
10856 // Restore the memory image of control word to original value
10857 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
10860 // Get the X86 opcode to use.
10862 switch (MI->getOpcode()) {
10863 default: llvm_unreachable("illegal opcode!");
10864 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
10865 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
10866 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
10867 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
10868 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
10869 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
10870 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
10871 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
10872 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
10876 MachineOperand &Op = MI->getOperand(0);
10878 AM.BaseType = X86AddressMode::RegBase;
10879 AM.Base.Reg = Op.getReg();
10881 AM.BaseType = X86AddressMode::FrameIndexBase;
10882 AM.Base.FrameIndex = Op.getIndex();
10884 Op = MI->getOperand(1);
10886 AM.Scale = Op.getImm();
10887 Op = MI->getOperand(2);
10889 AM.IndexReg = Op.getImm();
10890 Op = MI->getOperand(3);
10891 if (Op.isGlobal()) {
10892 AM.GV = Op.getGlobal();
10894 AM.Disp = Op.getImm();
10896 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
10897 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
10899 // Reload the original control word now.
10900 addFrameReference(BuildMI(*BB, MI, DL,
10901 TII->get(X86::FLDCW16m)), CWFrameIdx);
10903 MI->eraseFromParent(); // The pseudo instruction is gone now.
10906 // String/text processing lowering.
10907 case X86::PCMPISTRM128REG:
10908 case X86::VPCMPISTRM128REG:
10909 return EmitPCMP(MI, BB, 3, false /* in-mem */);
10910 case X86::PCMPISTRM128MEM:
10911 case X86::VPCMPISTRM128MEM:
10912 return EmitPCMP(MI, BB, 3, true /* in-mem */);
10913 case X86::PCMPESTRM128REG:
10914 case X86::VPCMPESTRM128REG:
10915 return EmitPCMP(MI, BB, 5, false /* in mem */);
10916 case X86::PCMPESTRM128MEM:
10917 case X86::VPCMPESTRM128MEM:
10918 return EmitPCMP(MI, BB, 5, true /* in mem */);
10920 // Thread synchronization.
10922 return EmitMonitor(MI, BB);
10924 return EmitMwait(MI, BB);
10926 // Atomic Lowering.
10927 case X86::ATOMAND32:
10928 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
10929 X86::AND32ri, X86::MOV32rm,
10931 X86::NOT32r, X86::EAX,
10932 X86::GR32RegisterClass);
10933 case X86::ATOMOR32:
10934 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
10935 X86::OR32ri, X86::MOV32rm,
10937 X86::NOT32r, X86::EAX,
10938 X86::GR32RegisterClass);
10939 case X86::ATOMXOR32:
10940 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
10941 X86::XOR32ri, X86::MOV32rm,
10943 X86::NOT32r, X86::EAX,
10944 X86::GR32RegisterClass);
10945 case X86::ATOMNAND32:
10946 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
10947 X86::AND32ri, X86::MOV32rm,
10949 X86::NOT32r, X86::EAX,
10950 X86::GR32RegisterClass, true);
10951 case X86::ATOMMIN32:
10952 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
10953 case X86::ATOMMAX32:
10954 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
10955 case X86::ATOMUMIN32:
10956 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
10957 case X86::ATOMUMAX32:
10958 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
10960 case X86::ATOMAND16:
10961 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
10962 X86::AND16ri, X86::MOV16rm,
10964 X86::NOT16r, X86::AX,
10965 X86::GR16RegisterClass);
10966 case X86::ATOMOR16:
10967 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
10968 X86::OR16ri, X86::MOV16rm,
10970 X86::NOT16r, X86::AX,
10971 X86::GR16RegisterClass);
10972 case X86::ATOMXOR16:
10973 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
10974 X86::XOR16ri, X86::MOV16rm,
10976 X86::NOT16r, X86::AX,
10977 X86::GR16RegisterClass);
10978 case X86::ATOMNAND16:
10979 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
10980 X86::AND16ri, X86::MOV16rm,
10982 X86::NOT16r, X86::AX,
10983 X86::GR16RegisterClass, true);
10984 case X86::ATOMMIN16:
10985 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
10986 case X86::ATOMMAX16:
10987 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
10988 case X86::ATOMUMIN16:
10989 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
10990 case X86::ATOMUMAX16:
10991 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
10993 case X86::ATOMAND8:
10994 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
10995 X86::AND8ri, X86::MOV8rm,
10997 X86::NOT8r, X86::AL,
10998 X86::GR8RegisterClass);
11000 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
11001 X86::OR8ri, X86::MOV8rm,
11003 X86::NOT8r, X86::AL,
11004 X86::GR8RegisterClass);
11005 case X86::ATOMXOR8:
11006 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
11007 X86::XOR8ri, X86::MOV8rm,
11009 X86::NOT8r, X86::AL,
11010 X86::GR8RegisterClass);
11011 case X86::ATOMNAND8:
11012 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
11013 X86::AND8ri, X86::MOV8rm,
11015 X86::NOT8r, X86::AL,
11016 X86::GR8RegisterClass, true);
11017 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
11018 // This group is for 64-bit host.
11019 case X86::ATOMAND64:
11020 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
11021 X86::AND64ri32, X86::MOV64rm,
11023 X86::NOT64r, X86::RAX,
11024 X86::GR64RegisterClass);
11025 case X86::ATOMOR64:
11026 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
11027 X86::OR64ri32, X86::MOV64rm,
11029 X86::NOT64r, X86::RAX,
11030 X86::GR64RegisterClass);
11031 case X86::ATOMXOR64:
11032 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
11033 X86::XOR64ri32, X86::MOV64rm,
11035 X86::NOT64r, X86::RAX,
11036 X86::GR64RegisterClass);
11037 case X86::ATOMNAND64:
11038 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
11039 X86::AND64ri32, X86::MOV64rm,
11041 X86::NOT64r, X86::RAX,
11042 X86::GR64RegisterClass, true);
11043 case X86::ATOMMIN64:
11044 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
11045 case X86::ATOMMAX64:
11046 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
11047 case X86::ATOMUMIN64:
11048 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
11049 case X86::ATOMUMAX64:
11050 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
11052 // This group does 64-bit operations on a 32-bit host.
11053 case X86::ATOMAND6432:
11054 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11055 X86::AND32rr, X86::AND32rr,
11056 X86::AND32ri, X86::AND32ri,
11058 case X86::ATOMOR6432:
11059 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11060 X86::OR32rr, X86::OR32rr,
11061 X86::OR32ri, X86::OR32ri,
11063 case X86::ATOMXOR6432:
11064 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11065 X86::XOR32rr, X86::XOR32rr,
11066 X86::XOR32ri, X86::XOR32ri,
11068 case X86::ATOMNAND6432:
11069 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11070 X86::AND32rr, X86::AND32rr,
11071 X86::AND32ri, X86::AND32ri,
11073 case X86::ATOMADD6432:
11074 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11075 X86::ADD32rr, X86::ADC32rr,
11076 X86::ADD32ri, X86::ADC32ri,
11078 case X86::ATOMSUB6432:
11079 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11080 X86::SUB32rr, X86::SBB32rr,
11081 X86::SUB32ri, X86::SBB32ri,
11083 case X86::ATOMSWAP6432:
11084 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11085 X86::MOV32rr, X86::MOV32rr,
11086 X86::MOV32ri, X86::MOV32ri,
11088 case X86::VASTART_SAVE_XMM_REGS:
11089 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
11091 case X86::VAARG_64:
11092 return EmitVAARG64WithCustomInserter(MI, BB);
11096 //===----------------------------------------------------------------------===//
11097 // X86 Optimization Hooks
11098 //===----------------------------------------------------------------------===//
11100 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
11104 const SelectionDAG &DAG,
11105 unsigned Depth) const {
11106 unsigned Opc = Op.getOpcode();
11107 assert((Opc >= ISD::BUILTIN_OP_END ||
11108 Opc == ISD::INTRINSIC_WO_CHAIN ||
11109 Opc == ISD::INTRINSIC_W_CHAIN ||
11110 Opc == ISD::INTRINSIC_VOID) &&
11111 "Should use MaskedValueIsZero if you don't know whether Op"
11112 " is a target node!");
11114 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
11128 // These nodes' second result is a boolean.
11129 if (Op.getResNo() == 0)
11132 case X86ISD::SETCC:
11133 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
11134 Mask.getBitWidth() - 1);
11139 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
11140 unsigned Depth) const {
11141 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
11142 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
11143 return Op.getValueType().getScalarType().getSizeInBits();
11149 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
11150 /// node is a GlobalAddress + offset.
11151 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
11152 const GlobalValue* &GA,
11153 int64_t &Offset) const {
11154 if (N->getOpcode() == X86ISD::Wrapper) {
11155 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
11156 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
11157 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
11161 return TargetLowering::isGAPlusOffset(N, GA, Offset);
11164 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
11165 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
11166 TargetLowering::DAGCombinerInfo &DCI) {
11167 DebugLoc dl = N->getDebugLoc();
11168 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
11169 SDValue V1 = SVOp->getOperand(0);
11170 SDValue V2 = SVOp->getOperand(1);
11171 EVT VT = SVOp->getValueType(0);
11173 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
11174 V2.getOpcode() == ISD::CONCAT_VECTORS) {
11178 // V UNDEF BUILD_VECTOR UNDEF
11180 // CONCAT_VECTOR CONCAT_VECTOR
11183 // RESULT: V + zero extended
11185 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
11186 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
11187 V1.getOperand(1).getOpcode() != ISD::UNDEF)
11190 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
11193 // To match the shuffle mask, the first half of the mask should
11194 // be exactly the first vector, and all the rest a splat with the
11195 // first element of the second one.
11196 int NumElems = VT.getVectorNumElements();
11197 for (int i = 0; i < NumElems/2; ++i)
11198 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
11199 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
11202 // Emit a zeroed vector and insert the desired subvector on its
11204 SDValue Zeros = getZeroVector(VT, true /* HasSSE2 */, DAG, dl);
11205 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0),
11206 DAG.getConstant(0, MVT::i32), DAG, dl);
11207 return DCI.CombineTo(N, InsV);
11213 /// PerformShuffleCombine - Performs several different shuffle combines.
11214 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
11215 TargetLowering::DAGCombinerInfo &DCI) {
11216 DebugLoc dl = N->getDebugLoc();
11217 EVT VT = N->getValueType(0);
11219 // Don't create instructions with illegal types after legalize types has run.
11220 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11221 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
11224 // Only handle pure VECTOR_SHUFFLE nodes.
11225 if (VT.getSizeInBits() == 256 && N->getOpcode() == ISD::VECTOR_SHUFFLE)
11226 return PerformShuffleCombine256(N, DAG, DCI);
11228 // Only handle 128 wide vector from here on.
11229 if (VT.getSizeInBits() != 128)
11232 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
11233 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
11234 // consecutive, non-overlapping, and in the right order.
11235 SmallVector<SDValue, 16> Elts;
11236 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
11237 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
11239 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
11242 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
11243 /// generation and convert it from being a bunch of shuffles and extracts
11244 /// to a simple store and scalar loads to extract the elements.
11245 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
11246 const TargetLowering &TLI) {
11247 SDValue InputVector = N->getOperand(0);
11249 // Only operate on vectors of 4 elements, where the alternative shuffling
11250 // gets to be more expensive.
11251 if (InputVector.getValueType() != MVT::v4i32)
11254 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
11255 // single use which is a sign-extend or zero-extend, and all elements are
11257 SmallVector<SDNode *, 4> Uses;
11258 unsigned ExtractedElements = 0;
11259 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
11260 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
11261 if (UI.getUse().getResNo() != InputVector.getResNo())
11264 SDNode *Extract = *UI;
11265 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
11268 if (Extract->getValueType(0) != MVT::i32)
11270 if (!Extract->hasOneUse())
11272 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
11273 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
11275 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
11278 // Record which element was extracted.
11279 ExtractedElements |=
11280 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
11282 Uses.push_back(Extract);
11285 // If not all the elements were used, this may not be worthwhile.
11286 if (ExtractedElements != 15)
11289 // Ok, we've now decided to do the transformation.
11290 DebugLoc dl = InputVector.getDebugLoc();
11292 // Store the value to a temporary stack slot.
11293 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
11294 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
11295 MachinePointerInfo(), false, false, 0);
11297 // Replace each use (extract) with a load of the appropriate element.
11298 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
11299 UE = Uses.end(); UI != UE; ++UI) {
11300 SDNode *Extract = *UI;
11302 // cOMpute the element's address.
11303 SDValue Idx = Extract->getOperand(1);
11305 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
11306 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
11307 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
11309 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
11310 StackPtr, OffsetVal);
11312 // Load the scalar.
11313 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
11314 ScalarAddr, MachinePointerInfo(),
11317 // Replace the exact with the load.
11318 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
11321 // The replacement was made in place; don't return anything.
11325 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
11326 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
11327 const X86Subtarget *Subtarget) {
11328 DebugLoc DL = N->getDebugLoc();
11329 SDValue Cond = N->getOperand(0);
11330 // Get the LHS/RHS of the select.
11331 SDValue LHS = N->getOperand(1);
11332 SDValue RHS = N->getOperand(2);
11334 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
11335 // instructions match the semantics of the common C idiom x<y?x:y but not
11336 // x<=y?x:y, because of how they handle negative zero (which can be
11337 // ignored in unsafe-math mode).
11338 if (Subtarget->hasSSE2() &&
11339 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
11340 Cond.getOpcode() == ISD::SETCC) {
11341 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
11343 unsigned Opcode = 0;
11344 // Check for x CC y ? x : y.
11345 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
11346 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
11350 // Converting this to a min would handle NaNs incorrectly, and swapping
11351 // the operands would cause it to handle comparisons between positive
11352 // and negative zero incorrectly.
11353 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
11354 if (!UnsafeFPMath &&
11355 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
11357 std::swap(LHS, RHS);
11359 Opcode = X86ISD::FMIN;
11362 // Converting this to a min would handle comparisons between positive
11363 // and negative zero incorrectly.
11364 if (!UnsafeFPMath &&
11365 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
11367 Opcode = X86ISD::FMIN;
11370 // Converting this to a min would handle both negative zeros and NaNs
11371 // incorrectly, but we can swap the operands to fix both.
11372 std::swap(LHS, RHS);
11376 Opcode = X86ISD::FMIN;
11380 // Converting this to a max would handle comparisons between positive
11381 // and negative zero incorrectly.
11382 if (!UnsafeFPMath &&
11383 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
11385 Opcode = X86ISD::FMAX;
11388 // Converting this to a max would handle NaNs incorrectly, and swapping
11389 // the operands would cause it to handle comparisons between positive
11390 // and negative zero incorrectly.
11391 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
11392 if (!UnsafeFPMath &&
11393 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
11395 std::swap(LHS, RHS);
11397 Opcode = X86ISD::FMAX;
11400 // Converting this to a max would handle both negative zeros and NaNs
11401 // incorrectly, but we can swap the operands to fix both.
11402 std::swap(LHS, RHS);
11406 Opcode = X86ISD::FMAX;
11409 // Check for x CC y ? y : x -- a min/max with reversed arms.
11410 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
11411 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
11415 // Converting this to a min would handle comparisons between positive
11416 // and negative zero incorrectly, and swapping the operands would
11417 // cause it to handle NaNs incorrectly.
11418 if (!UnsafeFPMath &&
11419 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
11420 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11422 std::swap(LHS, RHS);
11424 Opcode = X86ISD::FMIN;
11427 // Converting this to a min would handle NaNs incorrectly.
11428 if (!UnsafeFPMath &&
11429 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
11431 Opcode = X86ISD::FMIN;
11434 // Converting this to a min would handle both negative zeros and NaNs
11435 // incorrectly, but we can swap the operands to fix both.
11436 std::swap(LHS, RHS);
11440 Opcode = X86ISD::FMIN;
11444 // Converting this to a max would handle NaNs incorrectly.
11445 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11447 Opcode = X86ISD::FMAX;
11450 // Converting this to a max would handle comparisons between positive
11451 // and negative zero incorrectly, and swapping the operands would
11452 // cause it to handle NaNs incorrectly.
11453 if (!UnsafeFPMath &&
11454 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
11455 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11457 std::swap(LHS, RHS);
11459 Opcode = X86ISD::FMAX;
11462 // Converting this to a max would handle both negative zeros and NaNs
11463 // incorrectly, but we can swap the operands to fix both.
11464 std::swap(LHS, RHS);
11468 Opcode = X86ISD::FMAX;
11474 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
11477 // If this is a select between two integer constants, try to do some
11479 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
11480 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
11481 // Don't do this for crazy integer types.
11482 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
11483 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
11484 // so that TrueC (the true value) is larger than FalseC.
11485 bool NeedsCondInvert = false;
11487 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
11488 // Efficiently invertible.
11489 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
11490 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
11491 isa<ConstantSDNode>(Cond.getOperand(1))))) {
11492 NeedsCondInvert = true;
11493 std::swap(TrueC, FalseC);
11496 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
11497 if (FalseC->getAPIntValue() == 0 &&
11498 TrueC->getAPIntValue().isPowerOf2()) {
11499 if (NeedsCondInvert) // Invert the condition if needed.
11500 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11501 DAG.getConstant(1, Cond.getValueType()));
11503 // Zero extend the condition if needed.
11504 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
11506 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
11507 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
11508 DAG.getConstant(ShAmt, MVT::i8));
11511 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
11512 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
11513 if (NeedsCondInvert) // Invert the condition if needed.
11514 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11515 DAG.getConstant(1, Cond.getValueType()));
11517 // Zero extend the condition if needed.
11518 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
11519 FalseC->getValueType(0), Cond);
11520 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11521 SDValue(FalseC, 0));
11524 // Optimize cases that will turn into an LEA instruction. This requires
11525 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
11526 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
11527 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
11528 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
11530 bool isFastMultiplier = false;
11532 switch ((unsigned char)Diff) {
11534 case 1: // result = add base, cond
11535 case 2: // result = lea base( , cond*2)
11536 case 3: // result = lea base(cond, cond*2)
11537 case 4: // result = lea base( , cond*4)
11538 case 5: // result = lea base(cond, cond*4)
11539 case 8: // result = lea base( , cond*8)
11540 case 9: // result = lea base(cond, cond*8)
11541 isFastMultiplier = true;
11546 if (isFastMultiplier) {
11547 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
11548 if (NeedsCondInvert) // Invert the condition if needed.
11549 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11550 DAG.getConstant(1, Cond.getValueType()));
11552 // Zero extend the condition if needed.
11553 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
11555 // Scale the condition by the difference.
11557 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
11558 DAG.getConstant(Diff, Cond.getValueType()));
11560 // Add the base if non-zero.
11561 if (FalseC->getAPIntValue() != 0)
11562 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11563 SDValue(FalseC, 0));
11573 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
11574 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
11575 TargetLowering::DAGCombinerInfo &DCI) {
11576 DebugLoc DL = N->getDebugLoc();
11578 // If the flag operand isn't dead, don't touch this CMOV.
11579 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
11582 SDValue FalseOp = N->getOperand(0);
11583 SDValue TrueOp = N->getOperand(1);
11584 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
11585 SDValue Cond = N->getOperand(3);
11586 if (CC == X86::COND_E || CC == X86::COND_NE) {
11587 switch (Cond.getOpcode()) {
11591 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
11592 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
11593 return (CC == X86::COND_E) ? FalseOp : TrueOp;
11597 // If this is a select between two integer constants, try to do some
11598 // optimizations. Note that the operands are ordered the opposite of SELECT
11600 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
11601 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
11602 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
11603 // larger than FalseC (the false value).
11604 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
11605 CC = X86::GetOppositeBranchCondition(CC);
11606 std::swap(TrueC, FalseC);
11609 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
11610 // This is efficient for any integer data type (including i8/i16) and
11612 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
11613 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11614 DAG.getConstant(CC, MVT::i8), Cond);
11616 // Zero extend the condition if needed.
11617 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
11619 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
11620 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
11621 DAG.getConstant(ShAmt, MVT::i8));
11622 if (N->getNumValues() == 2) // Dead flag value?
11623 return DCI.CombineTo(N, Cond, SDValue());
11627 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
11628 // for any integer data type, including i8/i16.
11629 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
11630 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11631 DAG.getConstant(CC, MVT::i8), Cond);
11633 // Zero extend the condition if needed.
11634 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
11635 FalseC->getValueType(0), Cond);
11636 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11637 SDValue(FalseC, 0));
11639 if (N->getNumValues() == 2) // Dead flag value?
11640 return DCI.CombineTo(N, Cond, SDValue());
11644 // Optimize cases that will turn into an LEA instruction. This requires
11645 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
11646 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
11647 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
11648 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
11650 bool isFastMultiplier = false;
11652 switch ((unsigned char)Diff) {
11654 case 1: // result = add base, cond
11655 case 2: // result = lea base( , cond*2)
11656 case 3: // result = lea base(cond, cond*2)
11657 case 4: // result = lea base( , cond*4)
11658 case 5: // result = lea base(cond, cond*4)
11659 case 8: // result = lea base( , cond*8)
11660 case 9: // result = lea base(cond, cond*8)
11661 isFastMultiplier = true;
11666 if (isFastMultiplier) {
11667 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
11668 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11669 DAG.getConstant(CC, MVT::i8), Cond);
11670 // Zero extend the condition if needed.
11671 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
11673 // Scale the condition by the difference.
11675 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
11676 DAG.getConstant(Diff, Cond.getValueType()));
11678 // Add the base if non-zero.
11679 if (FalseC->getAPIntValue() != 0)
11680 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11681 SDValue(FalseC, 0));
11682 if (N->getNumValues() == 2) // Dead flag value?
11683 return DCI.CombineTo(N, Cond, SDValue());
11693 /// PerformMulCombine - Optimize a single multiply with constant into two
11694 /// in order to implement it with two cheaper instructions, e.g.
11695 /// LEA + SHL, LEA + LEA.
11696 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
11697 TargetLowering::DAGCombinerInfo &DCI) {
11698 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
11701 EVT VT = N->getValueType(0);
11702 if (VT != MVT::i64)
11705 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
11708 uint64_t MulAmt = C->getZExtValue();
11709 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
11712 uint64_t MulAmt1 = 0;
11713 uint64_t MulAmt2 = 0;
11714 if ((MulAmt % 9) == 0) {
11716 MulAmt2 = MulAmt / 9;
11717 } else if ((MulAmt % 5) == 0) {
11719 MulAmt2 = MulAmt / 5;
11720 } else if ((MulAmt % 3) == 0) {
11722 MulAmt2 = MulAmt / 3;
11725 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
11726 DebugLoc DL = N->getDebugLoc();
11728 if (isPowerOf2_64(MulAmt2) &&
11729 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
11730 // If second multiplifer is pow2, issue it first. We want the multiply by
11731 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
11733 std::swap(MulAmt1, MulAmt2);
11736 if (isPowerOf2_64(MulAmt1))
11737 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
11738 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
11740 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
11741 DAG.getConstant(MulAmt1, VT));
11743 if (isPowerOf2_64(MulAmt2))
11744 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
11745 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
11747 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
11748 DAG.getConstant(MulAmt2, VT));
11750 // Do not add new nodes to DAG combiner worklist.
11751 DCI.CombineTo(N, NewMul, false);
11756 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
11757 SDValue N0 = N->getOperand(0);
11758 SDValue N1 = N->getOperand(1);
11759 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
11760 EVT VT = N0.getValueType();
11762 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
11763 // since the result of setcc_c is all zero's or all ones.
11764 if (N1C && N0.getOpcode() == ISD::AND &&
11765 N0.getOperand(1).getOpcode() == ISD::Constant) {
11766 SDValue N00 = N0.getOperand(0);
11767 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
11768 ((N00.getOpcode() == ISD::ANY_EXTEND ||
11769 N00.getOpcode() == ISD::ZERO_EXTEND) &&
11770 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
11771 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
11772 APInt ShAmt = N1C->getAPIntValue();
11773 Mask = Mask.shl(ShAmt);
11775 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
11776 N00, DAG.getConstant(Mask, VT));
11783 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
11785 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
11786 const X86Subtarget *Subtarget) {
11787 EVT VT = N->getValueType(0);
11788 if (!VT.isVector() && VT.isInteger() &&
11789 N->getOpcode() == ISD::SHL)
11790 return PerformSHLCombine(N, DAG);
11792 // On X86 with SSE2 support, we can transform this to a vector shift if
11793 // all elements are shifted by the same amount. We can't do this in legalize
11794 // because the a constant vector is typically transformed to a constant pool
11795 // so we have no knowledge of the shift amount.
11796 if (!Subtarget->hasSSE2())
11799 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
11802 SDValue ShAmtOp = N->getOperand(1);
11803 EVT EltVT = VT.getVectorElementType();
11804 DebugLoc DL = N->getDebugLoc();
11805 SDValue BaseShAmt = SDValue();
11806 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
11807 unsigned NumElts = VT.getVectorNumElements();
11809 for (; i != NumElts; ++i) {
11810 SDValue Arg = ShAmtOp.getOperand(i);
11811 if (Arg.getOpcode() == ISD::UNDEF) continue;
11815 for (; i != NumElts; ++i) {
11816 SDValue Arg = ShAmtOp.getOperand(i);
11817 if (Arg.getOpcode() == ISD::UNDEF) continue;
11818 if (Arg != BaseShAmt) {
11822 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
11823 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
11824 SDValue InVec = ShAmtOp.getOperand(0);
11825 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
11826 unsigned NumElts = InVec.getValueType().getVectorNumElements();
11828 for (; i != NumElts; ++i) {
11829 SDValue Arg = InVec.getOperand(i);
11830 if (Arg.getOpcode() == ISD::UNDEF) continue;
11834 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
11835 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
11836 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
11837 if (C->getZExtValue() == SplatIdx)
11838 BaseShAmt = InVec.getOperand(1);
11841 if (BaseShAmt.getNode() == 0)
11842 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
11843 DAG.getIntPtrConstant(0));
11847 // The shift amount is an i32.
11848 if (EltVT.bitsGT(MVT::i32))
11849 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
11850 else if (EltVT.bitsLT(MVT::i32))
11851 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
11853 // The shift amount is identical so we can do a vector shift.
11854 SDValue ValOp = N->getOperand(0);
11855 switch (N->getOpcode()) {
11857 llvm_unreachable("Unknown shift opcode!");
11860 if (VT == MVT::v2i64)
11861 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11862 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
11864 if (VT == MVT::v4i32)
11865 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11866 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
11868 if (VT == MVT::v8i16)
11869 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11870 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
11874 if (VT == MVT::v4i32)
11875 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11876 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
11878 if (VT == MVT::v8i16)
11879 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11880 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
11884 if (VT == MVT::v2i64)
11885 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11886 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
11888 if (VT == MVT::v4i32)
11889 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11890 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
11892 if (VT == MVT::v8i16)
11893 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11894 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
11902 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
11903 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
11904 // and friends. Likewise for OR -> CMPNEQSS.
11905 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
11906 TargetLowering::DAGCombinerInfo &DCI,
11907 const X86Subtarget *Subtarget) {
11910 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
11911 // we're requiring SSE2 for both.
11912 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
11913 SDValue N0 = N->getOperand(0);
11914 SDValue N1 = N->getOperand(1);
11915 SDValue CMP0 = N0->getOperand(1);
11916 SDValue CMP1 = N1->getOperand(1);
11917 DebugLoc DL = N->getDebugLoc();
11919 // The SETCCs should both refer to the same CMP.
11920 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
11923 SDValue CMP00 = CMP0->getOperand(0);
11924 SDValue CMP01 = CMP0->getOperand(1);
11925 EVT VT = CMP00.getValueType();
11927 if (VT == MVT::f32 || VT == MVT::f64) {
11928 bool ExpectingFlags = false;
11929 // Check for any users that want flags:
11930 for (SDNode::use_iterator UI = N->use_begin(),
11932 !ExpectingFlags && UI != UE; ++UI)
11933 switch (UI->getOpcode()) {
11938 ExpectingFlags = true;
11940 case ISD::CopyToReg:
11941 case ISD::SIGN_EXTEND:
11942 case ISD::ZERO_EXTEND:
11943 case ISD::ANY_EXTEND:
11947 if (!ExpectingFlags) {
11948 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
11949 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
11951 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
11952 X86::CondCode tmp = cc0;
11957 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
11958 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
11959 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
11960 X86ISD::NodeType NTOperator = is64BitFP ?
11961 X86ISD::FSETCCsd : X86ISD::FSETCCss;
11962 // FIXME: need symbolic constants for these magic numbers.
11963 // See X86ATTInstPrinter.cpp:printSSECC().
11964 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
11965 SDValue OnesOrZeroesF = DAG.getNode(NTOperator, DL, MVT::f32, CMP00, CMP01,
11966 DAG.getConstant(x86cc, MVT::i8));
11967 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, MVT::i32,
11969 SDValue ANDed = DAG.getNode(ISD::AND, DL, MVT::i32, OnesOrZeroesI,
11970 DAG.getConstant(1, MVT::i32));
11971 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
11972 return OneBitOfTruth;
11980 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
11981 TargetLowering::DAGCombinerInfo &DCI,
11982 const X86Subtarget *Subtarget) {
11983 if (DCI.isBeforeLegalizeOps())
11986 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
11990 // Want to form ANDNP nodes:
11991 // 1) In the hopes of then easily combining them with OR and AND nodes
11992 // to form PBLEND/PSIGN.
11993 // 2) To match ANDN packed intrinsics
11994 EVT VT = N->getValueType(0);
11995 if (VT != MVT::v2i64 && VT != MVT::v4i64)
11998 SDValue N0 = N->getOperand(0);
11999 SDValue N1 = N->getOperand(1);
12000 DebugLoc DL = N->getDebugLoc();
12002 // Check LHS for vnot
12003 if (N0.getOpcode() == ISD::XOR &&
12004 ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
12005 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
12007 // Check RHS for vnot
12008 if (N1.getOpcode() == ISD::XOR &&
12009 ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
12010 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
12015 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
12016 TargetLowering::DAGCombinerInfo &DCI,
12017 const X86Subtarget *Subtarget) {
12018 if (DCI.isBeforeLegalizeOps())
12021 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
12025 EVT VT = N->getValueType(0);
12026 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64 && VT != MVT::v2i64)
12029 SDValue N0 = N->getOperand(0);
12030 SDValue N1 = N->getOperand(1);
12032 // look for psign/blend
12033 if (Subtarget->hasSSSE3()) {
12034 if (VT == MVT::v2i64) {
12035 // Canonicalize pandn to RHS
12036 if (N0.getOpcode() == X86ISD::ANDNP)
12038 // or (and (m, x), (pandn m, y))
12039 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
12040 SDValue Mask = N1.getOperand(0);
12041 SDValue X = N1.getOperand(1);
12043 if (N0.getOperand(0) == Mask)
12044 Y = N0.getOperand(1);
12045 if (N0.getOperand(1) == Mask)
12046 Y = N0.getOperand(0);
12048 // Check to see if the mask appeared in both the AND and ANDNP and
12052 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
12053 if (Mask.getOpcode() != ISD::BITCAST ||
12054 X.getOpcode() != ISD::BITCAST ||
12055 Y.getOpcode() != ISD::BITCAST)
12058 // Look through mask bitcast.
12059 Mask = Mask.getOperand(0);
12060 EVT MaskVT = Mask.getValueType();
12062 // Validate that the Mask operand is a vector sra node. The sra node
12063 // will be an intrinsic.
12064 if (Mask.getOpcode() != ISD::INTRINSIC_WO_CHAIN)
12067 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
12068 // there is no psrai.b
12069 switch (cast<ConstantSDNode>(Mask.getOperand(0))->getZExtValue()) {
12070 case Intrinsic::x86_sse2_psrai_w:
12071 case Intrinsic::x86_sse2_psrai_d:
12073 default: return SDValue();
12076 // Check that the SRA is all signbits.
12077 SDValue SraC = Mask.getOperand(2);
12078 unsigned SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
12079 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
12080 if ((SraAmt + 1) != EltBits)
12083 DebugLoc DL = N->getDebugLoc();
12085 // Now we know we at least have a plendvb with the mask val. See if
12086 // we can form a psignb/w/d.
12087 // psign = x.type == y.type == mask.type && y = sub(0, x);
12088 X = X.getOperand(0);
12089 Y = Y.getOperand(0);
12090 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
12091 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
12092 X.getValueType() == MaskVT && X.getValueType() == Y.getValueType()){
12095 case 8: Opc = X86ISD::PSIGNB; break;
12096 case 16: Opc = X86ISD::PSIGNW; break;
12097 case 32: Opc = X86ISD::PSIGND; break;
12101 SDValue Sign = DAG.getNode(Opc, DL, MaskVT, X, Mask.getOperand(1));
12102 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Sign);
12105 // PBLENDVB only available on SSE 4.1
12106 if (!Subtarget->hasSSE41())
12109 X = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, X);
12110 Y = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Y);
12111 Mask = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Mask);
12112 Mask = DAG.getNode(X86ISD::PBLENDVB, DL, MVT::v16i8, X, Y, Mask);
12113 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Mask);
12118 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
12119 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
12121 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
12123 if (!N0.hasOneUse() || !N1.hasOneUse())
12126 SDValue ShAmt0 = N0.getOperand(1);
12127 if (ShAmt0.getValueType() != MVT::i8)
12129 SDValue ShAmt1 = N1.getOperand(1);
12130 if (ShAmt1.getValueType() != MVT::i8)
12132 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
12133 ShAmt0 = ShAmt0.getOperand(0);
12134 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
12135 ShAmt1 = ShAmt1.getOperand(0);
12137 DebugLoc DL = N->getDebugLoc();
12138 unsigned Opc = X86ISD::SHLD;
12139 SDValue Op0 = N0.getOperand(0);
12140 SDValue Op1 = N1.getOperand(0);
12141 if (ShAmt0.getOpcode() == ISD::SUB) {
12142 Opc = X86ISD::SHRD;
12143 std::swap(Op0, Op1);
12144 std::swap(ShAmt0, ShAmt1);
12147 unsigned Bits = VT.getSizeInBits();
12148 if (ShAmt1.getOpcode() == ISD::SUB) {
12149 SDValue Sum = ShAmt1.getOperand(0);
12150 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
12151 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
12152 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
12153 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
12154 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
12155 return DAG.getNode(Opc, DL, VT,
12157 DAG.getNode(ISD::TRUNCATE, DL,
12160 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
12161 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
12163 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
12164 return DAG.getNode(Opc, DL, VT,
12165 N0.getOperand(0), N1.getOperand(0),
12166 DAG.getNode(ISD::TRUNCATE, DL,
12173 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
12174 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
12175 const X86Subtarget *Subtarget) {
12176 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
12177 // the FP state in cases where an emms may be missing.
12178 // A preferable solution to the general problem is to figure out the right
12179 // places to insert EMMS. This qualifies as a quick hack.
12181 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
12182 StoreSDNode *St = cast<StoreSDNode>(N);
12183 EVT VT = St->getValue().getValueType();
12184 if (VT.getSizeInBits() != 64)
12187 const Function *F = DAG.getMachineFunction().getFunction();
12188 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
12189 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
12190 && Subtarget->hasSSE2();
12191 if ((VT.isVector() ||
12192 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
12193 isa<LoadSDNode>(St->getValue()) &&
12194 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
12195 St->getChain().hasOneUse() && !St->isVolatile()) {
12196 SDNode* LdVal = St->getValue().getNode();
12197 LoadSDNode *Ld = 0;
12198 int TokenFactorIndex = -1;
12199 SmallVector<SDValue, 8> Ops;
12200 SDNode* ChainVal = St->getChain().getNode();
12201 // Must be a store of a load. We currently handle two cases: the load
12202 // is a direct child, and it's under an intervening TokenFactor. It is
12203 // possible to dig deeper under nested TokenFactors.
12204 if (ChainVal == LdVal)
12205 Ld = cast<LoadSDNode>(St->getChain());
12206 else if (St->getValue().hasOneUse() &&
12207 ChainVal->getOpcode() == ISD::TokenFactor) {
12208 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
12209 if (ChainVal->getOperand(i).getNode() == LdVal) {
12210 TokenFactorIndex = i;
12211 Ld = cast<LoadSDNode>(St->getValue());
12213 Ops.push_back(ChainVal->getOperand(i));
12217 if (!Ld || !ISD::isNormalLoad(Ld))
12220 // If this is not the MMX case, i.e. we are just turning i64 load/store
12221 // into f64 load/store, avoid the transformation if there are multiple
12222 // uses of the loaded value.
12223 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
12226 DebugLoc LdDL = Ld->getDebugLoc();
12227 DebugLoc StDL = N->getDebugLoc();
12228 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
12229 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
12231 if (Subtarget->is64Bit() || F64IsLegal) {
12232 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
12233 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
12234 Ld->getPointerInfo(), Ld->isVolatile(),
12235 Ld->isNonTemporal(), Ld->getAlignment());
12236 SDValue NewChain = NewLd.getValue(1);
12237 if (TokenFactorIndex != -1) {
12238 Ops.push_back(NewChain);
12239 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
12242 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
12243 St->getPointerInfo(),
12244 St->isVolatile(), St->isNonTemporal(),
12245 St->getAlignment());
12248 // Otherwise, lower to two pairs of 32-bit loads / stores.
12249 SDValue LoAddr = Ld->getBasePtr();
12250 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
12251 DAG.getConstant(4, MVT::i32));
12253 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
12254 Ld->getPointerInfo(),
12255 Ld->isVolatile(), Ld->isNonTemporal(),
12256 Ld->getAlignment());
12257 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
12258 Ld->getPointerInfo().getWithOffset(4),
12259 Ld->isVolatile(), Ld->isNonTemporal(),
12260 MinAlign(Ld->getAlignment(), 4));
12262 SDValue NewChain = LoLd.getValue(1);
12263 if (TokenFactorIndex != -1) {
12264 Ops.push_back(LoLd);
12265 Ops.push_back(HiLd);
12266 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
12270 LoAddr = St->getBasePtr();
12271 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
12272 DAG.getConstant(4, MVT::i32));
12274 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
12275 St->getPointerInfo(),
12276 St->isVolatile(), St->isNonTemporal(),
12277 St->getAlignment());
12278 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
12279 St->getPointerInfo().getWithOffset(4),
12281 St->isNonTemporal(),
12282 MinAlign(St->getAlignment(), 4));
12283 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
12288 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
12289 /// X86ISD::FXOR nodes.
12290 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
12291 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
12292 // F[X]OR(0.0, x) -> x
12293 // F[X]OR(x, 0.0) -> x
12294 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
12295 if (C->getValueAPF().isPosZero())
12296 return N->getOperand(1);
12297 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
12298 if (C->getValueAPF().isPosZero())
12299 return N->getOperand(0);
12303 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
12304 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
12305 // FAND(0.0, x) -> 0.0
12306 // FAND(x, 0.0) -> 0.0
12307 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
12308 if (C->getValueAPF().isPosZero())
12309 return N->getOperand(0);
12310 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
12311 if (C->getValueAPF().isPosZero())
12312 return N->getOperand(1);
12316 static SDValue PerformBTCombine(SDNode *N,
12318 TargetLowering::DAGCombinerInfo &DCI) {
12319 // BT ignores high bits in the bit index operand.
12320 SDValue Op1 = N->getOperand(1);
12321 if (Op1.hasOneUse()) {
12322 unsigned BitWidth = Op1.getValueSizeInBits();
12323 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
12324 APInt KnownZero, KnownOne;
12325 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
12326 !DCI.isBeforeLegalizeOps());
12327 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12328 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
12329 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
12330 DCI.CommitTargetLoweringOpt(TLO);
12335 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
12336 SDValue Op = N->getOperand(0);
12337 if (Op.getOpcode() == ISD::BITCAST)
12338 Op = Op.getOperand(0);
12339 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
12340 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
12341 VT.getVectorElementType().getSizeInBits() ==
12342 OpVT.getVectorElementType().getSizeInBits()) {
12343 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
12348 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
12349 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
12350 // (and (i32 x86isd::setcc_carry), 1)
12351 // This eliminates the zext. This transformation is necessary because
12352 // ISD::SETCC is always legalized to i8.
12353 DebugLoc dl = N->getDebugLoc();
12354 SDValue N0 = N->getOperand(0);
12355 EVT VT = N->getValueType(0);
12356 if (N0.getOpcode() == ISD::AND &&
12358 N0.getOperand(0).hasOneUse()) {
12359 SDValue N00 = N0.getOperand(0);
12360 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
12362 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
12363 if (!C || C->getZExtValue() != 1)
12365 return DAG.getNode(ISD::AND, dl, VT,
12366 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
12367 N00.getOperand(0), N00.getOperand(1)),
12368 DAG.getConstant(1, VT));
12374 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
12375 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG) {
12376 unsigned X86CC = N->getConstantOperandVal(0);
12377 SDValue EFLAG = N->getOperand(1);
12378 DebugLoc DL = N->getDebugLoc();
12380 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
12381 // a zext and produces an all-ones bit which is more useful than 0/1 in some
12383 if (X86CC == X86::COND_B)
12384 return DAG.getNode(ISD::AND, DL, MVT::i8,
12385 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
12386 DAG.getConstant(X86CC, MVT::i8), EFLAG),
12387 DAG.getConstant(1, MVT::i8));
12392 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
12393 const X86TargetLowering *XTLI) {
12394 SDValue Op0 = N->getOperand(0);
12395 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
12396 // a 32-bit target where SSE doesn't support i64->FP operations.
12397 if (Op0.getOpcode() == ISD::LOAD) {
12398 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
12399 EVT VT = Ld->getValueType(0);
12400 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
12401 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
12402 !XTLI->getSubtarget()->is64Bit() &&
12403 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
12404 SDValue FILDChain = XTLI->BuildFILD(SDValue(N, 0), Ld->getValueType(0),
12405 Ld->getChain(), Op0, DAG);
12406 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
12413 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
12414 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
12415 X86TargetLowering::DAGCombinerInfo &DCI) {
12416 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
12417 // the result is either zero or one (depending on the input carry bit).
12418 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
12419 if (X86::isZeroNode(N->getOperand(0)) &&
12420 X86::isZeroNode(N->getOperand(1)) &&
12421 // We don't have a good way to replace an EFLAGS use, so only do this when
12423 SDValue(N, 1).use_empty()) {
12424 DebugLoc DL = N->getDebugLoc();
12425 EVT VT = N->getValueType(0);
12426 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
12427 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
12428 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
12429 DAG.getConstant(X86::COND_B,MVT::i8),
12431 DAG.getConstant(1, VT));
12432 return DCI.CombineTo(N, Res1, CarryOut);
12438 // fold (add Y, (sete X, 0)) -> adc 0, Y
12439 // (add Y, (setne X, 0)) -> sbb -1, Y
12440 // (sub (sete X, 0), Y) -> sbb 0, Y
12441 // (sub (setne X, 0), Y) -> adc -1, Y
12442 static SDValue OptimizeConditonalInDecrement(SDNode *N, SelectionDAG &DAG) {
12443 DebugLoc DL = N->getDebugLoc();
12445 // Look through ZExts.
12446 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
12447 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
12450 SDValue SetCC = Ext.getOperand(0);
12451 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
12454 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
12455 if (CC != X86::COND_E && CC != X86::COND_NE)
12458 SDValue Cmp = SetCC.getOperand(1);
12459 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
12460 !X86::isZeroNode(Cmp.getOperand(1)) ||
12461 !Cmp.getOperand(0).getValueType().isInteger())
12464 SDValue CmpOp0 = Cmp.getOperand(0);
12465 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
12466 DAG.getConstant(1, CmpOp0.getValueType()));
12468 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
12469 if (CC == X86::COND_NE)
12470 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
12471 DL, OtherVal.getValueType(), OtherVal,
12472 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
12473 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
12474 DL, OtherVal.getValueType(), OtherVal,
12475 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
12478 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
12479 DAGCombinerInfo &DCI) const {
12480 SelectionDAG &DAG = DCI.DAG;
12481 switch (N->getOpcode()) {
12483 case ISD::EXTRACT_VECTOR_ELT:
12484 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
12485 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
12486 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
12488 case ISD::SUB: return OptimizeConditonalInDecrement(N, DAG);
12489 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
12490 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
12493 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
12494 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
12495 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
12496 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
12497 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, this);
12499 case X86ISD::FOR: return PerformFORCombine(N, DAG);
12500 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
12501 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
12502 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
12503 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
12504 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG);
12505 case X86ISD::SHUFPS: // Handle all target specific shuffles
12506 case X86ISD::SHUFPD:
12507 case X86ISD::PALIGN:
12508 case X86ISD::PUNPCKHBW:
12509 case X86ISD::PUNPCKHWD:
12510 case X86ISD::PUNPCKHDQ:
12511 case X86ISD::PUNPCKHQDQ:
12512 case X86ISD::UNPCKHPS:
12513 case X86ISD::UNPCKHPD:
12514 case X86ISD::PUNPCKLBW:
12515 case X86ISD::PUNPCKLWD:
12516 case X86ISD::PUNPCKLDQ:
12517 case X86ISD::PUNPCKLQDQ:
12518 case X86ISD::UNPCKLPS:
12519 case X86ISD::UNPCKLPD:
12520 case X86ISD::VUNPCKLPSY:
12521 case X86ISD::VUNPCKLPDY:
12522 case X86ISD::MOVHLPS:
12523 case X86ISD::MOVLHPS:
12524 case X86ISD::PSHUFD:
12525 case X86ISD::PSHUFHW:
12526 case X86ISD::PSHUFLW:
12527 case X86ISD::MOVSS:
12528 case X86ISD::MOVSD:
12529 case X86ISD::VPERMIL:
12530 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI);
12536 /// isTypeDesirableForOp - Return true if the target has native support for
12537 /// the specified value type and it is 'desirable' to use the type for the
12538 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
12539 /// instruction encodings are longer and some i16 instructions are slow.
12540 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
12541 if (!isTypeLegal(VT))
12543 if (VT != MVT::i16)
12550 case ISD::SIGN_EXTEND:
12551 case ISD::ZERO_EXTEND:
12552 case ISD::ANY_EXTEND:
12565 /// IsDesirableToPromoteOp - This method query the target whether it is
12566 /// beneficial for dag combiner to promote the specified node. If true, it
12567 /// should return the desired promotion type by reference.
12568 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
12569 EVT VT = Op.getValueType();
12570 if (VT != MVT::i16)
12573 bool Promote = false;
12574 bool Commute = false;
12575 switch (Op.getOpcode()) {
12578 LoadSDNode *LD = cast<LoadSDNode>(Op);
12579 // If the non-extending load has a single use and it's not live out, then it
12580 // might be folded.
12581 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
12582 Op.hasOneUse()*/) {
12583 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12584 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
12585 // The only case where we'd want to promote LOAD (rather then it being
12586 // promoted as an operand is when it's only use is liveout.
12587 if (UI->getOpcode() != ISD::CopyToReg)
12594 case ISD::SIGN_EXTEND:
12595 case ISD::ZERO_EXTEND:
12596 case ISD::ANY_EXTEND:
12601 SDValue N0 = Op.getOperand(0);
12602 // Look out for (store (shl (load), x)).
12603 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
12616 SDValue N0 = Op.getOperand(0);
12617 SDValue N1 = Op.getOperand(1);
12618 if (!Commute && MayFoldLoad(N1))
12620 // Avoid disabling potential load folding opportunities.
12621 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
12623 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
12633 //===----------------------------------------------------------------------===//
12634 // X86 Inline Assembly Support
12635 //===----------------------------------------------------------------------===//
12637 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
12638 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
12640 std::string AsmStr = IA->getAsmString();
12642 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
12643 SmallVector<StringRef, 4> AsmPieces;
12644 SplitString(AsmStr, AsmPieces, ";\n");
12646 switch (AsmPieces.size()) {
12647 default: return false;
12649 AsmStr = AsmPieces[0];
12651 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
12653 // FIXME: this should verify that we are targeting a 486 or better. If not,
12654 // we will turn this bswap into something that will be lowered to logical ops
12655 // instead of emitting the bswap asm. For now, we don't support 486 or lower
12656 // so don't worry about this.
12658 if (AsmPieces.size() == 2 &&
12659 (AsmPieces[0] == "bswap" ||
12660 AsmPieces[0] == "bswapq" ||
12661 AsmPieces[0] == "bswapl") &&
12662 (AsmPieces[1] == "$0" ||
12663 AsmPieces[1] == "${0:q}")) {
12664 // No need to check constraints, nothing other than the equivalent of
12665 // "=r,0" would be valid here.
12666 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12667 if (!Ty || Ty->getBitWidth() % 16 != 0)
12669 return IntrinsicLowering::LowerToByteSwap(CI);
12671 // rorw $$8, ${0:w} --> llvm.bswap.i16
12672 if (CI->getType()->isIntegerTy(16) &&
12673 AsmPieces.size() == 3 &&
12674 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
12675 AsmPieces[1] == "$$8," &&
12676 AsmPieces[2] == "${0:w}" &&
12677 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
12679 const std::string &ConstraintsStr = IA->getConstraintString();
12680 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
12681 std::sort(AsmPieces.begin(), AsmPieces.end());
12682 if (AsmPieces.size() == 4 &&
12683 AsmPieces[0] == "~{cc}" &&
12684 AsmPieces[1] == "~{dirflag}" &&
12685 AsmPieces[2] == "~{flags}" &&
12686 AsmPieces[3] == "~{fpsr}") {
12687 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12688 if (!Ty || Ty->getBitWidth() % 16 != 0)
12690 return IntrinsicLowering::LowerToByteSwap(CI);
12695 if (CI->getType()->isIntegerTy(32) &&
12696 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
12697 SmallVector<StringRef, 4> Words;
12698 SplitString(AsmPieces[0], Words, " \t,");
12699 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
12700 Words[2] == "${0:w}") {
12702 SplitString(AsmPieces[1], Words, " \t,");
12703 if (Words.size() == 3 && Words[0] == "rorl" && Words[1] == "$$16" &&
12704 Words[2] == "$0") {
12706 SplitString(AsmPieces[2], Words, " \t,");
12707 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
12708 Words[2] == "${0:w}") {
12710 const std::string &ConstraintsStr = IA->getConstraintString();
12711 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
12712 std::sort(AsmPieces.begin(), AsmPieces.end());
12713 if (AsmPieces.size() == 4 &&
12714 AsmPieces[0] == "~{cc}" &&
12715 AsmPieces[1] == "~{dirflag}" &&
12716 AsmPieces[2] == "~{flags}" &&
12717 AsmPieces[3] == "~{fpsr}") {
12718 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12719 if (!Ty || Ty->getBitWidth() % 16 != 0)
12721 return IntrinsicLowering::LowerToByteSwap(CI);
12728 if (CI->getType()->isIntegerTy(64)) {
12729 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
12730 if (Constraints.size() >= 2 &&
12731 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
12732 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
12733 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
12734 SmallVector<StringRef, 4> Words;
12735 SplitString(AsmPieces[0], Words, " \t");
12736 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
12738 SplitString(AsmPieces[1], Words, " \t");
12739 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
12741 SplitString(AsmPieces[2], Words, " \t,");
12742 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
12743 Words[2] == "%edx") {
12744 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12745 if (!Ty || Ty->getBitWidth() % 16 != 0)
12747 return IntrinsicLowering::LowerToByteSwap(CI);
12760 /// getConstraintType - Given a constraint letter, return the type of
12761 /// constraint it is for this target.
12762 X86TargetLowering::ConstraintType
12763 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
12764 if (Constraint.size() == 1) {
12765 switch (Constraint[0]) {
12776 return C_RegisterClass;
12800 return TargetLowering::getConstraintType(Constraint);
12803 /// Examine constraint type and operand type and determine a weight value.
12804 /// This object must already have been set up with the operand type
12805 /// and the current alternative constraint selected.
12806 TargetLowering::ConstraintWeight
12807 X86TargetLowering::getSingleConstraintMatchWeight(
12808 AsmOperandInfo &info, const char *constraint) const {
12809 ConstraintWeight weight = CW_Invalid;
12810 Value *CallOperandVal = info.CallOperandVal;
12811 // If we don't have a value, we can't do a match,
12812 // but allow it at the lowest weight.
12813 if (CallOperandVal == NULL)
12815 Type *type = CallOperandVal->getType();
12816 // Look at the constraint type.
12817 switch (*constraint) {
12819 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
12830 if (CallOperandVal->getType()->isIntegerTy())
12831 weight = CW_SpecificReg;
12836 if (type->isFloatingPointTy())
12837 weight = CW_SpecificReg;
12840 if (type->isX86_MMXTy() && Subtarget->hasMMX())
12841 weight = CW_SpecificReg;
12845 if ((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasXMM())
12846 weight = CW_Register;
12849 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
12850 if (C->getZExtValue() <= 31)
12851 weight = CW_Constant;
12855 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12856 if (C->getZExtValue() <= 63)
12857 weight = CW_Constant;
12861 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12862 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
12863 weight = CW_Constant;
12867 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12868 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
12869 weight = CW_Constant;
12873 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12874 if (C->getZExtValue() <= 3)
12875 weight = CW_Constant;
12879 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12880 if (C->getZExtValue() <= 0xff)
12881 weight = CW_Constant;
12886 if (dyn_cast<ConstantFP>(CallOperandVal)) {
12887 weight = CW_Constant;
12891 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12892 if ((C->getSExtValue() >= -0x80000000LL) &&
12893 (C->getSExtValue() <= 0x7fffffffLL))
12894 weight = CW_Constant;
12898 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12899 if (C->getZExtValue() <= 0xffffffff)
12900 weight = CW_Constant;
12907 /// LowerXConstraint - try to replace an X constraint, which matches anything,
12908 /// with another that has more specific requirements based on the type of the
12909 /// corresponding operand.
12910 const char *X86TargetLowering::
12911 LowerXConstraint(EVT ConstraintVT) const {
12912 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
12913 // 'f' like normal targets.
12914 if (ConstraintVT.isFloatingPoint()) {
12915 if (Subtarget->hasXMMInt())
12917 if (Subtarget->hasXMM())
12921 return TargetLowering::LowerXConstraint(ConstraintVT);
12924 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
12925 /// vector. If it is invalid, don't add anything to Ops.
12926 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
12927 std::string &Constraint,
12928 std::vector<SDValue>&Ops,
12929 SelectionDAG &DAG) const {
12930 SDValue Result(0, 0);
12932 // Only support length 1 constraints for now.
12933 if (Constraint.length() > 1) return;
12935 char ConstraintLetter = Constraint[0];
12936 switch (ConstraintLetter) {
12939 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12940 if (C->getZExtValue() <= 31) {
12941 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12947 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12948 if (C->getZExtValue() <= 63) {
12949 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12955 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12956 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
12957 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12963 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12964 if (C->getZExtValue() <= 255) {
12965 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12971 // 32-bit signed value
12972 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12973 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
12974 C->getSExtValue())) {
12975 // Widen to 64 bits here to get it sign extended.
12976 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
12979 // FIXME gcc accepts some relocatable values here too, but only in certain
12980 // memory models; it's complicated.
12985 // 32-bit unsigned value
12986 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12987 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
12988 C->getZExtValue())) {
12989 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12993 // FIXME gcc accepts some relocatable values here too, but only in certain
12994 // memory models; it's complicated.
12998 // Literal immediates are always ok.
12999 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
13000 // Widen to 64 bits here to get it sign extended.
13001 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
13005 // In any sort of PIC mode addresses need to be computed at runtime by
13006 // adding in a register or some sort of table lookup. These can't
13007 // be used as immediates.
13008 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
13011 // If we are in non-pic codegen mode, we allow the address of a global (with
13012 // an optional displacement) to be used with 'i'.
13013 GlobalAddressSDNode *GA = 0;
13014 int64_t Offset = 0;
13016 // Match either (GA), (GA+C), (GA+C1+C2), etc.
13018 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
13019 Offset += GA->getOffset();
13021 } else if (Op.getOpcode() == ISD::ADD) {
13022 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
13023 Offset += C->getZExtValue();
13024 Op = Op.getOperand(0);
13027 } else if (Op.getOpcode() == ISD::SUB) {
13028 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
13029 Offset += -C->getZExtValue();
13030 Op = Op.getOperand(0);
13035 // Otherwise, this isn't something we can handle, reject it.
13039 const GlobalValue *GV = GA->getGlobal();
13040 // If we require an extra load to get this address, as in PIC mode, we
13041 // can't accept it.
13042 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
13043 getTargetMachine())))
13046 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
13047 GA->getValueType(0), Offset);
13052 if (Result.getNode()) {
13053 Ops.push_back(Result);
13056 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
13059 std::pair<unsigned, const TargetRegisterClass*>
13060 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
13062 // First, see if this is a constraint that directly corresponds to an LLVM
13064 if (Constraint.size() == 1) {
13065 // GCC Constraint Letters
13066 switch (Constraint[0]) {
13068 // TODO: Slight differences here in allocation order and leaving
13069 // RIP in the class. Do they matter any more here than they do
13070 // in the normal allocation?
13071 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
13072 if (Subtarget->is64Bit()) {
13073 if (VT == MVT::i32 || VT == MVT::f32)
13074 return std::make_pair(0U, X86::GR32RegisterClass);
13075 else if (VT == MVT::i16)
13076 return std::make_pair(0U, X86::GR16RegisterClass);
13077 else if (VT == MVT::i8 || VT == MVT::i1)
13078 return std::make_pair(0U, X86::GR8RegisterClass);
13079 else if (VT == MVT::i64 || VT == MVT::f64)
13080 return std::make_pair(0U, X86::GR64RegisterClass);
13083 // 32-bit fallthrough
13084 case 'Q': // Q_REGS
13085 if (VT == MVT::i32 || VT == MVT::f32)
13086 return std::make_pair(0U, X86::GR32_ABCDRegisterClass);
13087 else if (VT == MVT::i16)
13088 return std::make_pair(0U, X86::GR16_ABCDRegisterClass);
13089 else if (VT == MVT::i8 || VT == MVT::i1)
13090 return std::make_pair(0U, X86::GR8_ABCD_LRegisterClass);
13091 else if (VT == MVT::i64)
13092 return std::make_pair(0U, X86::GR64_ABCDRegisterClass);
13094 case 'r': // GENERAL_REGS
13095 case 'l': // INDEX_REGS
13096 if (VT == MVT::i8 || VT == MVT::i1)
13097 return std::make_pair(0U, X86::GR8RegisterClass);
13098 if (VT == MVT::i16)
13099 return std::make_pair(0U, X86::GR16RegisterClass);
13100 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
13101 return std::make_pair(0U, X86::GR32RegisterClass);
13102 return std::make_pair(0U, X86::GR64RegisterClass);
13103 case 'R': // LEGACY_REGS
13104 if (VT == MVT::i8 || VT == MVT::i1)
13105 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
13106 if (VT == MVT::i16)
13107 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
13108 if (VT == MVT::i32 || !Subtarget->is64Bit())
13109 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
13110 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
13111 case 'f': // FP Stack registers.
13112 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
13113 // value to the correct fpstack register class.
13114 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
13115 return std::make_pair(0U, X86::RFP32RegisterClass);
13116 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
13117 return std::make_pair(0U, X86::RFP64RegisterClass);
13118 return std::make_pair(0U, X86::RFP80RegisterClass);
13119 case 'y': // MMX_REGS if MMX allowed.
13120 if (!Subtarget->hasMMX()) break;
13121 return std::make_pair(0U, X86::VR64RegisterClass);
13122 case 'Y': // SSE_REGS if SSE2 allowed
13123 if (!Subtarget->hasXMMInt()) break;
13125 case 'x': // SSE_REGS if SSE1 allowed
13126 if (!Subtarget->hasXMM()) break;
13128 switch (VT.getSimpleVT().SimpleTy) {
13130 // Scalar SSE types.
13133 return std::make_pair(0U, X86::FR32RegisterClass);
13136 return std::make_pair(0U, X86::FR64RegisterClass);
13144 return std::make_pair(0U, X86::VR128RegisterClass);
13150 // Use the default implementation in TargetLowering to convert the register
13151 // constraint into a member of a register class.
13152 std::pair<unsigned, const TargetRegisterClass*> Res;
13153 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
13155 // Not found as a standard register?
13156 if (Res.second == 0) {
13157 // Map st(0) -> st(7) -> ST0
13158 if (Constraint.size() == 7 && Constraint[0] == '{' &&
13159 tolower(Constraint[1]) == 's' &&
13160 tolower(Constraint[2]) == 't' &&
13161 Constraint[3] == '(' &&
13162 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
13163 Constraint[5] == ')' &&
13164 Constraint[6] == '}') {
13166 Res.first = X86::ST0+Constraint[4]-'0';
13167 Res.second = X86::RFP80RegisterClass;
13171 // GCC allows "st(0)" to be called just plain "st".
13172 if (StringRef("{st}").equals_lower(Constraint)) {
13173 Res.first = X86::ST0;
13174 Res.second = X86::RFP80RegisterClass;
13179 if (StringRef("{flags}").equals_lower(Constraint)) {
13180 Res.first = X86::EFLAGS;
13181 Res.second = X86::CCRRegisterClass;
13185 // 'A' means EAX + EDX.
13186 if (Constraint == "A") {
13187 Res.first = X86::EAX;
13188 Res.second = X86::GR32_ADRegisterClass;
13194 // Otherwise, check to see if this is a register class of the wrong value
13195 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
13196 // turn into {ax},{dx}.
13197 if (Res.second->hasType(VT))
13198 return Res; // Correct type already, nothing to do.
13200 // All of the single-register GCC register classes map their values onto
13201 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
13202 // really want an 8-bit or 32-bit register, map to the appropriate register
13203 // class and return the appropriate register.
13204 if (Res.second == X86::GR16RegisterClass) {
13205 if (VT == MVT::i8) {
13206 unsigned DestReg = 0;
13207 switch (Res.first) {
13209 case X86::AX: DestReg = X86::AL; break;
13210 case X86::DX: DestReg = X86::DL; break;
13211 case X86::CX: DestReg = X86::CL; break;
13212 case X86::BX: DestReg = X86::BL; break;
13215 Res.first = DestReg;
13216 Res.second = X86::GR8RegisterClass;
13218 } else if (VT == MVT::i32) {
13219 unsigned DestReg = 0;
13220 switch (Res.first) {
13222 case X86::AX: DestReg = X86::EAX; break;
13223 case X86::DX: DestReg = X86::EDX; break;
13224 case X86::CX: DestReg = X86::ECX; break;
13225 case X86::BX: DestReg = X86::EBX; break;
13226 case X86::SI: DestReg = X86::ESI; break;
13227 case X86::DI: DestReg = X86::EDI; break;
13228 case X86::BP: DestReg = X86::EBP; break;
13229 case X86::SP: DestReg = X86::ESP; break;
13232 Res.first = DestReg;
13233 Res.second = X86::GR32RegisterClass;
13235 } else if (VT == MVT::i64) {
13236 unsigned DestReg = 0;
13237 switch (Res.first) {
13239 case X86::AX: DestReg = X86::RAX; break;
13240 case X86::DX: DestReg = X86::RDX; break;
13241 case X86::CX: DestReg = X86::RCX; break;
13242 case X86::BX: DestReg = X86::RBX; break;
13243 case X86::SI: DestReg = X86::RSI; break;
13244 case X86::DI: DestReg = X86::RDI; break;
13245 case X86::BP: DestReg = X86::RBP; break;
13246 case X86::SP: DestReg = X86::RSP; break;
13249 Res.first = DestReg;
13250 Res.second = X86::GR64RegisterClass;
13253 } else if (Res.second == X86::FR32RegisterClass ||
13254 Res.second == X86::FR64RegisterClass ||
13255 Res.second == X86::VR128RegisterClass) {
13256 // Handle references to XMM physical registers that got mapped into the
13257 // wrong class. This can happen with constraints like {xmm0} where the
13258 // target independent register mapper will just pick the first match it can
13259 // find, ignoring the required type.
13260 if (VT == MVT::f32)
13261 Res.second = X86::FR32RegisterClass;
13262 else if (VT == MVT::f64)
13263 Res.second = X86::FR64RegisterClass;
13264 else if (X86::VR128RegisterClass->hasType(VT))
13265 Res.second = X86::VR128RegisterClass;