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"
54 #include "llvm/Target/TargetOptions.h"
56 using namespace dwarf;
58 STATISTIC(NumTailCalls, "Number of tail calls");
60 // Forward declarations.
61 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
64 static SDValue Insert128BitVector(SDValue Result,
70 static SDValue Extract128BitVector(SDValue Vec,
75 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
76 /// sets things up to match to an AVX VEXTRACTF128 instruction or a
77 /// simple subregister reference. Idx is an index in the 128 bits we
78 /// want. It need not be aligned to a 128-bit bounday. That makes
79 /// lowering EXTRACT_VECTOR_ELT operations easier.
80 static SDValue Extract128BitVector(SDValue Vec,
84 EVT VT = Vec.getValueType();
85 assert(VT.getSizeInBits() == 256 && "Unexpected vector size!");
86 EVT ElVT = VT.getVectorElementType();
87 int Factor = VT.getSizeInBits()/128;
88 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
89 VT.getVectorNumElements()/Factor);
91 // Extract from UNDEF is UNDEF.
92 if (Vec.getOpcode() == ISD::UNDEF)
93 return DAG.getNode(ISD::UNDEF, dl, ResultVT);
95 if (isa<ConstantSDNode>(Idx)) {
96 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
98 // Extract the relevant 128 bits. Generate an EXTRACT_SUBVECTOR
99 // we can match to VEXTRACTF128.
100 unsigned ElemsPerChunk = 128 / ElVT.getSizeInBits();
102 // This is the index of the first element of the 128-bit chunk
104 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / 128)
107 SDValue VecIdx = DAG.getConstant(NormalizedIdxVal, MVT::i32);
108 SDValue Result = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec,
117 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
118 /// sets things up to match to an AVX VINSERTF128 instruction or a
119 /// simple superregister reference. Idx is an index in the 128 bits
120 /// we want. It need not be aligned to a 128-bit bounday. That makes
121 /// lowering INSERT_VECTOR_ELT operations easier.
122 static SDValue Insert128BitVector(SDValue Result,
127 if (isa<ConstantSDNode>(Idx)) {
128 EVT VT = Vec.getValueType();
129 assert(VT.getSizeInBits() == 128 && "Unexpected vector size!");
131 EVT ElVT = VT.getVectorElementType();
132 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
133 EVT ResultVT = Result.getValueType();
135 // Insert the relevant 128 bits.
136 unsigned ElemsPerChunk = 128/ElVT.getSizeInBits();
138 // This is the index of the first element of the 128-bit chunk
140 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/128)
143 SDValue VecIdx = DAG.getConstant(NormalizedIdxVal, MVT::i32);
144 Result = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec,
152 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
153 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
154 bool is64Bit = Subtarget->is64Bit();
156 if (Subtarget->isTargetEnvMacho()) {
158 return new X8664_MachoTargetObjectFile();
159 return new TargetLoweringObjectFileMachO();
162 if (Subtarget->isTargetELF())
163 return new TargetLoweringObjectFileELF();
164 if (Subtarget->isTargetCOFF() && !Subtarget->isTargetEnvMacho())
165 return new TargetLoweringObjectFileCOFF();
166 llvm_unreachable("unknown subtarget type");
169 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
170 : TargetLowering(TM, createTLOF(TM)) {
171 Subtarget = &TM.getSubtarget<X86Subtarget>();
172 X86ScalarSSEf64 = Subtarget->hasXMMInt() || Subtarget->hasAVX();
173 X86ScalarSSEf32 = Subtarget->hasXMM() || Subtarget->hasAVX();
174 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
176 RegInfo = TM.getRegisterInfo();
177 TD = getTargetData();
179 // Set up the TargetLowering object.
180 static MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
182 // X86 is weird, it always uses i8 for shift amounts and setcc results.
183 setBooleanContents(ZeroOrOneBooleanContent);
185 // For 64-bit since we have so many registers use the ILP scheduler, for
186 // 32-bit code use the register pressure specific scheduling.
187 if (Subtarget->is64Bit())
188 setSchedulingPreference(Sched::ILP);
190 setSchedulingPreference(Sched::RegPressure);
191 setStackPointerRegisterToSaveRestore(X86StackPtr);
193 if (Subtarget->isTargetWindows() && !Subtarget->isTargetCygMing()) {
194 // Setup Windows compiler runtime calls.
195 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
196 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
197 setLibcallName(RTLIB::SREM_I64, "_allrem");
198 setLibcallName(RTLIB::UREM_I64, "_aullrem");
199 setLibcallName(RTLIB::MUL_I64, "_allmul");
200 setLibcallName(RTLIB::FPTOUINT_F64_I64, "_ftol2");
201 setLibcallName(RTLIB::FPTOUINT_F32_I64, "_ftol2");
202 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
203 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
204 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
205 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
206 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
207 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::C);
208 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::C);
211 if (Subtarget->isTargetDarwin()) {
212 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
213 setUseUnderscoreSetJmp(false);
214 setUseUnderscoreLongJmp(false);
215 } else if (Subtarget->isTargetMingw()) {
216 // MS runtime is weird: it exports _setjmp, but longjmp!
217 setUseUnderscoreSetJmp(true);
218 setUseUnderscoreLongJmp(false);
220 setUseUnderscoreSetJmp(true);
221 setUseUnderscoreLongJmp(true);
224 // Set up the register classes.
225 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
226 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
227 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
228 if (Subtarget->is64Bit())
229 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
231 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
233 // We don't accept any truncstore of integer registers.
234 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
235 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
236 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
237 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
238 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
239 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
241 // SETOEQ and SETUNE require checking two conditions.
242 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
243 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
244 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
245 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
246 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
247 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
249 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
251 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
252 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
253 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
255 if (Subtarget->is64Bit()) {
256 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
257 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
258 } else if (!UseSoftFloat) {
259 // We have an algorithm for SSE2->double, and we turn this into a
260 // 64-bit FILD followed by conditional FADD for other targets.
261 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
262 // We have an algorithm for SSE2, and we turn this into a 64-bit
263 // FILD for other targets.
264 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
267 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
269 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
270 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
273 // SSE has no i16 to fp conversion, only i32
274 if (X86ScalarSSEf32) {
275 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
276 // f32 and f64 cases are Legal, f80 case is not
277 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
279 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
280 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
283 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
284 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
287 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
288 // are Legal, f80 is custom lowered.
289 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
290 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
292 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
294 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
295 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
297 if (X86ScalarSSEf32) {
298 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
299 // f32 and f64 cases are Legal, f80 case is not
300 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
302 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
303 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
306 // Handle FP_TO_UINT by promoting the destination to a larger signed
308 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
309 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
310 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
312 if (Subtarget->is64Bit()) {
313 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
314 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
315 } else if (!UseSoftFloat) {
316 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
317 // Expand FP_TO_UINT into a select.
318 // FIXME: We would like to use a Custom expander here eventually to do
319 // the optimal thing for SSE vs. the default expansion in the legalizer.
320 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
322 // With SSE3 we can use fisttpll to convert to a signed i64; without
323 // SSE, we're stuck with a fistpll.
324 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
327 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
328 if (!X86ScalarSSEf64) {
329 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
330 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
331 if (Subtarget->is64Bit()) {
332 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
333 // Without SSE, i64->f64 goes through memory.
334 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
338 // Scalar integer divide and remainder are lowered to use operations that
339 // produce two results, to match the available instructions. This exposes
340 // the two-result form to trivial CSE, which is able to combine x/y and x%y
341 // into a single instruction.
343 // Scalar integer multiply-high is also lowered to use two-result
344 // operations, to match the available instructions. However, plain multiply
345 // (low) operations are left as Legal, as there are single-result
346 // instructions for this in x86. Using the two-result multiply instructions
347 // when both high and low results are needed must be arranged by dagcombine.
348 for (unsigned i = 0, e = 4; i != e; ++i) {
350 setOperationAction(ISD::MULHS, VT, Expand);
351 setOperationAction(ISD::MULHU, VT, Expand);
352 setOperationAction(ISD::SDIV, VT, Expand);
353 setOperationAction(ISD::UDIV, VT, Expand);
354 setOperationAction(ISD::SREM, VT, Expand);
355 setOperationAction(ISD::UREM, VT, Expand);
357 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
358 setOperationAction(ISD::ADDC, VT, Custom);
359 setOperationAction(ISD::ADDE, VT, Custom);
360 setOperationAction(ISD::SUBC, VT, Custom);
361 setOperationAction(ISD::SUBE, VT, Custom);
364 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
365 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
366 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
367 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
368 if (Subtarget->is64Bit())
369 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
370 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
371 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
372 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
373 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
374 setOperationAction(ISD::FREM , MVT::f32 , Expand);
375 setOperationAction(ISD::FREM , MVT::f64 , Expand);
376 setOperationAction(ISD::FREM , MVT::f80 , Expand);
377 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
379 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
380 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
381 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
382 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
383 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
384 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
385 if (Subtarget->is64Bit()) {
386 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
387 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
390 if (Subtarget->hasPOPCNT()) {
391 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
393 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
394 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
395 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
396 if (Subtarget->is64Bit())
397 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
400 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
401 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
403 // These should be promoted to a larger select which is supported.
404 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
405 // X86 wants to expand cmov itself.
406 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
407 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
408 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
409 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
410 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
411 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
412 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
413 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
414 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
415 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
416 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
417 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
418 if (Subtarget->is64Bit()) {
419 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
420 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
422 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
425 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
426 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
427 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
428 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
429 if (Subtarget->is64Bit())
430 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
431 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
432 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
433 if (Subtarget->is64Bit()) {
434 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
435 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
436 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
437 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
438 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
440 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
441 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
442 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
443 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
444 if (Subtarget->is64Bit()) {
445 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
446 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
447 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
450 if (Subtarget->hasXMM())
451 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
453 setOperationAction(ISD::MEMBARRIER , MVT::Other, Custom);
454 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
456 // On X86 and X86-64, atomic operations are lowered to locked instructions.
457 // Locked instructions, in turn, have implicit fence semantics (all memory
458 // operations are flushed before issuing the locked instruction, and they
459 // are not buffered), so we can fold away the common pattern of
460 // fence-atomic-fence.
461 setShouldFoldAtomicFences(true);
463 // Expand certain atomics
464 for (unsigned i = 0, e = 4; i != e; ++i) {
466 setOperationAction(ISD::ATOMIC_CMP_SWAP, VT, Custom);
467 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
468 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
471 if (!Subtarget->is64Bit()) {
472 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Custom);
473 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
474 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
475 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
476 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
477 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
478 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
479 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
482 if (Subtarget->hasCmpxchg16b()) {
483 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i128, Custom);
486 // FIXME - use subtarget debug flags
487 if (!Subtarget->isTargetDarwin() &&
488 !Subtarget->isTargetELF() &&
489 !Subtarget->isTargetCygMing()) {
490 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
493 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
494 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
495 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
496 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
497 if (Subtarget->is64Bit()) {
498 setExceptionPointerRegister(X86::RAX);
499 setExceptionSelectorRegister(X86::RDX);
501 setExceptionPointerRegister(X86::EAX);
502 setExceptionSelectorRegister(X86::EDX);
504 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
505 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
507 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
509 setOperationAction(ISD::TRAP, MVT::Other, Legal);
511 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
512 setOperationAction(ISD::VASTART , MVT::Other, Custom);
513 setOperationAction(ISD::VAEND , MVT::Other, Expand);
514 if (Subtarget->is64Bit()) {
515 setOperationAction(ISD::VAARG , MVT::Other, Custom);
516 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
518 setOperationAction(ISD::VAARG , MVT::Other, Expand);
519 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
522 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
523 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
524 setOperationAction(ISD::DYNAMIC_STACKALLOC,
525 (Subtarget->is64Bit() ? MVT::i64 : MVT::i32),
526 ((Subtarget->isTargetCOFF()
527 && !Subtarget->isTargetEnvMacho()) ||
528 EnableSegmentedStacks
531 if (!UseSoftFloat && X86ScalarSSEf64) {
532 // f32 and f64 use SSE.
533 // Set up the FP register classes.
534 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
535 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
537 // Use ANDPD to simulate FABS.
538 setOperationAction(ISD::FABS , MVT::f64, Custom);
539 setOperationAction(ISD::FABS , MVT::f32, Custom);
541 // Use XORP to simulate FNEG.
542 setOperationAction(ISD::FNEG , MVT::f64, Custom);
543 setOperationAction(ISD::FNEG , MVT::f32, Custom);
545 // Use ANDPD and ORPD to simulate FCOPYSIGN.
546 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
547 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
549 // Lower this to FGETSIGNx86 plus an AND.
550 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
551 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
553 // We don't support sin/cos/fmod
554 setOperationAction(ISD::FSIN , MVT::f64, Expand);
555 setOperationAction(ISD::FCOS , MVT::f64, Expand);
556 setOperationAction(ISD::FSIN , MVT::f32, Expand);
557 setOperationAction(ISD::FCOS , MVT::f32, Expand);
559 // Expand FP immediates into loads from the stack, except for the special
561 addLegalFPImmediate(APFloat(+0.0)); // xorpd
562 addLegalFPImmediate(APFloat(+0.0f)); // xorps
563 } else if (!UseSoftFloat && X86ScalarSSEf32) {
564 // Use SSE for f32, x87 for f64.
565 // Set up the FP register classes.
566 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
567 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
569 // Use ANDPS to simulate FABS.
570 setOperationAction(ISD::FABS , MVT::f32, Custom);
572 // Use XORP to simulate FNEG.
573 setOperationAction(ISD::FNEG , MVT::f32, Custom);
575 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
577 // Use ANDPS and ORPS to simulate FCOPYSIGN.
578 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
579 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
581 // We don't support sin/cos/fmod
582 setOperationAction(ISD::FSIN , MVT::f32, Expand);
583 setOperationAction(ISD::FCOS , MVT::f32, Expand);
585 // Special cases we handle for FP constants.
586 addLegalFPImmediate(APFloat(+0.0f)); // xorps
587 addLegalFPImmediate(APFloat(+0.0)); // FLD0
588 addLegalFPImmediate(APFloat(+1.0)); // FLD1
589 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
590 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
593 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
594 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
596 } else if (!UseSoftFloat) {
597 // f32 and f64 in x87.
598 // Set up the FP register classes.
599 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
600 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
602 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
603 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
604 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
605 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
608 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
609 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
611 addLegalFPImmediate(APFloat(+0.0)); // FLD0
612 addLegalFPImmediate(APFloat(+1.0)); // FLD1
613 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
614 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
615 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
616 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
617 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
618 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
621 // We don't support FMA.
622 setOperationAction(ISD::FMA, MVT::f64, Expand);
623 setOperationAction(ISD::FMA, MVT::f32, Expand);
625 // Long double always uses X87.
627 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
628 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
629 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
631 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
632 addLegalFPImmediate(TmpFlt); // FLD0
634 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
637 APFloat TmpFlt2(+1.0);
638 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
640 addLegalFPImmediate(TmpFlt2); // FLD1
641 TmpFlt2.changeSign();
642 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
646 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
647 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
650 setOperationAction(ISD::FMA, MVT::f80, Expand);
653 // Always use a library call for pow.
654 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
655 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
656 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
658 setOperationAction(ISD::FLOG, MVT::f80, Expand);
659 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
660 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
661 setOperationAction(ISD::FEXP, MVT::f80, Expand);
662 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
664 // First set operation action for all vector types to either promote
665 // (for widening) or expand (for scalarization). Then we will selectively
666 // turn on ones that can be effectively codegen'd.
667 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
668 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
669 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
670 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
671 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
672 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
673 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
674 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
675 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
676 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
677 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
678 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
679 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
680 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
681 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
682 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
683 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
684 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
685 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
686 setOperationAction(ISD::INSERT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
687 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
688 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
689 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
690 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
691 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
692 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
693 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
694 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
695 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
696 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
697 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
698 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
699 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
700 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
701 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
702 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
703 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
704 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
705 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
706 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
707 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
708 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
709 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
710 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
711 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
712 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
713 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
714 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
715 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
716 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
717 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
718 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
719 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
720 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
721 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
722 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
723 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
724 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
725 setTruncStoreAction((MVT::SimpleValueType)VT,
726 (MVT::SimpleValueType)InnerVT, Expand);
727 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
728 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
729 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
732 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
733 // with -msoft-float, disable use of MMX as well.
734 if (!UseSoftFloat && Subtarget->hasMMX()) {
735 addRegisterClass(MVT::x86mmx, X86::VR64RegisterClass);
736 // No operations on x86mmx supported, everything uses intrinsics.
739 // MMX-sized vectors (other than x86mmx) are expected to be expanded
740 // into smaller operations.
741 setOperationAction(ISD::MULHS, MVT::v8i8, Expand);
742 setOperationAction(ISD::MULHS, MVT::v4i16, Expand);
743 setOperationAction(ISD::MULHS, MVT::v2i32, Expand);
744 setOperationAction(ISD::MULHS, MVT::v1i64, Expand);
745 setOperationAction(ISD::AND, MVT::v8i8, Expand);
746 setOperationAction(ISD::AND, MVT::v4i16, Expand);
747 setOperationAction(ISD::AND, MVT::v2i32, Expand);
748 setOperationAction(ISD::AND, MVT::v1i64, Expand);
749 setOperationAction(ISD::OR, MVT::v8i8, Expand);
750 setOperationAction(ISD::OR, MVT::v4i16, Expand);
751 setOperationAction(ISD::OR, MVT::v2i32, Expand);
752 setOperationAction(ISD::OR, MVT::v1i64, Expand);
753 setOperationAction(ISD::XOR, MVT::v8i8, Expand);
754 setOperationAction(ISD::XOR, MVT::v4i16, Expand);
755 setOperationAction(ISD::XOR, MVT::v2i32, Expand);
756 setOperationAction(ISD::XOR, MVT::v1i64, Expand);
757 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Expand);
758 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Expand);
759 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Expand);
760 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Expand);
761 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
762 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
763 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
764 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
765 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
766 setOperationAction(ISD::BITCAST, MVT::v8i8, Expand);
767 setOperationAction(ISD::BITCAST, MVT::v4i16, Expand);
768 setOperationAction(ISD::BITCAST, MVT::v2i32, Expand);
769 setOperationAction(ISD::BITCAST, MVT::v1i64, Expand);
771 if (!UseSoftFloat && Subtarget->hasXMM()) {
772 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
774 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
775 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
776 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
777 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
778 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
779 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
780 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
781 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
782 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
783 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
784 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
785 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
788 if (!UseSoftFloat && Subtarget->hasXMMInt()) {
789 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
791 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
792 // registers cannot be used even for integer operations.
793 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
794 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
795 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
796 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
798 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
799 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
800 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
801 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
802 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
803 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
804 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
805 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
806 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
807 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
808 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
809 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
810 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
811 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
812 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
813 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
815 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
816 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
817 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
818 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
820 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
821 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
822 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
823 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
824 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
826 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
827 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
828 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
829 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
830 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
832 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
833 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
834 EVT VT = (MVT::SimpleValueType)i;
835 // Do not attempt to custom lower non-power-of-2 vectors
836 if (!isPowerOf2_32(VT.getVectorNumElements()))
838 // Do not attempt to custom lower non-128-bit vectors
839 if (!VT.is128BitVector())
841 setOperationAction(ISD::BUILD_VECTOR,
842 VT.getSimpleVT().SimpleTy, Custom);
843 setOperationAction(ISD::VECTOR_SHUFFLE,
844 VT.getSimpleVT().SimpleTy, Custom);
845 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
846 VT.getSimpleVT().SimpleTy, Custom);
849 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
850 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
851 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
852 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
853 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
854 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
856 if (Subtarget->is64Bit()) {
857 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
858 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
861 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
862 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
863 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
866 // Do not attempt to promote non-128-bit vectors
867 if (!VT.is128BitVector())
870 setOperationAction(ISD::AND, SVT, Promote);
871 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
872 setOperationAction(ISD::OR, SVT, Promote);
873 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
874 setOperationAction(ISD::XOR, SVT, Promote);
875 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
876 setOperationAction(ISD::LOAD, SVT, Promote);
877 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
878 setOperationAction(ISD::SELECT, SVT, Promote);
879 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
882 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
884 // Custom lower v2i64 and v2f64 selects.
885 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
886 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
887 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
888 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
890 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
891 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
894 if (Subtarget->hasSSE41() || Subtarget->hasAVX()) {
895 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
896 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
897 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
898 setOperationAction(ISD::FRINT, MVT::f32, Legal);
899 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
900 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
901 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
902 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
903 setOperationAction(ISD::FRINT, MVT::f64, Legal);
904 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
906 // FIXME: Do we need to handle scalar-to-vector here?
907 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
909 // Can turn SHL into an integer multiply.
910 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
911 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
913 // i8 and i16 vectors are custom , because the source register and source
914 // source memory operand types are not the same width. f32 vectors are
915 // custom since the immediate controlling the insert encodes additional
917 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
918 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
919 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
920 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
922 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
923 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
924 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
925 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
927 if (Subtarget->is64Bit()) {
928 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
929 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
933 if (Subtarget->hasSSE2() || Subtarget->hasAVX()) {
934 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
935 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
936 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
937 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
939 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
940 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
941 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
943 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
944 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
947 if (Subtarget->hasSSE42() || Subtarget->hasAVX())
948 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
950 if (!UseSoftFloat && Subtarget->hasAVX()) {
951 addRegisterClass(MVT::v32i8, X86::VR256RegisterClass);
952 addRegisterClass(MVT::v16i16, X86::VR256RegisterClass);
953 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
954 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
955 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
956 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
958 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
959 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
960 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
962 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
963 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
964 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
965 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
966 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
967 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
969 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
970 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
971 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
972 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
973 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
974 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
976 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
977 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
978 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
980 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f64, Custom);
981 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i64, Custom);
982 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
983 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
984 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i8, Custom);
985 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i16, Custom);
987 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
988 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
989 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
990 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
992 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
993 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
994 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
995 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
997 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
998 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1000 setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
1001 setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
1002 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
1003 setOperationAction(ISD::VSETCC, MVT::v4i64, Custom);
1005 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1006 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1007 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1009 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1010 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1011 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1012 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1014 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1015 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1016 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1017 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1019 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1020 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1021 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1022 // Don't lower v32i8 because there is no 128-bit byte mul
1024 // Custom lower several nodes for 256-bit types.
1025 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1026 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
1027 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
1030 // Extract subvector is special because the value type
1031 // (result) is 128-bit but the source is 256-bit wide.
1032 if (VT.is128BitVector())
1033 setOperationAction(ISD::EXTRACT_SUBVECTOR, SVT, Custom);
1035 // Do not attempt to custom lower other non-256-bit vectors
1036 if (!VT.is256BitVector())
1039 setOperationAction(ISD::BUILD_VECTOR, SVT, Custom);
1040 setOperationAction(ISD::VECTOR_SHUFFLE, SVT, Custom);
1041 setOperationAction(ISD::INSERT_VECTOR_ELT, SVT, Custom);
1042 setOperationAction(ISD::EXTRACT_VECTOR_ELT, SVT, Custom);
1043 setOperationAction(ISD::SCALAR_TO_VECTOR, SVT, Custom);
1044 setOperationAction(ISD::INSERT_SUBVECTOR, SVT, Custom);
1047 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1048 for (unsigned i = (unsigned)MVT::v32i8; i != (unsigned)MVT::v4i64; ++i) {
1049 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
1052 // Do not attempt to promote non-256-bit vectors
1053 if (!VT.is256BitVector())
1056 setOperationAction(ISD::AND, SVT, Promote);
1057 AddPromotedToType (ISD::AND, SVT, MVT::v4i64);
1058 setOperationAction(ISD::OR, SVT, Promote);
1059 AddPromotedToType (ISD::OR, SVT, MVT::v4i64);
1060 setOperationAction(ISD::XOR, SVT, Promote);
1061 AddPromotedToType (ISD::XOR, SVT, MVT::v4i64);
1062 setOperationAction(ISD::LOAD, SVT, Promote);
1063 AddPromotedToType (ISD::LOAD, SVT, MVT::v4i64);
1064 setOperationAction(ISD::SELECT, SVT, Promote);
1065 AddPromotedToType (ISD::SELECT, SVT, MVT::v4i64);
1069 // SIGN_EXTEND_INREGs are evaluated by the extend type. Handle the expansion
1070 // of this type with custom code.
1071 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1072 VT != (unsigned)MVT::LAST_VECTOR_VALUETYPE; VT++) {
1073 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT, Custom);
1076 // We want to custom lower some of our intrinsics.
1077 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1080 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1081 // handle type legalization for these operations here.
1083 // FIXME: We really should do custom legalization for addition and
1084 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1085 // than generic legalization for 64-bit multiplication-with-overflow, though.
1086 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1087 // Add/Sub/Mul with overflow operations are custom lowered.
1089 setOperationAction(ISD::SADDO, VT, Custom);
1090 setOperationAction(ISD::UADDO, VT, Custom);
1091 setOperationAction(ISD::SSUBO, VT, Custom);
1092 setOperationAction(ISD::USUBO, VT, Custom);
1093 setOperationAction(ISD::SMULO, VT, Custom);
1094 setOperationAction(ISD::UMULO, VT, Custom);
1097 // There are no 8-bit 3-address imul/mul instructions
1098 setOperationAction(ISD::SMULO, MVT::i8, Expand);
1099 setOperationAction(ISD::UMULO, MVT::i8, Expand);
1101 if (!Subtarget->is64Bit()) {
1102 // These libcalls are not available in 32-bit.
1103 setLibcallName(RTLIB::SHL_I128, 0);
1104 setLibcallName(RTLIB::SRL_I128, 0);
1105 setLibcallName(RTLIB::SRA_I128, 0);
1108 // We have target-specific dag combine patterns for the following nodes:
1109 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1110 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1111 setTargetDAGCombine(ISD::BUILD_VECTOR);
1112 setTargetDAGCombine(ISD::SELECT);
1113 setTargetDAGCombine(ISD::SHL);
1114 setTargetDAGCombine(ISD::SRA);
1115 setTargetDAGCombine(ISD::SRL);
1116 setTargetDAGCombine(ISD::OR);
1117 setTargetDAGCombine(ISD::AND);
1118 setTargetDAGCombine(ISD::ADD);
1119 setTargetDAGCombine(ISD::SUB);
1120 setTargetDAGCombine(ISD::STORE);
1121 setTargetDAGCombine(ISD::ZERO_EXTEND);
1122 setTargetDAGCombine(ISD::SINT_TO_FP);
1123 if (Subtarget->is64Bit())
1124 setTargetDAGCombine(ISD::MUL);
1126 computeRegisterProperties();
1128 // On Darwin, -Os means optimize for size without hurting performance,
1129 // do not reduce the limit.
1130 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1131 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1132 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1133 maxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1134 maxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1135 maxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1136 setPrefLoopAlignment(16);
1137 benefitFromCodePlacementOpt = true;
1139 setPrefFunctionAlignment(4);
1143 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1148 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1149 /// the desired ByVal argument alignment.
1150 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1153 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1154 if (VTy->getBitWidth() == 128)
1156 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1157 unsigned EltAlign = 0;
1158 getMaxByValAlign(ATy->getElementType(), EltAlign);
1159 if (EltAlign > MaxAlign)
1160 MaxAlign = EltAlign;
1161 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1162 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1163 unsigned EltAlign = 0;
1164 getMaxByValAlign(STy->getElementType(i), EltAlign);
1165 if (EltAlign > MaxAlign)
1166 MaxAlign = EltAlign;
1174 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1175 /// function arguments in the caller parameter area. For X86, aggregates
1176 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1177 /// are at 4-byte boundaries.
1178 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1179 if (Subtarget->is64Bit()) {
1180 // Max of 8 and alignment of type.
1181 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1188 if (Subtarget->hasXMM())
1189 getMaxByValAlign(Ty, Align);
1193 /// getOptimalMemOpType - Returns the target specific optimal type for load
1194 /// and store operations as a result of memset, memcpy, and memmove
1195 /// lowering. If DstAlign is zero that means it's safe to destination
1196 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1197 /// means there isn't a need to check it against alignment requirement,
1198 /// probably because the source does not need to be loaded. If
1199 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1200 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1201 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1202 /// constant so it does not need to be loaded.
1203 /// It returns EVT::Other if the type should be determined using generic
1204 /// target-independent logic.
1206 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1207 unsigned DstAlign, unsigned SrcAlign,
1208 bool NonScalarIntSafe,
1210 MachineFunction &MF) const {
1211 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1212 // linux. This is because the stack realignment code can't handle certain
1213 // cases like PR2962. This should be removed when PR2962 is fixed.
1214 const Function *F = MF.getFunction();
1215 if (NonScalarIntSafe &&
1216 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1218 (Subtarget->isUnalignedMemAccessFast() ||
1219 ((DstAlign == 0 || DstAlign >= 16) &&
1220 (SrcAlign == 0 || SrcAlign >= 16))) &&
1221 Subtarget->getStackAlignment() >= 16) {
1222 if (Subtarget->hasSSE2())
1224 if (Subtarget->hasSSE1())
1226 } else if (!MemcpyStrSrc && Size >= 8 &&
1227 !Subtarget->is64Bit() &&
1228 Subtarget->getStackAlignment() >= 8 &&
1229 Subtarget->hasXMMInt()) {
1230 // Do not use f64 to lower memcpy if source is string constant. It's
1231 // better to use i32 to avoid the loads.
1235 if (Subtarget->is64Bit() && Size >= 8)
1240 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1241 /// current function. The returned value is a member of the
1242 /// MachineJumpTableInfo::JTEntryKind enum.
1243 unsigned X86TargetLowering::getJumpTableEncoding() const {
1244 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1246 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1247 Subtarget->isPICStyleGOT())
1248 return MachineJumpTableInfo::EK_Custom32;
1250 // Otherwise, use the normal jump table encoding heuristics.
1251 return TargetLowering::getJumpTableEncoding();
1255 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1256 const MachineBasicBlock *MBB,
1257 unsigned uid,MCContext &Ctx) const{
1258 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1259 Subtarget->isPICStyleGOT());
1260 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1262 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1263 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1266 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1268 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1269 SelectionDAG &DAG) const {
1270 if (!Subtarget->is64Bit())
1271 // This doesn't have DebugLoc associated with it, but is not really the
1272 // same as a Register.
1273 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1277 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1278 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1280 const MCExpr *X86TargetLowering::
1281 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1282 MCContext &Ctx) const {
1283 // X86-64 uses RIP relative addressing based on the jump table label.
1284 if (Subtarget->isPICStyleRIPRel())
1285 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1287 // Otherwise, the reference is relative to the PIC base.
1288 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1291 // FIXME: Why this routine is here? Move to RegInfo!
1292 std::pair<const TargetRegisterClass*, uint8_t>
1293 X86TargetLowering::findRepresentativeClass(EVT VT) const{
1294 const TargetRegisterClass *RRC = 0;
1296 switch (VT.getSimpleVT().SimpleTy) {
1298 return TargetLowering::findRepresentativeClass(VT);
1299 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1300 RRC = (Subtarget->is64Bit()
1301 ? X86::GR64RegisterClass : X86::GR32RegisterClass);
1304 RRC = X86::VR64RegisterClass;
1306 case MVT::f32: case MVT::f64:
1307 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1308 case MVT::v4f32: case MVT::v2f64:
1309 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1311 RRC = X86::VR128RegisterClass;
1314 return std::make_pair(RRC, Cost);
1317 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1318 unsigned &Offset) const {
1319 if (!Subtarget->isTargetLinux())
1322 if (Subtarget->is64Bit()) {
1323 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1325 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1338 //===----------------------------------------------------------------------===//
1339 // Return Value Calling Convention Implementation
1340 //===----------------------------------------------------------------------===//
1342 #include "X86GenCallingConv.inc"
1345 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1346 MachineFunction &MF, bool isVarArg,
1347 const SmallVectorImpl<ISD::OutputArg> &Outs,
1348 LLVMContext &Context) const {
1349 SmallVector<CCValAssign, 16> RVLocs;
1350 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1352 return CCInfo.CheckReturn(Outs, RetCC_X86);
1356 X86TargetLowering::LowerReturn(SDValue Chain,
1357 CallingConv::ID CallConv, bool isVarArg,
1358 const SmallVectorImpl<ISD::OutputArg> &Outs,
1359 const SmallVectorImpl<SDValue> &OutVals,
1360 DebugLoc dl, SelectionDAG &DAG) const {
1361 MachineFunction &MF = DAG.getMachineFunction();
1362 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1364 SmallVector<CCValAssign, 16> RVLocs;
1365 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1366 RVLocs, *DAG.getContext());
1367 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1369 // Add the regs to the liveout set for the function.
1370 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1371 for (unsigned i = 0; i != RVLocs.size(); ++i)
1372 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1373 MRI.addLiveOut(RVLocs[i].getLocReg());
1377 SmallVector<SDValue, 6> RetOps;
1378 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1379 // Operand #1 = Bytes To Pop
1380 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1383 // Copy the result values into the output registers.
1384 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1385 CCValAssign &VA = RVLocs[i];
1386 assert(VA.isRegLoc() && "Can only return in registers!");
1387 SDValue ValToCopy = OutVals[i];
1388 EVT ValVT = ValToCopy.getValueType();
1390 // If this is x86-64, and we disabled SSE, we can't return FP values,
1391 // or SSE or MMX vectors.
1392 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1393 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1394 (Subtarget->is64Bit() && !Subtarget->hasXMM())) {
1395 report_fatal_error("SSE register return with SSE disabled");
1397 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1398 // llvm-gcc has never done it right and no one has noticed, so this
1399 // should be OK for now.
1400 if (ValVT == MVT::f64 &&
1401 (Subtarget->is64Bit() && !Subtarget->hasXMMInt()))
1402 report_fatal_error("SSE2 register return with SSE2 disabled");
1404 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1405 // the RET instruction and handled by the FP Stackifier.
1406 if (VA.getLocReg() == X86::ST0 ||
1407 VA.getLocReg() == X86::ST1) {
1408 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1409 // change the value to the FP stack register class.
1410 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1411 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1412 RetOps.push_back(ValToCopy);
1413 // Don't emit a copytoreg.
1417 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1418 // which is returned in RAX / RDX.
1419 if (Subtarget->is64Bit()) {
1420 if (ValVT == MVT::x86mmx) {
1421 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1422 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
1423 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1425 // If we don't have SSE2 available, convert to v4f32 so the generated
1426 // register is legal.
1427 if (!Subtarget->hasSSE2())
1428 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
1433 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1434 Flag = Chain.getValue(1);
1437 // The x86-64 ABI for returning structs by value requires that we copy
1438 // the sret argument into %rax for the return. We saved the argument into
1439 // a virtual register in the entry block, so now we copy the value out
1441 if (Subtarget->is64Bit() &&
1442 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1443 MachineFunction &MF = DAG.getMachineFunction();
1444 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1445 unsigned Reg = FuncInfo->getSRetReturnReg();
1447 "SRetReturnReg should have been set in LowerFormalArguments().");
1448 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1450 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1451 Flag = Chain.getValue(1);
1453 // RAX now acts like a return value.
1454 MRI.addLiveOut(X86::RAX);
1457 RetOps[0] = Chain; // Update chain.
1459 // Add the flag if we have it.
1461 RetOps.push_back(Flag);
1463 return DAG.getNode(X86ISD::RET_FLAG, dl,
1464 MVT::Other, &RetOps[0], RetOps.size());
1467 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N) const {
1468 if (N->getNumValues() != 1)
1470 if (!N->hasNUsesOfValue(1, 0))
1473 SDNode *Copy = *N->use_begin();
1474 if (Copy->getOpcode() != ISD::CopyToReg &&
1475 Copy->getOpcode() != ISD::FP_EXTEND)
1478 bool HasRet = false;
1479 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1481 if (UI->getOpcode() != X86ISD::RET_FLAG)
1490 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
1491 ISD::NodeType ExtendKind) const {
1493 // TODO: Is this also valid on 32-bit?
1494 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
1495 ReturnMVT = MVT::i8;
1497 ReturnMVT = MVT::i32;
1499 EVT MinVT = getRegisterType(Context, ReturnMVT);
1500 return VT.bitsLT(MinVT) ? MinVT : VT;
1503 /// LowerCallResult - Lower the result values of a call into the
1504 /// appropriate copies out of appropriate physical registers.
1507 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1508 CallingConv::ID CallConv, bool isVarArg,
1509 const SmallVectorImpl<ISD::InputArg> &Ins,
1510 DebugLoc dl, SelectionDAG &DAG,
1511 SmallVectorImpl<SDValue> &InVals) const {
1513 // Assign locations to each value returned by this call.
1514 SmallVector<CCValAssign, 16> RVLocs;
1515 bool Is64Bit = Subtarget->is64Bit();
1516 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1517 getTargetMachine(), RVLocs, *DAG.getContext());
1518 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1520 // Copy all of the result registers out of their specified physreg.
1521 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1522 CCValAssign &VA = RVLocs[i];
1523 EVT CopyVT = VA.getValVT();
1525 // If this is x86-64, and we disabled SSE, we can't return FP values
1526 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1527 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasXMM())) {
1528 report_fatal_error("SSE register return with SSE disabled");
1533 // If this is a call to a function that returns an fp value on the floating
1534 // point stack, we must guarantee the the value is popped from the stack, so
1535 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1536 // if the return value is not used. We use the FpPOP_RETVAL instruction
1538 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1539 // If we prefer to use the value in xmm registers, copy it out as f80 and
1540 // use a truncate to move it from fp stack reg to xmm reg.
1541 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1542 SDValue Ops[] = { Chain, InFlag };
1543 Chain = SDValue(DAG.getMachineNode(X86::FpPOP_RETVAL, dl, CopyVT,
1544 MVT::Other, MVT::Glue, Ops, 2), 1);
1545 Val = Chain.getValue(0);
1547 // Round the f80 to the right size, which also moves it to the appropriate
1549 if (CopyVT != VA.getValVT())
1550 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1551 // This truncation won't change the value.
1552 DAG.getIntPtrConstant(1));
1554 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1555 CopyVT, InFlag).getValue(1);
1556 Val = Chain.getValue(0);
1558 InFlag = Chain.getValue(2);
1559 InVals.push_back(Val);
1566 //===----------------------------------------------------------------------===//
1567 // C & StdCall & Fast Calling Convention implementation
1568 //===----------------------------------------------------------------------===//
1569 // StdCall calling convention seems to be standard for many Windows' API
1570 // routines and around. It differs from C calling convention just a little:
1571 // callee should clean up the stack, not caller. Symbols should be also
1572 // decorated in some fancy way :) It doesn't support any vector arguments.
1573 // For info on fast calling convention see Fast Calling Convention (tail call)
1574 // implementation LowerX86_32FastCCCallTo.
1576 /// CallIsStructReturn - Determines whether a call uses struct return
1578 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1582 return Outs[0].Flags.isSRet();
1585 /// ArgsAreStructReturn - Determines whether a function uses struct
1586 /// return semantics.
1588 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1592 return Ins[0].Flags.isSRet();
1595 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1596 /// by "Src" to address "Dst" with size and alignment information specified by
1597 /// the specific parameter attribute. The copy will be passed as a byval
1598 /// function parameter.
1600 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1601 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1603 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1605 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1606 /*isVolatile*/false, /*AlwaysInline=*/true,
1607 MachinePointerInfo(), MachinePointerInfo());
1610 /// IsTailCallConvention - Return true if the calling convention is one that
1611 /// supports tail call optimization.
1612 static bool IsTailCallConvention(CallingConv::ID CC) {
1613 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1616 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1617 if (!CI->isTailCall())
1621 CallingConv::ID CalleeCC = CS.getCallingConv();
1622 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
1628 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1629 /// a tailcall target by changing its ABI.
1630 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1631 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1635 X86TargetLowering::LowerMemArgument(SDValue Chain,
1636 CallingConv::ID CallConv,
1637 const SmallVectorImpl<ISD::InputArg> &Ins,
1638 DebugLoc dl, SelectionDAG &DAG,
1639 const CCValAssign &VA,
1640 MachineFrameInfo *MFI,
1642 // Create the nodes corresponding to a load from this parameter slot.
1643 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1644 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1645 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1648 // If value is passed by pointer we have address passed instead of the value
1650 if (VA.getLocInfo() == CCValAssign::Indirect)
1651 ValVT = VA.getLocVT();
1653 ValVT = VA.getValVT();
1655 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1656 // changed with more analysis.
1657 // In case of tail call optimization mark all arguments mutable. Since they
1658 // could be overwritten by lowering of arguments in case of a tail call.
1659 if (Flags.isByVal()) {
1660 unsigned Bytes = Flags.getByValSize();
1661 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
1662 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
1663 return DAG.getFrameIndex(FI, getPointerTy());
1665 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1666 VA.getLocMemOffset(), isImmutable);
1667 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1668 return DAG.getLoad(ValVT, dl, Chain, FIN,
1669 MachinePointerInfo::getFixedStack(FI),
1675 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1676 CallingConv::ID CallConv,
1678 const SmallVectorImpl<ISD::InputArg> &Ins,
1681 SmallVectorImpl<SDValue> &InVals)
1683 MachineFunction &MF = DAG.getMachineFunction();
1684 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1686 const Function* Fn = MF.getFunction();
1687 if (Fn->hasExternalLinkage() &&
1688 Subtarget->isTargetCygMing() &&
1689 Fn->getName() == "main")
1690 FuncInfo->setForceFramePointer(true);
1692 MachineFrameInfo *MFI = MF.getFrameInfo();
1693 bool Is64Bit = Subtarget->is64Bit();
1694 bool IsWin64 = Subtarget->isTargetWin64();
1696 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1697 "Var args not supported with calling convention fastcc or ghc");
1699 // Assign locations to all of the incoming arguments.
1700 SmallVector<CCValAssign, 16> ArgLocs;
1701 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1702 ArgLocs, *DAG.getContext());
1704 // Allocate shadow area for Win64
1706 CCInfo.AllocateStack(32, 8);
1709 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
1711 unsigned LastVal = ~0U;
1713 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1714 CCValAssign &VA = ArgLocs[i];
1715 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1717 assert(VA.getValNo() != LastVal &&
1718 "Don't support value assigned to multiple locs yet");
1719 LastVal = VA.getValNo();
1721 if (VA.isRegLoc()) {
1722 EVT RegVT = VA.getLocVT();
1723 TargetRegisterClass *RC = NULL;
1724 if (RegVT == MVT::i32)
1725 RC = X86::GR32RegisterClass;
1726 else if (Is64Bit && RegVT == MVT::i64)
1727 RC = X86::GR64RegisterClass;
1728 else if (RegVT == MVT::f32)
1729 RC = X86::FR32RegisterClass;
1730 else if (RegVT == MVT::f64)
1731 RC = X86::FR64RegisterClass;
1732 else if (RegVT.isVector() && RegVT.getSizeInBits() == 256)
1733 RC = X86::VR256RegisterClass;
1734 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1735 RC = X86::VR128RegisterClass;
1736 else if (RegVT == MVT::x86mmx)
1737 RC = X86::VR64RegisterClass;
1739 llvm_unreachable("Unknown argument type!");
1741 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1742 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1744 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1745 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1747 if (VA.getLocInfo() == CCValAssign::SExt)
1748 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1749 DAG.getValueType(VA.getValVT()));
1750 else if (VA.getLocInfo() == CCValAssign::ZExt)
1751 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1752 DAG.getValueType(VA.getValVT()));
1753 else if (VA.getLocInfo() == CCValAssign::BCvt)
1754 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
1756 if (VA.isExtInLoc()) {
1757 // Handle MMX values passed in XMM regs.
1758 if (RegVT.isVector()) {
1759 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(),
1762 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1765 assert(VA.isMemLoc());
1766 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1769 // If value is passed via pointer - do a load.
1770 if (VA.getLocInfo() == CCValAssign::Indirect)
1771 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
1772 MachinePointerInfo(), false, false, 0);
1774 InVals.push_back(ArgValue);
1777 // The x86-64 ABI for returning structs by value requires that we copy
1778 // the sret argument into %rax for the return. Save the argument into
1779 // a virtual register so that we can access it from the return points.
1780 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1781 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1782 unsigned Reg = FuncInfo->getSRetReturnReg();
1784 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1785 FuncInfo->setSRetReturnReg(Reg);
1787 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1788 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1791 unsigned StackSize = CCInfo.getNextStackOffset();
1792 // Align stack specially for tail calls.
1793 if (FuncIsMadeTailCallSafe(CallConv))
1794 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1796 // If the function takes variable number of arguments, make a frame index for
1797 // the start of the first vararg value... for expansion of llvm.va_start.
1799 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1800 CallConv != CallingConv::X86_ThisCall)) {
1801 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1804 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1806 // FIXME: We should really autogenerate these arrays
1807 static const unsigned GPR64ArgRegsWin64[] = {
1808 X86::RCX, X86::RDX, X86::R8, X86::R9
1810 static const unsigned GPR64ArgRegs64Bit[] = {
1811 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1813 static const unsigned XMMArgRegs64Bit[] = {
1814 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1815 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1817 const unsigned *GPR64ArgRegs;
1818 unsigned NumXMMRegs = 0;
1821 // The XMM registers which might contain var arg parameters are shadowed
1822 // in their paired GPR. So we only need to save the GPR to their home
1824 TotalNumIntRegs = 4;
1825 GPR64ArgRegs = GPR64ArgRegsWin64;
1827 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1828 GPR64ArgRegs = GPR64ArgRegs64Bit;
1830 NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs64Bit, TotalNumXMMRegs);
1832 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1835 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1836 assert(!(NumXMMRegs && !Subtarget->hasXMM()) &&
1837 "SSE register cannot be used when SSE is disabled!");
1838 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1839 "SSE register cannot be used when SSE is disabled!");
1840 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasXMM())
1841 // Kernel mode asks for SSE to be disabled, so don't push them
1843 TotalNumXMMRegs = 0;
1846 const TargetFrameLowering &TFI = *getTargetMachine().getFrameLowering();
1847 // Get to the caller-allocated home save location. Add 8 to account
1848 // for the return address.
1849 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
1850 FuncInfo->setRegSaveFrameIndex(
1851 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
1852 // Fixup to set vararg frame on shadow area (4 x i64).
1854 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
1856 // For X86-64, if there are vararg parameters that are passed via
1857 // registers, then we must store them to their spots on the stack so they
1858 // may be loaded by deferencing the result of va_next.
1859 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1860 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1861 FuncInfo->setRegSaveFrameIndex(
1862 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1866 // Store the integer parameter registers.
1867 SmallVector<SDValue, 8> MemOps;
1868 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1870 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1871 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1872 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1873 DAG.getIntPtrConstant(Offset));
1874 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1875 X86::GR64RegisterClass);
1876 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1878 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1879 MachinePointerInfo::getFixedStack(
1880 FuncInfo->getRegSaveFrameIndex(), Offset),
1882 MemOps.push_back(Store);
1886 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1887 // Now store the XMM (fp + vector) parameter registers.
1888 SmallVector<SDValue, 11> SaveXMMOps;
1889 SaveXMMOps.push_back(Chain);
1891 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1892 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1893 SaveXMMOps.push_back(ALVal);
1895 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1896 FuncInfo->getRegSaveFrameIndex()));
1897 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1898 FuncInfo->getVarArgsFPOffset()));
1900 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1901 unsigned VReg = MF.addLiveIn(XMMArgRegs64Bit[NumXMMRegs],
1902 X86::VR128RegisterClass);
1903 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1904 SaveXMMOps.push_back(Val);
1906 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1908 &SaveXMMOps[0], SaveXMMOps.size()));
1911 if (!MemOps.empty())
1912 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1913 &MemOps[0], MemOps.size());
1917 // Some CCs need callee pop.
1918 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, GuaranteedTailCallOpt)) {
1919 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1921 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1922 // If this is an sret function, the return should pop the hidden pointer.
1923 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1924 FuncInfo->setBytesToPopOnReturn(4);
1928 // RegSaveFrameIndex is X86-64 only.
1929 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1930 if (CallConv == CallingConv::X86_FastCall ||
1931 CallConv == CallingConv::X86_ThisCall)
1932 // fastcc functions can't have varargs.
1933 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1936 FuncInfo->setArgumentStackSize(StackSize);
1942 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1943 SDValue StackPtr, SDValue Arg,
1944 DebugLoc dl, SelectionDAG &DAG,
1945 const CCValAssign &VA,
1946 ISD::ArgFlagsTy Flags) const {
1947 unsigned LocMemOffset = VA.getLocMemOffset();
1948 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1949 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1950 if (Flags.isByVal())
1951 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1953 return DAG.getStore(Chain, dl, Arg, PtrOff,
1954 MachinePointerInfo::getStack(LocMemOffset),
1958 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1959 /// optimization is performed and it is required.
1961 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1962 SDValue &OutRetAddr, SDValue Chain,
1963 bool IsTailCall, bool Is64Bit,
1964 int FPDiff, DebugLoc dl) const {
1965 // Adjust the Return address stack slot.
1966 EVT VT = getPointerTy();
1967 OutRetAddr = getReturnAddressFrameIndex(DAG);
1969 // Load the "old" Return address.
1970 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
1972 return SDValue(OutRetAddr.getNode(), 1);
1975 /// EmitTailCallStoreRetAddr - Emit a store of the return address if tail call
1976 /// optimization is performed and it is required (FPDiff!=0).
1978 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1979 SDValue Chain, SDValue RetAddrFrIdx,
1980 bool Is64Bit, int FPDiff, DebugLoc dl) {
1981 // Store the return address to the appropriate stack slot.
1982 if (!FPDiff) return Chain;
1983 // Calculate the new stack slot for the return address.
1984 int SlotSize = Is64Bit ? 8 : 4;
1985 int NewReturnAddrFI =
1986 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1987 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1988 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1989 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1990 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
1996 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1997 CallingConv::ID CallConv, bool isVarArg,
1999 const SmallVectorImpl<ISD::OutputArg> &Outs,
2000 const SmallVectorImpl<SDValue> &OutVals,
2001 const SmallVectorImpl<ISD::InputArg> &Ins,
2002 DebugLoc dl, SelectionDAG &DAG,
2003 SmallVectorImpl<SDValue> &InVals) const {
2004 MachineFunction &MF = DAG.getMachineFunction();
2005 bool Is64Bit = Subtarget->is64Bit();
2006 bool IsWin64 = Subtarget->isTargetWin64();
2007 bool IsStructRet = CallIsStructReturn(Outs);
2008 bool IsSibcall = false;
2011 // Check if it's really possible to do a tail call.
2012 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2013 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
2014 Outs, OutVals, Ins, DAG);
2016 // Sibcalls are automatically detected tailcalls which do not require
2018 if (!GuaranteedTailCallOpt && isTailCall)
2025 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2026 "Var args not supported with calling convention fastcc or ghc");
2028 // Analyze operands of the call, assigning locations to each operand.
2029 SmallVector<CCValAssign, 16> ArgLocs;
2030 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
2031 ArgLocs, *DAG.getContext());
2033 // Allocate shadow area for Win64
2035 CCInfo.AllocateStack(32, 8);
2038 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2040 // Get a count of how many bytes are to be pushed on the stack.
2041 unsigned NumBytes = CCInfo.getNextStackOffset();
2043 // This is a sibcall. The memory operands are available in caller's
2044 // own caller's stack.
2046 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
2047 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2050 if (isTailCall && !IsSibcall) {
2051 // Lower arguments at fp - stackoffset + fpdiff.
2052 unsigned NumBytesCallerPushed =
2053 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
2054 FPDiff = NumBytesCallerPushed - NumBytes;
2056 // Set the delta of movement of the returnaddr stackslot.
2057 // But only set if delta is greater than previous delta.
2058 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
2059 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
2063 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2065 SDValue RetAddrFrIdx;
2066 // Load return address for tail calls.
2067 if (isTailCall && FPDiff)
2068 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2069 Is64Bit, FPDiff, dl);
2071 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2072 SmallVector<SDValue, 8> MemOpChains;
2075 // Walk the register/memloc assignments, inserting copies/loads. In the case
2076 // of tail call optimization arguments are handle later.
2077 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2078 CCValAssign &VA = ArgLocs[i];
2079 EVT RegVT = VA.getLocVT();
2080 SDValue Arg = OutVals[i];
2081 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2082 bool isByVal = Flags.isByVal();
2084 // Promote the value if needed.
2085 switch (VA.getLocInfo()) {
2086 default: llvm_unreachable("Unknown loc info!");
2087 case CCValAssign::Full: break;
2088 case CCValAssign::SExt:
2089 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2091 case CCValAssign::ZExt:
2092 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2094 case CCValAssign::AExt:
2095 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
2096 // Special case: passing MMX values in XMM registers.
2097 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2098 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2099 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2101 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2103 case CCValAssign::BCvt:
2104 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2106 case CCValAssign::Indirect: {
2107 // Store the argument.
2108 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2109 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2110 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2111 MachinePointerInfo::getFixedStack(FI),
2118 if (VA.isRegLoc()) {
2119 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2120 if (isVarArg && IsWin64) {
2121 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2122 // shadow reg if callee is a varargs function.
2123 unsigned ShadowReg = 0;
2124 switch (VA.getLocReg()) {
2125 case X86::XMM0: ShadowReg = X86::RCX; break;
2126 case X86::XMM1: ShadowReg = X86::RDX; break;
2127 case X86::XMM2: ShadowReg = X86::R8; break;
2128 case X86::XMM3: ShadowReg = X86::R9; break;
2131 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2133 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2134 assert(VA.isMemLoc());
2135 if (StackPtr.getNode() == 0)
2136 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
2137 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2138 dl, DAG, VA, Flags));
2142 if (!MemOpChains.empty())
2143 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2144 &MemOpChains[0], MemOpChains.size());
2146 // Build a sequence of copy-to-reg nodes chained together with token chain
2147 // and flag operands which copy the outgoing args into registers.
2149 // Tail call byval lowering might overwrite argument registers so in case of
2150 // tail call optimization the copies to registers are lowered later.
2152 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2153 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2154 RegsToPass[i].second, InFlag);
2155 InFlag = Chain.getValue(1);
2158 if (Subtarget->isPICStyleGOT()) {
2159 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2162 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
2163 DAG.getNode(X86ISD::GlobalBaseReg,
2164 DebugLoc(), getPointerTy()),
2166 InFlag = Chain.getValue(1);
2168 // If we are tail calling and generating PIC/GOT style code load the
2169 // address of the callee into ECX. The value in ecx is used as target of
2170 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2171 // for tail calls on PIC/GOT architectures. Normally we would just put the
2172 // address of GOT into ebx and then call target@PLT. But for tail calls
2173 // ebx would be restored (since ebx is callee saved) before jumping to the
2176 // Note: The actual moving to ECX is done further down.
2177 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2178 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2179 !G->getGlobal()->hasProtectedVisibility())
2180 Callee = LowerGlobalAddress(Callee, DAG);
2181 else if (isa<ExternalSymbolSDNode>(Callee))
2182 Callee = LowerExternalSymbol(Callee, DAG);
2186 if (Is64Bit && isVarArg && !IsWin64) {
2187 // From AMD64 ABI document:
2188 // For calls that may call functions that use varargs or stdargs
2189 // (prototype-less calls or calls to functions containing ellipsis (...) in
2190 // the declaration) %al is used as hidden argument to specify the number
2191 // of SSE registers used. The contents of %al do not need to match exactly
2192 // the number of registers, but must be an ubound on the number of SSE
2193 // registers used and is in the range 0 - 8 inclusive.
2195 // Count the number of XMM registers allocated.
2196 static const unsigned XMMArgRegs[] = {
2197 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2198 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2200 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2201 assert((Subtarget->hasXMM() || !NumXMMRegs)
2202 && "SSE registers cannot be used when SSE is disabled");
2204 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2205 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2206 InFlag = Chain.getValue(1);
2210 // For tail calls lower the arguments to the 'real' stack slot.
2212 // Force all the incoming stack arguments to be loaded from the stack
2213 // before any new outgoing arguments are stored to the stack, because the
2214 // outgoing stack slots may alias the incoming argument stack slots, and
2215 // the alias isn't otherwise explicit. This is slightly more conservative
2216 // than necessary, because it means that each store effectively depends
2217 // on every argument instead of just those arguments it would clobber.
2218 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2220 SmallVector<SDValue, 8> MemOpChains2;
2223 // Do not flag preceding copytoreg stuff together with the following stuff.
2225 if (GuaranteedTailCallOpt) {
2226 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2227 CCValAssign &VA = ArgLocs[i];
2230 assert(VA.isMemLoc());
2231 SDValue Arg = OutVals[i];
2232 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2233 // Create frame index.
2234 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2235 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2236 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2237 FIN = DAG.getFrameIndex(FI, getPointerTy());
2239 if (Flags.isByVal()) {
2240 // Copy relative to framepointer.
2241 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2242 if (StackPtr.getNode() == 0)
2243 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2245 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2247 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2251 // Store relative to framepointer.
2252 MemOpChains2.push_back(
2253 DAG.getStore(ArgChain, dl, Arg, FIN,
2254 MachinePointerInfo::getFixedStack(FI),
2260 if (!MemOpChains2.empty())
2261 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2262 &MemOpChains2[0], MemOpChains2.size());
2264 // Copy arguments to their registers.
2265 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2266 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2267 RegsToPass[i].second, InFlag);
2268 InFlag = Chain.getValue(1);
2272 // Store the return address to the appropriate stack slot.
2273 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2277 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2278 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2279 // In the 64-bit large code model, we have to make all calls
2280 // through a register, since the call instruction's 32-bit
2281 // pc-relative offset may not be large enough to hold the whole
2283 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2284 // If the callee is a GlobalAddress node (quite common, every direct call
2285 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2288 // We should use extra load for direct calls to dllimported functions in
2290 const GlobalValue *GV = G->getGlobal();
2291 if (!GV->hasDLLImportLinkage()) {
2292 unsigned char OpFlags = 0;
2293 bool ExtraLoad = false;
2294 unsigned WrapperKind = ISD::DELETED_NODE;
2296 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2297 // external symbols most go through the PLT in PIC mode. If the symbol
2298 // has hidden or protected visibility, or if it is static or local, then
2299 // we don't need to use the PLT - we can directly call it.
2300 if (Subtarget->isTargetELF() &&
2301 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2302 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2303 OpFlags = X86II::MO_PLT;
2304 } else if (Subtarget->isPICStyleStubAny() &&
2305 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2306 (!Subtarget->getTargetTriple().isMacOSX() ||
2307 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2308 // PC-relative references to external symbols should go through $stub,
2309 // unless we're building with the leopard linker or later, which
2310 // automatically synthesizes these stubs.
2311 OpFlags = X86II::MO_DARWIN_STUB;
2312 } else if (Subtarget->isPICStyleRIPRel() &&
2313 isa<Function>(GV) &&
2314 cast<Function>(GV)->hasFnAttr(Attribute::NonLazyBind)) {
2315 // If the function is marked as non-lazy, generate an indirect call
2316 // which loads from the GOT directly. This avoids runtime overhead
2317 // at the cost of eager binding (and one extra byte of encoding).
2318 OpFlags = X86II::MO_GOTPCREL;
2319 WrapperKind = X86ISD::WrapperRIP;
2323 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2324 G->getOffset(), OpFlags);
2326 // Add a wrapper if needed.
2327 if (WrapperKind != ISD::DELETED_NODE)
2328 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
2329 // Add extra indirection if needed.
2331 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
2332 MachinePointerInfo::getGOT(),
2335 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2336 unsigned char OpFlags = 0;
2338 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
2339 // external symbols should go through the PLT.
2340 if (Subtarget->isTargetELF() &&
2341 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2342 OpFlags = X86II::MO_PLT;
2343 } else if (Subtarget->isPICStyleStubAny() &&
2344 (!Subtarget->getTargetTriple().isMacOSX() ||
2345 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2346 // PC-relative references to external symbols should go through $stub,
2347 // unless we're building with the leopard linker or later, which
2348 // automatically synthesizes these stubs.
2349 OpFlags = X86II::MO_DARWIN_STUB;
2352 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2356 // Returns a chain & a flag for retval copy to use.
2357 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2358 SmallVector<SDValue, 8> Ops;
2360 if (!IsSibcall && isTailCall) {
2361 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2362 DAG.getIntPtrConstant(0, true), InFlag);
2363 InFlag = Chain.getValue(1);
2366 Ops.push_back(Chain);
2367 Ops.push_back(Callee);
2370 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2372 // Add argument registers to the end of the list so that they are known live
2374 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2375 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2376 RegsToPass[i].second.getValueType()));
2378 // Add an implicit use GOT pointer in EBX.
2379 if (!isTailCall && Subtarget->isPICStyleGOT())
2380 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2382 // Add an implicit use of AL for non-Windows x86 64-bit vararg functions.
2383 if (Is64Bit && isVarArg && !IsWin64)
2384 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2386 if (InFlag.getNode())
2387 Ops.push_back(InFlag);
2391 //// If this is the first return lowered for this function, add the regs
2392 //// to the liveout set for the function.
2393 // This isn't right, although it's probably harmless on x86; liveouts
2394 // should be computed from returns not tail calls. Consider a void
2395 // function making a tail call to a function returning int.
2396 return DAG.getNode(X86ISD::TC_RETURN, dl,
2397 NodeTys, &Ops[0], Ops.size());
2400 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2401 InFlag = Chain.getValue(1);
2403 // Create the CALLSEQ_END node.
2404 unsigned NumBytesForCalleeToPush;
2405 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, GuaranteedTailCallOpt))
2406 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2407 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2408 // If this is a call to a struct-return function, the callee
2409 // pops the hidden struct pointer, so we have to push it back.
2410 // This is common for Darwin/X86, Linux & Mingw32 targets.
2411 NumBytesForCalleeToPush = 4;
2413 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2415 // Returns a flag for retval copy to use.
2417 Chain = DAG.getCALLSEQ_END(Chain,
2418 DAG.getIntPtrConstant(NumBytes, true),
2419 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2422 InFlag = Chain.getValue(1);
2425 // Handle result values, copying them out of physregs into vregs that we
2427 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2428 Ins, dl, DAG, InVals);
2432 //===----------------------------------------------------------------------===//
2433 // Fast Calling Convention (tail call) implementation
2434 //===----------------------------------------------------------------------===//
2436 // Like std call, callee cleans arguments, convention except that ECX is
2437 // reserved for storing the tail called function address. Only 2 registers are
2438 // free for argument passing (inreg). Tail call optimization is performed
2440 // * tailcallopt is enabled
2441 // * caller/callee are fastcc
2442 // On X86_64 architecture with GOT-style position independent code only local
2443 // (within module) calls are supported at the moment.
2444 // To keep the stack aligned according to platform abi the function
2445 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2446 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2447 // If a tail called function callee has more arguments than the caller the
2448 // caller needs to make sure that there is room to move the RETADDR to. This is
2449 // achieved by reserving an area the size of the argument delta right after the
2450 // original REtADDR, but before the saved framepointer or the spilled registers
2451 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2463 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2464 /// for a 16 byte align requirement.
2466 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2467 SelectionDAG& DAG) const {
2468 MachineFunction &MF = DAG.getMachineFunction();
2469 const TargetMachine &TM = MF.getTarget();
2470 const TargetFrameLowering &TFI = *TM.getFrameLowering();
2471 unsigned StackAlignment = TFI.getStackAlignment();
2472 uint64_t AlignMask = StackAlignment - 1;
2473 int64_t Offset = StackSize;
2474 uint64_t SlotSize = TD->getPointerSize();
2475 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2476 // Number smaller than 12 so just add the difference.
2477 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2479 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2480 Offset = ((~AlignMask) & Offset) + StackAlignment +
2481 (StackAlignment-SlotSize);
2486 /// MatchingStackOffset - Return true if the given stack call argument is
2487 /// already available in the same position (relatively) of the caller's
2488 /// incoming argument stack.
2490 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2491 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2492 const X86InstrInfo *TII) {
2493 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2495 if (Arg.getOpcode() == ISD::CopyFromReg) {
2496 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2497 if (!TargetRegisterInfo::isVirtualRegister(VR))
2499 MachineInstr *Def = MRI->getVRegDef(VR);
2502 if (!Flags.isByVal()) {
2503 if (!TII->isLoadFromStackSlot(Def, FI))
2506 unsigned Opcode = Def->getOpcode();
2507 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2508 Def->getOperand(1).isFI()) {
2509 FI = Def->getOperand(1).getIndex();
2510 Bytes = Flags.getByValSize();
2514 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2515 if (Flags.isByVal())
2516 // ByVal argument is passed in as a pointer but it's now being
2517 // dereferenced. e.g.
2518 // define @foo(%struct.X* %A) {
2519 // tail call @bar(%struct.X* byval %A)
2522 SDValue Ptr = Ld->getBasePtr();
2523 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2526 FI = FINode->getIndex();
2527 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
2528 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
2529 FI = FINode->getIndex();
2530 Bytes = Flags.getByValSize();
2534 assert(FI != INT_MAX);
2535 if (!MFI->isFixedObjectIndex(FI))
2537 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2540 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2541 /// for tail call optimization. Targets which want to do tail call
2542 /// optimization should implement this function.
2544 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2545 CallingConv::ID CalleeCC,
2547 bool isCalleeStructRet,
2548 bool isCallerStructRet,
2549 const SmallVectorImpl<ISD::OutputArg> &Outs,
2550 const SmallVectorImpl<SDValue> &OutVals,
2551 const SmallVectorImpl<ISD::InputArg> &Ins,
2552 SelectionDAG& DAG) const {
2553 if (!IsTailCallConvention(CalleeCC) &&
2554 CalleeCC != CallingConv::C)
2557 // If -tailcallopt is specified, make fastcc functions tail-callable.
2558 const MachineFunction &MF = DAG.getMachineFunction();
2559 const Function *CallerF = DAG.getMachineFunction().getFunction();
2560 CallingConv::ID CallerCC = CallerF->getCallingConv();
2561 bool CCMatch = CallerCC == CalleeCC;
2563 if (GuaranteedTailCallOpt) {
2564 if (IsTailCallConvention(CalleeCC) && CCMatch)
2569 // Look for obvious safe cases to perform tail call optimization that do not
2570 // require ABI changes. This is what gcc calls sibcall.
2572 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2573 // emit a special epilogue.
2574 if (RegInfo->needsStackRealignment(MF))
2577 // Also avoid sibcall optimization if either caller or callee uses struct
2578 // return semantics.
2579 if (isCalleeStructRet || isCallerStructRet)
2582 // An stdcall caller is expected to clean up its arguments; the callee
2583 // isn't going to do that.
2584 if (!CCMatch && CallerCC==CallingConv::X86_StdCall)
2587 // Do not sibcall optimize vararg calls unless all arguments are passed via
2589 if (isVarArg && !Outs.empty()) {
2591 // Optimizing for varargs on Win64 is unlikely to be safe without
2592 // additional testing.
2593 if (Subtarget->isTargetWin64())
2596 SmallVector<CCValAssign, 16> ArgLocs;
2597 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2598 getTargetMachine(), ArgLocs, *DAG.getContext());
2600 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2601 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
2602 if (!ArgLocs[i].isRegLoc())
2606 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2607 // Therefore if it's not used by the call it is not safe to optimize this into
2609 bool Unused = false;
2610 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2617 SmallVector<CCValAssign, 16> RVLocs;
2618 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(),
2619 getTargetMachine(), RVLocs, *DAG.getContext());
2620 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2621 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2622 CCValAssign &VA = RVLocs[i];
2623 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2628 // If the calling conventions do not match, then we'd better make sure the
2629 // results are returned in the same way as what the caller expects.
2631 SmallVector<CCValAssign, 16> RVLocs1;
2632 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
2633 getTargetMachine(), RVLocs1, *DAG.getContext());
2634 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2636 SmallVector<CCValAssign, 16> RVLocs2;
2637 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
2638 getTargetMachine(), RVLocs2, *DAG.getContext());
2639 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2641 if (RVLocs1.size() != RVLocs2.size())
2643 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2644 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2646 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2648 if (RVLocs1[i].isRegLoc()) {
2649 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2652 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2658 // If the callee takes no arguments then go on to check the results of the
2660 if (!Outs.empty()) {
2661 // Check if stack adjustment is needed. For now, do not do this if any
2662 // argument is passed on the stack.
2663 SmallVector<CCValAssign, 16> ArgLocs;
2664 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2665 getTargetMachine(), ArgLocs, *DAG.getContext());
2667 // Allocate shadow area for Win64
2668 if (Subtarget->isTargetWin64()) {
2669 CCInfo.AllocateStack(32, 8);
2672 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2673 if (CCInfo.getNextStackOffset()) {
2674 MachineFunction &MF = DAG.getMachineFunction();
2675 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2678 // Check if the arguments are already laid out in the right way as
2679 // the caller's fixed stack objects.
2680 MachineFrameInfo *MFI = MF.getFrameInfo();
2681 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2682 const X86InstrInfo *TII =
2683 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2684 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2685 CCValAssign &VA = ArgLocs[i];
2686 SDValue Arg = OutVals[i];
2687 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2688 if (VA.getLocInfo() == CCValAssign::Indirect)
2690 if (!VA.isRegLoc()) {
2691 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2698 // If the tailcall address may be in a register, then make sure it's
2699 // possible to register allocate for it. In 32-bit, the call address can
2700 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2701 // callee-saved registers are restored. These happen to be the same
2702 // registers used to pass 'inreg' arguments so watch out for those.
2703 if (!Subtarget->is64Bit() &&
2704 !isa<GlobalAddressSDNode>(Callee) &&
2705 !isa<ExternalSymbolSDNode>(Callee)) {
2706 unsigned NumInRegs = 0;
2707 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2708 CCValAssign &VA = ArgLocs[i];
2711 unsigned Reg = VA.getLocReg();
2714 case X86::EAX: case X86::EDX: case X86::ECX:
2715 if (++NumInRegs == 3)
2727 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2728 return X86::createFastISel(funcInfo);
2732 //===----------------------------------------------------------------------===//
2733 // Other Lowering Hooks
2734 //===----------------------------------------------------------------------===//
2736 static bool MayFoldLoad(SDValue Op) {
2737 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
2740 static bool MayFoldIntoStore(SDValue Op) {
2741 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
2744 static bool isTargetShuffle(unsigned Opcode) {
2746 default: return false;
2747 case X86ISD::PSHUFD:
2748 case X86ISD::PSHUFHW:
2749 case X86ISD::PSHUFLW:
2750 case X86ISD::SHUFPD:
2751 case X86ISD::PALIGN:
2752 case X86ISD::SHUFPS:
2753 case X86ISD::MOVLHPS:
2754 case X86ISD::MOVLHPD:
2755 case X86ISD::MOVHLPS:
2756 case X86ISD::MOVLPS:
2757 case X86ISD::MOVLPD:
2758 case X86ISD::MOVSHDUP:
2759 case X86ISD::MOVSLDUP:
2760 case X86ISD::MOVDDUP:
2763 case X86ISD::UNPCKLPS:
2764 case X86ISD::UNPCKLPD:
2765 case X86ISD::VUNPCKLPSY:
2766 case X86ISD::VUNPCKLPDY:
2767 case X86ISD::PUNPCKLWD:
2768 case X86ISD::PUNPCKLBW:
2769 case X86ISD::PUNPCKLDQ:
2770 case X86ISD::PUNPCKLQDQ:
2771 case X86ISD::UNPCKHPS:
2772 case X86ISD::UNPCKHPD:
2773 case X86ISD::VUNPCKHPSY:
2774 case X86ISD::VUNPCKHPDY:
2775 case X86ISD::PUNPCKHWD:
2776 case X86ISD::PUNPCKHBW:
2777 case X86ISD::PUNPCKHDQ:
2778 case X86ISD::PUNPCKHQDQ:
2779 case X86ISD::VPERMILPS:
2780 case X86ISD::VPERMILPSY:
2781 case X86ISD::VPERMILPD:
2782 case X86ISD::VPERMILPDY:
2783 case X86ISD::VPERM2F128:
2789 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2790 SDValue V1, SelectionDAG &DAG) {
2792 default: llvm_unreachable("Unknown x86 shuffle node");
2793 case X86ISD::MOVSHDUP:
2794 case X86ISD::MOVSLDUP:
2795 case X86ISD::MOVDDUP:
2796 return DAG.getNode(Opc, dl, VT, V1);
2802 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2803 SDValue V1, unsigned TargetMask, SelectionDAG &DAG) {
2805 default: llvm_unreachable("Unknown x86 shuffle node");
2806 case X86ISD::PSHUFD:
2807 case X86ISD::PSHUFHW:
2808 case X86ISD::PSHUFLW:
2809 case X86ISD::VPERMILPS:
2810 case X86ISD::VPERMILPSY:
2811 case X86ISD::VPERMILPD:
2812 case X86ISD::VPERMILPDY:
2813 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
2819 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2820 SDValue V1, SDValue V2, unsigned TargetMask, SelectionDAG &DAG) {
2822 default: llvm_unreachable("Unknown x86 shuffle node");
2823 case X86ISD::PALIGN:
2824 case X86ISD::SHUFPD:
2825 case X86ISD::SHUFPS:
2826 case X86ISD::VPERM2F128:
2827 return DAG.getNode(Opc, dl, VT, V1, V2,
2828 DAG.getConstant(TargetMask, MVT::i8));
2833 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2834 SDValue V1, SDValue V2, SelectionDAG &DAG) {
2836 default: llvm_unreachable("Unknown x86 shuffle node");
2837 case X86ISD::MOVLHPS:
2838 case X86ISD::MOVLHPD:
2839 case X86ISD::MOVHLPS:
2840 case X86ISD::MOVLPS:
2841 case X86ISD::MOVLPD:
2844 case X86ISD::UNPCKLPS:
2845 case X86ISD::UNPCKLPD:
2846 case X86ISD::VUNPCKLPSY:
2847 case X86ISD::VUNPCKLPDY:
2848 case X86ISD::PUNPCKLWD:
2849 case X86ISD::PUNPCKLBW:
2850 case X86ISD::PUNPCKLDQ:
2851 case X86ISD::PUNPCKLQDQ:
2852 case X86ISD::UNPCKHPS:
2853 case X86ISD::UNPCKHPD:
2854 case X86ISD::VUNPCKHPSY:
2855 case X86ISD::VUNPCKHPDY:
2856 case X86ISD::PUNPCKHWD:
2857 case X86ISD::PUNPCKHBW:
2858 case X86ISD::PUNPCKHDQ:
2859 case X86ISD::PUNPCKHQDQ:
2860 return DAG.getNode(Opc, dl, VT, V1, V2);
2865 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2866 MachineFunction &MF = DAG.getMachineFunction();
2867 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2868 int ReturnAddrIndex = FuncInfo->getRAIndex();
2870 if (ReturnAddrIndex == 0) {
2871 // Set up a frame object for the return address.
2872 uint64_t SlotSize = TD->getPointerSize();
2873 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2875 FuncInfo->setRAIndex(ReturnAddrIndex);
2878 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2882 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2883 bool hasSymbolicDisplacement) {
2884 // Offset should fit into 32 bit immediate field.
2885 if (!isInt<32>(Offset))
2888 // If we don't have a symbolic displacement - we don't have any extra
2890 if (!hasSymbolicDisplacement)
2893 // FIXME: Some tweaks might be needed for medium code model.
2894 if (M != CodeModel::Small && M != CodeModel::Kernel)
2897 // For small code model we assume that latest object is 16MB before end of 31
2898 // bits boundary. We may also accept pretty large negative constants knowing
2899 // that all objects are in the positive half of address space.
2900 if (M == CodeModel::Small && Offset < 16*1024*1024)
2903 // For kernel code model we know that all object resist in the negative half
2904 // of 32bits address space. We may not accept negative offsets, since they may
2905 // be just off and we may accept pretty large positive ones.
2906 if (M == CodeModel::Kernel && Offset > 0)
2912 /// isCalleePop - Determines whether the callee is required to pop its
2913 /// own arguments. Callee pop is necessary to support tail calls.
2914 bool X86::isCalleePop(CallingConv::ID CallingConv,
2915 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
2919 switch (CallingConv) {
2922 case CallingConv::X86_StdCall:
2924 case CallingConv::X86_FastCall:
2926 case CallingConv::X86_ThisCall:
2928 case CallingConv::Fast:
2930 case CallingConv::GHC:
2935 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2936 /// specific condition code, returning the condition code and the LHS/RHS of the
2937 /// comparison to make.
2938 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2939 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2941 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2942 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2943 // X > -1 -> X == 0, jump !sign.
2944 RHS = DAG.getConstant(0, RHS.getValueType());
2945 return X86::COND_NS;
2946 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2947 // X < 0 -> X == 0, jump on sign.
2949 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2951 RHS = DAG.getConstant(0, RHS.getValueType());
2952 return X86::COND_LE;
2956 switch (SetCCOpcode) {
2957 default: llvm_unreachable("Invalid integer condition!");
2958 case ISD::SETEQ: return X86::COND_E;
2959 case ISD::SETGT: return X86::COND_G;
2960 case ISD::SETGE: return X86::COND_GE;
2961 case ISD::SETLT: return X86::COND_L;
2962 case ISD::SETLE: return X86::COND_LE;
2963 case ISD::SETNE: return X86::COND_NE;
2964 case ISD::SETULT: return X86::COND_B;
2965 case ISD::SETUGT: return X86::COND_A;
2966 case ISD::SETULE: return X86::COND_BE;
2967 case ISD::SETUGE: return X86::COND_AE;
2971 // First determine if it is required or is profitable to flip the operands.
2973 // If LHS is a foldable load, but RHS is not, flip the condition.
2974 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
2975 !ISD::isNON_EXTLoad(RHS.getNode())) {
2976 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2977 std::swap(LHS, RHS);
2980 switch (SetCCOpcode) {
2986 std::swap(LHS, RHS);
2990 // On a floating point condition, the flags are set as follows:
2992 // 0 | 0 | 0 | X > Y
2993 // 0 | 0 | 1 | X < Y
2994 // 1 | 0 | 0 | X == Y
2995 // 1 | 1 | 1 | unordered
2996 switch (SetCCOpcode) {
2997 default: llvm_unreachable("Condcode should be pre-legalized away");
2999 case ISD::SETEQ: return X86::COND_E;
3000 case ISD::SETOLT: // flipped
3002 case ISD::SETGT: return X86::COND_A;
3003 case ISD::SETOLE: // flipped
3005 case ISD::SETGE: return X86::COND_AE;
3006 case ISD::SETUGT: // flipped
3008 case ISD::SETLT: return X86::COND_B;
3009 case ISD::SETUGE: // flipped
3011 case ISD::SETLE: return X86::COND_BE;
3013 case ISD::SETNE: return X86::COND_NE;
3014 case ISD::SETUO: return X86::COND_P;
3015 case ISD::SETO: return X86::COND_NP;
3017 case ISD::SETUNE: return X86::COND_INVALID;
3021 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
3022 /// code. Current x86 isa includes the following FP cmov instructions:
3023 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3024 static bool hasFPCMov(unsigned X86CC) {
3040 /// isFPImmLegal - Returns true if the target can instruction select the
3041 /// specified FP immediate natively. If false, the legalizer will
3042 /// materialize the FP immediate as a load from a constant pool.
3043 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3044 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3045 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3051 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3052 /// the specified range (L, H].
3053 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3054 return (Val < 0) || (Val >= Low && Val < Hi);
3057 /// isUndefOrInRange - Return true if every element in Mask, begining
3058 /// from position Pos and ending in Pos+Size, falls within the specified
3059 /// range (L, L+Pos]. or is undef.
3060 static bool isUndefOrInRange(const SmallVectorImpl<int> &Mask,
3061 int Pos, int Size, int Low, int Hi) {
3062 for (int i = Pos, e = Pos+Size; i != e; ++i)
3063 if (!isUndefOrInRange(Mask[i], Low, Hi))
3068 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3069 /// specified value.
3070 static bool isUndefOrEqual(int Val, int CmpVal) {
3071 if (Val < 0 || Val == CmpVal)
3076 /// isSequentialOrUndefInRange - Return true if every element in Mask, begining
3077 /// from position Pos and ending in Pos+Size, falls within the specified
3078 /// sequential range (L, L+Pos]. or is undef.
3079 static bool isSequentialOrUndefInRange(const SmallVectorImpl<int> &Mask,
3080 int Pos, int Size, int Low) {
3081 for (int i = Pos, e = Pos+Size; i != e; ++i, ++Low)
3082 if (!isUndefOrEqual(Mask[i], Low))
3087 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
3088 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
3089 /// the second operand.
3090 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3091 if (VT == MVT::v4f32 || VT == MVT::v4i32 )
3092 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
3093 if (VT == MVT::v2f64 || VT == MVT::v2i64)
3094 return (Mask[0] < 2 && Mask[1] < 2);
3098 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
3099 SmallVector<int, 8> M;
3101 return ::isPSHUFDMask(M, N->getValueType(0));
3104 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
3105 /// is suitable for input to PSHUFHW.
3106 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3107 if (VT != MVT::v8i16)
3110 // Lower quadword copied in order or undef.
3111 for (int i = 0; i != 4; ++i)
3112 if (Mask[i] >= 0 && Mask[i] != i)
3115 // Upper quadword shuffled.
3116 for (int i = 4; i != 8; ++i)
3117 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
3123 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
3124 SmallVector<int, 8> M;
3126 return ::isPSHUFHWMask(M, N->getValueType(0));
3129 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
3130 /// is suitable for input to PSHUFLW.
3131 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3132 if (VT != MVT::v8i16)
3135 // Upper quadword copied in order.
3136 for (int i = 4; i != 8; ++i)
3137 if (Mask[i] >= 0 && Mask[i] != i)
3140 // Lower quadword shuffled.
3141 for (int i = 0; i != 4; ++i)
3148 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
3149 SmallVector<int, 8> M;
3151 return ::isPSHUFLWMask(M, N->getValueType(0));
3154 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
3155 /// is suitable for input to PALIGNR.
3156 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
3158 int i, e = VT.getVectorNumElements();
3159 if (VT.getSizeInBits() != 128 && VT.getSizeInBits() != 64)
3162 // Do not handle v2i64 / v2f64 shuffles with palignr.
3163 if (e < 4 || !hasSSSE3)
3166 for (i = 0; i != e; ++i)
3170 // All undef, not a palignr.
3174 // Make sure we're shifting in the right direction.
3178 int s = Mask[i] - i;
3180 // Check the rest of the elements to see if they are consecutive.
3181 for (++i; i != e; ++i) {
3183 if (m >= 0 && m != s+i)
3189 /// isVSHUFPSYMask - Return true if the specified VECTOR_SHUFFLE operand
3190 /// specifies a shuffle of elements that is suitable for input to 256-bit
3192 static bool isVSHUFPSYMask(const SmallVectorImpl<int> &Mask, EVT VT,
3193 const X86Subtarget *Subtarget) {
3194 int NumElems = VT.getVectorNumElements();
3196 if (!Subtarget->hasAVX() || VT.getSizeInBits() != 256)
3202 // VSHUFPSY divides the resulting vector into 4 chunks.
3203 // The sources are also splitted into 4 chunks, and each destination
3204 // chunk must come from a different source chunk.
3206 // SRC1 => X7 X6 X5 X4 X3 X2 X1 X0
3207 // SRC2 => Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y9
3209 // DST => Y7..Y4, Y7..Y4, X7..X4, X7..X4,
3210 // Y3..Y0, Y3..Y0, X3..X0, X3..X0
3212 int QuarterSize = NumElems/4;
3213 int HalfSize = QuarterSize*2;
3214 for (int i = 0; i < QuarterSize; ++i)
3215 if (!isUndefOrInRange(Mask[i], 0, HalfSize))
3217 for (int i = QuarterSize; i < QuarterSize*2; ++i)
3218 if (!isUndefOrInRange(Mask[i], NumElems, NumElems+HalfSize))
3221 // The mask of the second half must be the same as the first but with
3222 // the appropriate offsets. This works in the same way as VPERMILPS
3223 // works with masks.
3224 for (int i = QuarterSize*2; i < QuarterSize*3; ++i) {
3225 if (!isUndefOrInRange(Mask[i], HalfSize, NumElems))
3227 int FstHalfIdx = i-HalfSize;
3228 if (Mask[FstHalfIdx] < 0)
3230 if (!isUndefOrEqual(Mask[i], Mask[FstHalfIdx]+HalfSize))
3233 for (int i = QuarterSize*3; i < NumElems; ++i) {
3234 if (!isUndefOrInRange(Mask[i], NumElems+HalfSize, NumElems*2))
3236 int FstHalfIdx = i-HalfSize;
3237 if (Mask[FstHalfIdx] < 0)
3239 if (!isUndefOrEqual(Mask[i], Mask[FstHalfIdx]+HalfSize))
3247 /// getShuffleVSHUFPSYImmediate - Return the appropriate immediate to shuffle
3248 /// the specified VECTOR_MASK mask with VSHUFPSY instruction.
3249 static unsigned getShuffleVSHUFPSYImmediate(SDNode *N) {
3250 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3251 EVT VT = SVOp->getValueType(0);
3252 int NumElems = VT.getVectorNumElements();
3254 assert(NumElems == 8 && VT.getSizeInBits() == 256 &&
3255 "Only supports v8i32 and v8f32 types");
3257 int HalfSize = NumElems/2;
3259 for (int i = 0; i != NumElems ; ++i) {
3260 if (SVOp->getMaskElt(i) < 0)
3262 // The mask of the first half must be equal to the second one.
3263 unsigned Shamt = (i%HalfSize)*2;
3264 unsigned Elt = SVOp->getMaskElt(i) % HalfSize;
3265 Mask |= Elt << Shamt;
3271 /// isVSHUFPDYMask - Return true if the specified VECTOR_SHUFFLE operand
3272 /// specifies a shuffle of elements that is suitable for input to 256-bit
3273 /// VSHUFPDY. This shuffle doesn't have the same restriction as the PS
3274 /// version and the mask of the second half isn't binded with the first
3276 static bool isVSHUFPDYMask(const SmallVectorImpl<int> &Mask, EVT VT,
3277 const X86Subtarget *Subtarget) {
3278 int NumElems = VT.getVectorNumElements();
3280 if (!Subtarget->hasAVX() || VT.getSizeInBits() != 256)
3286 // VSHUFPSY divides the resulting vector into 4 chunks.
3287 // The sources are also splitted into 4 chunks, and each destination
3288 // chunk must come from a different source chunk.
3290 // SRC1 => X3 X2 X1 X0
3291 // SRC2 => Y3 Y2 Y1 Y0
3293 // DST => Y2..Y3, X2..X3, Y1..Y0, X1..X0
3295 int QuarterSize = NumElems/4;
3296 int HalfSize = QuarterSize*2;
3297 for (int i = 0; i < QuarterSize; ++i)
3298 if (!isUndefOrInRange(Mask[i], 0, HalfSize))
3300 for (int i = QuarterSize; i < QuarterSize*2; ++i)
3301 if (!isUndefOrInRange(Mask[i], NumElems, NumElems+HalfSize))
3303 for (int i = QuarterSize*2; i < QuarterSize*3; ++i)
3304 if (!isUndefOrInRange(Mask[i], HalfSize, NumElems))
3306 for (int i = QuarterSize*3; i < NumElems; ++i)
3307 if (!isUndefOrInRange(Mask[i], NumElems+HalfSize, NumElems*2))
3313 /// getShuffleVSHUFPDYImmediate - Return the appropriate immediate to shuffle
3314 /// the specified VECTOR_MASK mask with VSHUFPDY instruction.
3315 static unsigned getShuffleVSHUFPDYImmediate(SDNode *N) {
3316 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3317 EVT VT = SVOp->getValueType(0);
3318 int NumElems = VT.getVectorNumElements();
3320 assert(NumElems == 4 && VT.getSizeInBits() == 256 &&
3321 "Only supports v4i64 and v4f64 types");
3323 int HalfSize = NumElems/2;
3325 for (int i = 0; i != NumElems ; ++i) {
3326 if (SVOp->getMaskElt(i) < 0)
3328 int Elt = SVOp->getMaskElt(i) % HalfSize;
3335 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
3336 /// specifies a shuffle of elements that is suitable for input to 128-bit
3337 /// SHUFPS and SHUFPD.
3338 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3339 int NumElems = VT.getVectorNumElements();
3341 if (VT.getSizeInBits() != 128)
3344 if (NumElems != 2 && NumElems != 4)
3347 int Half = NumElems / 2;
3348 for (int i = 0; i < Half; ++i)
3349 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3351 for (int i = Half; i < NumElems; ++i)
3352 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3358 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
3359 SmallVector<int, 8> M;
3361 return ::isSHUFPMask(M, N->getValueType(0));
3364 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
3365 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
3366 /// half elements to come from vector 1 (which would equal the dest.) and
3367 /// the upper half to come from vector 2.
3368 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3369 int NumElems = VT.getVectorNumElements();
3371 if (NumElems != 2 && NumElems != 4)
3374 int Half = NumElems / 2;
3375 for (int i = 0; i < Half; ++i)
3376 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3378 for (int i = Half; i < NumElems; ++i)
3379 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3384 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
3385 SmallVector<int, 8> M;
3387 return isCommutedSHUFPMask(M, N->getValueType(0));
3390 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3391 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3392 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
3393 EVT VT = N->getValueType(0);
3394 unsigned NumElems = VT.getVectorNumElements();
3396 if (VT.getSizeInBits() != 128)
3402 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3403 return isUndefOrEqual(N->getMaskElt(0), 6) &&
3404 isUndefOrEqual(N->getMaskElt(1), 7) &&
3405 isUndefOrEqual(N->getMaskElt(2), 2) &&
3406 isUndefOrEqual(N->getMaskElt(3), 3);
3409 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3410 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3412 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
3413 EVT VT = N->getValueType(0);
3414 unsigned NumElems = VT.getVectorNumElements();
3416 if (VT.getSizeInBits() != 128)
3422 return isUndefOrEqual(N->getMaskElt(0), 2) &&
3423 isUndefOrEqual(N->getMaskElt(1), 3) &&
3424 isUndefOrEqual(N->getMaskElt(2), 2) &&
3425 isUndefOrEqual(N->getMaskElt(3), 3);
3428 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3429 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3430 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
3431 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3433 if (NumElems != 2 && NumElems != 4)
3436 for (unsigned i = 0; i < NumElems/2; ++i)
3437 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
3440 for (unsigned i = NumElems/2; i < NumElems; ++i)
3441 if (!isUndefOrEqual(N->getMaskElt(i), i))
3447 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3448 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3449 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
3450 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3452 if ((NumElems != 2 && NumElems != 4)
3453 || N->getValueType(0).getSizeInBits() > 128)
3456 for (unsigned i = 0; i < NumElems/2; ++i)
3457 if (!isUndefOrEqual(N->getMaskElt(i), i))
3460 for (unsigned i = 0; i < NumElems/2; ++i)
3461 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
3467 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3468 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3469 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3470 bool V2IsSplat = false) {
3471 int NumElts = VT.getVectorNumElements();
3473 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3474 "Unsupported vector type for unpckh");
3476 if (VT.getSizeInBits() == 256 && NumElts != 4 && NumElts != 8)
3479 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3480 // independently on 128-bit lanes.
3481 unsigned NumLanes = VT.getSizeInBits()/128;
3482 unsigned NumLaneElts = NumElts/NumLanes;
3485 unsigned End = NumLaneElts;
3486 for (unsigned s = 0; s < NumLanes; ++s) {
3487 for (unsigned i = Start, j = s * NumLaneElts;
3491 int BitI1 = Mask[i+1];
3492 if (!isUndefOrEqual(BitI, j))
3495 if (!isUndefOrEqual(BitI1, NumElts))
3498 if (!isUndefOrEqual(BitI1, j + NumElts))
3502 // Process the next 128 bits.
3503 Start += NumLaneElts;
3510 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3511 SmallVector<int, 8> M;
3513 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
3516 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3517 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3518 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3519 bool V2IsSplat = false) {
3520 int NumElts = VT.getVectorNumElements();
3522 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3523 "Unsupported vector type for unpckh");
3525 if (VT.getSizeInBits() == 256 && NumElts != 4 && NumElts != 8)
3528 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3529 // independently on 128-bit lanes.
3530 unsigned NumLanes = VT.getSizeInBits()/128;
3531 unsigned NumLaneElts = NumElts/NumLanes;
3534 unsigned End = NumLaneElts;
3535 for (unsigned l = 0; l != NumLanes; ++l) {
3536 for (unsigned i = Start, j = (l*NumLaneElts)+NumLaneElts/2;
3537 i != End; i += 2, ++j) {
3539 int BitI1 = Mask[i+1];
3540 if (!isUndefOrEqual(BitI, j))
3543 if (isUndefOrEqual(BitI1, NumElts))
3546 if (!isUndefOrEqual(BitI1, j+NumElts))
3550 // Process the next 128 bits.
3551 Start += NumLaneElts;
3557 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3558 SmallVector<int, 8> M;
3560 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3563 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3564 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3566 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3567 int NumElems = VT.getVectorNumElements();
3568 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3571 // For 256-bit i64/f64, use MOVDDUPY instead, so reject the matching pattern
3572 // FIXME: Need a better way to get rid of this, there's no latency difference
3573 // between UNPCKLPD and MOVDDUP, the later should always be checked first and
3574 // the former later. We should also remove the "_undef" special mask.
3575 if (NumElems == 4 && VT.getSizeInBits() == 256)
3578 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3579 // independently on 128-bit lanes.
3580 unsigned NumLanes = VT.getSizeInBits() / 128;
3581 unsigned NumLaneElts = NumElems / NumLanes;
3583 for (unsigned s = 0; s < NumLanes; ++s) {
3584 for (unsigned i = s * NumLaneElts, j = s * NumLaneElts;
3585 i != NumLaneElts * (s + 1);
3588 int BitI1 = Mask[i+1];
3590 if (!isUndefOrEqual(BitI, j))
3592 if (!isUndefOrEqual(BitI1, j))
3600 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3601 SmallVector<int, 8> M;
3603 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3606 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3607 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3609 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3610 int NumElems = VT.getVectorNumElements();
3611 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3614 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3616 int BitI1 = Mask[i+1];
3617 if (!isUndefOrEqual(BitI, j))
3619 if (!isUndefOrEqual(BitI1, j))
3625 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3626 SmallVector<int, 8> M;
3628 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3631 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3632 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3633 /// MOVSD, and MOVD, i.e. setting the lowest element.
3634 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3635 if (VT.getVectorElementType().getSizeInBits() < 32)
3638 int NumElts = VT.getVectorNumElements();
3640 if (!isUndefOrEqual(Mask[0], NumElts))
3643 for (int i = 1; i < NumElts; ++i)
3644 if (!isUndefOrEqual(Mask[i], i))
3650 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3651 SmallVector<int, 8> M;
3653 return ::isMOVLMask(M, N->getValueType(0));
3656 /// isVPERM2F128Mask - Match 256-bit shuffles where the elements are considered
3657 /// as permutations between 128-bit chunks or halves. As an example: this
3659 /// vector_shuffle <4, 5, 6, 7, 12, 13, 14, 15>
3660 /// The first half comes from the second half of V1 and the second half from the
3661 /// the second half of V2.
3662 static bool isVPERM2F128Mask(const SmallVectorImpl<int> &Mask, EVT VT,
3663 const X86Subtarget *Subtarget) {
3664 if (!Subtarget->hasAVX() || VT.getSizeInBits() != 256)
3667 // The shuffle result is divided into half A and half B. In total the two
3668 // sources have 4 halves, namely: C, D, E, F. The final values of A and
3669 // B must come from C, D, E or F.
3670 int HalfSize = VT.getVectorNumElements()/2;
3671 bool MatchA = false, MatchB = false;
3673 // Check if A comes from one of C, D, E, F.
3674 for (int Half = 0; Half < 4; ++Half) {
3675 if (isSequentialOrUndefInRange(Mask, 0, HalfSize, Half*HalfSize)) {
3681 // Check if B comes from one of C, D, E, F.
3682 for (int Half = 0; Half < 4; ++Half) {
3683 if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, Half*HalfSize)) {
3689 return MatchA && MatchB;
3692 /// getShuffleVPERM2F128Immediate - Return the appropriate immediate to shuffle
3693 /// the specified VECTOR_MASK mask with VPERM2F128 instructions.
3694 static unsigned getShuffleVPERM2F128Immediate(SDNode *N) {
3695 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3696 EVT VT = SVOp->getValueType(0);
3698 int HalfSize = VT.getVectorNumElements()/2;
3700 int FstHalf = 0, SndHalf = 0;
3701 for (int i = 0; i < HalfSize; ++i) {
3702 if (SVOp->getMaskElt(i) > 0) {
3703 FstHalf = SVOp->getMaskElt(i)/HalfSize;
3707 for (int i = HalfSize; i < HalfSize*2; ++i) {
3708 if (SVOp->getMaskElt(i) > 0) {
3709 SndHalf = SVOp->getMaskElt(i)/HalfSize;
3714 return (FstHalf | (SndHalf << 4));
3717 /// isVPERMILPDMask - Return true if the specified VECTOR_SHUFFLE operand
3718 /// specifies a shuffle of elements that is suitable for input to VPERMILPD*.
3719 /// Note that VPERMIL mask matching is different depending whether theunderlying
3720 /// type is 32 or 64. In the VPERMILPS the high half of the mask should point
3721 /// to the same elements of the low, but to the higher half of the source.
3722 /// In VPERMILPD the two lanes could be shuffled independently of each other
3723 /// with the same restriction that lanes can't be crossed.
3724 static bool isVPERMILPDMask(const SmallVectorImpl<int> &Mask, EVT VT,
3725 const X86Subtarget *Subtarget) {
3726 int NumElts = VT.getVectorNumElements();
3727 int NumLanes = VT.getSizeInBits()/128;
3729 if (!Subtarget->hasAVX())
3732 // Match any permutation of 128-bit vector with 64-bit types
3733 if (NumLanes == 1 && NumElts != 2)
3736 // Only match 256-bit with 32 types
3737 if (VT.getSizeInBits() == 256 && NumElts != 4)
3740 // The mask on the high lane is independent of the low. Both can match
3741 // any element in inside its own lane, but can't cross.
3742 int LaneSize = NumElts/NumLanes;
3743 for (int l = 0; l < NumLanes; ++l)
3744 for (int i = l*LaneSize; i < LaneSize*(l+1); ++i) {
3745 int LaneStart = l*LaneSize;
3746 if (!isUndefOrInRange(Mask[i], LaneStart, LaneStart+LaneSize))
3753 /// isVPERMILPSMask - Return true if the specified VECTOR_SHUFFLE operand
3754 /// specifies a shuffle of elements that is suitable for input to VPERMILPS*.
3755 /// Note that VPERMIL mask matching is different depending whether theunderlying
3756 /// type is 32 or 64. In the VPERMILPS the high half of the mask should point
3757 /// to the same elements of the low, but to the higher half of the source.
3758 /// In VPERMILPD the two lanes could be shuffled independently of each other
3759 /// with the same restriction that lanes can't be crossed.
3760 static bool isVPERMILPSMask(const SmallVectorImpl<int> &Mask, EVT VT,
3761 const X86Subtarget *Subtarget) {
3762 unsigned NumElts = VT.getVectorNumElements();
3763 unsigned NumLanes = VT.getSizeInBits()/128;
3765 if (!Subtarget->hasAVX())
3768 // Match any permutation of 128-bit vector with 32-bit types
3769 if (NumLanes == 1 && NumElts != 4)
3772 // Only match 256-bit with 32 types
3773 if (VT.getSizeInBits() == 256 && NumElts != 8)
3776 // The mask on the high lane should be the same as the low. Actually,
3777 // they can differ if any of the corresponding index in a lane is undef
3778 // and the other stays in range.
3779 int LaneSize = NumElts/NumLanes;
3780 for (int i = 0; i < LaneSize; ++i) {
3781 int HighElt = i+LaneSize;
3782 bool HighValid = isUndefOrInRange(Mask[HighElt], LaneSize, NumElts);
3783 bool LowValid = isUndefOrInRange(Mask[i], 0, LaneSize);
3785 if (!HighValid || !LowValid)
3787 if (Mask[i] < 0 || Mask[HighElt] < 0)
3789 if (Mask[HighElt]-Mask[i] != LaneSize)
3796 /// getShuffleVPERMILPSImmediate - Return the appropriate immediate to shuffle
3797 /// the specified VECTOR_MASK mask with VPERMILPS* instructions.
3798 static unsigned getShuffleVPERMILPSImmediate(SDNode *N) {
3799 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3800 EVT VT = SVOp->getValueType(0);
3802 int NumElts = VT.getVectorNumElements();
3803 int NumLanes = VT.getSizeInBits()/128;
3804 int LaneSize = NumElts/NumLanes;
3806 // Although the mask is equal for both lanes do it twice to get the cases
3807 // where a mask will match because the same mask element is undef on the
3808 // first half but valid on the second. This would get pathological cases
3809 // such as: shuffle <u, 0, 1, 2, 4, 4, 5, 6>, which is completely valid.
3811 for (int l = 0; l < NumLanes; ++l) {
3812 for (int i = 0; i < LaneSize; ++i) {
3813 int MaskElt = SVOp->getMaskElt(i+(l*LaneSize));
3816 if (MaskElt >= LaneSize)
3817 MaskElt -= LaneSize;
3818 Mask |= MaskElt << (i*2);
3825 /// getShuffleVPERMILPDImmediate - Return the appropriate immediate to shuffle
3826 /// the specified VECTOR_MASK mask with VPERMILPD* instructions.
3827 static unsigned getShuffleVPERMILPDImmediate(SDNode *N) {
3828 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3829 EVT VT = SVOp->getValueType(0);
3831 int NumElts = VT.getVectorNumElements();
3832 int NumLanes = VT.getSizeInBits()/128;
3835 int LaneSize = NumElts/NumLanes;
3836 for (int l = 0; l < NumLanes; ++l)
3837 for (int i = l*LaneSize; i < LaneSize*(l+1); ++i) {
3838 int MaskElt = SVOp->getMaskElt(i);
3841 Mask |= (MaskElt-l*LaneSize) << i;
3847 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3848 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3849 /// element of vector 2 and the other elements to come from vector 1 in order.
3850 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3851 bool V2IsSplat = false, bool V2IsUndef = false) {
3852 int NumOps = VT.getVectorNumElements();
3853 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3856 if (!isUndefOrEqual(Mask[0], 0))
3859 for (int i = 1; i < NumOps; ++i)
3860 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3861 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3862 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3868 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3869 bool V2IsUndef = false) {
3870 SmallVector<int, 8> M;
3872 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3875 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3876 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3877 /// Masks to match: <1, 1, 3, 3> or <1, 1, 3, 3, 5, 5, 7, 7>
3878 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N,
3879 const X86Subtarget *Subtarget) {
3880 if (!Subtarget->hasSSE3() && !Subtarget->hasAVX())
3883 // The second vector must be undef
3884 if (N->getOperand(1).getOpcode() != ISD::UNDEF)
3887 EVT VT = N->getValueType(0);
3888 unsigned NumElems = VT.getVectorNumElements();
3890 if ((VT.getSizeInBits() == 128 && NumElems != 4) ||
3891 (VT.getSizeInBits() == 256 && NumElems != 8))
3894 // "i+1" is the value the indexed mask element must have
3895 for (unsigned i = 0; i < NumElems; i += 2)
3896 if (!isUndefOrEqual(N->getMaskElt(i), i+1) ||
3897 !isUndefOrEqual(N->getMaskElt(i+1), i+1))
3903 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3904 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3905 /// Masks to match: <0, 0, 2, 2> or <0, 0, 2, 2, 4, 4, 6, 6>
3906 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N,
3907 const X86Subtarget *Subtarget) {
3908 if (!Subtarget->hasSSE3() && !Subtarget->hasAVX())
3911 // The second vector must be undef
3912 if (N->getOperand(1).getOpcode() != ISD::UNDEF)
3915 EVT VT = N->getValueType(0);
3916 unsigned NumElems = VT.getVectorNumElements();
3918 if ((VT.getSizeInBits() == 128 && NumElems != 4) ||
3919 (VT.getSizeInBits() == 256 && NumElems != 8))
3922 // "i" is the value the indexed mask element must have
3923 for (unsigned i = 0; i < NumElems; i += 2)
3924 if (!isUndefOrEqual(N->getMaskElt(i), i) ||
3925 !isUndefOrEqual(N->getMaskElt(i+1), i))
3931 /// isMOVDDUPYMask - Return true if the specified VECTOR_SHUFFLE operand
3932 /// specifies a shuffle of elements that is suitable for input to 256-bit
3933 /// version of MOVDDUP.
3934 static bool isMOVDDUPYMask(ShuffleVectorSDNode *N,
3935 const X86Subtarget *Subtarget) {
3936 EVT VT = N->getValueType(0);
3937 int NumElts = VT.getVectorNumElements();
3938 bool V2IsUndef = N->getOperand(1).getOpcode() == ISD::UNDEF;
3940 if (!Subtarget->hasAVX() || VT.getSizeInBits() != 256 ||
3941 !V2IsUndef || NumElts != 4)
3944 for (int i = 0; i != NumElts/2; ++i)
3945 if (!isUndefOrEqual(N->getMaskElt(i), 0))
3947 for (int i = NumElts/2; i != NumElts; ++i)
3948 if (!isUndefOrEqual(N->getMaskElt(i), NumElts/2))
3953 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3954 /// specifies a shuffle of elements that is suitable for input to 128-bit
3955 /// version of MOVDDUP.
3956 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3957 EVT VT = N->getValueType(0);
3959 if (VT.getSizeInBits() != 128)
3962 int e = VT.getVectorNumElements() / 2;
3963 for (int i = 0; i < e; ++i)
3964 if (!isUndefOrEqual(N->getMaskElt(i), i))
3966 for (int i = 0; i < e; ++i)
3967 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3972 /// isVEXTRACTF128Index - Return true if the specified
3973 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
3974 /// suitable for input to VEXTRACTF128.
3975 bool X86::isVEXTRACTF128Index(SDNode *N) {
3976 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3979 // The index should be aligned on a 128-bit boundary.
3981 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3983 unsigned VL = N->getValueType(0).getVectorNumElements();
3984 unsigned VBits = N->getValueType(0).getSizeInBits();
3985 unsigned ElSize = VBits / VL;
3986 bool Result = (Index * ElSize) % 128 == 0;
3991 /// isVINSERTF128Index - Return true if the specified INSERT_SUBVECTOR
3992 /// operand specifies a subvector insert that is suitable for input to
3994 bool X86::isVINSERTF128Index(SDNode *N) {
3995 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3998 // The index should be aligned on a 128-bit boundary.
4000 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4002 unsigned VL = N->getValueType(0).getVectorNumElements();
4003 unsigned VBits = N->getValueType(0).getSizeInBits();
4004 unsigned ElSize = VBits / VL;
4005 bool Result = (Index * ElSize) % 128 == 0;
4010 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
4011 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
4012 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
4013 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
4014 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
4016 unsigned Shift = (NumOperands == 4) ? 2 : 1;
4018 for (int i = 0; i < NumOperands; ++i) {
4019 int Val = SVOp->getMaskElt(NumOperands-i-1);
4020 if (Val < 0) Val = 0;
4021 if (Val >= NumOperands) Val -= NumOperands;
4023 if (i != NumOperands - 1)
4029 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
4030 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
4031 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
4032 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
4034 // 8 nodes, but we only care about the last 4.
4035 for (unsigned i = 7; i >= 4; --i) {
4036 int Val = SVOp->getMaskElt(i);
4045 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
4046 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
4047 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
4048 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
4050 // 8 nodes, but we only care about the first 4.
4051 for (int i = 3; i >= 0; --i) {
4052 int Val = SVOp->getMaskElt(i);
4061 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
4062 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
4063 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
4064 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
4065 EVT VVT = N->getValueType(0);
4066 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
4070 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
4071 Val = SVOp->getMaskElt(i);
4075 assert(Val - i > 0 && "PALIGNR imm should be positive");
4076 return (Val - i) * EltSize;
4079 /// getExtractVEXTRACTF128Immediate - Return the appropriate immediate
4080 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
4082 unsigned X86::getExtractVEXTRACTF128Immediate(SDNode *N) {
4083 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4084 llvm_unreachable("Illegal extract subvector for VEXTRACTF128");
4087 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4089 EVT VecVT = N->getOperand(0).getValueType();
4090 EVT ElVT = VecVT.getVectorElementType();
4092 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
4093 return Index / NumElemsPerChunk;
4096 /// getInsertVINSERTF128Immediate - Return the appropriate immediate
4097 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
4099 unsigned X86::getInsertVINSERTF128Immediate(SDNode *N) {
4100 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4101 llvm_unreachable("Illegal insert subvector for VINSERTF128");
4104 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4106 EVT VecVT = N->getValueType(0);
4107 EVT ElVT = VecVT.getVectorElementType();
4109 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
4110 return Index / NumElemsPerChunk;
4113 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
4115 bool X86::isZeroNode(SDValue Elt) {
4116 return ((isa<ConstantSDNode>(Elt) &&
4117 cast<ConstantSDNode>(Elt)->isNullValue()) ||
4118 (isa<ConstantFPSDNode>(Elt) &&
4119 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
4122 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
4123 /// their permute mask.
4124 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
4125 SelectionDAG &DAG) {
4126 EVT VT = SVOp->getValueType(0);
4127 unsigned NumElems = VT.getVectorNumElements();
4128 SmallVector<int, 8> MaskVec;
4130 for (unsigned i = 0; i != NumElems; ++i) {
4131 int idx = SVOp->getMaskElt(i);
4133 MaskVec.push_back(idx);
4134 else if (idx < (int)NumElems)
4135 MaskVec.push_back(idx + NumElems);
4137 MaskVec.push_back(idx - NumElems);
4139 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
4140 SVOp->getOperand(0), &MaskVec[0]);
4143 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
4144 /// the two vector operands have swapped position.
4145 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
4146 unsigned NumElems = VT.getVectorNumElements();
4147 for (unsigned i = 0; i != NumElems; ++i) {
4151 else if (idx < (int)NumElems)
4152 Mask[i] = idx + NumElems;
4154 Mask[i] = idx - NumElems;
4158 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
4159 /// match movhlps. The lower half elements should come from upper half of
4160 /// V1 (and in order), and the upper half elements should come from the upper
4161 /// half of V2 (and in order).
4162 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
4163 EVT VT = Op->getValueType(0);
4164 if (VT.getSizeInBits() != 128)
4166 if (VT.getVectorNumElements() != 4)
4168 for (unsigned i = 0, e = 2; i != e; ++i)
4169 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
4171 for (unsigned i = 2; i != 4; ++i)
4172 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
4177 /// isScalarLoadToVector - Returns true if the node is a scalar load that
4178 /// is promoted to a vector. It also returns the LoadSDNode by reference if
4180 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
4181 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
4183 N = N->getOperand(0).getNode();
4184 if (!ISD::isNON_EXTLoad(N))
4187 *LD = cast<LoadSDNode>(N);
4191 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
4192 /// match movlp{s|d}. The lower half elements should come from lower half of
4193 /// V1 (and in order), and the upper half elements should come from the upper
4194 /// half of V2 (and in order). And since V1 will become the source of the
4195 /// MOVLP, it must be either a vector load or a scalar load to vector.
4196 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
4197 ShuffleVectorSDNode *Op) {
4198 EVT VT = Op->getValueType(0);
4199 if (VT.getSizeInBits() != 128)
4202 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
4204 // Is V2 is a vector load, don't do this transformation. We will try to use
4205 // load folding shufps op.
4206 if (ISD::isNON_EXTLoad(V2))
4209 unsigned NumElems = VT.getVectorNumElements();
4211 if (NumElems != 2 && NumElems != 4)
4213 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
4214 if (!isUndefOrEqual(Op->getMaskElt(i), i))
4216 for (unsigned i = NumElems/2; i != NumElems; ++i)
4217 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
4222 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
4224 static bool isSplatVector(SDNode *N) {
4225 if (N->getOpcode() != ISD::BUILD_VECTOR)
4228 SDValue SplatValue = N->getOperand(0);
4229 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
4230 if (N->getOperand(i) != SplatValue)
4235 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
4236 /// to an zero vector.
4237 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
4238 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
4239 SDValue V1 = N->getOperand(0);
4240 SDValue V2 = N->getOperand(1);
4241 unsigned NumElems = N->getValueType(0).getVectorNumElements();
4242 for (unsigned i = 0; i != NumElems; ++i) {
4243 int Idx = N->getMaskElt(i);
4244 if (Idx >= (int)NumElems) {
4245 unsigned Opc = V2.getOpcode();
4246 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
4248 if (Opc != ISD::BUILD_VECTOR ||
4249 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
4251 } else if (Idx >= 0) {
4252 unsigned Opc = V1.getOpcode();
4253 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
4255 if (Opc != ISD::BUILD_VECTOR ||
4256 !X86::isZeroNode(V1.getOperand(Idx)))
4263 /// getZeroVector - Returns a vector of specified type with all zero elements.
4265 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
4267 assert(VT.isVector() && "Expected a vector type");
4269 // Always build SSE zero vectors as <4 x i32> bitcasted
4270 // to their dest type. This ensures they get CSE'd.
4272 if (VT.getSizeInBits() == 128) { // SSE
4273 if (HasSSE2) { // SSE2
4274 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
4275 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4277 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
4278 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4280 } else if (VT.getSizeInBits() == 256) { // AVX
4281 // 256-bit logic and arithmetic instructions in AVX are
4282 // all floating-point, no support for integer ops. Default
4283 // to emitting fp zeroed vectors then.
4284 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
4285 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4286 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
4288 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4291 /// getOnesVector - Returns a vector of specified type with all bits set.
4292 /// Always build ones vectors as <4 x i32>. For 256-bit types, use two
4293 /// <4 x i32> inserted in a <8 x i32> appropriately. Then bitcast to their
4294 /// original type, ensuring they get CSE'd.
4295 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
4296 assert(VT.isVector() && "Expected a vector type");
4297 assert((VT.is128BitVector() || VT.is256BitVector())
4298 && "Expected a 128-bit or 256-bit vector type");
4300 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
4301 SDValue Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
4302 Cst, Cst, Cst, Cst);
4304 if (VT.is256BitVector()) {
4305 SDValue InsV = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, MVT::v8i32),
4306 Vec, DAG.getConstant(0, MVT::i32), DAG, dl);
4307 Vec = Insert128BitVector(InsV, Vec,
4308 DAG.getConstant(4 /* NumElems/2 */, MVT::i32), DAG, dl);
4311 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4314 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
4315 /// that point to V2 points to its first element.
4316 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4317 EVT VT = SVOp->getValueType(0);
4318 unsigned NumElems = VT.getVectorNumElements();
4320 bool Changed = false;
4321 SmallVector<int, 8> MaskVec;
4322 SVOp->getMask(MaskVec);
4324 for (unsigned i = 0; i != NumElems; ++i) {
4325 if (MaskVec[i] > (int)NumElems) {
4326 MaskVec[i] = NumElems;
4331 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
4332 SVOp->getOperand(1), &MaskVec[0]);
4333 return SDValue(SVOp, 0);
4336 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
4337 /// operation of specified width.
4338 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
4340 unsigned NumElems = VT.getVectorNumElements();
4341 SmallVector<int, 8> Mask;
4342 Mask.push_back(NumElems);
4343 for (unsigned i = 1; i != NumElems; ++i)
4345 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4348 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
4349 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
4351 unsigned NumElems = VT.getVectorNumElements();
4352 SmallVector<int, 8> Mask;
4353 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4355 Mask.push_back(i + NumElems);
4357 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4360 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
4361 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
4363 unsigned NumElems = VT.getVectorNumElements();
4364 unsigned Half = NumElems/2;
4365 SmallVector<int, 8> Mask;
4366 for (unsigned i = 0; i != Half; ++i) {
4367 Mask.push_back(i + Half);
4368 Mask.push_back(i + NumElems + Half);
4370 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4373 // PromoteSplati8i16 - All i16 and i8 vector types can't be used directly by
4374 // a generic shuffle instruction because the target has no such instructions.
4375 // Generate shuffles which repeat i16 and i8 several times until they can be
4376 // represented by v4f32 and then be manipulated by target suported shuffles.
4377 static SDValue PromoteSplati8i16(SDValue V, SelectionDAG &DAG, int &EltNo) {
4378 EVT VT = V.getValueType();
4379 int NumElems = VT.getVectorNumElements();
4380 DebugLoc dl = V.getDebugLoc();
4382 while (NumElems > 4) {
4383 if (EltNo < NumElems/2) {
4384 V = getUnpackl(DAG, dl, VT, V, V);
4386 V = getUnpackh(DAG, dl, VT, V, V);
4387 EltNo -= NumElems/2;
4394 /// getLegalSplat - Generate a legal splat with supported x86 shuffles
4395 static SDValue getLegalSplat(SelectionDAG &DAG, SDValue V, int EltNo) {
4396 EVT VT = V.getValueType();
4397 DebugLoc dl = V.getDebugLoc();
4398 assert((VT.getSizeInBits() == 128 || VT.getSizeInBits() == 256)
4399 && "Vector size not supported");
4401 if (VT.getSizeInBits() == 128) {
4402 V = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V);
4403 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
4404 V = DAG.getVectorShuffle(MVT::v4f32, dl, V, DAG.getUNDEF(MVT::v4f32),
4407 // To use VPERMILPS to splat scalars, the second half of indicies must
4408 // refer to the higher part, which is a duplication of the lower one,
4409 // because VPERMILPS can only handle in-lane permutations.
4410 int SplatMask[8] = { EltNo, EltNo, EltNo, EltNo,
4411 EltNo+4, EltNo+4, EltNo+4, EltNo+4 };
4413 V = DAG.getNode(ISD::BITCAST, dl, MVT::v8f32, V);
4414 V = DAG.getVectorShuffle(MVT::v8f32, dl, V, DAG.getUNDEF(MVT::v8f32),
4418 return DAG.getNode(ISD::BITCAST, dl, VT, V);
4421 /// PromoteSplat - Splat is promoted to target supported vector shuffles.
4422 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
4423 EVT SrcVT = SV->getValueType(0);
4424 SDValue V1 = SV->getOperand(0);
4425 DebugLoc dl = SV->getDebugLoc();
4427 int EltNo = SV->getSplatIndex();
4428 int NumElems = SrcVT.getVectorNumElements();
4429 unsigned Size = SrcVT.getSizeInBits();
4431 assert(((Size == 128 && NumElems > 4) || Size == 256) &&
4432 "Unknown how to promote splat for type");
4434 // Extract the 128-bit part containing the splat element and update
4435 // the splat element index when it refers to the higher register.
4437 unsigned Idx = (EltNo > NumElems/2) ? NumElems/2 : 0;
4438 V1 = Extract128BitVector(V1, DAG.getConstant(Idx, MVT::i32), DAG, dl);
4440 EltNo -= NumElems/2;
4443 // All i16 and i8 vector types can't be used directly by a generic shuffle
4444 // instruction because the target has no such instruction. Generate shuffles
4445 // which repeat i16 and i8 several times until they fit in i32, and then can
4446 // be manipulated by target suported shuffles.
4447 EVT EltVT = SrcVT.getVectorElementType();
4448 if (EltVT == MVT::i8 || EltVT == MVT::i16)
4449 V1 = PromoteSplati8i16(V1, DAG, EltNo);
4451 // Recreate the 256-bit vector and place the same 128-bit vector
4452 // into the low and high part. This is necessary because we want
4453 // to use VPERM* to shuffle the vectors
4455 SDValue InsV = Insert128BitVector(DAG.getUNDEF(SrcVT), V1,
4456 DAG.getConstant(0, MVT::i32), DAG, dl);
4457 V1 = Insert128BitVector(InsV, V1,
4458 DAG.getConstant(NumElems/2, MVT::i32), DAG, dl);
4461 return getLegalSplat(DAG, V1, EltNo);
4464 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
4465 /// vector of zero or undef vector. This produces a shuffle where the low
4466 /// element of V2 is swizzled into the zero/undef vector, landing at element
4467 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4468 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4469 bool isZero, bool HasSSE2,
4470 SelectionDAG &DAG) {
4471 EVT VT = V2.getValueType();
4473 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
4474 unsigned NumElems = VT.getVectorNumElements();
4475 SmallVector<int, 16> MaskVec;
4476 for (unsigned i = 0; i != NumElems; ++i)
4477 // If this is the insertion idx, put the low elt of V2 here.
4478 MaskVec.push_back(i == Idx ? NumElems : i);
4479 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
4482 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4483 /// element of the result of the vector shuffle.
4484 static SDValue getShuffleScalarElt(SDNode *N, int Index, SelectionDAG &DAG,
4487 return SDValue(); // Limit search depth.
4489 SDValue V = SDValue(N, 0);
4490 EVT VT = V.getValueType();
4491 unsigned Opcode = V.getOpcode();
4493 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4494 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4495 Index = SV->getMaskElt(Index);
4498 return DAG.getUNDEF(VT.getVectorElementType());
4500 int NumElems = VT.getVectorNumElements();
4501 SDValue NewV = (Index < NumElems) ? SV->getOperand(0) : SV->getOperand(1);
4502 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG, Depth+1);
4505 // Recurse into target specific vector shuffles to find scalars.
4506 if (isTargetShuffle(Opcode)) {
4507 int NumElems = VT.getVectorNumElements();
4508 SmallVector<unsigned, 16> ShuffleMask;
4512 case X86ISD::SHUFPS:
4513 case X86ISD::SHUFPD:
4514 ImmN = N->getOperand(N->getNumOperands()-1);
4515 DecodeSHUFPSMask(NumElems,
4516 cast<ConstantSDNode>(ImmN)->getZExtValue(),
4519 case X86ISD::PUNPCKHBW:
4520 case X86ISD::PUNPCKHWD:
4521 case X86ISD::PUNPCKHDQ:
4522 case X86ISD::PUNPCKHQDQ:
4523 DecodePUNPCKHMask(NumElems, ShuffleMask);
4525 case X86ISD::UNPCKHPS:
4526 case X86ISD::UNPCKHPD:
4527 case X86ISD::VUNPCKHPSY:
4528 case X86ISD::VUNPCKHPDY:
4529 DecodeUNPCKHPMask(NumElems, ShuffleMask);
4531 case X86ISD::PUNPCKLBW:
4532 case X86ISD::PUNPCKLWD:
4533 case X86ISD::PUNPCKLDQ:
4534 case X86ISD::PUNPCKLQDQ:
4535 DecodePUNPCKLMask(VT, ShuffleMask);
4537 case X86ISD::UNPCKLPS:
4538 case X86ISD::UNPCKLPD:
4539 case X86ISD::VUNPCKLPSY:
4540 case X86ISD::VUNPCKLPDY:
4541 DecodeUNPCKLPMask(VT, ShuffleMask);
4543 case X86ISD::MOVHLPS:
4544 DecodeMOVHLPSMask(NumElems, ShuffleMask);
4546 case X86ISD::MOVLHPS:
4547 DecodeMOVLHPSMask(NumElems, ShuffleMask);
4549 case X86ISD::PSHUFD:
4550 ImmN = N->getOperand(N->getNumOperands()-1);
4551 DecodePSHUFMask(NumElems,
4552 cast<ConstantSDNode>(ImmN)->getZExtValue(),
4555 case X86ISD::PSHUFHW:
4556 ImmN = N->getOperand(N->getNumOperands()-1);
4557 DecodePSHUFHWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
4560 case X86ISD::PSHUFLW:
4561 ImmN = N->getOperand(N->getNumOperands()-1);
4562 DecodePSHUFLWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
4566 case X86ISD::MOVSD: {
4567 // The index 0 always comes from the first element of the second source,
4568 // this is why MOVSS and MOVSD are used in the first place. The other
4569 // elements come from the other positions of the first source vector.
4570 unsigned OpNum = (Index == 0) ? 1 : 0;
4571 return getShuffleScalarElt(V.getOperand(OpNum).getNode(), Index, DAG,
4574 case X86ISD::VPERMILPS:
4575 ImmN = N->getOperand(N->getNumOperands()-1);
4576 DecodeVPERMILPSMask(4, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4579 case X86ISD::VPERMILPSY:
4580 ImmN = N->getOperand(N->getNumOperands()-1);
4581 DecodeVPERMILPSMask(8, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4584 case X86ISD::VPERMILPD:
4585 ImmN = N->getOperand(N->getNumOperands()-1);
4586 DecodeVPERMILPDMask(2, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4589 case X86ISD::VPERMILPDY:
4590 ImmN = N->getOperand(N->getNumOperands()-1);
4591 DecodeVPERMILPDMask(4, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4594 case X86ISD::VPERM2F128:
4595 ImmN = N->getOperand(N->getNumOperands()-1);
4596 DecodeVPERM2F128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4600 assert("not implemented for target shuffle node");
4604 Index = ShuffleMask[Index];
4606 return DAG.getUNDEF(VT.getVectorElementType());
4608 SDValue NewV = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
4609 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG,
4613 // Actual nodes that may contain scalar elements
4614 if (Opcode == ISD::BITCAST) {
4615 V = V.getOperand(0);
4616 EVT SrcVT = V.getValueType();
4617 unsigned NumElems = VT.getVectorNumElements();
4619 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4623 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4624 return (Index == 0) ? V.getOperand(0)
4625 : DAG.getUNDEF(VT.getVectorElementType());
4627 if (V.getOpcode() == ISD::BUILD_VECTOR)
4628 return V.getOperand(Index);
4633 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
4634 /// shuffle operation which come from a consecutively from a zero. The
4635 /// search can start in two different directions, from left or right.
4637 unsigned getNumOfConsecutiveZeros(SDNode *N, int NumElems,
4638 bool ZerosFromLeft, SelectionDAG &DAG) {
4641 while (i < NumElems) {
4642 unsigned Index = ZerosFromLeft ? i : NumElems-i-1;
4643 SDValue Elt = getShuffleScalarElt(N, Index, DAG, 0);
4644 if (!(Elt.getNode() &&
4645 (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt))))
4653 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies from MaskI to
4654 /// MaskE correspond consecutively to elements from one of the vector operands,
4655 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
4657 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp, int MaskI, int MaskE,
4658 int OpIdx, int NumElems, unsigned &OpNum) {
4659 bool SeenV1 = false;
4660 bool SeenV2 = false;
4662 for (int i = MaskI; i <= MaskE; ++i, ++OpIdx) {
4663 int Idx = SVOp->getMaskElt(i);
4664 // Ignore undef indicies
4673 // Only accept consecutive elements from the same vector
4674 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
4678 OpNum = SeenV1 ? 0 : 1;
4682 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
4683 /// logical left shift of a vector.
4684 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4685 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4686 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4687 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4688 false /* check zeros from right */, DAG);
4694 // Considering the elements in the mask that are not consecutive zeros,
4695 // check if they consecutively come from only one of the source vectors.
4697 // V1 = {X, A, B, C} 0
4699 // vector_shuffle V1, V2 <1, 2, 3, X>
4701 if (!isShuffleMaskConsecutive(SVOp,
4702 0, // Mask Start Index
4703 NumElems-NumZeros-1, // Mask End Index
4704 NumZeros, // Where to start looking in the src vector
4705 NumElems, // Number of elements in vector
4706 OpSrc)) // Which source operand ?
4711 ShVal = SVOp->getOperand(OpSrc);
4715 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
4716 /// logical left shift of a vector.
4717 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4718 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4719 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4720 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4721 true /* check zeros from left */, DAG);
4727 // Considering the elements in the mask that are not consecutive zeros,
4728 // check if they consecutively come from only one of the source vectors.
4730 // 0 { A, B, X, X } = V2
4732 // vector_shuffle V1, V2 <X, X, 4, 5>
4734 if (!isShuffleMaskConsecutive(SVOp,
4735 NumZeros, // Mask Start Index
4736 NumElems-1, // Mask End Index
4737 0, // Where to start looking in the src vector
4738 NumElems, // Number of elements in vector
4739 OpSrc)) // Which source operand ?
4744 ShVal = SVOp->getOperand(OpSrc);
4748 /// isVectorShift - Returns true if the shuffle can be implemented as a
4749 /// logical left or right shift of a vector.
4750 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4751 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4752 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
4753 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
4759 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
4761 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4762 unsigned NumNonZero, unsigned NumZero,
4764 const TargetLowering &TLI) {
4768 DebugLoc dl = Op.getDebugLoc();
4771 for (unsigned i = 0; i < 16; ++i) {
4772 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4773 if (ThisIsNonZero && First) {
4775 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4777 V = DAG.getUNDEF(MVT::v8i16);
4782 SDValue ThisElt(0, 0), LastElt(0, 0);
4783 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4784 if (LastIsNonZero) {
4785 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4786 MVT::i16, Op.getOperand(i-1));
4788 if (ThisIsNonZero) {
4789 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4790 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4791 ThisElt, DAG.getConstant(8, MVT::i8));
4793 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4797 if (ThisElt.getNode())
4798 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4799 DAG.getIntPtrConstant(i/2));
4803 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
4806 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
4808 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4809 unsigned NumNonZero, unsigned NumZero,
4811 const TargetLowering &TLI) {
4815 DebugLoc dl = Op.getDebugLoc();
4818 for (unsigned i = 0; i < 8; ++i) {
4819 bool isNonZero = (NonZeros & (1 << i)) != 0;
4823 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4825 V = DAG.getUNDEF(MVT::v8i16);
4828 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4829 MVT::v8i16, V, Op.getOperand(i),
4830 DAG.getIntPtrConstant(i));
4837 /// getVShift - Return a vector logical shift node.
4839 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4840 unsigned NumBits, SelectionDAG &DAG,
4841 const TargetLowering &TLI, DebugLoc dl) {
4842 EVT ShVT = MVT::v2i64;
4843 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
4844 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
4845 return DAG.getNode(ISD::BITCAST, dl, VT,
4846 DAG.getNode(Opc, dl, ShVT, SrcOp,
4847 DAG.getConstant(NumBits,
4848 TLI.getShiftAmountTy(SrcOp.getValueType()))));
4852 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
4853 SelectionDAG &DAG) const {
4855 // Check if the scalar load can be widened into a vector load. And if
4856 // the address is "base + cst" see if the cst can be "absorbed" into
4857 // the shuffle mask.
4858 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4859 SDValue Ptr = LD->getBasePtr();
4860 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4862 EVT PVT = LD->getValueType(0);
4863 if (PVT != MVT::i32 && PVT != MVT::f32)
4868 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4869 FI = FINode->getIndex();
4871 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4872 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4873 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4874 Offset = Ptr.getConstantOperandVal(1);
4875 Ptr = Ptr.getOperand(0);
4880 // FIXME: 256-bit vector instructions don't require a strict alignment,
4881 // improve this code to support it better.
4882 unsigned RequiredAlign = VT.getSizeInBits()/8;
4883 SDValue Chain = LD->getChain();
4884 // Make sure the stack object alignment is at least 16 or 32.
4885 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4886 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
4887 if (MFI->isFixedObjectIndex(FI)) {
4888 // Can't change the alignment. FIXME: It's possible to compute
4889 // the exact stack offset and reference FI + adjust offset instead.
4890 // If someone *really* cares about this. That's the way to implement it.
4893 MFI->setObjectAlignment(FI, RequiredAlign);
4897 // (Offset % 16 or 32) must be multiple of 4. Then address is then
4898 // Ptr + (Offset & ~15).
4901 if ((Offset % RequiredAlign) & 3)
4903 int64_t StartOffset = Offset & ~(RequiredAlign-1);
4905 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
4906 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4908 int EltNo = (Offset - StartOffset) >> 2;
4909 int NumElems = VT.getVectorNumElements();
4911 EVT CanonVT = VT.getSizeInBits() == 128 ? MVT::v4i32 : MVT::v8i32;
4912 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
4913 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
4914 LD->getPointerInfo().getWithOffset(StartOffset),
4917 // Canonicalize it to a v4i32 or v8i32 shuffle.
4918 SmallVector<int, 8> Mask;
4919 for (int i = 0; i < NumElems; ++i)
4920 Mask.push_back(EltNo);
4922 V1 = DAG.getNode(ISD::BITCAST, dl, CanonVT, V1);
4923 return DAG.getNode(ISD::BITCAST, dl, NVT,
4924 DAG.getVectorShuffle(CanonVT, dl, V1,
4925 DAG.getUNDEF(CanonVT),&Mask[0]));
4931 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
4932 /// vector of type 'VT', see if the elements can be replaced by a single large
4933 /// load which has the same value as a build_vector whose operands are 'elts'.
4935 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4937 /// FIXME: we'd also like to handle the case where the last elements are zero
4938 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4939 /// There's even a handy isZeroNode for that purpose.
4940 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
4941 DebugLoc &DL, SelectionDAG &DAG) {
4942 EVT EltVT = VT.getVectorElementType();
4943 unsigned NumElems = Elts.size();
4945 LoadSDNode *LDBase = NULL;
4946 unsigned LastLoadedElt = -1U;
4948 // For each element in the initializer, see if we've found a load or an undef.
4949 // If we don't find an initial load element, or later load elements are
4950 // non-consecutive, bail out.
4951 for (unsigned i = 0; i < NumElems; ++i) {
4952 SDValue Elt = Elts[i];
4954 if (!Elt.getNode() ||
4955 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4958 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4960 LDBase = cast<LoadSDNode>(Elt.getNode());
4964 if (Elt.getOpcode() == ISD::UNDEF)
4967 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4968 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
4973 // If we have found an entire vector of loads and undefs, then return a large
4974 // load of the entire vector width starting at the base pointer. If we found
4975 // consecutive loads for the low half, generate a vzext_load node.
4976 if (LastLoadedElt == NumElems - 1) {
4977 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
4978 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4979 LDBase->getPointerInfo(),
4980 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
4981 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4982 LDBase->getPointerInfo(),
4983 LDBase->isVolatile(), LDBase->isNonTemporal(),
4984 LDBase->getAlignment());
4985 } else if (NumElems == 4 && LastLoadedElt == 1 &&
4986 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
4987 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4988 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4989 SDValue ResNode = DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys,
4991 LDBase->getMemOperand());
4992 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
4998 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
4999 DebugLoc dl = Op.getDebugLoc();
5001 EVT VT = Op.getValueType();
5002 EVT ExtVT = VT.getVectorElementType();
5003 unsigned NumElems = Op.getNumOperands();
5005 // Vectors containing all zeros can be matched by pxor and xorps later
5006 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5007 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
5008 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
5009 if (Op.getValueType() == MVT::v4i32 ||
5010 Op.getValueType() == MVT::v8i32)
5013 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
5016 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
5017 // vectors or broken into v4i32 operations on 256-bit vectors.
5018 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5019 if (Op.getValueType() == MVT::v4i32)
5022 return getOnesVector(Op.getValueType(), DAG, dl);
5025 unsigned EVTBits = ExtVT.getSizeInBits();
5027 unsigned NumZero = 0;
5028 unsigned NumNonZero = 0;
5029 unsigned NonZeros = 0;
5030 bool IsAllConstants = true;
5031 SmallSet<SDValue, 8> Values;
5032 for (unsigned i = 0; i < NumElems; ++i) {
5033 SDValue Elt = Op.getOperand(i);
5034 if (Elt.getOpcode() == ISD::UNDEF)
5037 if (Elt.getOpcode() != ISD::Constant &&
5038 Elt.getOpcode() != ISD::ConstantFP)
5039 IsAllConstants = false;
5040 if (X86::isZeroNode(Elt))
5043 NonZeros |= (1 << i);
5048 // All undef vector. Return an UNDEF. All zero vectors were handled above.
5049 if (NumNonZero == 0)
5050 return DAG.getUNDEF(VT);
5052 // Special case for single non-zero, non-undef, element.
5053 if (NumNonZero == 1) {
5054 unsigned Idx = CountTrailingZeros_32(NonZeros);
5055 SDValue Item = Op.getOperand(Idx);
5057 // If this is an insertion of an i64 value on x86-32, and if the top bits of
5058 // the value are obviously zero, truncate the value to i32 and do the
5059 // insertion that way. Only do this if the value is non-constant or if the
5060 // value is a constant being inserted into element 0. It is cheaper to do
5061 // a constant pool load than it is to do a movd + shuffle.
5062 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
5063 (!IsAllConstants || Idx == 0)) {
5064 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
5066 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
5067 EVT VecVT = MVT::v4i32;
5068 unsigned VecElts = 4;
5070 // Truncate the value (which may itself be a constant) to i32, and
5071 // convert it to a vector with movd (S2V+shuffle to zero extend).
5072 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
5073 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
5074 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
5075 Subtarget->hasSSE2(), DAG);
5077 // Now we have our 32-bit value zero extended in the low element of
5078 // a vector. If Idx != 0, swizzle it into place.
5080 SmallVector<int, 4> Mask;
5081 Mask.push_back(Idx);
5082 for (unsigned i = 1; i != VecElts; ++i)
5084 Item = DAG.getVectorShuffle(VecVT, dl, Item,
5085 DAG.getUNDEF(Item.getValueType()),
5088 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Item);
5092 // If we have a constant or non-constant insertion into the low element of
5093 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
5094 // the rest of the elements. This will be matched as movd/movq/movss/movsd
5095 // depending on what the source datatype is.
5098 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5099 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
5100 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
5101 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5102 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
5103 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
5105 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
5106 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
5107 assert(VT.getSizeInBits() == 128 && "Expected an SSE value type!");
5108 EVT MiddleVT = MVT::v4i32;
5109 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
5110 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
5111 Subtarget->hasSSE2(), DAG);
5112 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
5116 // Is it a vector logical left shift?
5117 if (NumElems == 2 && Idx == 1 &&
5118 X86::isZeroNode(Op.getOperand(0)) &&
5119 !X86::isZeroNode(Op.getOperand(1))) {
5120 unsigned NumBits = VT.getSizeInBits();
5121 return getVShift(true, VT,
5122 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5123 VT, Op.getOperand(1)),
5124 NumBits/2, DAG, *this, dl);
5127 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
5130 // Otherwise, if this is a vector with i32 or f32 elements, and the element
5131 // is a non-constant being inserted into an element other than the low one,
5132 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
5133 // movd/movss) to move this into the low element, then shuffle it into
5135 if (EVTBits == 32) {
5136 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5138 // Turn it into a shuffle of zero and zero-extended scalar to vector.
5139 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
5140 Subtarget->hasSSE2(), DAG);
5141 SmallVector<int, 8> MaskVec;
5142 for (unsigned i = 0; i < NumElems; i++)
5143 MaskVec.push_back(i == Idx ? 0 : 1);
5144 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
5148 // Splat is obviously ok. Let legalizer expand it to a shuffle.
5149 if (Values.size() == 1) {
5150 if (EVTBits == 32) {
5151 // Instead of a shuffle like this:
5152 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
5153 // Check if it's possible to issue this instead.
5154 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
5155 unsigned Idx = CountTrailingZeros_32(NonZeros);
5156 SDValue Item = Op.getOperand(Idx);
5157 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
5158 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
5163 // A vector full of immediates; various special cases are already
5164 // handled, so this is best done with a single constant-pool load.
5168 // For AVX-length vectors, build the individual 128-bit pieces and use
5169 // shuffles to put them in place.
5170 if (VT.getSizeInBits() == 256 && !ISD::isBuildVectorAllZeros(Op.getNode())) {
5171 SmallVector<SDValue, 32> V;
5172 for (unsigned i = 0; i < NumElems; ++i)
5173 V.push_back(Op.getOperand(i));
5175 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
5177 // Build both the lower and upper subvector.
5178 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[0], NumElems/2);
5179 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[NumElems / 2],
5182 // Recreate the wider vector with the lower and upper part.
5183 SDValue Vec = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT), Lower,
5184 DAG.getConstant(0, MVT::i32), DAG, dl);
5185 return Insert128BitVector(Vec, Upper, DAG.getConstant(NumElems/2, MVT::i32),
5189 // Let legalizer expand 2-wide build_vectors.
5190 if (EVTBits == 64) {
5191 if (NumNonZero == 1) {
5192 // One half is zero or undef.
5193 unsigned Idx = CountTrailingZeros_32(NonZeros);
5194 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
5195 Op.getOperand(Idx));
5196 return getShuffleVectorZeroOrUndef(V2, Idx, true,
5197 Subtarget->hasSSE2(), DAG);
5202 // If element VT is < 32 bits, convert it to inserts into a zero vector.
5203 if (EVTBits == 8 && NumElems == 16) {
5204 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
5206 if (V.getNode()) return V;
5209 if (EVTBits == 16 && NumElems == 8) {
5210 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
5212 if (V.getNode()) return V;
5215 // If element VT is == 32 bits, turn it into a number of shuffles.
5216 SmallVector<SDValue, 8> V;
5218 if (NumElems == 4 && NumZero > 0) {
5219 for (unsigned i = 0; i < 4; ++i) {
5220 bool isZero = !(NonZeros & (1 << i));
5222 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
5224 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5227 for (unsigned i = 0; i < 2; ++i) {
5228 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
5231 V[i] = V[i*2]; // Must be a zero vector.
5234 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
5237 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
5240 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
5245 SmallVector<int, 8> MaskVec;
5246 bool Reverse = (NonZeros & 0x3) == 2;
5247 for (unsigned i = 0; i < 2; ++i)
5248 MaskVec.push_back(Reverse ? 1-i : i);
5249 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
5250 for (unsigned i = 0; i < 2; ++i)
5251 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
5252 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
5255 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
5256 // Check for a build vector of consecutive loads.
5257 for (unsigned i = 0; i < NumElems; ++i)
5258 V[i] = Op.getOperand(i);
5260 // Check for elements which are consecutive loads.
5261 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
5265 // For SSE 4.1, use insertps to put the high elements into the low element.
5266 if (getSubtarget()->hasSSE41()) {
5268 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
5269 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
5271 Result = DAG.getUNDEF(VT);
5273 for (unsigned i = 1; i < NumElems; ++i) {
5274 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
5275 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
5276 Op.getOperand(i), DAG.getIntPtrConstant(i));
5281 // Otherwise, expand into a number of unpckl*, start by extending each of
5282 // our (non-undef) elements to the full vector width with the element in the
5283 // bottom slot of the vector (which generates no code for SSE).
5284 for (unsigned i = 0; i < NumElems; ++i) {
5285 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
5286 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5288 V[i] = DAG.getUNDEF(VT);
5291 // Next, we iteratively mix elements, e.g. for v4f32:
5292 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
5293 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
5294 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
5295 unsigned EltStride = NumElems >> 1;
5296 while (EltStride != 0) {
5297 for (unsigned i = 0; i < EltStride; ++i) {
5298 // If V[i+EltStride] is undef and this is the first round of mixing,
5299 // then it is safe to just drop this shuffle: V[i] is already in the
5300 // right place, the one element (since it's the first round) being
5301 // inserted as undef can be dropped. This isn't safe for successive
5302 // rounds because they will permute elements within both vectors.
5303 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
5304 EltStride == NumElems/2)
5307 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
5316 // LowerMMXCONCAT_VECTORS - We support concatenate two MMX registers and place
5317 // them in a MMX register. This is better than doing a stack convert.
5318 static SDValue LowerMMXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5319 DebugLoc dl = Op.getDebugLoc();
5320 EVT ResVT = Op.getValueType();
5322 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
5323 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
5325 SDValue InVec = DAG.getNode(ISD::BITCAST,dl, MVT::v1i64, Op.getOperand(0));
5326 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
5327 InVec = Op.getOperand(1);
5328 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5329 unsigned NumElts = ResVT.getVectorNumElements();
5330 VecOp = DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
5331 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
5332 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
5334 InVec = DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, InVec);
5335 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
5336 Mask[0] = 0; Mask[1] = 2;
5337 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
5339 return DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
5342 // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
5343 // to create 256-bit vectors from two other 128-bit ones.
5344 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5345 DebugLoc dl = Op.getDebugLoc();
5346 EVT ResVT = Op.getValueType();
5348 assert(ResVT.getSizeInBits() == 256 && "Value type must be 256-bit wide");
5350 SDValue V1 = Op.getOperand(0);
5351 SDValue V2 = Op.getOperand(1);
5352 unsigned NumElems = ResVT.getVectorNumElements();
5354 SDValue V = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, ResVT), V1,
5355 DAG.getConstant(0, MVT::i32), DAG, dl);
5356 return Insert128BitVector(V, V2, DAG.getConstant(NumElems/2, MVT::i32),
5361 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
5362 EVT ResVT = Op.getValueType();
5364 assert(Op.getNumOperands() == 2);
5365 assert((ResVT.getSizeInBits() == 128 || ResVT.getSizeInBits() == 256) &&
5366 "Unsupported CONCAT_VECTORS for value type");
5368 // We support concatenate two MMX registers and place them in a MMX register.
5369 // This is better than doing a stack convert.
5370 if (ResVT.is128BitVector())
5371 return LowerMMXCONCAT_VECTORS(Op, DAG);
5373 // 256-bit AVX can use the vinsertf128 instruction to create 256-bit vectors
5374 // from two other 128-bit ones.
5375 return LowerAVXCONCAT_VECTORS(Op, DAG);
5378 // v8i16 shuffles - Prefer shuffles in the following order:
5379 // 1. [all] pshuflw, pshufhw, optional move
5380 // 2. [ssse3] 1 x pshufb
5381 // 3. [ssse3] 2 x pshufb + 1 x por
5382 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
5384 X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
5385 SelectionDAG &DAG) const {
5386 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5387 SDValue V1 = SVOp->getOperand(0);
5388 SDValue V2 = SVOp->getOperand(1);
5389 DebugLoc dl = SVOp->getDebugLoc();
5390 SmallVector<int, 8> MaskVals;
5392 // Determine if more than 1 of the words in each of the low and high quadwords
5393 // of the result come from the same quadword of one of the two inputs. Undef
5394 // mask values count as coming from any quadword, for better codegen.
5395 SmallVector<unsigned, 4> LoQuad(4);
5396 SmallVector<unsigned, 4> HiQuad(4);
5397 BitVector InputQuads(4);
5398 for (unsigned i = 0; i < 8; ++i) {
5399 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
5400 int EltIdx = SVOp->getMaskElt(i);
5401 MaskVals.push_back(EltIdx);
5410 InputQuads.set(EltIdx / 4);
5413 int BestLoQuad = -1;
5414 unsigned MaxQuad = 1;
5415 for (unsigned i = 0; i < 4; ++i) {
5416 if (LoQuad[i] > MaxQuad) {
5418 MaxQuad = LoQuad[i];
5422 int BestHiQuad = -1;
5424 for (unsigned i = 0; i < 4; ++i) {
5425 if (HiQuad[i] > MaxQuad) {
5427 MaxQuad = HiQuad[i];
5431 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
5432 // of the two input vectors, shuffle them into one input vector so only a
5433 // single pshufb instruction is necessary. If There are more than 2 input
5434 // quads, disable the next transformation since it does not help SSSE3.
5435 bool V1Used = InputQuads[0] || InputQuads[1];
5436 bool V2Used = InputQuads[2] || InputQuads[3];
5437 if (Subtarget->hasSSSE3()) {
5438 if (InputQuads.count() == 2 && V1Used && V2Used) {
5439 BestLoQuad = InputQuads.find_first();
5440 BestHiQuad = InputQuads.find_next(BestLoQuad);
5442 if (InputQuads.count() > 2) {
5448 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
5449 // the shuffle mask. If a quad is scored as -1, that means that it contains
5450 // words from all 4 input quadwords.
5452 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
5453 SmallVector<int, 8> MaskV;
5454 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
5455 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
5456 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
5457 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
5458 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
5459 NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
5461 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
5462 // source words for the shuffle, to aid later transformations.
5463 bool AllWordsInNewV = true;
5464 bool InOrder[2] = { true, true };
5465 for (unsigned i = 0; i != 8; ++i) {
5466 int idx = MaskVals[i];
5468 InOrder[i/4] = false;
5469 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
5471 AllWordsInNewV = false;
5475 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
5476 if (AllWordsInNewV) {
5477 for (int i = 0; i != 8; ++i) {
5478 int idx = MaskVals[i];
5481 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
5482 if ((idx != i) && idx < 4)
5484 if ((idx != i) && idx > 3)
5493 // If we've eliminated the use of V2, and the new mask is a pshuflw or
5494 // pshufhw, that's as cheap as it gets. Return the new shuffle.
5495 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
5496 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
5497 unsigned TargetMask = 0;
5498 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
5499 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
5500 TargetMask = pshufhw ? X86::getShufflePSHUFHWImmediate(NewV.getNode()):
5501 X86::getShufflePSHUFLWImmediate(NewV.getNode());
5502 V1 = NewV.getOperand(0);
5503 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
5507 // If we have SSSE3, and all words of the result are from 1 input vector,
5508 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
5509 // is present, fall back to case 4.
5510 if (Subtarget->hasSSSE3()) {
5511 SmallVector<SDValue,16> pshufbMask;
5513 // If we have elements from both input vectors, set the high bit of the
5514 // shuffle mask element to zero out elements that come from V2 in the V1
5515 // mask, and elements that come from V1 in the V2 mask, so that the two
5516 // results can be OR'd together.
5517 bool TwoInputs = V1Used && V2Used;
5518 for (unsigned i = 0; i != 8; ++i) {
5519 int EltIdx = MaskVals[i] * 2;
5520 if (TwoInputs && (EltIdx >= 16)) {
5521 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5522 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5525 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
5526 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
5528 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V1);
5529 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
5530 DAG.getNode(ISD::BUILD_VECTOR, dl,
5531 MVT::v16i8, &pshufbMask[0], 16));
5533 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5535 // Calculate the shuffle mask for the second input, shuffle it, and
5536 // OR it with the first shuffled input.
5538 for (unsigned i = 0; i != 8; ++i) {
5539 int EltIdx = MaskVals[i] * 2;
5541 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5542 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5545 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
5546 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
5548 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V2);
5549 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
5550 DAG.getNode(ISD::BUILD_VECTOR, dl,
5551 MVT::v16i8, &pshufbMask[0], 16));
5552 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
5553 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5556 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
5557 // and update MaskVals with new element order.
5558 BitVector InOrder(8);
5559 if (BestLoQuad >= 0) {
5560 SmallVector<int, 8> MaskV;
5561 for (int i = 0; i != 4; ++i) {
5562 int idx = MaskVals[i];
5564 MaskV.push_back(-1);
5566 } else if ((idx / 4) == BestLoQuad) {
5567 MaskV.push_back(idx & 3);
5570 MaskV.push_back(-1);
5573 for (unsigned i = 4; i != 8; ++i)
5575 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
5578 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
5579 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
5581 X86::getShufflePSHUFLWImmediate(NewV.getNode()),
5585 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
5586 // and update MaskVals with the new element order.
5587 if (BestHiQuad >= 0) {
5588 SmallVector<int, 8> MaskV;
5589 for (unsigned i = 0; i != 4; ++i)
5591 for (unsigned i = 4; i != 8; ++i) {
5592 int idx = MaskVals[i];
5594 MaskV.push_back(-1);
5596 } else if ((idx / 4) == BestHiQuad) {
5597 MaskV.push_back((idx & 3) + 4);
5600 MaskV.push_back(-1);
5603 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
5606 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
5607 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
5609 X86::getShufflePSHUFHWImmediate(NewV.getNode()),
5613 // In case BestHi & BestLo were both -1, which means each quadword has a word
5614 // from each of the four input quadwords, calculate the InOrder bitvector now
5615 // before falling through to the insert/extract cleanup.
5616 if (BestLoQuad == -1 && BestHiQuad == -1) {
5618 for (int i = 0; i != 8; ++i)
5619 if (MaskVals[i] < 0 || MaskVals[i] == i)
5623 // The other elements are put in the right place using pextrw and pinsrw.
5624 for (unsigned i = 0; i != 8; ++i) {
5627 int EltIdx = MaskVals[i];
5630 SDValue ExtOp = (EltIdx < 8)
5631 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
5632 DAG.getIntPtrConstant(EltIdx))
5633 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
5634 DAG.getIntPtrConstant(EltIdx - 8));
5635 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
5636 DAG.getIntPtrConstant(i));
5641 // v16i8 shuffles - Prefer shuffles in the following order:
5642 // 1. [ssse3] 1 x pshufb
5643 // 2. [ssse3] 2 x pshufb + 1 x por
5644 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
5646 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
5648 const X86TargetLowering &TLI) {
5649 SDValue V1 = SVOp->getOperand(0);
5650 SDValue V2 = SVOp->getOperand(1);
5651 DebugLoc dl = SVOp->getDebugLoc();
5652 SmallVector<int, 16> MaskVals;
5653 SVOp->getMask(MaskVals);
5655 // If we have SSSE3, case 1 is generated when all result bytes come from
5656 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
5657 // present, fall back to case 3.
5658 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
5661 for (unsigned i = 0; i < 16; ++i) {
5662 int EltIdx = MaskVals[i];
5671 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
5672 if (TLI.getSubtarget()->hasSSSE3()) {
5673 SmallVector<SDValue,16> pshufbMask;
5675 // If all result elements are from one input vector, then only translate
5676 // undef mask values to 0x80 (zero out result) in the pshufb mask.
5678 // Otherwise, we have elements from both input vectors, and must zero out
5679 // elements that come from V2 in the first mask, and V1 in the second mask
5680 // so that we can OR them together.
5681 bool TwoInputs = !(V1Only || V2Only);
5682 for (unsigned i = 0; i != 16; ++i) {
5683 int EltIdx = MaskVals[i];
5684 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
5685 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5688 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
5690 // If all the elements are from V2, assign it to V1 and return after
5691 // building the first pshufb.
5694 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
5695 DAG.getNode(ISD::BUILD_VECTOR, dl,
5696 MVT::v16i8, &pshufbMask[0], 16));
5700 // Calculate the shuffle mask for the second input, shuffle it, and
5701 // OR it with the first shuffled input.
5703 for (unsigned i = 0; i != 16; ++i) {
5704 int EltIdx = MaskVals[i];
5706 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5709 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
5711 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
5712 DAG.getNode(ISD::BUILD_VECTOR, dl,
5713 MVT::v16i8, &pshufbMask[0], 16));
5714 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
5717 // No SSSE3 - Calculate in place words and then fix all out of place words
5718 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
5719 // the 16 different words that comprise the two doublequadword input vectors.
5720 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5721 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
5722 SDValue NewV = V2Only ? V2 : V1;
5723 for (int i = 0; i != 8; ++i) {
5724 int Elt0 = MaskVals[i*2];
5725 int Elt1 = MaskVals[i*2+1];
5727 // This word of the result is all undef, skip it.
5728 if (Elt0 < 0 && Elt1 < 0)
5731 // This word of the result is already in the correct place, skip it.
5732 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
5734 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
5737 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
5738 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
5741 // If Elt0 and Elt1 are defined, are consecutive, and can be load
5742 // using a single extract together, load it and store it.
5743 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
5744 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
5745 DAG.getIntPtrConstant(Elt1 / 2));
5746 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5747 DAG.getIntPtrConstant(i));
5751 // If Elt1 is defined, extract it from the appropriate source. If the
5752 // source byte is not also odd, shift the extracted word left 8 bits
5753 // otherwise clear the bottom 8 bits if we need to do an or.
5755 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
5756 DAG.getIntPtrConstant(Elt1 / 2));
5757 if ((Elt1 & 1) == 0)
5758 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
5760 TLI.getShiftAmountTy(InsElt.getValueType())));
5762 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
5763 DAG.getConstant(0xFF00, MVT::i16));
5765 // If Elt0 is defined, extract it from the appropriate source. If the
5766 // source byte is not also even, shift the extracted word right 8 bits. If
5767 // Elt1 was also defined, OR the extracted values together before
5768 // inserting them in the result.
5770 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
5771 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
5772 if ((Elt0 & 1) != 0)
5773 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
5775 TLI.getShiftAmountTy(InsElt0.getValueType())));
5777 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
5778 DAG.getConstant(0x00FF, MVT::i16));
5779 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
5782 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5783 DAG.getIntPtrConstant(i));
5785 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
5788 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
5789 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
5790 /// done when every pair / quad of shuffle mask elements point to elements in
5791 /// the right sequence. e.g.
5792 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
5794 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
5795 SelectionDAG &DAG, DebugLoc dl) {
5796 EVT VT = SVOp->getValueType(0);
5797 SDValue V1 = SVOp->getOperand(0);
5798 SDValue V2 = SVOp->getOperand(1);
5799 unsigned NumElems = VT.getVectorNumElements();
5800 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
5802 switch (VT.getSimpleVT().SimpleTy) {
5803 default: assert(false && "Unexpected!");
5804 case MVT::v4f32: NewVT = MVT::v2f64; break;
5805 case MVT::v4i32: NewVT = MVT::v2i64; break;
5806 case MVT::v8i16: NewVT = MVT::v4i32; break;
5807 case MVT::v16i8: NewVT = MVT::v4i32; break;
5810 int Scale = NumElems / NewWidth;
5811 SmallVector<int, 8> MaskVec;
5812 for (unsigned i = 0; i < NumElems; i += Scale) {
5814 for (int j = 0; j < Scale; ++j) {
5815 int EltIdx = SVOp->getMaskElt(i+j);
5819 StartIdx = EltIdx - (EltIdx % Scale);
5820 if (EltIdx != StartIdx + j)
5824 MaskVec.push_back(-1);
5826 MaskVec.push_back(StartIdx / Scale);
5829 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
5830 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
5831 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
5834 /// getVZextMovL - Return a zero-extending vector move low node.
5836 static SDValue getVZextMovL(EVT VT, EVT OpVT,
5837 SDValue SrcOp, SelectionDAG &DAG,
5838 const X86Subtarget *Subtarget, DebugLoc dl) {
5839 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
5840 LoadSDNode *LD = NULL;
5841 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
5842 LD = dyn_cast<LoadSDNode>(SrcOp);
5844 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
5846 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
5847 if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
5848 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
5849 SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
5850 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
5852 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
5853 return DAG.getNode(ISD::BITCAST, dl, VT,
5854 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5855 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5863 return DAG.getNode(ISD::BITCAST, dl, VT,
5864 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5865 DAG.getNode(ISD::BITCAST, dl,
5869 /// areShuffleHalvesWithinDisjointLanes - Check whether each half of a vector
5870 /// shuffle node referes to only one lane in the sources.
5871 static bool areShuffleHalvesWithinDisjointLanes(ShuffleVectorSDNode *SVOp) {
5872 EVT VT = SVOp->getValueType(0);
5873 int NumElems = VT.getVectorNumElements();
5874 int HalfSize = NumElems/2;
5875 SmallVector<int, 16> M;
5877 bool MatchA = false, MatchB = false;
5879 for (int l = 0; l < NumElems*2; l += HalfSize) {
5880 if (isUndefOrInRange(M, 0, HalfSize, l, l+HalfSize)) {
5886 for (int l = 0; l < NumElems*2; l += HalfSize) {
5887 if (isUndefOrInRange(M, HalfSize, HalfSize, l, l+HalfSize)) {
5893 return MatchA && MatchB;
5896 /// LowerVECTOR_SHUFFLE_256 - Handle all 256-bit wide vectors shuffles
5897 /// which could not be matched by any known target speficic shuffle
5899 LowerVECTOR_SHUFFLE_256(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
5900 if (areShuffleHalvesWithinDisjointLanes(SVOp)) {
5901 // If each half of a vector shuffle node referes to only one lane in the
5902 // source vectors, extract each used 128-bit lane and shuffle them using
5903 // 128-bit shuffles. Then, concatenate the results. Otherwise leave
5904 // the work to the legalizer.
5905 DebugLoc dl = SVOp->getDebugLoc();
5906 EVT VT = SVOp->getValueType(0);
5907 int NumElems = VT.getVectorNumElements();
5908 int HalfSize = NumElems/2;
5910 // Extract the reference for each half
5911 int FstVecExtractIdx = 0, SndVecExtractIdx = 0;
5912 int FstVecOpNum = 0, SndVecOpNum = 0;
5913 for (int i = 0; i < HalfSize; ++i) {
5914 int Elt = SVOp->getMaskElt(i);
5915 if (SVOp->getMaskElt(i) < 0)
5917 FstVecOpNum = Elt/NumElems;
5918 FstVecExtractIdx = Elt % NumElems < HalfSize ? 0 : HalfSize;
5921 for (int i = HalfSize; i < NumElems; ++i) {
5922 int Elt = SVOp->getMaskElt(i);
5923 if (SVOp->getMaskElt(i) < 0)
5925 SndVecOpNum = Elt/NumElems;
5926 SndVecExtractIdx = Elt % NumElems < HalfSize ? 0 : HalfSize;
5930 // Extract the subvectors
5931 SDValue V1 = Extract128BitVector(SVOp->getOperand(FstVecOpNum),
5932 DAG.getConstant(FstVecExtractIdx, MVT::i32), DAG, dl);
5933 SDValue V2 = Extract128BitVector(SVOp->getOperand(SndVecOpNum),
5934 DAG.getConstant(SndVecExtractIdx, MVT::i32), DAG, dl);
5936 // Generate 128-bit shuffles
5937 SmallVector<int, 16> MaskV1, MaskV2;
5938 for (int i = 0; i < HalfSize; ++i) {
5939 int Elt = SVOp->getMaskElt(i);
5940 MaskV1.push_back(Elt < 0 ? Elt : Elt % HalfSize);
5942 for (int i = HalfSize; i < NumElems; ++i) {
5943 int Elt = SVOp->getMaskElt(i);
5944 MaskV2.push_back(Elt < 0 ? Elt : Elt % HalfSize);
5947 EVT NVT = V1.getValueType();
5948 V1 = DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &MaskV1[0]);
5949 V2 = DAG.getVectorShuffle(NVT, dl, V2, DAG.getUNDEF(NVT), &MaskV2[0]);
5951 // Concatenate the result back
5952 SDValue V = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT), V1,
5953 DAG.getConstant(0, MVT::i32), DAG, dl);
5954 return Insert128BitVector(V, V2, DAG.getConstant(NumElems/2, MVT::i32),
5961 /// LowerVECTOR_SHUFFLE_128v4 - Handle all 128-bit wide vectors with
5962 /// 4 elements, and match them with several different shuffle types.
5964 LowerVECTOR_SHUFFLE_128v4(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
5965 SDValue V1 = SVOp->getOperand(0);
5966 SDValue V2 = SVOp->getOperand(1);
5967 DebugLoc dl = SVOp->getDebugLoc();
5968 EVT VT = SVOp->getValueType(0);
5970 assert(VT.getSizeInBits() == 128 && "Unsupported vector size");
5972 SmallVector<std::pair<int, int>, 8> Locs;
5974 SmallVector<int, 8> Mask1(4U, -1);
5975 SmallVector<int, 8> PermMask;
5976 SVOp->getMask(PermMask);
5980 for (unsigned i = 0; i != 4; ++i) {
5981 int Idx = PermMask[i];
5983 Locs[i] = std::make_pair(-1, -1);
5985 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
5987 Locs[i] = std::make_pair(0, NumLo);
5991 Locs[i] = std::make_pair(1, NumHi);
5993 Mask1[2+NumHi] = Idx;
5999 if (NumLo <= 2 && NumHi <= 2) {
6000 // If no more than two elements come from either vector. This can be
6001 // implemented with two shuffles. First shuffle gather the elements.
6002 // The second shuffle, which takes the first shuffle as both of its
6003 // vector operands, put the elements into the right order.
6004 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
6006 SmallVector<int, 8> Mask2(4U, -1);
6008 for (unsigned i = 0; i != 4; ++i) {
6009 if (Locs[i].first == -1)
6012 unsigned Idx = (i < 2) ? 0 : 4;
6013 Idx += Locs[i].first * 2 + Locs[i].second;
6018 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
6019 } else if (NumLo == 3 || NumHi == 3) {
6020 // Otherwise, we must have three elements from one vector, call it X, and
6021 // one element from the other, call it Y. First, use a shufps to build an
6022 // intermediate vector with the one element from Y and the element from X
6023 // that will be in the same half in the final destination (the indexes don't
6024 // matter). Then, use a shufps to build the final vector, taking the half
6025 // containing the element from Y from the intermediate, and the other half
6028 // Normalize it so the 3 elements come from V1.
6029 CommuteVectorShuffleMask(PermMask, VT);
6033 // Find the element from V2.
6035 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
6036 int Val = PermMask[HiIndex];
6043 Mask1[0] = PermMask[HiIndex];
6045 Mask1[2] = PermMask[HiIndex^1];
6047 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
6050 Mask1[0] = PermMask[0];
6051 Mask1[1] = PermMask[1];
6052 Mask1[2] = HiIndex & 1 ? 6 : 4;
6053 Mask1[3] = HiIndex & 1 ? 4 : 6;
6054 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
6056 Mask1[0] = HiIndex & 1 ? 2 : 0;
6057 Mask1[1] = HiIndex & 1 ? 0 : 2;
6058 Mask1[2] = PermMask[2];
6059 Mask1[3] = PermMask[3];
6064 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
6068 // Break it into (shuffle shuffle_hi, shuffle_lo).
6071 SmallVector<int,8> LoMask(4U, -1);
6072 SmallVector<int,8> HiMask(4U, -1);
6074 SmallVector<int,8> *MaskPtr = &LoMask;
6075 unsigned MaskIdx = 0;
6078 for (unsigned i = 0; i != 4; ++i) {
6085 int Idx = PermMask[i];
6087 Locs[i] = std::make_pair(-1, -1);
6088 } else if (Idx < 4) {
6089 Locs[i] = std::make_pair(MaskIdx, LoIdx);
6090 (*MaskPtr)[LoIdx] = Idx;
6093 Locs[i] = std::make_pair(MaskIdx, HiIdx);
6094 (*MaskPtr)[HiIdx] = Idx;
6099 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
6100 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
6101 SmallVector<int, 8> MaskOps;
6102 for (unsigned i = 0; i != 4; ++i) {
6103 if (Locs[i].first == -1) {
6104 MaskOps.push_back(-1);
6106 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
6107 MaskOps.push_back(Idx);
6110 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
6113 static bool MayFoldVectorLoad(SDValue V) {
6114 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
6115 V = V.getOperand(0);
6116 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
6117 V = V.getOperand(0);
6123 // FIXME: the version above should always be used. Since there's
6124 // a bug where several vector shuffles can't be folded because the
6125 // DAG is not updated during lowering and a node claims to have two
6126 // uses while it only has one, use this version, and let isel match
6127 // another instruction if the load really happens to have more than
6128 // one use. Remove this version after this bug get fixed.
6129 // rdar://8434668, PR8156
6130 static bool RelaxedMayFoldVectorLoad(SDValue V) {
6131 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
6132 V = V.getOperand(0);
6133 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
6134 V = V.getOperand(0);
6135 if (ISD::isNormalLoad(V.getNode()))
6140 /// CanFoldShuffleIntoVExtract - Check if the current shuffle is used by
6141 /// a vector extract, and if both can be later optimized into a single load.
6142 /// This is done in visitEXTRACT_VECTOR_ELT and the conditions are checked
6143 /// here because otherwise a target specific shuffle node is going to be
6144 /// emitted for this shuffle, and the optimization not done.
6145 /// FIXME: This is probably not the best approach, but fix the problem
6146 /// until the right path is decided.
6148 bool CanXFormVExtractWithShuffleIntoLoad(SDValue V, SelectionDAG &DAG,
6149 const TargetLowering &TLI) {
6150 EVT VT = V.getValueType();
6151 ShuffleVectorSDNode *SVOp = dyn_cast<ShuffleVectorSDNode>(V);
6153 // Be sure that the vector shuffle is present in a pattern like this:
6154 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), c) -> (f32 load $addr)
6158 SDNode *N = *V.getNode()->use_begin();
6159 if (N->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
6162 SDValue EltNo = N->getOperand(1);
6163 if (!isa<ConstantSDNode>(EltNo))
6166 // If the bit convert changed the number of elements, it is unsafe
6167 // to examine the mask.
6168 bool HasShuffleIntoBitcast = false;
6169 if (V.getOpcode() == ISD::BITCAST) {
6170 EVT SrcVT = V.getOperand(0).getValueType();
6171 if (SrcVT.getVectorNumElements() != VT.getVectorNumElements())
6173 V = V.getOperand(0);
6174 HasShuffleIntoBitcast = true;
6177 // Select the input vector, guarding against out of range extract vector.
6178 unsigned NumElems = VT.getVectorNumElements();
6179 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
6180 int Idx = (Elt > NumElems) ? -1 : SVOp->getMaskElt(Elt);
6181 V = (Idx < (int)NumElems) ? V.getOperand(0) : V.getOperand(1);
6183 // Skip one more bit_convert if necessary
6184 if (V.getOpcode() == ISD::BITCAST)
6185 V = V.getOperand(0);
6187 if (ISD::isNormalLoad(V.getNode())) {
6188 // Is the original load suitable?
6189 LoadSDNode *LN0 = cast<LoadSDNode>(V);
6191 // FIXME: avoid the multi-use bug that is preventing lots of
6192 // of foldings to be detected, this is still wrong of course, but
6193 // give the temporary desired behavior, and if it happens that
6194 // the load has real more uses, during isel it will not fold, and
6195 // will generate poor code.
6196 if (!LN0 || LN0->isVolatile()) // || !LN0->hasOneUse()
6199 if (!HasShuffleIntoBitcast)
6202 // If there's a bitcast before the shuffle, check if the load type and
6203 // alignment is valid.
6204 unsigned Align = LN0->getAlignment();
6206 TLI.getTargetData()->getABITypeAlignment(
6207 VT.getTypeForEVT(*DAG.getContext()));
6209 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
6217 SDValue getMOVDDup(SDValue &Op, DebugLoc &dl, SDValue V1, SelectionDAG &DAG) {
6218 EVT VT = Op.getValueType();
6220 // Canonizalize to v2f64.
6221 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
6222 return DAG.getNode(ISD::BITCAST, dl, VT,
6223 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
6228 SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
6230 SDValue V1 = Op.getOperand(0);
6231 SDValue V2 = Op.getOperand(1);
6232 EVT VT = Op.getValueType();
6234 assert(VT != MVT::v2i64 && "unsupported shuffle type");
6236 if (HasSSE2 && VT == MVT::v2f64)
6237 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
6240 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V2, DAG);
6244 SDValue getMOVHighToLow(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG) {
6245 SDValue V1 = Op.getOperand(0);
6246 SDValue V2 = Op.getOperand(1);
6247 EVT VT = Op.getValueType();
6249 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
6250 "unsupported shuffle type");
6252 if (V2.getOpcode() == ISD::UNDEF)
6256 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
6259 static inline unsigned getSHUFPOpcode(EVT VT) {
6260 switch(VT.getSimpleVT().SimpleTy) {
6261 case MVT::v8i32: // Use fp unit for int unpack.
6263 case MVT::v4i32: // Use fp unit for int unpack.
6264 case MVT::v4f32: return X86ISD::SHUFPS;
6265 case MVT::v4i64: // Use fp unit for int unpack.
6267 case MVT::v2i64: // Use fp unit for int unpack.
6268 case MVT::v2f64: return X86ISD::SHUFPD;
6270 llvm_unreachable("Unknown type for shufp*");
6276 SDValue getMOVLP(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
6277 SDValue V1 = Op.getOperand(0);
6278 SDValue V2 = Op.getOperand(1);
6279 EVT VT = Op.getValueType();
6280 unsigned NumElems = VT.getVectorNumElements();
6282 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
6283 // operand of these instructions is only memory, so check if there's a
6284 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
6286 bool CanFoldLoad = false;
6288 // Trivial case, when V2 comes from a load.
6289 if (MayFoldVectorLoad(V2))
6292 // When V1 is a load, it can be folded later into a store in isel, example:
6293 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
6295 // (MOVLPSmr addr:$src1, VR128:$src2)
6296 // So, recognize this potential and also use MOVLPS or MOVLPD
6297 if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
6300 // Both of them can't be memory operations though.
6301 if (MayFoldVectorLoad(V1) && MayFoldVectorLoad(V2))
6302 CanFoldLoad = false;
6305 if (HasSSE2 && NumElems == 2)
6306 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
6309 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
6312 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6313 // movl and movlp will both match v2i64, but v2i64 is never matched by
6314 // movl earlier because we make it strict to avoid messing with the movlp load
6315 // folding logic (see the code above getMOVLP call). Match it here then,
6316 // this is horrible, but will stay like this until we move all shuffle
6317 // matching to x86 specific nodes. Note that for the 1st condition all
6318 // types are matched with movsd.
6319 if ((HasSSE2 && NumElems == 2) || !X86::isMOVLMask(SVOp))
6320 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
6322 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
6325 assert(VT != MVT::v4i32 && "unsupported shuffle type");
6327 // Invert the operand order and use SHUFPS to match it.
6328 return getTargetShuffleNode(getSHUFPOpcode(VT), dl, VT, V2, V1,
6329 X86::getShuffleSHUFImmediate(SVOp), DAG);
6332 static inline unsigned getUNPCKLOpcode(EVT VT) {
6333 switch(VT.getSimpleVT().SimpleTy) {
6334 case MVT::v4i32: return X86ISD::PUNPCKLDQ;
6335 case MVT::v2i64: return X86ISD::PUNPCKLQDQ;
6336 case MVT::v4f32: return X86ISD::UNPCKLPS;
6337 case MVT::v2f64: return X86ISD::UNPCKLPD;
6338 case MVT::v8i32: // Use fp unit for int unpack.
6339 case MVT::v8f32: return X86ISD::VUNPCKLPSY;
6340 case MVT::v4i64: // Use fp unit for int unpack.
6341 case MVT::v4f64: return X86ISD::VUNPCKLPDY;
6342 case MVT::v16i8: return X86ISD::PUNPCKLBW;
6343 case MVT::v8i16: return X86ISD::PUNPCKLWD;
6345 llvm_unreachable("Unknown type for unpckl");
6350 static inline unsigned getUNPCKHOpcode(EVT VT) {
6351 switch(VT.getSimpleVT().SimpleTy) {
6352 case MVT::v4i32: return X86ISD::PUNPCKHDQ;
6353 case MVT::v2i64: return X86ISD::PUNPCKHQDQ;
6354 case MVT::v4f32: return X86ISD::UNPCKHPS;
6355 case MVT::v2f64: return X86ISD::UNPCKHPD;
6356 case MVT::v8i32: // Use fp unit for int unpack.
6357 case MVT::v8f32: return X86ISD::VUNPCKHPSY;
6358 case MVT::v4i64: // Use fp unit for int unpack.
6359 case MVT::v4f64: return X86ISD::VUNPCKHPDY;
6360 case MVT::v16i8: return X86ISD::PUNPCKHBW;
6361 case MVT::v8i16: return X86ISD::PUNPCKHWD;
6363 llvm_unreachable("Unknown type for unpckh");
6368 static inline unsigned getVPERMILOpcode(EVT VT) {
6369 switch(VT.getSimpleVT().SimpleTy) {
6371 case MVT::v4f32: return X86ISD::VPERMILPS;
6373 case MVT::v2f64: return X86ISD::VPERMILPD;
6375 case MVT::v8f32: return X86ISD::VPERMILPSY;
6377 case MVT::v4f64: return X86ISD::VPERMILPDY;
6379 llvm_unreachable("Unknown type for vpermil");
6384 /// isVectorBroadcast - Check if the node chain is suitable to be xformed to
6385 /// a vbroadcast node. The nodes are suitable whenever we can fold a load coming
6386 /// from a 32 or 64 bit scalar. Update Op to the desired load to be folded.
6387 static bool isVectorBroadcast(SDValue &Op) {
6388 EVT VT = Op.getValueType();
6389 bool Is256 = VT.getSizeInBits() == 256;
6391 assert((VT.getSizeInBits() == 128 || Is256) &&
6392 "Unsupported type for vbroadcast node");
6395 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
6396 V = V.getOperand(0);
6398 if (Is256 && !(V.hasOneUse() &&
6399 V.getOpcode() == ISD::INSERT_SUBVECTOR &&
6400 V.getOperand(0).getOpcode() == ISD::UNDEF))
6404 V = V.getOperand(1);
6405 if (V.hasOneUse() && V.getOpcode() != ISD::SCALAR_TO_VECTOR)
6408 // Check the source scalar_to_vector type. 256-bit broadcasts are
6409 // supported for 32/64-bit sizes, while 128-bit ones are only supported
6410 // for 32-bit scalars.
6411 unsigned ScalarSize = V.getOperand(0).getValueType().getSizeInBits();
6412 if (ScalarSize != 32 && ScalarSize != 64)
6414 if (!Is256 && ScalarSize == 64)
6417 V = V.getOperand(0);
6418 if (!MayFoldLoad(V))
6421 // Return the load node
6427 SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG,
6428 const TargetLowering &TLI,
6429 const X86Subtarget *Subtarget) {
6430 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6431 EVT VT = Op.getValueType();
6432 DebugLoc dl = Op.getDebugLoc();
6433 SDValue V1 = Op.getOperand(0);
6434 SDValue V2 = Op.getOperand(1);
6436 if (isZeroShuffle(SVOp))
6437 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
6439 // Handle splat operations
6440 if (SVOp->isSplat()) {
6441 unsigned NumElem = VT.getVectorNumElements();
6442 int Size = VT.getSizeInBits();
6443 // Special case, this is the only place now where it's allowed to return
6444 // a vector_shuffle operation without using a target specific node, because
6445 // *hopefully* it will be optimized away by the dag combiner. FIXME: should
6446 // this be moved to DAGCombine instead?
6447 if (NumElem <= 4 && CanXFormVExtractWithShuffleIntoLoad(Op, DAG, TLI))
6450 // Use vbroadcast whenever the splat comes from a foldable load
6451 if (Subtarget->hasAVX() && isVectorBroadcast(V1))
6452 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, V1);
6454 // Handle splats by matching through known shuffle masks
6455 if ((Size == 128 && NumElem <= 4) ||
6456 (Size == 256 && NumElem < 8))
6459 // All remaning splats are promoted to target supported vector shuffles.
6460 return PromoteSplat(SVOp, DAG);
6463 // If the shuffle can be profitably rewritten as a narrower shuffle, then
6465 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
6466 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
6467 if (NewOp.getNode())
6468 return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
6469 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
6470 // FIXME: Figure out a cleaner way to do this.
6471 // Try to make use of movq to zero out the top part.
6472 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
6473 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
6474 if (NewOp.getNode()) {
6475 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
6476 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
6477 DAG, Subtarget, dl);
6479 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
6480 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
6481 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
6482 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
6483 DAG, Subtarget, dl);
6490 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
6491 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6492 SDValue V1 = Op.getOperand(0);
6493 SDValue V2 = Op.getOperand(1);
6494 EVT VT = Op.getValueType();
6495 DebugLoc dl = Op.getDebugLoc();
6496 unsigned NumElems = VT.getVectorNumElements();
6497 bool isMMX = VT.getSizeInBits() == 64;
6498 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
6499 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
6500 bool V1IsSplat = false;
6501 bool V2IsSplat = false;
6502 bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
6503 bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
6504 bool HasSSSE3 = Subtarget->hasSSSE3() || Subtarget->hasAVX();
6505 MachineFunction &MF = DAG.getMachineFunction();
6506 bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
6508 // Shuffle operations on MMX not supported.
6512 // Vector shuffle lowering takes 3 steps:
6514 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
6515 // narrowing and commutation of operands should be handled.
6516 // 2) Matching of shuffles with known shuffle masks to x86 target specific
6518 // 3) Rewriting of unmatched masks into new generic shuffle operations,
6519 // so the shuffle can be broken into other shuffles and the legalizer can
6520 // try the lowering again.
6522 // The general ideia is that no vector_shuffle operation should be left to
6523 // be matched during isel, all of them must be converted to a target specific
6526 // Normalize the input vectors. Here splats, zeroed vectors, profitable
6527 // narrowing and commutation of operands should be handled. The actual code
6528 // doesn't include all of those, work in progress...
6529 SDValue NewOp = NormalizeVectorShuffle(Op, DAG, *this, Subtarget);
6530 if (NewOp.getNode())
6533 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
6534 // unpckh_undef). Only use pshufd if speed is more important than size.
6535 if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp))
6536 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
6537 if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp))
6538 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
6540 if (X86::isMOVDDUPMask(SVOp) && HasSSE3 && V2IsUndef &&
6541 RelaxedMayFoldVectorLoad(V1))
6542 return getMOVDDup(Op, dl, V1, DAG);
6544 if (X86::isMOVHLPS_v_undef_Mask(SVOp))
6545 return getMOVHighToLow(Op, dl, DAG);
6547 // Use to match splats
6548 if (HasSSE2 && X86::isUNPCKHMask(SVOp) && V2IsUndef &&
6549 (VT == MVT::v2f64 || VT == MVT::v2i64))
6550 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
6552 if (X86::isPSHUFDMask(SVOp)) {
6553 // The actual implementation will match the mask in the if above and then
6554 // during isel it can match several different instructions, not only pshufd
6555 // as its name says, sad but true, emulate the behavior for now...
6556 if (X86::isMOVDDUPMask(SVOp) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
6557 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
6559 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
6561 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
6562 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
6564 return getTargetShuffleNode(getSHUFPOpcode(VT), dl, VT, V1, V1,
6568 // Check if this can be converted into a logical shift.
6569 bool isLeft = false;
6572 bool isShift = getSubtarget()->hasSSE2() &&
6573 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
6574 if (isShift && ShVal.hasOneUse()) {
6575 // If the shifted value has multiple uses, it may be cheaper to use
6576 // v_set0 + movlhps or movhlps, etc.
6577 EVT EltVT = VT.getVectorElementType();
6578 ShAmt *= EltVT.getSizeInBits();
6579 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
6582 if (X86::isMOVLMask(SVOp)) {
6585 if (ISD::isBuildVectorAllZeros(V1.getNode()))
6586 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
6587 if (!X86::isMOVLPMask(SVOp)) {
6588 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
6589 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
6591 if (VT == MVT::v4i32 || VT == MVT::v4f32)
6592 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
6596 // FIXME: fold these into legal mask.
6597 if (X86::isMOVLHPSMask(SVOp) && !X86::isUNPCKLMask(SVOp))
6598 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
6600 if (X86::isMOVHLPSMask(SVOp))
6601 return getMOVHighToLow(Op, dl, DAG);
6603 if (X86::isMOVSHDUPMask(SVOp, Subtarget))
6604 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
6606 if (X86::isMOVSLDUPMask(SVOp, Subtarget))
6607 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
6609 if (X86::isMOVLPMask(SVOp))
6610 return getMOVLP(Op, dl, DAG, HasSSE2);
6612 if (ShouldXformToMOVHLPS(SVOp) ||
6613 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
6614 return CommuteVectorShuffle(SVOp, DAG);
6617 // No better options. Use a vshl / vsrl.
6618 EVT EltVT = VT.getVectorElementType();
6619 ShAmt *= EltVT.getSizeInBits();
6620 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
6623 bool Commuted = false;
6624 // FIXME: This should also accept a bitcast of a splat? Be careful, not
6625 // 1,1,1,1 -> v8i16 though.
6626 V1IsSplat = isSplatVector(V1.getNode());
6627 V2IsSplat = isSplatVector(V2.getNode());
6629 // Canonicalize the splat or undef, if present, to be on the RHS.
6630 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
6631 Op = CommuteVectorShuffle(SVOp, DAG);
6632 SVOp = cast<ShuffleVectorSDNode>(Op);
6633 V1 = SVOp->getOperand(0);
6634 V2 = SVOp->getOperand(1);
6635 std::swap(V1IsSplat, V2IsSplat);
6636 std::swap(V1IsUndef, V2IsUndef);
6640 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
6641 // Shuffling low element of v1 into undef, just return v1.
6644 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
6645 // the instruction selector will not match, so get a canonical MOVL with
6646 // swapped operands to undo the commute.
6647 return getMOVL(DAG, dl, VT, V2, V1);
6650 if (X86::isUNPCKLMask(SVOp))
6651 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V2, DAG);
6653 if (X86::isUNPCKHMask(SVOp))
6654 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V2, DAG);
6657 // Normalize mask so all entries that point to V2 points to its first
6658 // element then try to match unpck{h|l} again. If match, return a
6659 // new vector_shuffle with the corrected mask.
6660 SDValue NewMask = NormalizeMask(SVOp, DAG);
6661 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
6662 if (NSVOp != SVOp) {
6663 if (X86::isUNPCKLMask(NSVOp, true)) {
6665 } else if (X86::isUNPCKHMask(NSVOp, true)) {
6672 // Commute is back and try unpck* again.
6673 // FIXME: this seems wrong.
6674 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
6675 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
6677 if (X86::isUNPCKLMask(NewSVOp))
6678 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V2, V1, DAG);
6680 if (X86::isUNPCKHMask(NewSVOp))
6681 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V2, V1, DAG);
6684 // Normalize the node to match x86 shuffle ops if needed
6685 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
6686 return CommuteVectorShuffle(SVOp, DAG);
6688 // The checks below are all present in isShuffleMaskLegal, but they are
6689 // inlined here right now to enable us to directly emit target specific
6690 // nodes, and remove one by one until they don't return Op anymore.
6691 SmallVector<int, 16> M;
6694 if (isPALIGNRMask(M, VT, HasSSSE3))
6695 return getTargetShuffleNode(X86ISD::PALIGN, dl, VT, V1, V2,
6696 X86::getShufflePALIGNRImmediate(SVOp),
6699 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
6700 SVOp->getSplatIndex() == 0 && V2IsUndef) {
6701 if (VT == MVT::v2f64)
6702 return getTargetShuffleNode(X86ISD::UNPCKLPD, dl, VT, V1, V1, DAG);
6703 if (VT == MVT::v2i64)
6704 return getTargetShuffleNode(X86ISD::PUNPCKLQDQ, dl, VT, V1, V1, DAG);
6707 if (isPSHUFHWMask(M, VT))
6708 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
6709 X86::getShufflePSHUFHWImmediate(SVOp),
6712 if (isPSHUFLWMask(M, VT))
6713 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
6714 X86::getShufflePSHUFLWImmediate(SVOp),
6717 if (isSHUFPMask(M, VT))
6718 return getTargetShuffleNode(getSHUFPOpcode(VT), dl, VT, V1, V2,
6719 X86::getShuffleSHUFImmediate(SVOp), DAG);
6721 if (X86::isUNPCKL_v_undef_Mask(SVOp))
6722 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
6723 if (X86::isUNPCKH_v_undef_Mask(SVOp))
6724 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
6726 //===--------------------------------------------------------------------===//
6727 // Generate target specific nodes for 128 or 256-bit shuffles only
6728 // supported in the AVX instruction set.
6731 // Handle VMOVDDUPY permutations
6732 if (isMOVDDUPYMask(SVOp, Subtarget))
6733 return getTargetShuffleNode(X86ISD::MOVDDUP, dl, VT, V1, DAG);
6735 // Handle VPERMILPS* permutations
6736 if (isVPERMILPSMask(M, VT, Subtarget))
6737 return getTargetShuffleNode(getVPERMILOpcode(VT), dl, VT, V1,
6738 getShuffleVPERMILPSImmediate(SVOp), DAG);
6740 // Handle VPERMILPD* permutations
6741 if (isVPERMILPDMask(M, VT, Subtarget))
6742 return getTargetShuffleNode(getVPERMILOpcode(VT), dl, VT, V1,
6743 getShuffleVPERMILPDImmediate(SVOp), DAG);
6745 // Handle VPERM2F128 permutations
6746 if (isVPERM2F128Mask(M, VT, Subtarget))
6747 return getTargetShuffleNode(X86ISD::VPERM2F128, dl, VT, V1, V2,
6748 getShuffleVPERM2F128Immediate(SVOp), DAG);
6750 // Handle VSHUFPSY permutations
6751 if (isVSHUFPSYMask(M, VT, Subtarget))
6752 return getTargetShuffleNode(getSHUFPOpcode(VT), dl, VT, V1, V2,
6753 getShuffleVSHUFPSYImmediate(SVOp), DAG);
6755 // Handle VSHUFPDY permutations
6756 if (isVSHUFPDYMask(M, VT, Subtarget))
6757 return getTargetShuffleNode(getSHUFPOpcode(VT), dl, VT, V1, V2,
6758 getShuffleVSHUFPDYImmediate(SVOp), DAG);
6760 //===--------------------------------------------------------------------===//
6761 // Since no target specific shuffle was selected for this generic one,
6762 // lower it into other known shuffles. FIXME: this isn't true yet, but
6763 // this is the plan.
6766 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
6767 if (VT == MVT::v8i16) {
6768 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
6769 if (NewOp.getNode())
6773 if (VT == MVT::v16i8) {
6774 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
6775 if (NewOp.getNode())
6779 // Handle all 128-bit wide vectors with 4 elements, and match them with
6780 // several different shuffle types.
6781 if (NumElems == 4 && VT.getSizeInBits() == 128)
6782 return LowerVECTOR_SHUFFLE_128v4(SVOp, DAG);
6784 // Handle general 256-bit shuffles
6785 if (VT.is256BitVector())
6786 return LowerVECTOR_SHUFFLE_256(SVOp, DAG);
6792 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
6793 SelectionDAG &DAG) const {
6794 EVT VT = Op.getValueType();
6795 DebugLoc dl = Op.getDebugLoc();
6797 if (Op.getOperand(0).getValueType().getSizeInBits() != 128)
6800 if (VT.getSizeInBits() == 8) {
6801 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
6802 Op.getOperand(0), Op.getOperand(1));
6803 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
6804 DAG.getValueType(VT));
6805 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6806 } else if (VT.getSizeInBits() == 16) {
6807 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6808 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
6810 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
6811 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6812 DAG.getNode(ISD::BITCAST, dl,
6816 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
6817 Op.getOperand(0), Op.getOperand(1));
6818 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
6819 DAG.getValueType(VT));
6820 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6821 } else if (VT == MVT::f32) {
6822 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
6823 // the result back to FR32 register. It's only worth matching if the
6824 // result has a single use which is a store or a bitcast to i32. And in
6825 // the case of a store, it's not worth it if the index is a constant 0,
6826 // because a MOVSSmr can be used instead, which is smaller and faster.
6827 if (!Op.hasOneUse())
6829 SDNode *User = *Op.getNode()->use_begin();
6830 if ((User->getOpcode() != ISD::STORE ||
6831 (isa<ConstantSDNode>(Op.getOperand(1)) &&
6832 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
6833 (User->getOpcode() != ISD::BITCAST ||
6834 User->getValueType(0) != MVT::i32))
6836 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6837 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
6840 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
6841 } else if (VT == MVT::i32) {
6842 // ExtractPS works with constant index.
6843 if (isa<ConstantSDNode>(Op.getOperand(1)))
6851 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
6852 SelectionDAG &DAG) const {
6853 if (!isa<ConstantSDNode>(Op.getOperand(1)))
6856 SDValue Vec = Op.getOperand(0);
6857 EVT VecVT = Vec.getValueType();
6859 // If this is a 256-bit vector result, first extract the 128-bit vector and
6860 // then extract the element from the 128-bit vector.
6861 if (VecVT.getSizeInBits() == 256) {
6862 DebugLoc dl = Op.getNode()->getDebugLoc();
6863 unsigned NumElems = VecVT.getVectorNumElements();
6864 SDValue Idx = Op.getOperand(1);
6865 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
6867 // Get the 128-bit vector.
6868 bool Upper = IdxVal >= NumElems/2;
6869 Vec = Extract128BitVector(Vec,
6870 DAG.getConstant(Upper ? NumElems/2 : 0, MVT::i32), DAG, dl);
6872 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
6873 Upper ? DAG.getConstant(IdxVal-NumElems/2, MVT::i32) : Idx);
6876 assert(Vec.getValueSizeInBits() <= 128 && "Unexpected vector length");
6878 if (Subtarget->hasSSE41() || Subtarget->hasAVX()) {
6879 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
6884 EVT VT = Op.getValueType();
6885 DebugLoc dl = Op.getDebugLoc();
6886 // TODO: handle v16i8.
6887 if (VT.getSizeInBits() == 16) {
6888 SDValue Vec = Op.getOperand(0);
6889 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6891 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
6892 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6893 DAG.getNode(ISD::BITCAST, dl,
6896 // Transform it so it match pextrw which produces a 32-bit result.
6897 EVT EltVT = MVT::i32;
6898 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
6899 Op.getOperand(0), Op.getOperand(1));
6900 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
6901 DAG.getValueType(VT));
6902 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6903 } else if (VT.getSizeInBits() == 32) {
6904 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6908 // SHUFPS the element to the lowest double word, then movss.
6909 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
6910 EVT VVT = Op.getOperand(0).getValueType();
6911 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
6912 DAG.getUNDEF(VVT), Mask);
6913 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
6914 DAG.getIntPtrConstant(0));
6915 } else if (VT.getSizeInBits() == 64) {
6916 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
6917 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
6918 // to match extract_elt for f64.
6919 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6923 // UNPCKHPD the element to the lowest double word, then movsd.
6924 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
6925 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
6926 int Mask[2] = { 1, -1 };
6927 EVT VVT = Op.getOperand(0).getValueType();
6928 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
6929 DAG.getUNDEF(VVT), Mask);
6930 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
6931 DAG.getIntPtrConstant(0));
6938 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
6939 SelectionDAG &DAG) const {
6940 EVT VT = Op.getValueType();
6941 EVT EltVT = VT.getVectorElementType();
6942 DebugLoc dl = Op.getDebugLoc();
6944 SDValue N0 = Op.getOperand(0);
6945 SDValue N1 = Op.getOperand(1);
6946 SDValue N2 = Op.getOperand(2);
6948 if (VT.getSizeInBits() == 256)
6951 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
6952 isa<ConstantSDNode>(N2)) {
6954 if (VT == MVT::v8i16)
6955 Opc = X86ISD::PINSRW;
6956 else if (VT == MVT::v16i8)
6957 Opc = X86ISD::PINSRB;
6959 Opc = X86ISD::PINSRB;
6961 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
6963 if (N1.getValueType() != MVT::i32)
6964 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
6965 if (N2.getValueType() != MVT::i32)
6966 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
6967 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
6968 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
6969 // Bits [7:6] of the constant are the source select. This will always be
6970 // zero here. The DAG Combiner may combine an extract_elt index into these
6971 // bits. For example (insert (extract, 3), 2) could be matched by putting
6972 // the '3' into bits [7:6] of X86ISD::INSERTPS.
6973 // Bits [5:4] of the constant are the destination select. This is the
6974 // value of the incoming immediate.
6975 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
6976 // combine either bitwise AND or insert of float 0.0 to set these bits.
6977 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
6978 // Create this as a scalar to vector..
6979 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
6980 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
6981 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
6982 // PINSR* works with constant index.
6989 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
6990 EVT VT = Op.getValueType();
6991 EVT EltVT = VT.getVectorElementType();
6993 DebugLoc dl = Op.getDebugLoc();
6994 SDValue N0 = Op.getOperand(0);
6995 SDValue N1 = Op.getOperand(1);
6996 SDValue N2 = Op.getOperand(2);
6998 // If this is a 256-bit vector result, first extract the 128-bit vector,
6999 // insert the element into the extracted half and then place it back.
7000 if (VT.getSizeInBits() == 256) {
7001 if (!isa<ConstantSDNode>(N2))
7004 // Get the desired 128-bit vector half.
7005 unsigned NumElems = VT.getVectorNumElements();
7006 unsigned IdxVal = cast<ConstantSDNode>(N2)->getZExtValue();
7007 bool Upper = IdxVal >= NumElems/2;
7008 SDValue Ins128Idx = DAG.getConstant(Upper ? NumElems/2 : 0, MVT::i32);
7009 SDValue V = Extract128BitVector(N0, Ins128Idx, DAG, dl);
7011 // Insert the element into the desired half.
7012 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V,
7013 N1, Upper ? DAG.getConstant(IdxVal-NumElems/2, MVT::i32) : N2);
7015 // Insert the changed part back to the 256-bit vector
7016 return Insert128BitVector(N0, V, Ins128Idx, DAG, dl);
7019 if (Subtarget->hasSSE41() || Subtarget->hasAVX())
7020 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
7022 if (EltVT == MVT::i8)
7025 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
7026 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
7027 // as its second argument.
7028 if (N1.getValueType() != MVT::i32)
7029 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
7030 if (N2.getValueType() != MVT::i32)
7031 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
7032 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
7038 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
7039 LLVMContext *Context = DAG.getContext();
7040 DebugLoc dl = Op.getDebugLoc();
7041 EVT OpVT = Op.getValueType();
7043 // If this is a 256-bit vector result, first insert into a 128-bit
7044 // vector and then insert into the 256-bit vector.
7045 if (OpVT.getSizeInBits() > 128) {
7046 // Insert into a 128-bit vector.
7047 EVT VT128 = EVT::getVectorVT(*Context,
7048 OpVT.getVectorElementType(),
7049 OpVT.getVectorNumElements() / 2);
7051 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
7053 // Insert the 128-bit vector.
7054 return Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, OpVT), Op,
7055 DAG.getConstant(0, MVT::i32),
7059 if (Op.getValueType() == MVT::v1i64 &&
7060 Op.getOperand(0).getValueType() == MVT::i64)
7061 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
7063 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
7064 assert(Op.getValueType().getSimpleVT().getSizeInBits() == 128 &&
7065 "Expected an SSE type!");
7066 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(),
7067 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
7070 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
7071 // a simple subregister reference or explicit instructions to grab
7072 // upper bits of a vector.
7074 X86TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
7075 if (Subtarget->hasAVX()) {
7076 DebugLoc dl = Op.getNode()->getDebugLoc();
7077 SDValue Vec = Op.getNode()->getOperand(0);
7078 SDValue Idx = Op.getNode()->getOperand(1);
7080 if (Op.getNode()->getValueType(0).getSizeInBits() == 128
7081 && Vec.getNode()->getValueType(0).getSizeInBits() == 256) {
7082 return Extract128BitVector(Vec, Idx, DAG, dl);
7088 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
7089 // simple superregister reference or explicit instructions to insert
7090 // the upper bits of a vector.
7092 X86TargetLowering::LowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
7093 if (Subtarget->hasAVX()) {
7094 DebugLoc dl = Op.getNode()->getDebugLoc();
7095 SDValue Vec = Op.getNode()->getOperand(0);
7096 SDValue SubVec = Op.getNode()->getOperand(1);
7097 SDValue Idx = Op.getNode()->getOperand(2);
7099 if (Op.getNode()->getValueType(0).getSizeInBits() == 256
7100 && SubVec.getNode()->getValueType(0).getSizeInBits() == 128) {
7101 return Insert128BitVector(Vec, SubVec, Idx, DAG, dl);
7107 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
7108 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
7109 // one of the above mentioned nodes. It has to be wrapped because otherwise
7110 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
7111 // be used to form addressing mode. These wrapped nodes will be selected
7114 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
7115 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
7117 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7119 unsigned char OpFlag = 0;
7120 unsigned WrapperKind = X86ISD::Wrapper;
7121 CodeModel::Model M = getTargetMachine().getCodeModel();
7123 if (Subtarget->isPICStyleRIPRel() &&
7124 (M == CodeModel::Small || M == CodeModel::Kernel))
7125 WrapperKind = X86ISD::WrapperRIP;
7126 else if (Subtarget->isPICStyleGOT())
7127 OpFlag = X86II::MO_GOTOFF;
7128 else if (Subtarget->isPICStyleStubPIC())
7129 OpFlag = X86II::MO_PIC_BASE_OFFSET;
7131 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
7133 CP->getOffset(), OpFlag);
7134 DebugLoc DL = CP->getDebugLoc();
7135 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7136 // With PIC, the address is actually $g + Offset.
7138 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7139 DAG.getNode(X86ISD::GlobalBaseReg,
7140 DebugLoc(), getPointerTy()),
7147 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
7148 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
7150 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7152 unsigned char OpFlag = 0;
7153 unsigned WrapperKind = X86ISD::Wrapper;
7154 CodeModel::Model M = getTargetMachine().getCodeModel();
7156 if (Subtarget->isPICStyleRIPRel() &&
7157 (M == CodeModel::Small || M == CodeModel::Kernel))
7158 WrapperKind = X86ISD::WrapperRIP;
7159 else if (Subtarget->isPICStyleGOT())
7160 OpFlag = X86II::MO_GOTOFF;
7161 else if (Subtarget->isPICStyleStubPIC())
7162 OpFlag = X86II::MO_PIC_BASE_OFFSET;
7164 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
7166 DebugLoc DL = JT->getDebugLoc();
7167 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7169 // With PIC, the address is actually $g + Offset.
7171 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7172 DAG.getNode(X86ISD::GlobalBaseReg,
7173 DebugLoc(), getPointerTy()),
7180 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
7181 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
7183 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7185 unsigned char OpFlag = 0;
7186 unsigned WrapperKind = X86ISD::Wrapper;
7187 CodeModel::Model M = getTargetMachine().getCodeModel();
7189 if (Subtarget->isPICStyleRIPRel() &&
7190 (M == CodeModel::Small || M == CodeModel::Kernel)) {
7191 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
7192 OpFlag = X86II::MO_GOTPCREL;
7193 WrapperKind = X86ISD::WrapperRIP;
7194 } else if (Subtarget->isPICStyleGOT()) {
7195 OpFlag = X86II::MO_GOT;
7196 } else if (Subtarget->isPICStyleStubPIC()) {
7197 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
7198 } else if (Subtarget->isPICStyleStubNoDynamic()) {
7199 OpFlag = X86II::MO_DARWIN_NONLAZY;
7202 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
7204 DebugLoc DL = Op.getDebugLoc();
7205 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7208 // With PIC, the address is actually $g + Offset.
7209 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
7210 !Subtarget->is64Bit()) {
7211 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7212 DAG.getNode(X86ISD::GlobalBaseReg,
7213 DebugLoc(), getPointerTy()),
7217 // For symbols that require a load from a stub to get the address, emit the
7219 if (isGlobalStubReference(OpFlag))
7220 Result = DAG.getLoad(getPointerTy(), DL, DAG.getEntryNode(), Result,
7221 MachinePointerInfo::getGOT(), false, false, 0);
7227 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
7228 // Create the TargetBlockAddressAddress node.
7229 unsigned char OpFlags =
7230 Subtarget->ClassifyBlockAddressReference();
7231 CodeModel::Model M = getTargetMachine().getCodeModel();
7232 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
7233 DebugLoc dl = Op.getDebugLoc();
7234 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
7235 /*isTarget=*/true, OpFlags);
7237 if (Subtarget->isPICStyleRIPRel() &&
7238 (M == CodeModel::Small || M == CodeModel::Kernel))
7239 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
7241 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
7243 // With PIC, the address is actually $g + Offset.
7244 if (isGlobalRelativeToPICBase(OpFlags)) {
7245 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7246 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
7254 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
7256 SelectionDAG &DAG) const {
7257 // Create the TargetGlobalAddress node, folding in the constant
7258 // offset if it is legal.
7259 unsigned char OpFlags =
7260 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
7261 CodeModel::Model M = getTargetMachine().getCodeModel();
7263 if (OpFlags == X86II::MO_NO_FLAG &&
7264 X86::isOffsetSuitableForCodeModel(Offset, M)) {
7265 // A direct static reference to a global.
7266 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
7269 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
7272 if (Subtarget->isPICStyleRIPRel() &&
7273 (M == CodeModel::Small || M == CodeModel::Kernel))
7274 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
7276 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
7278 // With PIC, the address is actually $g + Offset.
7279 if (isGlobalRelativeToPICBase(OpFlags)) {
7280 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7281 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
7285 // For globals that require a load from a stub to get the address, emit the
7287 if (isGlobalStubReference(OpFlags))
7288 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
7289 MachinePointerInfo::getGOT(), false, false, 0);
7291 // If there was a non-zero offset that we didn't fold, create an explicit
7294 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
7295 DAG.getConstant(Offset, getPointerTy()));
7301 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
7302 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
7303 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
7304 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
7308 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
7309 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
7310 unsigned char OperandFlags) {
7311 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7312 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
7313 DebugLoc dl = GA->getDebugLoc();
7314 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
7315 GA->getValueType(0),
7319 SDValue Ops[] = { Chain, TGA, *InFlag };
7320 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
7322 SDValue Ops[] = { Chain, TGA };
7323 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
7326 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
7327 MFI->setAdjustsStack(true);
7329 SDValue Flag = Chain.getValue(1);
7330 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
7333 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
7335 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
7338 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
7339 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
7340 DAG.getNode(X86ISD::GlobalBaseReg,
7341 DebugLoc(), PtrVT), InFlag);
7342 InFlag = Chain.getValue(1);
7344 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
7347 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
7349 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
7351 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
7352 X86::RAX, X86II::MO_TLSGD);
7355 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
7356 // "local exec" model.
7357 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
7358 const EVT PtrVT, TLSModel::Model model,
7360 DebugLoc dl = GA->getDebugLoc();
7362 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
7363 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
7364 is64Bit ? 257 : 256));
7366 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
7367 DAG.getIntPtrConstant(0),
7368 MachinePointerInfo(Ptr), false, false, 0);
7370 unsigned char OperandFlags = 0;
7371 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
7373 unsigned WrapperKind = X86ISD::Wrapper;
7374 if (model == TLSModel::LocalExec) {
7375 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
7376 } else if (is64Bit) {
7377 assert(model == TLSModel::InitialExec);
7378 OperandFlags = X86II::MO_GOTTPOFF;
7379 WrapperKind = X86ISD::WrapperRIP;
7381 assert(model == TLSModel::InitialExec);
7382 OperandFlags = X86II::MO_INDNTPOFF;
7385 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
7387 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
7388 GA->getValueType(0),
7389 GA->getOffset(), OperandFlags);
7390 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
7392 if (model == TLSModel::InitialExec)
7393 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
7394 MachinePointerInfo::getGOT(), false, false, 0);
7396 // The address of the thread local variable is the add of the thread
7397 // pointer with the offset of the variable.
7398 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
7402 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
7404 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
7405 const GlobalValue *GV = GA->getGlobal();
7407 if (Subtarget->isTargetELF()) {
7408 // TODO: implement the "local dynamic" model
7409 // TODO: implement the "initial exec"model for pic executables
7411 // If GV is an alias then use the aliasee for determining
7412 // thread-localness.
7413 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
7414 GV = GA->resolveAliasedGlobal(false);
7416 TLSModel::Model model
7417 = getTLSModel(GV, getTargetMachine().getRelocationModel());
7420 case TLSModel::GeneralDynamic:
7421 case TLSModel::LocalDynamic: // not implemented
7422 if (Subtarget->is64Bit())
7423 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
7424 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
7426 case TLSModel::InitialExec:
7427 case TLSModel::LocalExec:
7428 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
7429 Subtarget->is64Bit());
7431 } else if (Subtarget->isTargetDarwin()) {
7432 // Darwin only has one model of TLS. Lower to that.
7433 unsigned char OpFlag = 0;
7434 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
7435 X86ISD::WrapperRIP : X86ISD::Wrapper;
7437 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7439 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
7440 !Subtarget->is64Bit();
7442 OpFlag = X86II::MO_TLVP_PIC_BASE;
7444 OpFlag = X86II::MO_TLVP;
7445 DebugLoc DL = Op.getDebugLoc();
7446 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
7447 GA->getValueType(0),
7448 GA->getOffset(), OpFlag);
7449 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7451 // With PIC32, the address is actually $g + Offset.
7453 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7454 DAG.getNode(X86ISD::GlobalBaseReg,
7455 DebugLoc(), getPointerTy()),
7458 // Lowering the machine isd will make sure everything is in the right
7460 SDValue Chain = DAG.getEntryNode();
7461 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
7462 SDValue Args[] = { Chain, Offset };
7463 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args, 2);
7465 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
7466 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7467 MFI->setAdjustsStack(true);
7469 // And our return value (tls address) is in the standard call return value
7471 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
7472 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
7476 "TLS not implemented for this target.");
7478 llvm_unreachable("Unreachable");
7483 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values and
7484 /// take a 2 x i32 value to shift plus a shift amount.
7485 SDValue X86TargetLowering::LowerShiftParts(SDValue Op, SelectionDAG &DAG) const {
7486 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
7487 EVT VT = Op.getValueType();
7488 unsigned VTBits = VT.getSizeInBits();
7489 DebugLoc dl = Op.getDebugLoc();
7490 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
7491 SDValue ShOpLo = Op.getOperand(0);
7492 SDValue ShOpHi = Op.getOperand(1);
7493 SDValue ShAmt = Op.getOperand(2);
7494 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
7495 DAG.getConstant(VTBits - 1, MVT::i8))
7496 : DAG.getConstant(0, VT);
7499 if (Op.getOpcode() == ISD::SHL_PARTS) {
7500 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
7501 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
7503 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
7504 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
7507 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
7508 DAG.getConstant(VTBits, MVT::i8));
7509 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
7510 AndNode, DAG.getConstant(0, MVT::i8));
7513 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7514 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
7515 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
7517 if (Op.getOpcode() == ISD::SHL_PARTS) {
7518 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
7519 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
7521 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
7522 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
7525 SDValue Ops[2] = { Lo, Hi };
7526 return DAG.getMergeValues(Ops, 2, dl);
7529 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
7530 SelectionDAG &DAG) const {
7531 EVT SrcVT = Op.getOperand(0).getValueType();
7533 if (SrcVT.isVector())
7536 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
7537 "Unknown SINT_TO_FP to lower!");
7539 // These are really Legal; return the operand so the caller accepts it as
7541 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
7543 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
7544 Subtarget->is64Bit()) {
7548 DebugLoc dl = Op.getDebugLoc();
7549 unsigned Size = SrcVT.getSizeInBits()/8;
7550 MachineFunction &MF = DAG.getMachineFunction();
7551 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
7552 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7553 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7555 MachinePointerInfo::getFixedStack(SSFI),
7557 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
7560 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
7562 SelectionDAG &DAG) const {
7564 DebugLoc DL = Op.getDebugLoc();
7566 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
7568 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
7570 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
7572 unsigned ByteSize = SrcVT.getSizeInBits()/8;
7574 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
7575 MachineMemOperand *MMO;
7577 int SSFI = FI->getIndex();
7579 DAG.getMachineFunction()
7580 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7581 MachineMemOperand::MOLoad, ByteSize, ByteSize);
7583 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
7584 StackSlot = StackSlot.getOperand(1);
7586 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
7587 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
7589 Tys, Ops, array_lengthof(Ops),
7593 Chain = Result.getValue(1);
7594 SDValue InFlag = Result.getValue(2);
7596 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
7597 // shouldn't be necessary except that RFP cannot be live across
7598 // multiple blocks. When stackifier is fixed, they can be uncoupled.
7599 MachineFunction &MF = DAG.getMachineFunction();
7600 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
7601 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
7602 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7603 Tys = DAG.getVTList(MVT::Other);
7605 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
7607 MachineMemOperand *MMO =
7608 DAG.getMachineFunction()
7609 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7610 MachineMemOperand::MOStore, SSFISize, SSFISize);
7612 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
7613 Ops, array_lengthof(Ops),
7614 Op.getValueType(), MMO);
7615 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
7616 MachinePointerInfo::getFixedStack(SSFI),
7623 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
7624 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
7625 SelectionDAG &DAG) const {
7626 // This algorithm is not obvious. Here it is in C code, more or less:
7628 double uint64_to_double( uint32_t hi, uint32_t lo ) {
7629 static const __m128i exp = { 0x4330000045300000ULL, 0 };
7630 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
7632 // Copy ints to xmm registers.
7633 __m128i xh = _mm_cvtsi32_si128( hi );
7634 __m128i xl = _mm_cvtsi32_si128( lo );
7636 // Combine into low half of a single xmm register.
7637 __m128i x = _mm_unpacklo_epi32( xh, xl );
7641 // Merge in appropriate exponents to give the integer bits the right
7643 x = _mm_unpacklo_epi32( x, exp );
7645 // Subtract away the biases to deal with the IEEE-754 double precision
7647 d = _mm_sub_pd( (__m128d) x, bias );
7649 // All conversions up to here are exact. The correctly rounded result is
7650 // calculated using the current rounding mode using the following
7652 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
7653 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
7654 // store doesn't really need to be here (except
7655 // maybe to zero the other double)
7660 DebugLoc dl = Op.getDebugLoc();
7661 LLVMContext *Context = DAG.getContext();
7663 // Build some magic constants.
7664 std::vector<Constant*> CV0;
7665 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
7666 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
7667 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
7668 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
7669 Constant *C0 = ConstantVector::get(CV0);
7670 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
7672 std::vector<Constant*> CV1;
7674 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
7676 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
7677 Constant *C1 = ConstantVector::get(CV1);
7678 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
7680 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
7681 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7683 DAG.getIntPtrConstant(1)));
7684 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
7685 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7687 DAG.getIntPtrConstant(0)));
7688 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
7689 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
7690 MachinePointerInfo::getConstantPool(),
7692 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
7693 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck2);
7694 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
7695 MachinePointerInfo::getConstantPool(),
7697 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
7699 // Add the halves; easiest way is to swap them into another reg first.
7700 int ShufMask[2] = { 1, -1 };
7701 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
7702 DAG.getUNDEF(MVT::v2f64), ShufMask);
7703 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
7704 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
7705 DAG.getIntPtrConstant(0));
7708 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
7709 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
7710 SelectionDAG &DAG) const {
7711 DebugLoc dl = Op.getDebugLoc();
7712 // FP constant to bias correct the final result.
7713 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
7716 // Load the 32-bit value into an XMM register.
7717 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
7720 // Zero out the upper parts of the register.
7721 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget->hasSSE2(), DAG);
7723 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7724 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
7725 DAG.getIntPtrConstant(0));
7727 // Or the load with the bias.
7728 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
7729 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7730 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
7732 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7733 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
7734 MVT::v2f64, Bias)));
7735 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7736 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
7737 DAG.getIntPtrConstant(0));
7739 // Subtract the bias.
7740 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
7742 // Handle final rounding.
7743 EVT DestVT = Op.getValueType();
7745 if (DestVT.bitsLT(MVT::f64)) {
7746 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
7747 DAG.getIntPtrConstant(0));
7748 } else if (DestVT.bitsGT(MVT::f64)) {
7749 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
7752 // Handle final rounding.
7756 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
7757 SelectionDAG &DAG) const {
7758 SDValue N0 = Op.getOperand(0);
7759 DebugLoc dl = Op.getDebugLoc();
7761 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
7762 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
7763 // the optimization here.
7764 if (DAG.SignBitIsZero(N0))
7765 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
7767 EVT SrcVT = N0.getValueType();
7768 EVT DstVT = Op.getValueType();
7769 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
7770 return LowerUINT_TO_FP_i64(Op, DAG);
7771 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
7772 return LowerUINT_TO_FP_i32(Op, DAG);
7774 // Make a 64-bit buffer, and use it to build an FILD.
7775 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
7776 if (SrcVT == MVT::i32) {
7777 SDValue WordOff = DAG.getConstant(4, getPointerTy());
7778 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
7779 getPointerTy(), StackSlot, WordOff);
7780 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7781 StackSlot, MachinePointerInfo(),
7783 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
7784 OffsetSlot, MachinePointerInfo(),
7786 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
7790 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
7791 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7792 StackSlot, MachinePointerInfo(),
7794 // For i64 source, we need to add the appropriate power of 2 if the input
7795 // was negative. This is the same as the optimization in
7796 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
7797 // we must be careful to do the computation in x87 extended precision, not
7798 // in SSE. (The generic code can't know it's OK to do this, or how to.)
7799 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
7800 MachineMemOperand *MMO =
7801 DAG.getMachineFunction()
7802 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7803 MachineMemOperand::MOLoad, 8, 8);
7805 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
7806 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
7807 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops, 3,
7810 APInt FF(32, 0x5F800000ULL);
7812 // Check whether the sign bit is set.
7813 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
7814 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
7817 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
7818 SDValue FudgePtr = DAG.getConstantPool(
7819 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
7822 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
7823 SDValue Zero = DAG.getIntPtrConstant(0);
7824 SDValue Four = DAG.getIntPtrConstant(4);
7825 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
7827 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
7829 // Load the value out, extending it from f32 to f80.
7830 // FIXME: Avoid the extend by constructing the right constant pool?
7831 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
7832 FudgePtr, MachinePointerInfo::getConstantPool(),
7833 MVT::f32, false, false, 4);
7834 // Extend everything to 80 bits to force it to be done on x87.
7835 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
7836 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
7839 std::pair<SDValue,SDValue> X86TargetLowering::
7840 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
7841 DebugLoc DL = Op.getDebugLoc();
7843 EVT DstTy = Op.getValueType();
7846 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
7850 assert(DstTy.getSimpleVT() <= MVT::i64 &&
7851 DstTy.getSimpleVT() >= MVT::i16 &&
7852 "Unknown FP_TO_SINT to lower!");
7854 // These are really Legal.
7855 if (DstTy == MVT::i32 &&
7856 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
7857 return std::make_pair(SDValue(), SDValue());
7858 if (Subtarget->is64Bit() &&
7859 DstTy == MVT::i64 &&
7860 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
7861 return std::make_pair(SDValue(), SDValue());
7863 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
7865 MachineFunction &MF = DAG.getMachineFunction();
7866 unsigned MemSize = DstTy.getSizeInBits()/8;
7867 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
7868 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7873 switch (DstTy.getSimpleVT().SimpleTy) {
7874 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
7875 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
7876 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
7877 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
7880 SDValue Chain = DAG.getEntryNode();
7881 SDValue Value = Op.getOperand(0);
7882 EVT TheVT = Op.getOperand(0).getValueType();
7883 if (isScalarFPTypeInSSEReg(TheVT)) {
7884 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
7885 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
7886 MachinePointerInfo::getFixedStack(SSFI),
7888 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
7890 Chain, StackSlot, DAG.getValueType(TheVT)
7893 MachineMemOperand *MMO =
7894 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7895 MachineMemOperand::MOLoad, MemSize, MemSize);
7896 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, 3,
7898 Chain = Value.getValue(1);
7899 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
7900 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7903 MachineMemOperand *MMO =
7904 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7905 MachineMemOperand::MOStore, MemSize, MemSize);
7907 // Build the FP_TO_INT*_IN_MEM
7908 SDValue Ops[] = { Chain, Value, StackSlot };
7909 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
7910 Ops, 3, DstTy, MMO);
7912 return std::make_pair(FIST, StackSlot);
7915 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
7916 SelectionDAG &DAG) const {
7917 if (Op.getValueType().isVector())
7920 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
7921 SDValue FIST = Vals.first, StackSlot = Vals.second;
7922 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
7923 if (FIST.getNode() == 0) return Op;
7926 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
7927 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
7930 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
7931 SelectionDAG &DAG) const {
7932 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
7933 SDValue FIST = Vals.first, StackSlot = Vals.second;
7934 assert(FIST.getNode() && "Unexpected failure");
7937 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
7938 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
7941 SDValue X86TargetLowering::LowerFABS(SDValue Op,
7942 SelectionDAG &DAG) const {
7943 LLVMContext *Context = DAG.getContext();
7944 DebugLoc dl = Op.getDebugLoc();
7945 EVT VT = Op.getValueType();
7948 EltVT = VT.getVectorElementType();
7949 std::vector<Constant*> CV;
7950 if (EltVT == MVT::f64) {
7951 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
7955 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
7961 Constant *C = ConstantVector::get(CV);
7962 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7963 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7964 MachinePointerInfo::getConstantPool(),
7966 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
7969 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
7970 LLVMContext *Context = DAG.getContext();
7971 DebugLoc dl = Op.getDebugLoc();
7972 EVT VT = Op.getValueType();
7975 EltVT = VT.getVectorElementType();
7976 std::vector<Constant*> CV;
7977 if (EltVT == MVT::f64) {
7978 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
7982 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
7988 Constant *C = ConstantVector::get(CV);
7989 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7990 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7991 MachinePointerInfo::getConstantPool(),
7993 if (VT.isVector()) {
7994 return DAG.getNode(ISD::BITCAST, dl, VT,
7995 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
7996 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7998 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, Mask)));
8000 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
8004 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
8005 LLVMContext *Context = DAG.getContext();
8006 SDValue Op0 = Op.getOperand(0);
8007 SDValue Op1 = Op.getOperand(1);
8008 DebugLoc dl = Op.getDebugLoc();
8009 EVT VT = Op.getValueType();
8010 EVT SrcVT = Op1.getValueType();
8012 // If second operand is smaller, extend it first.
8013 if (SrcVT.bitsLT(VT)) {
8014 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
8017 // And if it is bigger, shrink it first.
8018 if (SrcVT.bitsGT(VT)) {
8019 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
8023 // At this point the operands and the result should have the same
8024 // type, and that won't be f80 since that is not custom lowered.
8026 // First get the sign bit of second operand.
8027 std::vector<Constant*> CV;
8028 if (SrcVT == MVT::f64) {
8029 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
8030 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
8032 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
8033 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
8034 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
8035 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
8037 Constant *C = ConstantVector::get(CV);
8038 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8039 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
8040 MachinePointerInfo::getConstantPool(),
8042 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
8044 // Shift sign bit right or left if the two operands have different types.
8045 if (SrcVT.bitsGT(VT)) {
8046 // Op0 is MVT::f32, Op1 is MVT::f64.
8047 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
8048 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
8049 DAG.getConstant(32, MVT::i32));
8050 SignBit = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, SignBit);
8051 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
8052 DAG.getIntPtrConstant(0));
8055 // Clear first operand sign bit.
8057 if (VT == MVT::f64) {
8058 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
8059 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
8061 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
8062 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
8063 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
8064 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
8066 C = ConstantVector::get(CV);
8067 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8068 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8069 MachinePointerInfo::getConstantPool(),
8071 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
8073 // Or the value with the sign bit.
8074 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
8077 SDValue X86TargetLowering::LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) const {
8078 SDValue N0 = Op.getOperand(0);
8079 DebugLoc dl = Op.getDebugLoc();
8080 EVT VT = Op.getValueType();
8082 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
8083 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
8084 DAG.getConstant(1, VT));
8085 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT));
8088 /// Emit nodes that will be selected as "test Op0,Op0", or something
8090 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
8091 SelectionDAG &DAG) const {
8092 DebugLoc dl = Op.getDebugLoc();
8094 // CF and OF aren't always set the way we want. Determine which
8095 // of these we need.
8096 bool NeedCF = false;
8097 bool NeedOF = false;
8100 case X86::COND_A: case X86::COND_AE:
8101 case X86::COND_B: case X86::COND_BE:
8104 case X86::COND_G: case X86::COND_GE:
8105 case X86::COND_L: case X86::COND_LE:
8106 case X86::COND_O: case X86::COND_NO:
8111 // See if we can use the EFLAGS value from the operand instead of
8112 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
8113 // we prove that the arithmetic won't overflow, we can't use OF or CF.
8114 if (Op.getResNo() != 0 || NeedOF || NeedCF)
8115 // Emit a CMP with 0, which is the TEST pattern.
8116 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
8117 DAG.getConstant(0, Op.getValueType()));
8119 unsigned Opcode = 0;
8120 unsigned NumOperands = 0;
8121 switch (Op.getNode()->getOpcode()) {
8123 // Due to an isel shortcoming, be conservative if this add is likely to be
8124 // selected as part of a load-modify-store instruction. When the root node
8125 // in a match is a store, isel doesn't know how to remap non-chain non-flag
8126 // uses of other nodes in the match, such as the ADD in this case. This
8127 // leads to the ADD being left around and reselected, with the result being
8128 // two adds in the output. Alas, even if none our users are stores, that
8129 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
8130 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
8131 // climbing the DAG back to the root, and it doesn't seem to be worth the
8133 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
8134 UE = Op.getNode()->use_end(); UI != UE; ++UI)
8135 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
8138 if (ConstantSDNode *C =
8139 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
8140 // An add of one will be selected as an INC.
8141 if (C->getAPIntValue() == 1) {
8142 Opcode = X86ISD::INC;
8147 // An add of negative one (subtract of one) will be selected as a DEC.
8148 if (C->getAPIntValue().isAllOnesValue()) {
8149 Opcode = X86ISD::DEC;
8155 // Otherwise use a regular EFLAGS-setting add.
8156 Opcode = X86ISD::ADD;
8160 // If the primary and result isn't used, don't bother using X86ISD::AND,
8161 // because a TEST instruction will be better.
8162 bool NonFlagUse = false;
8163 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
8164 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
8166 unsigned UOpNo = UI.getOperandNo();
8167 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
8168 // Look pass truncate.
8169 UOpNo = User->use_begin().getOperandNo();
8170 User = *User->use_begin();
8173 if (User->getOpcode() != ISD::BRCOND &&
8174 User->getOpcode() != ISD::SETCC &&
8175 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
8188 // Due to the ISEL shortcoming noted above, be conservative if this op is
8189 // likely to be selected as part of a load-modify-store instruction.
8190 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
8191 UE = Op.getNode()->use_end(); UI != UE; ++UI)
8192 if (UI->getOpcode() == ISD::STORE)
8195 // Otherwise use a regular EFLAGS-setting instruction.
8196 switch (Op.getNode()->getOpcode()) {
8197 default: llvm_unreachable("unexpected operator!");
8198 case ISD::SUB: Opcode = X86ISD::SUB; break;
8199 case ISD::OR: Opcode = X86ISD::OR; break;
8200 case ISD::XOR: Opcode = X86ISD::XOR; break;
8201 case ISD::AND: Opcode = X86ISD::AND; break;
8213 return SDValue(Op.getNode(), 1);
8220 // Emit a CMP with 0, which is the TEST pattern.
8221 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
8222 DAG.getConstant(0, Op.getValueType()));
8224 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
8225 SmallVector<SDValue, 4> Ops;
8226 for (unsigned i = 0; i != NumOperands; ++i)
8227 Ops.push_back(Op.getOperand(i));
8229 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
8230 DAG.ReplaceAllUsesWith(Op, New);
8231 return SDValue(New.getNode(), 1);
8234 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
8236 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
8237 SelectionDAG &DAG) const {
8238 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
8239 if (C->getAPIntValue() == 0)
8240 return EmitTest(Op0, X86CC, DAG);
8242 DebugLoc dl = Op0.getDebugLoc();
8243 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
8246 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
8247 /// if it's possible.
8248 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
8249 DebugLoc dl, SelectionDAG &DAG) const {
8250 SDValue Op0 = And.getOperand(0);
8251 SDValue Op1 = And.getOperand(1);
8252 if (Op0.getOpcode() == ISD::TRUNCATE)
8253 Op0 = Op0.getOperand(0);
8254 if (Op1.getOpcode() == ISD::TRUNCATE)
8255 Op1 = Op1.getOperand(0);
8258 if (Op1.getOpcode() == ISD::SHL)
8259 std::swap(Op0, Op1);
8260 if (Op0.getOpcode() == ISD::SHL) {
8261 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
8262 if (And00C->getZExtValue() == 1) {
8263 // If we looked past a truncate, check that it's only truncating away
8265 unsigned BitWidth = Op0.getValueSizeInBits();
8266 unsigned AndBitWidth = And.getValueSizeInBits();
8267 if (BitWidth > AndBitWidth) {
8268 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
8269 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
8270 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
8274 RHS = Op0.getOperand(1);
8276 } else if (Op1.getOpcode() == ISD::Constant) {
8277 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
8278 SDValue AndLHS = Op0;
8279 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
8280 LHS = AndLHS.getOperand(0);
8281 RHS = AndLHS.getOperand(1);
8285 if (LHS.getNode()) {
8286 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
8287 // instruction. Since the shift amount is in-range-or-undefined, we know
8288 // that doing a bittest on the i32 value is ok. We extend to i32 because
8289 // the encoding for the i16 version is larger than the i32 version.
8290 // Also promote i16 to i32 for performance / code size reason.
8291 if (LHS.getValueType() == MVT::i8 ||
8292 LHS.getValueType() == MVT::i16)
8293 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
8295 // If the operand types disagree, extend the shift amount to match. Since
8296 // BT ignores high bits (like shifts) we can use anyextend.
8297 if (LHS.getValueType() != RHS.getValueType())
8298 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
8300 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
8301 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
8302 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
8303 DAG.getConstant(Cond, MVT::i8), BT);
8309 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
8310 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
8311 SDValue Op0 = Op.getOperand(0);
8312 SDValue Op1 = Op.getOperand(1);
8313 DebugLoc dl = Op.getDebugLoc();
8314 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
8316 // Optimize to BT if possible.
8317 // Lower (X & (1 << N)) == 0 to BT(X, N).
8318 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
8319 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
8320 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
8321 Op1.getOpcode() == ISD::Constant &&
8322 cast<ConstantSDNode>(Op1)->isNullValue() &&
8323 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
8324 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
8325 if (NewSetCC.getNode())
8329 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
8331 if (Op1.getOpcode() == ISD::Constant &&
8332 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
8333 cast<ConstantSDNode>(Op1)->isNullValue()) &&
8334 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
8336 // If the input is a setcc, then reuse the input setcc or use a new one with
8337 // the inverted condition.
8338 if (Op0.getOpcode() == X86ISD::SETCC) {
8339 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
8340 bool Invert = (CC == ISD::SETNE) ^
8341 cast<ConstantSDNode>(Op1)->isNullValue();
8342 if (!Invert) return Op0;
8344 CCode = X86::GetOppositeBranchCondition(CCode);
8345 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
8346 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
8350 bool isFP = Op1.getValueType().isFloatingPoint();
8351 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
8352 if (X86CC == X86::COND_INVALID)
8355 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
8356 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
8357 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
8360 // Lower256IntVETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
8361 // ones, and then concatenate the result back.
8362 static SDValue Lower256IntVETCC(SDValue Op, SelectionDAG &DAG) {
8363 EVT VT = Op.getValueType();
8365 assert(VT.getSizeInBits() == 256 && Op.getOpcode() == ISD::VSETCC &&
8366 "Unsupported value type for operation");
8368 int NumElems = VT.getVectorNumElements();
8369 DebugLoc dl = Op.getDebugLoc();
8370 SDValue CC = Op.getOperand(2);
8371 SDValue Idx0 = DAG.getConstant(0, MVT::i32);
8372 SDValue Idx1 = DAG.getConstant(NumElems/2, MVT::i32);
8374 // Extract the LHS vectors
8375 SDValue LHS = Op.getOperand(0);
8376 SDValue LHS1 = Extract128BitVector(LHS, Idx0, DAG, dl);
8377 SDValue LHS2 = Extract128BitVector(LHS, Idx1, DAG, dl);
8379 // Extract the RHS vectors
8380 SDValue RHS = Op.getOperand(1);
8381 SDValue RHS1 = Extract128BitVector(RHS, Idx0, DAG, dl);
8382 SDValue RHS2 = Extract128BitVector(RHS, Idx1, DAG, dl);
8384 // Issue the operation on the smaller types and concatenate the result back
8385 MVT EltVT = VT.getVectorElementType().getSimpleVT();
8386 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
8387 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
8388 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
8389 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
8393 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
8395 SDValue Op0 = Op.getOperand(0);
8396 SDValue Op1 = Op.getOperand(1);
8397 SDValue CC = Op.getOperand(2);
8398 EVT VT = Op.getValueType();
8399 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
8400 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
8401 DebugLoc dl = Op.getDebugLoc();
8405 EVT EltVT = Op0.getValueType().getVectorElementType();
8406 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
8408 unsigned Opc = EltVT == MVT::f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
8411 switch (SetCCOpcode) {
8414 case ISD::SETEQ: SSECC = 0; break;
8416 case ISD::SETGT: Swap = true; // Fallthrough
8418 case ISD::SETOLT: SSECC = 1; break;
8420 case ISD::SETGE: Swap = true; // Fallthrough
8422 case ISD::SETOLE: SSECC = 2; break;
8423 case ISD::SETUO: SSECC = 3; break;
8425 case ISD::SETNE: SSECC = 4; break;
8426 case ISD::SETULE: Swap = true;
8427 case ISD::SETUGE: SSECC = 5; break;
8428 case ISD::SETULT: Swap = true;
8429 case ISD::SETUGT: SSECC = 6; break;
8430 case ISD::SETO: SSECC = 7; break;
8433 std::swap(Op0, Op1);
8435 // In the two special cases we can't handle, emit two comparisons.
8437 if (SetCCOpcode == ISD::SETUEQ) {
8439 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
8440 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
8441 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
8443 else if (SetCCOpcode == ISD::SETONE) {
8445 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
8446 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
8447 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
8449 llvm_unreachable("Illegal FP comparison");
8451 // Handle all other FP comparisons here.
8452 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
8455 // Break 256-bit integer vector compare into smaller ones.
8456 if (!isFP && VT.getSizeInBits() == 256)
8457 return Lower256IntVETCC(Op, DAG);
8459 // We are handling one of the integer comparisons here. Since SSE only has
8460 // GT and EQ comparisons for integer, swapping operands and multiple
8461 // operations may be required for some comparisons.
8462 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
8463 bool Swap = false, Invert = false, FlipSigns = false;
8465 switch (VT.getSimpleVT().SimpleTy) {
8467 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
8468 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
8469 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
8470 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
8473 switch (SetCCOpcode) {
8475 case ISD::SETNE: Invert = true;
8476 case ISD::SETEQ: Opc = EQOpc; break;
8477 case ISD::SETLT: Swap = true;
8478 case ISD::SETGT: Opc = GTOpc; break;
8479 case ISD::SETGE: Swap = true;
8480 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
8481 case ISD::SETULT: Swap = true;
8482 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
8483 case ISD::SETUGE: Swap = true;
8484 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
8487 std::swap(Op0, Op1);
8489 // Since SSE has no unsigned integer comparisons, we need to flip the sign
8490 // bits of the inputs before performing those operations.
8492 EVT EltVT = VT.getVectorElementType();
8493 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
8495 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
8496 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
8498 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
8499 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
8502 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
8504 // If the logical-not of the result is required, perform that now.
8506 Result = DAG.getNOT(dl, Result, VT);
8511 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
8512 static bool isX86LogicalCmp(SDValue Op) {
8513 unsigned Opc = Op.getNode()->getOpcode();
8514 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
8516 if (Op.getResNo() == 1 &&
8517 (Opc == X86ISD::ADD ||
8518 Opc == X86ISD::SUB ||
8519 Opc == X86ISD::ADC ||
8520 Opc == X86ISD::SBB ||
8521 Opc == X86ISD::SMUL ||
8522 Opc == X86ISD::UMUL ||
8523 Opc == X86ISD::INC ||
8524 Opc == X86ISD::DEC ||
8525 Opc == X86ISD::OR ||
8526 Opc == X86ISD::XOR ||
8527 Opc == X86ISD::AND))
8530 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
8536 static bool isZero(SDValue V) {
8537 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
8538 return C && C->isNullValue();
8541 static bool isAllOnes(SDValue V) {
8542 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
8543 return C && C->isAllOnesValue();
8546 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
8547 bool addTest = true;
8548 SDValue Cond = Op.getOperand(0);
8549 SDValue Op1 = Op.getOperand(1);
8550 SDValue Op2 = Op.getOperand(2);
8551 DebugLoc DL = Op.getDebugLoc();
8554 if (Cond.getOpcode() == ISD::SETCC) {
8555 SDValue NewCond = LowerSETCC(Cond, DAG);
8556 if (NewCond.getNode())
8560 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
8561 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
8562 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
8563 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
8564 if (Cond.getOpcode() == X86ISD::SETCC &&
8565 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
8566 isZero(Cond.getOperand(1).getOperand(1))) {
8567 SDValue Cmp = Cond.getOperand(1);
8569 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
8571 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
8572 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
8573 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
8575 SDValue CmpOp0 = Cmp.getOperand(0);
8576 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
8577 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
8579 SDValue Res = // Res = 0 or -1.
8580 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
8581 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
8583 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
8584 Res = DAG.getNOT(DL, Res, Res.getValueType());
8586 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
8587 if (N2C == 0 || !N2C->isNullValue())
8588 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
8593 // Look past (and (setcc_carry (cmp ...)), 1).
8594 if (Cond.getOpcode() == ISD::AND &&
8595 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
8596 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
8597 if (C && C->getAPIntValue() == 1)
8598 Cond = Cond.getOperand(0);
8601 // If condition flag is set by a X86ISD::CMP, then use it as the condition
8602 // setting operand in place of the X86ISD::SETCC.
8603 if (Cond.getOpcode() == X86ISD::SETCC ||
8604 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
8605 CC = Cond.getOperand(0);
8607 SDValue Cmp = Cond.getOperand(1);
8608 unsigned Opc = Cmp.getOpcode();
8609 EVT VT = Op.getValueType();
8611 bool IllegalFPCMov = false;
8612 if (VT.isFloatingPoint() && !VT.isVector() &&
8613 !isScalarFPTypeInSSEReg(VT)) // FPStack?
8614 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
8616 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
8617 Opc == X86ISD::BT) { // FIXME
8624 // Look pass the truncate.
8625 if (Cond.getOpcode() == ISD::TRUNCATE)
8626 Cond = Cond.getOperand(0);
8628 // We know the result of AND is compared against zero. Try to match
8630 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
8631 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
8632 if (NewSetCC.getNode()) {
8633 CC = NewSetCC.getOperand(0);
8634 Cond = NewSetCC.getOperand(1);
8641 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
8642 Cond = EmitTest(Cond, X86::COND_NE, DAG);
8645 // a < b ? -1 : 0 -> RES = ~setcc_carry
8646 // a < b ? 0 : -1 -> RES = setcc_carry
8647 // a >= b ? -1 : 0 -> RES = setcc_carry
8648 // a >= b ? 0 : -1 -> RES = ~setcc_carry
8649 if (Cond.getOpcode() == X86ISD::CMP) {
8650 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
8652 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
8653 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
8654 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
8655 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
8656 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
8657 return DAG.getNOT(DL, Res, Res.getValueType());
8662 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
8663 // condition is true.
8664 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
8665 SDValue Ops[] = { Op2, Op1, CC, Cond };
8666 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops, array_lengthof(Ops));
8669 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
8670 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
8671 // from the AND / OR.
8672 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
8673 Opc = Op.getOpcode();
8674 if (Opc != ISD::OR && Opc != ISD::AND)
8676 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
8677 Op.getOperand(0).hasOneUse() &&
8678 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
8679 Op.getOperand(1).hasOneUse());
8682 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
8683 // 1 and that the SETCC node has a single use.
8684 static bool isXor1OfSetCC(SDValue Op) {
8685 if (Op.getOpcode() != ISD::XOR)
8687 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
8688 if (N1C && N1C->getAPIntValue() == 1) {
8689 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
8690 Op.getOperand(0).hasOneUse();
8695 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
8696 bool addTest = true;
8697 SDValue Chain = Op.getOperand(0);
8698 SDValue Cond = Op.getOperand(1);
8699 SDValue Dest = Op.getOperand(2);
8700 DebugLoc dl = Op.getDebugLoc();
8703 if (Cond.getOpcode() == ISD::SETCC) {
8704 SDValue NewCond = LowerSETCC(Cond, DAG);
8705 if (NewCond.getNode())
8709 // FIXME: LowerXALUO doesn't handle these!!
8710 else if (Cond.getOpcode() == X86ISD::ADD ||
8711 Cond.getOpcode() == X86ISD::SUB ||
8712 Cond.getOpcode() == X86ISD::SMUL ||
8713 Cond.getOpcode() == X86ISD::UMUL)
8714 Cond = LowerXALUO(Cond, DAG);
8717 // Look pass (and (setcc_carry (cmp ...)), 1).
8718 if (Cond.getOpcode() == ISD::AND &&
8719 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
8720 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
8721 if (C && C->getAPIntValue() == 1)
8722 Cond = Cond.getOperand(0);
8725 // If condition flag is set by a X86ISD::CMP, then use it as the condition
8726 // setting operand in place of the X86ISD::SETCC.
8727 if (Cond.getOpcode() == X86ISD::SETCC ||
8728 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
8729 CC = Cond.getOperand(0);
8731 SDValue Cmp = Cond.getOperand(1);
8732 unsigned Opc = Cmp.getOpcode();
8733 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
8734 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
8738 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
8742 // These can only come from an arithmetic instruction with overflow,
8743 // e.g. SADDO, UADDO.
8744 Cond = Cond.getNode()->getOperand(1);
8751 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
8752 SDValue Cmp = Cond.getOperand(0).getOperand(1);
8753 if (CondOpc == ISD::OR) {
8754 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
8755 // two branches instead of an explicit OR instruction with a
8757 if (Cmp == Cond.getOperand(1).getOperand(1) &&
8758 isX86LogicalCmp(Cmp)) {
8759 CC = Cond.getOperand(0).getOperand(0);
8760 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8761 Chain, Dest, CC, Cmp);
8762 CC = Cond.getOperand(1).getOperand(0);
8766 } else { // ISD::AND
8767 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
8768 // two branches instead of an explicit AND instruction with a
8769 // separate test. However, we only do this if this block doesn't
8770 // have a fall-through edge, because this requires an explicit
8771 // jmp when the condition is false.
8772 if (Cmp == Cond.getOperand(1).getOperand(1) &&
8773 isX86LogicalCmp(Cmp) &&
8774 Op.getNode()->hasOneUse()) {
8775 X86::CondCode CCode =
8776 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
8777 CCode = X86::GetOppositeBranchCondition(CCode);
8778 CC = DAG.getConstant(CCode, MVT::i8);
8779 SDNode *User = *Op.getNode()->use_begin();
8780 // Look for an unconditional branch following this conditional branch.
8781 // We need this because we need to reverse the successors in order
8782 // to implement FCMP_OEQ.
8783 if (User->getOpcode() == ISD::BR) {
8784 SDValue FalseBB = User->getOperand(1);
8786 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
8787 assert(NewBR == User);
8791 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8792 Chain, Dest, CC, Cmp);
8793 X86::CondCode CCode =
8794 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
8795 CCode = X86::GetOppositeBranchCondition(CCode);
8796 CC = DAG.getConstant(CCode, MVT::i8);
8802 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
8803 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
8804 // It should be transformed during dag combiner except when the condition
8805 // is set by a arithmetics with overflow node.
8806 X86::CondCode CCode =
8807 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
8808 CCode = X86::GetOppositeBranchCondition(CCode);
8809 CC = DAG.getConstant(CCode, MVT::i8);
8810 Cond = Cond.getOperand(0).getOperand(1);
8816 // Look pass the truncate.
8817 if (Cond.getOpcode() == ISD::TRUNCATE)
8818 Cond = Cond.getOperand(0);
8820 // We know the result of AND is compared against zero. Try to match
8822 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
8823 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
8824 if (NewSetCC.getNode()) {
8825 CC = NewSetCC.getOperand(0);
8826 Cond = NewSetCC.getOperand(1);
8833 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
8834 Cond = EmitTest(Cond, X86::COND_NE, DAG);
8836 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8837 Chain, Dest, CC, Cond);
8841 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
8842 // Calls to _alloca is needed to probe the stack when allocating more than 4k
8843 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
8844 // that the guard pages used by the OS virtual memory manager are allocated in
8845 // correct sequence.
8847 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
8848 SelectionDAG &DAG) const {
8849 assert((Subtarget->isTargetCygMing() || Subtarget->isTargetWindows() ||
8850 EnableSegmentedStacks) &&
8851 "This should be used only on Windows targets or when segmented stacks "
8853 assert(!Subtarget->isTargetEnvMacho());
8854 DebugLoc dl = Op.getDebugLoc();
8857 SDValue Chain = Op.getOperand(0);
8858 SDValue Size = Op.getOperand(1);
8859 // FIXME: Ensure alignment here
8861 bool Is64Bit = Subtarget->is64Bit();
8862 EVT SPTy = Is64Bit ? MVT::i64 : MVT::i32;
8864 if (EnableSegmentedStacks) {
8865 MachineFunction &MF = DAG.getMachineFunction();
8866 MachineRegisterInfo &MRI = MF.getRegInfo();
8869 // The 64 bit implementation of segmented stacks needs to clobber both r10
8870 // r11. This makes it impossible to use it along with nested paramenters.
8871 const Function *F = MF.getFunction();
8873 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
8875 if (I->hasNestAttr())
8876 report_fatal_error("Cannot use segmented stacks with functions that "
8877 "have nested arguments.");
8880 const TargetRegisterClass *AddrRegClass =
8881 getRegClassFor(Subtarget->is64Bit() ? MVT::i64:MVT::i32);
8882 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
8883 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
8884 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
8885 DAG.getRegister(Vreg, SPTy));
8886 SDValue Ops1[2] = { Value, Chain };
8887 return DAG.getMergeValues(Ops1, 2, dl);
8890 unsigned Reg = (Subtarget->is64Bit() ? X86::RAX : X86::EAX);
8892 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
8893 Flag = Chain.getValue(1);
8894 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
8896 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
8897 Flag = Chain.getValue(1);
8899 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
8901 SDValue Ops1[2] = { Chain.getValue(0), Chain };
8902 return DAG.getMergeValues(Ops1, 2, dl);
8906 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
8907 MachineFunction &MF = DAG.getMachineFunction();
8908 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
8910 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
8911 DebugLoc DL = Op.getDebugLoc();
8913 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
8914 // vastart just stores the address of the VarArgsFrameIndex slot into the
8915 // memory location argument.
8916 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
8918 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
8919 MachinePointerInfo(SV), false, false, 0);
8923 // gp_offset (0 - 6 * 8)
8924 // fp_offset (48 - 48 + 8 * 16)
8925 // overflow_arg_area (point to parameters coming in memory).
8927 SmallVector<SDValue, 8> MemOps;
8928 SDValue FIN = Op.getOperand(1);
8930 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
8931 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
8933 FIN, MachinePointerInfo(SV), false, false, 0);
8934 MemOps.push_back(Store);
8937 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8938 FIN, DAG.getIntPtrConstant(4));
8939 Store = DAG.getStore(Op.getOperand(0), DL,
8940 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
8942 FIN, MachinePointerInfo(SV, 4), false, false, 0);
8943 MemOps.push_back(Store);
8945 // Store ptr to overflow_arg_area
8946 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8947 FIN, DAG.getIntPtrConstant(4));
8948 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
8950 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
8951 MachinePointerInfo(SV, 8),
8953 MemOps.push_back(Store);
8955 // Store ptr to reg_save_area.
8956 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8957 FIN, DAG.getIntPtrConstant(8));
8958 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
8960 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
8961 MachinePointerInfo(SV, 16), false, false, 0);
8962 MemOps.push_back(Store);
8963 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
8964 &MemOps[0], MemOps.size());
8967 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
8968 assert(Subtarget->is64Bit() &&
8969 "LowerVAARG only handles 64-bit va_arg!");
8970 assert((Subtarget->isTargetLinux() ||
8971 Subtarget->isTargetDarwin()) &&
8972 "Unhandled target in LowerVAARG");
8973 assert(Op.getNode()->getNumOperands() == 4);
8974 SDValue Chain = Op.getOperand(0);
8975 SDValue SrcPtr = Op.getOperand(1);
8976 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
8977 unsigned Align = Op.getConstantOperandVal(3);
8978 DebugLoc dl = Op.getDebugLoc();
8980 EVT ArgVT = Op.getNode()->getValueType(0);
8981 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
8982 uint32_t ArgSize = getTargetData()->getTypeAllocSize(ArgTy);
8985 // Decide which area this value should be read from.
8986 // TODO: Implement the AMD64 ABI in its entirety. This simple
8987 // selection mechanism works only for the basic types.
8988 if (ArgVT == MVT::f80) {
8989 llvm_unreachable("va_arg for f80 not yet implemented");
8990 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
8991 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
8992 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
8993 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
8995 llvm_unreachable("Unhandled argument type in LowerVAARG");
8999 // Sanity Check: Make sure using fp_offset makes sense.
9000 assert(!UseSoftFloat &&
9001 !(DAG.getMachineFunction()
9002 .getFunction()->hasFnAttr(Attribute::NoImplicitFloat)) &&
9003 Subtarget->hasXMM());
9006 // Insert VAARG_64 node into the DAG
9007 // VAARG_64 returns two values: Variable Argument Address, Chain
9008 SmallVector<SDValue, 11> InstOps;
9009 InstOps.push_back(Chain);
9010 InstOps.push_back(SrcPtr);
9011 InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
9012 InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
9013 InstOps.push_back(DAG.getConstant(Align, MVT::i32));
9014 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
9015 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
9016 VTs, &InstOps[0], InstOps.size(),
9018 MachinePointerInfo(SV),
9023 Chain = VAARG.getValue(1);
9025 // Load the next argument and return it
9026 return DAG.getLoad(ArgVT, dl,
9029 MachinePointerInfo(),
9033 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
9034 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
9035 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
9036 SDValue Chain = Op.getOperand(0);
9037 SDValue DstPtr = Op.getOperand(1);
9038 SDValue SrcPtr = Op.getOperand(2);
9039 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
9040 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
9041 DebugLoc DL = Op.getDebugLoc();
9043 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
9044 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
9046 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
9050 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
9051 DebugLoc dl = Op.getDebugLoc();
9052 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9054 default: return SDValue(); // Don't custom lower most intrinsics.
9055 // Comparison intrinsics.
9056 case Intrinsic::x86_sse_comieq_ss:
9057 case Intrinsic::x86_sse_comilt_ss:
9058 case Intrinsic::x86_sse_comile_ss:
9059 case Intrinsic::x86_sse_comigt_ss:
9060 case Intrinsic::x86_sse_comige_ss:
9061 case Intrinsic::x86_sse_comineq_ss:
9062 case Intrinsic::x86_sse_ucomieq_ss:
9063 case Intrinsic::x86_sse_ucomilt_ss:
9064 case Intrinsic::x86_sse_ucomile_ss:
9065 case Intrinsic::x86_sse_ucomigt_ss:
9066 case Intrinsic::x86_sse_ucomige_ss:
9067 case Intrinsic::x86_sse_ucomineq_ss:
9068 case Intrinsic::x86_sse2_comieq_sd:
9069 case Intrinsic::x86_sse2_comilt_sd:
9070 case Intrinsic::x86_sse2_comile_sd:
9071 case Intrinsic::x86_sse2_comigt_sd:
9072 case Intrinsic::x86_sse2_comige_sd:
9073 case Intrinsic::x86_sse2_comineq_sd:
9074 case Intrinsic::x86_sse2_ucomieq_sd:
9075 case Intrinsic::x86_sse2_ucomilt_sd:
9076 case Intrinsic::x86_sse2_ucomile_sd:
9077 case Intrinsic::x86_sse2_ucomigt_sd:
9078 case Intrinsic::x86_sse2_ucomige_sd:
9079 case Intrinsic::x86_sse2_ucomineq_sd: {
9081 ISD::CondCode CC = ISD::SETCC_INVALID;
9084 case Intrinsic::x86_sse_comieq_ss:
9085 case Intrinsic::x86_sse2_comieq_sd:
9089 case Intrinsic::x86_sse_comilt_ss:
9090 case Intrinsic::x86_sse2_comilt_sd:
9094 case Intrinsic::x86_sse_comile_ss:
9095 case Intrinsic::x86_sse2_comile_sd:
9099 case Intrinsic::x86_sse_comigt_ss:
9100 case Intrinsic::x86_sse2_comigt_sd:
9104 case Intrinsic::x86_sse_comige_ss:
9105 case Intrinsic::x86_sse2_comige_sd:
9109 case Intrinsic::x86_sse_comineq_ss:
9110 case Intrinsic::x86_sse2_comineq_sd:
9114 case Intrinsic::x86_sse_ucomieq_ss:
9115 case Intrinsic::x86_sse2_ucomieq_sd:
9116 Opc = X86ISD::UCOMI;
9119 case Intrinsic::x86_sse_ucomilt_ss:
9120 case Intrinsic::x86_sse2_ucomilt_sd:
9121 Opc = X86ISD::UCOMI;
9124 case Intrinsic::x86_sse_ucomile_ss:
9125 case Intrinsic::x86_sse2_ucomile_sd:
9126 Opc = X86ISD::UCOMI;
9129 case Intrinsic::x86_sse_ucomigt_ss:
9130 case Intrinsic::x86_sse2_ucomigt_sd:
9131 Opc = X86ISD::UCOMI;
9134 case Intrinsic::x86_sse_ucomige_ss:
9135 case Intrinsic::x86_sse2_ucomige_sd:
9136 Opc = X86ISD::UCOMI;
9139 case Intrinsic::x86_sse_ucomineq_ss:
9140 case Intrinsic::x86_sse2_ucomineq_sd:
9141 Opc = X86ISD::UCOMI;
9146 SDValue LHS = Op.getOperand(1);
9147 SDValue RHS = Op.getOperand(2);
9148 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
9149 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
9150 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
9151 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
9152 DAG.getConstant(X86CC, MVT::i8), Cond);
9153 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
9155 // ptest and testp intrinsics. The intrinsic these come from are designed to
9156 // return an integer value, not just an instruction so lower it to the ptest
9157 // or testp pattern and a setcc for the result.
9158 case Intrinsic::x86_sse41_ptestz:
9159 case Intrinsic::x86_sse41_ptestc:
9160 case Intrinsic::x86_sse41_ptestnzc:
9161 case Intrinsic::x86_avx_ptestz_256:
9162 case Intrinsic::x86_avx_ptestc_256:
9163 case Intrinsic::x86_avx_ptestnzc_256:
9164 case Intrinsic::x86_avx_vtestz_ps:
9165 case Intrinsic::x86_avx_vtestc_ps:
9166 case Intrinsic::x86_avx_vtestnzc_ps:
9167 case Intrinsic::x86_avx_vtestz_pd:
9168 case Intrinsic::x86_avx_vtestc_pd:
9169 case Intrinsic::x86_avx_vtestnzc_pd:
9170 case Intrinsic::x86_avx_vtestz_ps_256:
9171 case Intrinsic::x86_avx_vtestc_ps_256:
9172 case Intrinsic::x86_avx_vtestnzc_ps_256:
9173 case Intrinsic::x86_avx_vtestz_pd_256:
9174 case Intrinsic::x86_avx_vtestc_pd_256:
9175 case Intrinsic::x86_avx_vtestnzc_pd_256: {
9176 bool IsTestPacked = false;
9179 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
9180 case Intrinsic::x86_avx_vtestz_ps:
9181 case Intrinsic::x86_avx_vtestz_pd:
9182 case Intrinsic::x86_avx_vtestz_ps_256:
9183 case Intrinsic::x86_avx_vtestz_pd_256:
9184 IsTestPacked = true; // Fallthrough
9185 case Intrinsic::x86_sse41_ptestz:
9186 case Intrinsic::x86_avx_ptestz_256:
9188 X86CC = X86::COND_E;
9190 case Intrinsic::x86_avx_vtestc_ps:
9191 case Intrinsic::x86_avx_vtestc_pd:
9192 case Intrinsic::x86_avx_vtestc_ps_256:
9193 case Intrinsic::x86_avx_vtestc_pd_256:
9194 IsTestPacked = true; // Fallthrough
9195 case Intrinsic::x86_sse41_ptestc:
9196 case Intrinsic::x86_avx_ptestc_256:
9198 X86CC = X86::COND_B;
9200 case Intrinsic::x86_avx_vtestnzc_ps:
9201 case Intrinsic::x86_avx_vtestnzc_pd:
9202 case Intrinsic::x86_avx_vtestnzc_ps_256:
9203 case Intrinsic::x86_avx_vtestnzc_pd_256:
9204 IsTestPacked = true; // Fallthrough
9205 case Intrinsic::x86_sse41_ptestnzc:
9206 case Intrinsic::x86_avx_ptestnzc_256:
9208 X86CC = X86::COND_A;
9212 SDValue LHS = Op.getOperand(1);
9213 SDValue RHS = Op.getOperand(2);
9214 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
9215 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
9216 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
9217 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
9218 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
9221 // Fix vector shift instructions where the last operand is a non-immediate
9223 case Intrinsic::x86_sse2_pslli_w:
9224 case Intrinsic::x86_sse2_pslli_d:
9225 case Intrinsic::x86_sse2_pslli_q:
9226 case Intrinsic::x86_sse2_psrli_w:
9227 case Intrinsic::x86_sse2_psrli_d:
9228 case Intrinsic::x86_sse2_psrli_q:
9229 case Intrinsic::x86_sse2_psrai_w:
9230 case Intrinsic::x86_sse2_psrai_d:
9231 case Intrinsic::x86_mmx_pslli_w:
9232 case Intrinsic::x86_mmx_pslli_d:
9233 case Intrinsic::x86_mmx_pslli_q:
9234 case Intrinsic::x86_mmx_psrli_w:
9235 case Intrinsic::x86_mmx_psrli_d:
9236 case Intrinsic::x86_mmx_psrli_q:
9237 case Intrinsic::x86_mmx_psrai_w:
9238 case Intrinsic::x86_mmx_psrai_d: {
9239 SDValue ShAmt = Op.getOperand(2);
9240 if (isa<ConstantSDNode>(ShAmt))
9243 unsigned NewIntNo = 0;
9244 EVT ShAmtVT = MVT::v4i32;
9246 case Intrinsic::x86_sse2_pslli_w:
9247 NewIntNo = Intrinsic::x86_sse2_psll_w;
9249 case Intrinsic::x86_sse2_pslli_d:
9250 NewIntNo = Intrinsic::x86_sse2_psll_d;
9252 case Intrinsic::x86_sse2_pslli_q:
9253 NewIntNo = Intrinsic::x86_sse2_psll_q;
9255 case Intrinsic::x86_sse2_psrli_w:
9256 NewIntNo = Intrinsic::x86_sse2_psrl_w;
9258 case Intrinsic::x86_sse2_psrli_d:
9259 NewIntNo = Intrinsic::x86_sse2_psrl_d;
9261 case Intrinsic::x86_sse2_psrli_q:
9262 NewIntNo = Intrinsic::x86_sse2_psrl_q;
9264 case Intrinsic::x86_sse2_psrai_w:
9265 NewIntNo = Intrinsic::x86_sse2_psra_w;
9267 case Intrinsic::x86_sse2_psrai_d:
9268 NewIntNo = Intrinsic::x86_sse2_psra_d;
9271 ShAmtVT = MVT::v2i32;
9273 case Intrinsic::x86_mmx_pslli_w:
9274 NewIntNo = Intrinsic::x86_mmx_psll_w;
9276 case Intrinsic::x86_mmx_pslli_d:
9277 NewIntNo = Intrinsic::x86_mmx_psll_d;
9279 case Intrinsic::x86_mmx_pslli_q:
9280 NewIntNo = Intrinsic::x86_mmx_psll_q;
9282 case Intrinsic::x86_mmx_psrli_w:
9283 NewIntNo = Intrinsic::x86_mmx_psrl_w;
9285 case Intrinsic::x86_mmx_psrli_d:
9286 NewIntNo = Intrinsic::x86_mmx_psrl_d;
9288 case Intrinsic::x86_mmx_psrli_q:
9289 NewIntNo = Intrinsic::x86_mmx_psrl_q;
9291 case Intrinsic::x86_mmx_psrai_w:
9292 NewIntNo = Intrinsic::x86_mmx_psra_w;
9294 case Intrinsic::x86_mmx_psrai_d:
9295 NewIntNo = Intrinsic::x86_mmx_psra_d;
9297 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
9303 // The vector shift intrinsics with scalars uses 32b shift amounts but
9304 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
9308 ShOps[1] = DAG.getConstant(0, MVT::i32);
9309 if (ShAmtVT == MVT::v4i32) {
9310 ShOps[2] = DAG.getUNDEF(MVT::i32);
9311 ShOps[3] = DAG.getUNDEF(MVT::i32);
9312 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
9314 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
9315 // FIXME this must be lowered to get rid of the invalid type.
9318 EVT VT = Op.getValueType();
9319 ShAmt = DAG.getNode(ISD::BITCAST, dl, VT, ShAmt);
9320 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9321 DAG.getConstant(NewIntNo, MVT::i32),
9322 Op.getOperand(1), ShAmt);
9327 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
9328 SelectionDAG &DAG) const {
9329 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
9330 MFI->setReturnAddressIsTaken(true);
9332 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9333 DebugLoc dl = Op.getDebugLoc();
9336 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
9338 DAG.getConstant(TD->getPointerSize(),
9339 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
9340 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
9341 DAG.getNode(ISD::ADD, dl, getPointerTy(),
9343 MachinePointerInfo(), false, false, 0);
9346 // Just load the return address.
9347 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
9348 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
9349 RetAddrFI, MachinePointerInfo(), false, false, 0);
9352 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
9353 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
9354 MFI->setFrameAddressIsTaken(true);
9356 EVT VT = Op.getValueType();
9357 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
9358 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9359 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
9360 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
9362 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
9363 MachinePointerInfo(),
9368 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
9369 SelectionDAG &DAG) const {
9370 return DAG.getIntPtrConstant(2*TD->getPointerSize());
9373 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
9374 MachineFunction &MF = DAG.getMachineFunction();
9375 SDValue Chain = Op.getOperand(0);
9376 SDValue Offset = Op.getOperand(1);
9377 SDValue Handler = Op.getOperand(2);
9378 DebugLoc dl = Op.getDebugLoc();
9380 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
9381 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
9383 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
9385 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
9386 DAG.getIntPtrConstant(TD->getPointerSize()));
9387 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
9388 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
9390 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
9391 MF.getRegInfo().addLiveOut(StoreAddrReg);
9393 return DAG.getNode(X86ISD::EH_RETURN, dl,
9395 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
9398 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
9399 SelectionDAG &DAG) const {
9400 SDValue Root = Op.getOperand(0);
9401 SDValue Trmp = Op.getOperand(1); // trampoline
9402 SDValue FPtr = Op.getOperand(2); // nested function
9403 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
9404 DebugLoc dl = Op.getDebugLoc();
9406 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
9408 if (Subtarget->is64Bit()) {
9409 SDValue OutChains[6];
9411 // Large code-model.
9412 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
9413 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
9415 const unsigned char N86R10 = X86_MC::getX86RegNum(X86::R10);
9416 const unsigned char N86R11 = X86_MC::getX86RegNum(X86::R11);
9418 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
9420 // Load the pointer to the nested function into R11.
9421 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
9422 SDValue Addr = Trmp;
9423 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
9424 Addr, MachinePointerInfo(TrmpAddr),
9427 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
9428 DAG.getConstant(2, MVT::i64));
9429 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
9430 MachinePointerInfo(TrmpAddr, 2),
9433 // Load the 'nest' parameter value into R10.
9434 // R10 is specified in X86CallingConv.td
9435 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
9436 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
9437 DAG.getConstant(10, MVT::i64));
9438 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
9439 Addr, MachinePointerInfo(TrmpAddr, 10),
9442 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
9443 DAG.getConstant(12, MVT::i64));
9444 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
9445 MachinePointerInfo(TrmpAddr, 12),
9448 // Jump to the nested function.
9449 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
9450 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
9451 DAG.getConstant(20, MVT::i64));
9452 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
9453 Addr, MachinePointerInfo(TrmpAddr, 20),
9456 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
9457 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
9458 DAG.getConstant(22, MVT::i64));
9459 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
9460 MachinePointerInfo(TrmpAddr, 22),
9464 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
9465 return DAG.getMergeValues(Ops, 2, dl);
9467 const Function *Func =
9468 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
9469 CallingConv::ID CC = Func->getCallingConv();
9474 llvm_unreachable("Unsupported calling convention");
9475 case CallingConv::C:
9476 case CallingConv::X86_StdCall: {
9477 // Pass 'nest' parameter in ECX.
9478 // Must be kept in sync with X86CallingConv.td
9481 // Check that ECX wasn't needed by an 'inreg' parameter.
9482 FunctionType *FTy = Func->getFunctionType();
9483 const AttrListPtr &Attrs = Func->getAttributes();
9485 if (!Attrs.isEmpty() && !Func->isVarArg()) {
9486 unsigned InRegCount = 0;
9489 for (FunctionType::param_iterator I = FTy->param_begin(),
9490 E = FTy->param_end(); I != E; ++I, ++Idx)
9491 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
9492 // FIXME: should only count parameters that are lowered to integers.
9493 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
9495 if (InRegCount > 2) {
9496 report_fatal_error("Nest register in use - reduce number of inreg"
9502 case CallingConv::X86_FastCall:
9503 case CallingConv::X86_ThisCall:
9504 case CallingConv::Fast:
9505 // Pass 'nest' parameter in EAX.
9506 // Must be kept in sync with X86CallingConv.td
9511 SDValue OutChains[4];
9514 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
9515 DAG.getConstant(10, MVT::i32));
9516 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
9518 // This is storing the opcode for MOV32ri.
9519 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
9520 const unsigned char N86Reg = X86_MC::getX86RegNum(NestReg);
9521 OutChains[0] = DAG.getStore(Root, dl,
9522 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
9523 Trmp, MachinePointerInfo(TrmpAddr),
9526 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
9527 DAG.getConstant(1, MVT::i32));
9528 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
9529 MachinePointerInfo(TrmpAddr, 1),
9532 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
9533 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
9534 DAG.getConstant(5, MVT::i32));
9535 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
9536 MachinePointerInfo(TrmpAddr, 5),
9539 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
9540 DAG.getConstant(6, MVT::i32));
9541 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
9542 MachinePointerInfo(TrmpAddr, 6),
9546 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
9547 return DAG.getMergeValues(Ops, 2, dl);
9551 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
9552 SelectionDAG &DAG) const {
9554 The rounding mode is in bits 11:10 of FPSR, and has the following
9561 FLT_ROUNDS, on the other hand, expects the following:
9568 To perform the conversion, we do:
9569 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
9572 MachineFunction &MF = DAG.getMachineFunction();
9573 const TargetMachine &TM = MF.getTarget();
9574 const TargetFrameLowering &TFI = *TM.getFrameLowering();
9575 unsigned StackAlignment = TFI.getStackAlignment();
9576 EVT VT = Op.getValueType();
9577 DebugLoc DL = Op.getDebugLoc();
9579 // Save FP Control Word to stack slot
9580 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
9581 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
9584 MachineMemOperand *MMO =
9585 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
9586 MachineMemOperand::MOStore, 2, 2);
9588 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
9589 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
9590 DAG.getVTList(MVT::Other),
9591 Ops, 2, MVT::i16, MMO);
9593 // Load FP Control Word from stack slot
9594 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
9595 MachinePointerInfo(), false, false, 0);
9597 // Transform as necessary
9599 DAG.getNode(ISD::SRL, DL, MVT::i16,
9600 DAG.getNode(ISD::AND, DL, MVT::i16,
9601 CWD, DAG.getConstant(0x800, MVT::i16)),
9602 DAG.getConstant(11, MVT::i8));
9604 DAG.getNode(ISD::SRL, DL, MVT::i16,
9605 DAG.getNode(ISD::AND, DL, MVT::i16,
9606 CWD, DAG.getConstant(0x400, MVT::i16)),
9607 DAG.getConstant(9, MVT::i8));
9610 DAG.getNode(ISD::AND, DL, MVT::i16,
9611 DAG.getNode(ISD::ADD, DL, MVT::i16,
9612 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
9613 DAG.getConstant(1, MVT::i16)),
9614 DAG.getConstant(3, MVT::i16));
9617 return DAG.getNode((VT.getSizeInBits() < 16 ?
9618 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
9621 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
9622 EVT VT = Op.getValueType();
9624 unsigned NumBits = VT.getSizeInBits();
9625 DebugLoc dl = Op.getDebugLoc();
9627 Op = Op.getOperand(0);
9628 if (VT == MVT::i8) {
9629 // Zero extend to i32 since there is not an i8 bsr.
9631 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
9634 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
9635 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
9636 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
9638 // If src is zero (i.e. bsr sets ZF), returns NumBits.
9641 DAG.getConstant(NumBits+NumBits-1, OpVT),
9642 DAG.getConstant(X86::COND_E, MVT::i8),
9645 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
9647 // Finally xor with NumBits-1.
9648 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
9651 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
9655 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
9656 EVT VT = Op.getValueType();
9658 unsigned NumBits = VT.getSizeInBits();
9659 DebugLoc dl = Op.getDebugLoc();
9661 Op = Op.getOperand(0);
9662 if (VT == MVT::i8) {
9664 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
9667 // Issue a bsf (scan bits forward) which also sets EFLAGS.
9668 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
9669 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
9671 // If src is zero (i.e. bsf sets ZF), returns NumBits.
9674 DAG.getConstant(NumBits, OpVT),
9675 DAG.getConstant(X86::COND_E, MVT::i8),
9678 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
9681 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
9685 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
9686 // ones, and then concatenate the result back.
9687 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
9688 EVT VT = Op.getValueType();
9690 assert(VT.getSizeInBits() == 256 && VT.isInteger() &&
9691 "Unsupported value type for operation");
9693 int NumElems = VT.getVectorNumElements();
9694 DebugLoc dl = Op.getDebugLoc();
9695 SDValue Idx0 = DAG.getConstant(0, MVT::i32);
9696 SDValue Idx1 = DAG.getConstant(NumElems/2, MVT::i32);
9698 // Extract the LHS vectors
9699 SDValue LHS = Op.getOperand(0);
9700 SDValue LHS1 = Extract128BitVector(LHS, Idx0, DAG, dl);
9701 SDValue LHS2 = Extract128BitVector(LHS, Idx1, DAG, dl);
9703 // Extract the RHS vectors
9704 SDValue RHS = Op.getOperand(1);
9705 SDValue RHS1 = Extract128BitVector(RHS, Idx0, DAG, dl);
9706 SDValue RHS2 = Extract128BitVector(RHS, Idx1, DAG, dl);
9708 MVT EltVT = VT.getVectorElementType().getSimpleVT();
9709 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
9711 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
9712 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
9713 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
9716 SDValue X86TargetLowering::LowerADD(SDValue Op, SelectionDAG &DAG) const {
9717 assert(Op.getValueType().getSizeInBits() == 256 &&
9718 Op.getValueType().isInteger() &&
9719 "Only handle AVX 256-bit vector integer operation");
9720 return Lower256IntArith(Op, DAG);
9723 SDValue X86TargetLowering::LowerSUB(SDValue Op, SelectionDAG &DAG) const {
9724 assert(Op.getValueType().getSizeInBits() == 256 &&
9725 Op.getValueType().isInteger() &&
9726 "Only handle AVX 256-bit vector integer operation");
9727 return Lower256IntArith(Op, DAG);
9730 SDValue X86TargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
9731 EVT VT = Op.getValueType();
9733 // Decompose 256-bit ops into smaller 128-bit ops.
9734 if (VT.getSizeInBits() == 256)
9735 return Lower256IntArith(Op, DAG);
9737 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
9738 DebugLoc dl = Op.getDebugLoc();
9740 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
9741 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
9742 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
9743 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
9744 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
9746 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
9747 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
9748 // return AloBlo + AloBhi + AhiBlo;
9750 SDValue A = Op.getOperand(0);
9751 SDValue B = Op.getOperand(1);
9753 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9754 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9755 A, DAG.getConstant(32, MVT::i32));
9756 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9757 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9758 B, DAG.getConstant(32, MVT::i32));
9759 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9760 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
9762 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9763 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
9765 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9766 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
9768 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9769 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9770 AloBhi, DAG.getConstant(32, MVT::i32));
9771 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9772 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9773 AhiBlo, DAG.getConstant(32, MVT::i32));
9774 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
9775 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
9779 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
9781 EVT VT = Op.getValueType();
9782 DebugLoc dl = Op.getDebugLoc();
9783 SDValue R = Op.getOperand(0);
9784 SDValue Amt = Op.getOperand(1);
9785 LLVMContext *Context = DAG.getContext();
9787 if (!(Subtarget->hasSSE2() || Subtarget->hasAVX()))
9790 // Decompose 256-bit shifts into smaller 128-bit shifts.
9791 if (VT.getSizeInBits() == 256) {
9792 int NumElems = VT.getVectorNumElements();
9793 MVT EltVT = VT.getVectorElementType().getSimpleVT();
9794 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
9796 // Extract the two vectors
9797 SDValue V1 = Extract128BitVector(R, DAG.getConstant(0, MVT::i32), DAG, dl);
9798 SDValue V2 = Extract128BitVector(R, DAG.getConstant(NumElems/2, MVT::i32),
9801 // Recreate the shift amount vectors
9803 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
9804 // Constant shift amount
9805 SmallVector<SDValue, 4> Amt1Csts;
9806 SmallVector<SDValue, 4> Amt2Csts;
9807 for (int i = 0; i < NumElems/2; ++i)
9808 Amt1Csts.push_back(Amt->getOperand(i));
9809 for (int i = NumElems/2; i < NumElems; ++i)
9810 Amt2Csts.push_back(Amt->getOperand(i));
9812 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT,
9813 &Amt1Csts[0], NumElems/2);
9814 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT,
9815 &Amt2Csts[0], NumElems/2);
9817 // Variable shift amount
9818 Amt1 = Extract128BitVector(Amt, DAG.getConstant(0, MVT::i32), DAG, dl);
9819 Amt2 = Extract128BitVector(Amt, DAG.getConstant(NumElems/2, MVT::i32),
9823 // Issue new vector shifts for the smaller types
9824 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
9825 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
9827 // Concatenate the result back
9828 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
9831 // Optimize shl/srl/sra with constant shift amount.
9832 if (isSplatVector(Amt.getNode())) {
9833 SDValue SclrAmt = Amt->getOperand(0);
9834 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(SclrAmt)) {
9835 uint64_t ShiftAmt = C->getZExtValue();
9837 if (VT == MVT::v2i64 && Op.getOpcode() == ISD::SHL)
9838 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9839 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9840 R, DAG.getConstant(ShiftAmt, MVT::i32));
9842 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SHL)
9843 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9844 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9845 R, DAG.getConstant(ShiftAmt, MVT::i32));
9847 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SHL)
9848 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9849 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9850 R, DAG.getConstant(ShiftAmt, MVT::i32));
9852 if (VT == MVT::v2i64 && Op.getOpcode() == ISD::SRL)
9853 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9854 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9855 R, DAG.getConstant(ShiftAmt, MVT::i32));
9857 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SRL)
9858 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9859 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9860 R, DAG.getConstant(ShiftAmt, MVT::i32));
9862 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SRL)
9863 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9864 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9865 R, DAG.getConstant(ShiftAmt, MVT::i32));
9867 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SRA)
9868 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9869 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9870 R, DAG.getConstant(ShiftAmt, MVT::i32));
9872 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SRA)
9873 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9874 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9875 R, DAG.getConstant(ShiftAmt, MVT::i32));
9879 // Lower SHL with variable shift amount.
9880 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
9881 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9882 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9883 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
9885 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
9887 std::vector<Constant*> CV(4, CI);
9888 Constant *C = ConstantVector::get(CV);
9889 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9890 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9891 MachinePointerInfo::getConstantPool(),
9894 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
9895 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
9896 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
9897 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
9899 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
9901 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9902 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9903 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
9905 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
9906 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
9908 std::vector<Constant*> CVM1(16, CM1);
9909 std::vector<Constant*> CVM2(16, CM2);
9910 Constant *C = ConstantVector::get(CVM1);
9911 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9912 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9913 MachinePointerInfo::getConstantPool(),
9916 // r = pblendv(r, psllw(r & (char16)15, 4), a);
9917 M = DAG.getNode(ISD::AND, dl, VT, R, M);
9918 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9919 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
9920 DAG.getConstant(4, MVT::i32));
9921 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
9923 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
9925 C = ConstantVector::get(CVM2);
9926 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9927 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9928 MachinePointerInfo::getConstantPool(),
9931 // r = pblendv(r, psllw(r & (char16)63, 2), a);
9932 M = DAG.getNode(ISD::AND, dl, VT, R, M);
9933 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9934 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
9935 DAG.getConstant(2, MVT::i32));
9936 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
9938 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
9940 // return pblendv(r, r+r, a);
9941 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT,
9942 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
9948 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
9949 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
9950 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
9951 // looks for this combo and may remove the "setcc" instruction if the "setcc"
9952 // has only one use.
9953 SDNode *N = Op.getNode();
9954 SDValue LHS = N->getOperand(0);
9955 SDValue RHS = N->getOperand(1);
9956 unsigned BaseOp = 0;
9958 DebugLoc DL = Op.getDebugLoc();
9959 switch (Op.getOpcode()) {
9960 default: llvm_unreachable("Unknown ovf instruction!");
9962 // A subtract of one will be selected as a INC. Note that INC doesn't
9963 // set CF, so we can't do this for UADDO.
9964 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
9966 BaseOp = X86ISD::INC;
9970 BaseOp = X86ISD::ADD;
9974 BaseOp = X86ISD::ADD;
9978 // A subtract of one will be selected as a DEC. Note that DEC doesn't
9979 // set CF, so we can't do this for USUBO.
9980 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
9982 BaseOp = X86ISD::DEC;
9986 BaseOp = X86ISD::SUB;
9990 BaseOp = X86ISD::SUB;
9994 BaseOp = X86ISD::SMUL;
9997 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
9998 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
10000 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
10003 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10004 DAG.getConstant(X86::COND_O, MVT::i32),
10005 SDValue(Sum.getNode(), 2));
10007 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
10011 // Also sets EFLAGS.
10012 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
10013 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
10016 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
10017 DAG.getConstant(Cond, MVT::i32),
10018 SDValue(Sum.getNode(), 1));
10020 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
10023 SDValue X86TargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const{
10024 DebugLoc dl = Op.getDebugLoc();
10025 SDNode* Node = Op.getNode();
10026 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
10027 EVT VT = Node->getValueType(0);
10029 if (Subtarget->hasSSE2() && VT.isVector()) {
10030 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
10031 ExtraVT.getScalarType().getSizeInBits();
10032 SDValue ShAmt = DAG.getConstant(BitsDiff, MVT::i32);
10034 unsigned SHLIntrinsicsID = 0;
10035 unsigned SRAIntrinsicsID = 0;
10036 switch (VT.getSimpleVT().SimpleTy) {
10040 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_q;
10041 SRAIntrinsicsID = 0;
10045 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_d;
10046 SRAIntrinsicsID = Intrinsic::x86_sse2_psrai_d;
10050 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_w;
10051 SRAIntrinsicsID = Intrinsic::x86_sse2_psrai_w;
10056 SDValue Tmp1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
10057 DAG.getConstant(SHLIntrinsicsID, MVT::i32),
10058 Node->getOperand(0), ShAmt);
10060 // In case of 1 bit sext, no need to shr
10061 if (ExtraVT.getScalarType().getSizeInBits() == 1) return Tmp1;
10063 if (SRAIntrinsicsID) {
10064 Tmp1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
10065 DAG.getConstant(SRAIntrinsicsID, MVT::i32),
10075 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
10076 DebugLoc dl = Op.getDebugLoc();
10078 // Go ahead and emit the fence on x86-64 even if we asked for no-sse2.
10079 // There isn't any reason to disable it if the target processor supports it.
10080 if (!Subtarget->hasSSE2() && !Subtarget->is64Bit()) {
10081 SDValue Chain = Op.getOperand(0);
10082 SDValue Zero = DAG.getConstant(0, MVT::i32);
10084 DAG.getRegister(X86::ESP, MVT::i32), // Base
10085 DAG.getTargetConstant(1, MVT::i8), // Scale
10086 DAG.getRegister(0, MVT::i32), // Index
10087 DAG.getTargetConstant(0, MVT::i32), // Disp
10088 DAG.getRegister(0, MVT::i32), // Segment.
10093 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
10094 array_lengthof(Ops));
10095 return SDValue(Res, 0);
10098 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
10100 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
10102 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10103 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
10104 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
10105 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
10107 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
10108 if (!Op1 && !Op2 && !Op3 && Op4)
10109 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
10111 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
10112 if (Op1 && !Op2 && !Op3 && !Op4)
10113 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
10115 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
10117 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
10120 SDValue X86TargetLowering::LowerATOMIC_FENCE(SDValue Op,
10121 SelectionDAG &DAG) const {
10122 DebugLoc dl = Op.getDebugLoc();
10123 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
10124 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
10125 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
10126 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
10128 // The only fence that needs an instruction is a sequentially-consistent
10129 // cross-thread fence.
10130 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
10131 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
10132 // no-sse2). There isn't any reason to disable it if the target processor
10134 if (Subtarget->hasSSE2() || Subtarget->is64Bit())
10135 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
10137 SDValue Chain = Op.getOperand(0);
10138 SDValue Zero = DAG.getConstant(0, MVT::i32);
10140 DAG.getRegister(X86::ESP, MVT::i32), // Base
10141 DAG.getTargetConstant(1, MVT::i8), // Scale
10142 DAG.getRegister(0, MVT::i32), // Index
10143 DAG.getTargetConstant(0, MVT::i32), // Disp
10144 DAG.getRegister(0, MVT::i32), // Segment.
10149 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
10150 array_lengthof(Ops));
10151 return SDValue(Res, 0);
10154 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
10155 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
10159 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
10160 EVT T = Op.getValueType();
10161 DebugLoc DL = Op.getDebugLoc();
10164 switch(T.getSimpleVT().SimpleTy) {
10166 assert(false && "Invalid value type!");
10167 case MVT::i8: Reg = X86::AL; size = 1; break;
10168 case MVT::i16: Reg = X86::AX; size = 2; break;
10169 case MVT::i32: Reg = X86::EAX; size = 4; break;
10171 assert(Subtarget->is64Bit() && "Node not type legal!");
10172 Reg = X86::RAX; size = 8;
10175 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
10176 Op.getOperand(2), SDValue());
10177 SDValue Ops[] = { cpIn.getValue(0),
10180 DAG.getTargetConstant(size, MVT::i8),
10181 cpIn.getValue(1) };
10182 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
10183 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
10184 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
10187 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
10191 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
10192 SelectionDAG &DAG) const {
10193 assert(Subtarget->is64Bit() && "Result not type legalized?");
10194 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
10195 SDValue TheChain = Op.getOperand(0);
10196 DebugLoc dl = Op.getDebugLoc();
10197 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
10198 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
10199 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
10201 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
10202 DAG.getConstant(32, MVT::i8));
10204 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
10207 return DAG.getMergeValues(Ops, 2, dl);
10210 SDValue X86TargetLowering::LowerBITCAST(SDValue Op,
10211 SelectionDAG &DAG) const {
10212 EVT SrcVT = Op.getOperand(0).getValueType();
10213 EVT DstVT = Op.getValueType();
10214 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
10215 Subtarget->hasMMX() && "Unexpected custom BITCAST");
10216 assert((DstVT == MVT::i64 ||
10217 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
10218 "Unexpected custom BITCAST");
10219 // i64 <=> MMX conversions are Legal.
10220 if (SrcVT==MVT::i64 && DstVT.isVector())
10222 if (DstVT==MVT::i64 && SrcVT.isVector())
10224 // MMX <=> MMX conversions are Legal.
10225 if (SrcVT.isVector() && DstVT.isVector())
10227 // All other conversions need to be expanded.
10231 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
10232 SDNode *Node = Op.getNode();
10233 DebugLoc dl = Node->getDebugLoc();
10234 EVT T = Node->getValueType(0);
10235 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
10236 DAG.getConstant(0, T), Node->getOperand(2));
10237 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
10238 cast<AtomicSDNode>(Node)->getMemoryVT(),
10239 Node->getOperand(0),
10240 Node->getOperand(1), negOp,
10241 cast<AtomicSDNode>(Node)->getSrcValue(),
10242 cast<AtomicSDNode>(Node)->getAlignment(),
10243 cast<AtomicSDNode>(Node)->getOrdering(),
10244 cast<AtomicSDNode>(Node)->getSynchScope());
10247 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
10248 SDNode *Node = Op.getNode();
10249 DebugLoc dl = Node->getDebugLoc();
10250 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
10252 // Convert seq_cst store -> xchg
10253 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
10254 // FIXME: On 32-bit, store -> fist or movq would be more efficient
10255 // (The only way to get a 16-byte store is cmpxchg16b)
10256 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
10257 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
10258 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
10259 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
10260 cast<AtomicSDNode>(Node)->getMemoryVT(),
10261 Node->getOperand(0),
10262 Node->getOperand(1), Node->getOperand(2),
10263 cast<AtomicSDNode>(Node)->getMemOperand(),
10264 cast<AtomicSDNode>(Node)->getOrdering(),
10265 cast<AtomicSDNode>(Node)->getSynchScope());
10266 return Swap.getValue(1);
10268 // Other atomic stores have a simple pattern.
10272 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
10273 EVT VT = Op.getNode()->getValueType(0);
10275 // Let legalize expand this if it isn't a legal type yet.
10276 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
10279 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
10282 bool ExtraOp = false;
10283 switch (Op.getOpcode()) {
10284 default: assert(0 && "Invalid code");
10285 case ISD::ADDC: Opc = X86ISD::ADD; break;
10286 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
10287 case ISD::SUBC: Opc = X86ISD::SUB; break;
10288 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
10292 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
10294 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
10295 Op.getOperand(1), Op.getOperand(2));
10298 /// LowerOperation - Provide custom lowering hooks for some operations.
10300 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
10301 switch (Op.getOpcode()) {
10302 default: llvm_unreachable("Should not custom lower this!");
10303 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op,DAG);
10304 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
10305 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op,DAG);
10306 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
10307 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
10308 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
10309 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
10310 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
10311 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
10312 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
10313 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
10314 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
10315 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, DAG);
10316 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
10317 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
10318 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
10319 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
10320 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
10321 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
10322 case ISD::SHL_PARTS:
10323 case ISD::SRA_PARTS:
10324 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
10325 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
10326 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
10327 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
10328 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
10329 case ISD::FABS: return LowerFABS(Op, DAG);
10330 case ISD::FNEG: return LowerFNEG(Op, DAG);
10331 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
10332 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
10333 case ISD::SETCC: return LowerSETCC(Op, DAG);
10334 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
10335 case ISD::SELECT: return LowerSELECT(Op, DAG);
10336 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
10337 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
10338 case ISD::VASTART: return LowerVASTART(Op, DAG);
10339 case ISD::VAARG: return LowerVAARG(Op, DAG);
10340 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
10341 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
10342 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
10343 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
10344 case ISD::FRAME_TO_ARGS_OFFSET:
10345 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
10346 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
10347 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
10348 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
10349 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
10350 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
10351 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
10352 case ISD::MUL: return LowerMUL(Op, DAG);
10355 case ISD::SHL: return LowerShift(Op, DAG);
10361 case ISD::UMULO: return LowerXALUO(Op, DAG);
10362 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
10363 case ISD::BITCAST: return LowerBITCAST(Op, DAG);
10367 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
10368 case ISD::ADD: return LowerADD(Op, DAG);
10369 case ISD::SUB: return LowerSUB(Op, DAG);
10373 static void ReplaceATOMIC_LOAD(SDNode *Node,
10374 SmallVectorImpl<SDValue> &Results,
10375 SelectionDAG &DAG) {
10376 DebugLoc dl = Node->getDebugLoc();
10377 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
10379 // Convert wide load -> cmpxchg8b/cmpxchg16b
10380 // FIXME: On 32-bit, load -> fild or movq would be more efficient
10381 // (The only way to get a 16-byte load is cmpxchg16b)
10382 // FIXME: 16-byte ATOMIC_CMP_SWAP isn't actually hooked up at the moment.
10383 SDValue Zero = DAG.getConstant(0, VT);
10384 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, dl, VT,
10385 Node->getOperand(0),
10386 Node->getOperand(1), Zero, Zero,
10387 cast<AtomicSDNode>(Node)->getMemOperand(),
10388 cast<AtomicSDNode>(Node)->getOrdering(),
10389 cast<AtomicSDNode>(Node)->getSynchScope());
10390 Results.push_back(Swap.getValue(0));
10391 Results.push_back(Swap.getValue(1));
10394 void X86TargetLowering::
10395 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
10396 SelectionDAG &DAG, unsigned NewOp) const {
10397 EVT T = Node->getValueType(0);
10398 DebugLoc dl = Node->getDebugLoc();
10399 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
10401 SDValue Chain = Node->getOperand(0);
10402 SDValue In1 = Node->getOperand(1);
10403 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
10404 Node->getOperand(2), DAG.getIntPtrConstant(0));
10405 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
10406 Node->getOperand(2), DAG.getIntPtrConstant(1));
10407 SDValue Ops[] = { Chain, In1, In2L, In2H };
10408 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
10410 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
10411 cast<MemSDNode>(Node)->getMemOperand());
10412 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
10413 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
10414 Results.push_back(Result.getValue(2));
10417 /// ReplaceNodeResults - Replace a node with an illegal result type
10418 /// with a new node built out of custom code.
10419 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
10420 SmallVectorImpl<SDValue>&Results,
10421 SelectionDAG &DAG) const {
10422 DebugLoc dl = N->getDebugLoc();
10423 switch (N->getOpcode()) {
10425 assert(false && "Do not know how to custom type legalize this operation!");
10427 case ISD::SIGN_EXTEND_INREG:
10432 // We don't want to expand or promote these.
10434 case ISD::FP_TO_SINT: {
10435 std::pair<SDValue,SDValue> Vals =
10436 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
10437 SDValue FIST = Vals.first, StackSlot = Vals.second;
10438 if (FIST.getNode() != 0) {
10439 EVT VT = N->getValueType(0);
10440 // Return a load from the stack slot.
10441 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
10442 MachinePointerInfo(), false, false, 0));
10446 case ISD::READCYCLECOUNTER: {
10447 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
10448 SDValue TheChain = N->getOperand(0);
10449 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
10450 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
10452 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
10454 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
10455 SDValue Ops[] = { eax, edx };
10456 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
10457 Results.push_back(edx.getValue(1));
10460 case ISD::ATOMIC_CMP_SWAP: {
10461 EVT T = N->getValueType(0);
10462 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
10463 bool Regs64bit = T == MVT::i128;
10464 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
10465 SDValue cpInL, cpInH;
10466 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
10467 DAG.getConstant(0, HalfT));
10468 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
10469 DAG.getConstant(1, HalfT));
10470 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
10471 Regs64bit ? X86::RAX : X86::EAX,
10473 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
10474 Regs64bit ? X86::RDX : X86::EDX,
10475 cpInH, cpInL.getValue(1));
10476 SDValue swapInL, swapInH;
10477 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
10478 DAG.getConstant(0, HalfT));
10479 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
10480 DAG.getConstant(1, HalfT));
10481 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
10482 Regs64bit ? X86::RBX : X86::EBX,
10483 swapInL, cpInH.getValue(1));
10484 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
10485 Regs64bit ? X86::RCX : X86::ECX,
10486 swapInH, swapInL.getValue(1));
10487 SDValue Ops[] = { swapInH.getValue(0),
10489 swapInH.getValue(1) };
10490 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
10491 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
10492 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
10493 X86ISD::LCMPXCHG8_DAG;
10494 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys,
10496 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
10497 Regs64bit ? X86::RAX : X86::EAX,
10498 HalfT, Result.getValue(1));
10499 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
10500 Regs64bit ? X86::RDX : X86::EDX,
10501 HalfT, cpOutL.getValue(2));
10502 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
10503 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF, 2));
10504 Results.push_back(cpOutH.getValue(1));
10507 case ISD::ATOMIC_LOAD_ADD:
10508 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
10510 case ISD::ATOMIC_LOAD_AND:
10511 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
10513 case ISD::ATOMIC_LOAD_NAND:
10514 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
10516 case ISD::ATOMIC_LOAD_OR:
10517 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
10519 case ISD::ATOMIC_LOAD_SUB:
10520 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
10522 case ISD::ATOMIC_LOAD_XOR:
10523 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
10525 case ISD::ATOMIC_SWAP:
10526 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
10528 case ISD::ATOMIC_LOAD:
10529 ReplaceATOMIC_LOAD(N, Results, DAG);
10533 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
10535 default: return NULL;
10536 case X86ISD::BSF: return "X86ISD::BSF";
10537 case X86ISD::BSR: return "X86ISD::BSR";
10538 case X86ISD::SHLD: return "X86ISD::SHLD";
10539 case X86ISD::SHRD: return "X86ISD::SHRD";
10540 case X86ISD::FAND: return "X86ISD::FAND";
10541 case X86ISD::FOR: return "X86ISD::FOR";
10542 case X86ISD::FXOR: return "X86ISD::FXOR";
10543 case X86ISD::FSRL: return "X86ISD::FSRL";
10544 case X86ISD::FILD: return "X86ISD::FILD";
10545 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
10546 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
10547 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
10548 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
10549 case X86ISD::FLD: return "X86ISD::FLD";
10550 case X86ISD::FST: return "X86ISD::FST";
10551 case X86ISD::CALL: return "X86ISD::CALL";
10552 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
10553 case X86ISD::BT: return "X86ISD::BT";
10554 case X86ISD::CMP: return "X86ISD::CMP";
10555 case X86ISD::COMI: return "X86ISD::COMI";
10556 case X86ISD::UCOMI: return "X86ISD::UCOMI";
10557 case X86ISD::SETCC: return "X86ISD::SETCC";
10558 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
10559 case X86ISD::FSETCCsd: return "X86ISD::FSETCCsd";
10560 case X86ISD::FSETCCss: return "X86ISD::FSETCCss";
10561 case X86ISD::CMOV: return "X86ISD::CMOV";
10562 case X86ISD::BRCOND: return "X86ISD::BRCOND";
10563 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
10564 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
10565 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
10566 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
10567 case X86ISD::Wrapper: return "X86ISD::Wrapper";
10568 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
10569 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
10570 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
10571 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
10572 case X86ISD::PINSRB: return "X86ISD::PINSRB";
10573 case X86ISD::PINSRW: return "X86ISD::PINSRW";
10574 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
10575 case X86ISD::ANDNP: return "X86ISD::ANDNP";
10576 case X86ISD::PSIGNB: return "X86ISD::PSIGNB";
10577 case X86ISD::PSIGNW: return "X86ISD::PSIGNW";
10578 case X86ISD::PSIGND: return "X86ISD::PSIGND";
10579 case X86ISD::PBLENDVB: return "X86ISD::PBLENDVB";
10580 case X86ISD::FMAX: return "X86ISD::FMAX";
10581 case X86ISD::FMIN: return "X86ISD::FMIN";
10582 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
10583 case X86ISD::FRCP: return "X86ISD::FRCP";
10584 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
10585 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
10586 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
10587 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
10588 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
10589 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
10590 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
10591 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
10592 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
10593 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
10594 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
10595 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
10596 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
10597 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
10598 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
10599 case X86ISD::VSHL: return "X86ISD::VSHL";
10600 case X86ISD::VSRL: return "X86ISD::VSRL";
10601 case X86ISD::CMPPD: return "X86ISD::CMPPD";
10602 case X86ISD::CMPPS: return "X86ISD::CMPPS";
10603 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
10604 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
10605 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
10606 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
10607 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
10608 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
10609 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
10610 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
10611 case X86ISD::ADD: return "X86ISD::ADD";
10612 case X86ISD::SUB: return "X86ISD::SUB";
10613 case X86ISD::ADC: return "X86ISD::ADC";
10614 case X86ISD::SBB: return "X86ISD::SBB";
10615 case X86ISD::SMUL: return "X86ISD::SMUL";
10616 case X86ISD::UMUL: return "X86ISD::UMUL";
10617 case X86ISD::INC: return "X86ISD::INC";
10618 case X86ISD::DEC: return "X86ISD::DEC";
10619 case X86ISD::OR: return "X86ISD::OR";
10620 case X86ISD::XOR: return "X86ISD::XOR";
10621 case X86ISD::AND: return "X86ISD::AND";
10622 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
10623 case X86ISD::PTEST: return "X86ISD::PTEST";
10624 case X86ISD::TESTP: return "X86ISD::TESTP";
10625 case X86ISD::PALIGN: return "X86ISD::PALIGN";
10626 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
10627 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
10628 case X86ISD::PSHUFHW_LD: return "X86ISD::PSHUFHW_LD";
10629 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
10630 case X86ISD::PSHUFLW_LD: return "X86ISD::PSHUFLW_LD";
10631 case X86ISD::SHUFPS: return "X86ISD::SHUFPS";
10632 case X86ISD::SHUFPD: return "X86ISD::SHUFPD";
10633 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
10634 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
10635 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
10636 case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
10637 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
10638 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
10639 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
10640 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
10641 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
10642 case X86ISD::MOVSHDUP_LD: return "X86ISD::MOVSHDUP_LD";
10643 case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
10644 case X86ISD::MOVSD: return "X86ISD::MOVSD";
10645 case X86ISD::MOVSS: return "X86ISD::MOVSS";
10646 case X86ISD::UNPCKLPS: return "X86ISD::UNPCKLPS";
10647 case X86ISD::UNPCKLPD: return "X86ISD::UNPCKLPD";
10648 case X86ISD::VUNPCKLPDY: return "X86ISD::VUNPCKLPDY";
10649 case X86ISD::UNPCKHPS: return "X86ISD::UNPCKHPS";
10650 case X86ISD::UNPCKHPD: return "X86ISD::UNPCKHPD";
10651 case X86ISD::PUNPCKLBW: return "X86ISD::PUNPCKLBW";
10652 case X86ISD::PUNPCKLWD: return "X86ISD::PUNPCKLWD";
10653 case X86ISD::PUNPCKLDQ: return "X86ISD::PUNPCKLDQ";
10654 case X86ISD::PUNPCKLQDQ: return "X86ISD::PUNPCKLQDQ";
10655 case X86ISD::PUNPCKHBW: return "X86ISD::PUNPCKHBW";
10656 case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
10657 case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
10658 case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
10659 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
10660 case X86ISD::VPERMILPS: return "X86ISD::VPERMILPS";
10661 case X86ISD::VPERMILPSY: return "X86ISD::VPERMILPSY";
10662 case X86ISD::VPERMILPD: return "X86ISD::VPERMILPD";
10663 case X86ISD::VPERMILPDY: return "X86ISD::VPERMILPDY";
10664 case X86ISD::VPERM2F128: return "X86ISD::VPERM2F128";
10665 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
10666 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
10667 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
10668 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
10669 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
10673 // isLegalAddressingMode - Return true if the addressing mode represented
10674 // by AM is legal for this target, for a load/store of the specified type.
10675 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
10677 // X86 supports extremely general addressing modes.
10678 CodeModel::Model M = getTargetMachine().getCodeModel();
10679 Reloc::Model R = getTargetMachine().getRelocationModel();
10681 // X86 allows a sign-extended 32-bit immediate field as a displacement.
10682 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
10687 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
10689 // If a reference to this global requires an extra load, we can't fold it.
10690 if (isGlobalStubReference(GVFlags))
10693 // If BaseGV requires a register for the PIC base, we cannot also have a
10694 // BaseReg specified.
10695 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
10698 // If lower 4G is not available, then we must use rip-relative addressing.
10699 if ((M != CodeModel::Small || R != Reloc::Static) &&
10700 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
10704 switch (AM.Scale) {
10710 // These scales always work.
10715 // These scales are formed with basereg+scalereg. Only accept if there is
10720 default: // Other stuff never works.
10728 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
10729 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10731 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
10732 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
10733 if (NumBits1 <= NumBits2)
10738 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
10739 if (!VT1.isInteger() || !VT2.isInteger())
10741 unsigned NumBits1 = VT1.getSizeInBits();
10742 unsigned NumBits2 = VT2.getSizeInBits();
10743 if (NumBits1 <= NumBits2)
10748 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
10749 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
10750 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
10753 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
10754 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
10755 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
10758 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
10759 // i16 instructions are longer (0x66 prefix) and potentially slower.
10760 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
10763 /// isShuffleMaskLegal - Targets can use this to indicate that they only
10764 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
10765 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
10766 /// are assumed to be legal.
10768 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
10770 // Very little shuffling can be done for 64-bit vectors right now.
10771 if (VT.getSizeInBits() == 64)
10772 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
10774 // FIXME: pshufb, blends, shifts.
10775 return (VT.getVectorNumElements() == 2 ||
10776 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
10777 isMOVLMask(M, VT) ||
10778 isSHUFPMask(M, VT) ||
10779 isPSHUFDMask(M, VT) ||
10780 isPSHUFHWMask(M, VT) ||
10781 isPSHUFLWMask(M, VT) ||
10782 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
10783 isUNPCKLMask(M, VT) ||
10784 isUNPCKHMask(M, VT) ||
10785 isUNPCKL_v_undef_Mask(M, VT) ||
10786 isUNPCKH_v_undef_Mask(M, VT));
10790 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
10792 unsigned NumElts = VT.getVectorNumElements();
10793 // FIXME: This collection of masks seems suspect.
10796 if (NumElts == 4 && VT.getSizeInBits() == 128) {
10797 return (isMOVLMask(Mask, VT) ||
10798 isCommutedMOVLMask(Mask, VT, true) ||
10799 isSHUFPMask(Mask, VT) ||
10800 isCommutedSHUFPMask(Mask, VT));
10805 //===----------------------------------------------------------------------===//
10806 // X86 Scheduler Hooks
10807 //===----------------------------------------------------------------------===//
10809 // private utility function
10810 MachineBasicBlock *
10811 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
10812 MachineBasicBlock *MBB,
10819 TargetRegisterClass *RC,
10820 bool invSrc) const {
10821 // For the atomic bitwise operator, we generate
10824 // ld t1 = [bitinstr.addr]
10825 // op t2 = t1, [bitinstr.val]
10827 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
10829 // fallthrough -->nextMBB
10830 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10831 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10832 MachineFunction::iterator MBBIter = MBB;
10835 /// First build the CFG
10836 MachineFunction *F = MBB->getParent();
10837 MachineBasicBlock *thisMBB = MBB;
10838 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
10839 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
10840 F->insert(MBBIter, newMBB);
10841 F->insert(MBBIter, nextMBB);
10843 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
10844 nextMBB->splice(nextMBB->begin(), thisMBB,
10845 llvm::next(MachineBasicBlock::iterator(bInstr)),
10847 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10849 // Update thisMBB to fall through to newMBB
10850 thisMBB->addSuccessor(newMBB);
10852 // newMBB jumps to itself and fall through to nextMBB
10853 newMBB->addSuccessor(nextMBB);
10854 newMBB->addSuccessor(newMBB);
10856 // Insert instructions into newMBB based on incoming instruction
10857 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
10858 "unexpected number of operands");
10859 DebugLoc dl = bInstr->getDebugLoc();
10860 MachineOperand& destOper = bInstr->getOperand(0);
10861 MachineOperand* argOpers[2 + X86::AddrNumOperands];
10862 int numArgs = bInstr->getNumOperands() - 1;
10863 for (int i=0; i < numArgs; ++i)
10864 argOpers[i] = &bInstr->getOperand(i+1);
10866 // x86 address has 4 operands: base, index, scale, and displacement
10867 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
10868 int valArgIndx = lastAddrIndx + 1;
10870 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
10871 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
10872 for (int i=0; i <= lastAddrIndx; ++i)
10873 (*MIB).addOperand(*argOpers[i]);
10875 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
10877 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
10882 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
10883 assert((argOpers[valArgIndx]->isReg() ||
10884 argOpers[valArgIndx]->isImm()) &&
10885 "invalid operand");
10886 if (argOpers[valArgIndx]->isReg())
10887 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
10889 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
10891 (*MIB).addOperand(*argOpers[valArgIndx]);
10893 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
10896 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
10897 for (int i=0; i <= lastAddrIndx; ++i)
10898 (*MIB).addOperand(*argOpers[i]);
10900 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
10901 (*MIB).setMemRefs(bInstr->memoperands_begin(),
10902 bInstr->memoperands_end());
10904 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
10905 MIB.addReg(EAXreg);
10908 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
10910 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
10914 // private utility function: 64 bit atomics on 32 bit host.
10915 MachineBasicBlock *
10916 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
10917 MachineBasicBlock *MBB,
10922 bool invSrc) const {
10923 // For the atomic bitwise operator, we generate
10924 // thisMBB (instructions are in pairs, except cmpxchg8b)
10925 // ld t1,t2 = [bitinstr.addr]
10927 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
10928 // op t5, t6 <- out1, out2, [bitinstr.val]
10929 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
10930 // mov ECX, EBX <- t5, t6
10931 // mov EAX, EDX <- t1, t2
10932 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
10933 // mov t3, t4 <- EAX, EDX
10935 // result in out1, out2
10936 // fallthrough -->nextMBB
10938 const TargetRegisterClass *RC = X86::GR32RegisterClass;
10939 const unsigned LoadOpc = X86::MOV32rm;
10940 const unsigned NotOpc = X86::NOT32r;
10941 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10942 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10943 MachineFunction::iterator MBBIter = MBB;
10946 /// First build the CFG
10947 MachineFunction *F = MBB->getParent();
10948 MachineBasicBlock *thisMBB = MBB;
10949 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
10950 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
10951 F->insert(MBBIter, newMBB);
10952 F->insert(MBBIter, nextMBB);
10954 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
10955 nextMBB->splice(nextMBB->begin(), thisMBB,
10956 llvm::next(MachineBasicBlock::iterator(bInstr)),
10958 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10960 // Update thisMBB to fall through to newMBB
10961 thisMBB->addSuccessor(newMBB);
10963 // newMBB jumps to itself and fall through to nextMBB
10964 newMBB->addSuccessor(nextMBB);
10965 newMBB->addSuccessor(newMBB);
10967 DebugLoc dl = bInstr->getDebugLoc();
10968 // Insert instructions into newMBB based on incoming instruction
10969 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
10970 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
10971 "unexpected number of operands");
10972 MachineOperand& dest1Oper = bInstr->getOperand(0);
10973 MachineOperand& dest2Oper = bInstr->getOperand(1);
10974 MachineOperand* argOpers[2 + X86::AddrNumOperands];
10975 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
10976 argOpers[i] = &bInstr->getOperand(i+2);
10978 // We use some of the operands multiple times, so conservatively just
10979 // clear any kill flags that might be present.
10980 if (argOpers[i]->isReg() && argOpers[i]->isUse())
10981 argOpers[i]->setIsKill(false);
10984 // x86 address has 5 operands: base, index, scale, displacement, and segment.
10985 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
10987 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
10988 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
10989 for (int i=0; i <= lastAddrIndx; ++i)
10990 (*MIB).addOperand(*argOpers[i]);
10991 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
10992 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
10993 // add 4 to displacement.
10994 for (int i=0; i <= lastAddrIndx-2; ++i)
10995 (*MIB).addOperand(*argOpers[i]);
10996 MachineOperand newOp3 = *(argOpers[3]);
10997 if (newOp3.isImm())
10998 newOp3.setImm(newOp3.getImm()+4);
11000 newOp3.setOffset(newOp3.getOffset()+4);
11001 (*MIB).addOperand(newOp3);
11002 (*MIB).addOperand(*argOpers[lastAddrIndx]);
11004 // t3/4 are defined later, at the bottom of the loop
11005 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
11006 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
11007 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
11008 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
11009 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
11010 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
11012 // The subsequent operations should be using the destination registers of
11013 //the PHI instructions.
11015 t1 = F->getRegInfo().createVirtualRegister(RC);
11016 t2 = F->getRegInfo().createVirtualRegister(RC);
11017 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
11018 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
11020 t1 = dest1Oper.getReg();
11021 t2 = dest2Oper.getReg();
11024 int valArgIndx = lastAddrIndx + 1;
11025 assert((argOpers[valArgIndx]->isReg() ||
11026 argOpers[valArgIndx]->isImm()) &&
11027 "invalid operand");
11028 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
11029 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
11030 if (argOpers[valArgIndx]->isReg())
11031 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
11033 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
11034 if (regOpcL != X86::MOV32rr)
11036 (*MIB).addOperand(*argOpers[valArgIndx]);
11037 assert(argOpers[valArgIndx + 1]->isReg() ==
11038 argOpers[valArgIndx]->isReg());
11039 assert(argOpers[valArgIndx + 1]->isImm() ==
11040 argOpers[valArgIndx]->isImm());
11041 if (argOpers[valArgIndx + 1]->isReg())
11042 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
11044 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
11045 if (regOpcH != X86::MOV32rr)
11047 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
11049 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
11051 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
11054 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
11056 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
11059 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
11060 for (int i=0; i <= lastAddrIndx; ++i)
11061 (*MIB).addOperand(*argOpers[i]);
11063 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
11064 (*MIB).setMemRefs(bInstr->memoperands_begin(),
11065 bInstr->memoperands_end());
11067 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
11068 MIB.addReg(X86::EAX);
11069 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
11070 MIB.addReg(X86::EDX);
11073 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
11075 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
11079 // private utility function
11080 MachineBasicBlock *
11081 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
11082 MachineBasicBlock *MBB,
11083 unsigned cmovOpc) const {
11084 // For the atomic min/max operator, we generate
11087 // ld t1 = [min/max.addr]
11088 // mov t2 = [min/max.val]
11090 // cmov[cond] t2 = t1
11092 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
11094 // fallthrough -->nextMBB
11096 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11097 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
11098 MachineFunction::iterator MBBIter = MBB;
11101 /// First build the CFG
11102 MachineFunction *F = MBB->getParent();
11103 MachineBasicBlock *thisMBB = MBB;
11104 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
11105 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
11106 F->insert(MBBIter, newMBB);
11107 F->insert(MBBIter, nextMBB);
11109 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
11110 nextMBB->splice(nextMBB->begin(), thisMBB,
11111 llvm::next(MachineBasicBlock::iterator(mInstr)),
11113 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
11115 // Update thisMBB to fall through to newMBB
11116 thisMBB->addSuccessor(newMBB);
11118 // newMBB jumps to newMBB and fall through to nextMBB
11119 newMBB->addSuccessor(nextMBB);
11120 newMBB->addSuccessor(newMBB);
11122 DebugLoc dl = mInstr->getDebugLoc();
11123 // Insert instructions into newMBB based on incoming instruction
11124 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
11125 "unexpected number of operands");
11126 MachineOperand& destOper = mInstr->getOperand(0);
11127 MachineOperand* argOpers[2 + X86::AddrNumOperands];
11128 int numArgs = mInstr->getNumOperands() - 1;
11129 for (int i=0; i < numArgs; ++i)
11130 argOpers[i] = &mInstr->getOperand(i+1);
11132 // x86 address has 4 operands: base, index, scale, and displacement
11133 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
11134 int valArgIndx = lastAddrIndx + 1;
11136 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
11137 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
11138 for (int i=0; i <= lastAddrIndx; ++i)
11139 (*MIB).addOperand(*argOpers[i]);
11141 // We only support register and immediate values
11142 assert((argOpers[valArgIndx]->isReg() ||
11143 argOpers[valArgIndx]->isImm()) &&
11144 "invalid operand");
11146 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
11147 if (argOpers[valArgIndx]->isReg())
11148 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
11150 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
11151 (*MIB).addOperand(*argOpers[valArgIndx]);
11153 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
11156 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
11161 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
11162 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
11166 // Cmp and exchange if none has modified the memory location
11167 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
11168 for (int i=0; i <= lastAddrIndx; ++i)
11169 (*MIB).addOperand(*argOpers[i]);
11171 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
11172 (*MIB).setMemRefs(mInstr->memoperands_begin(),
11173 mInstr->memoperands_end());
11175 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
11176 MIB.addReg(X86::EAX);
11179 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
11181 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
11185 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
11186 // or XMM0_V32I8 in AVX all of this code can be replaced with that
11187 // in the .td file.
11188 MachineBasicBlock *
11189 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
11190 unsigned numArgs, bool memArg) const {
11191 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
11192 "Target must have SSE4.2 or AVX features enabled");
11194 DebugLoc dl = MI->getDebugLoc();
11195 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11197 if (!Subtarget->hasAVX()) {
11199 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
11201 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
11204 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
11206 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
11209 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
11210 for (unsigned i = 0; i < numArgs; ++i) {
11211 MachineOperand &Op = MI->getOperand(i+1);
11212 if (!(Op.isReg() && Op.isImplicit()))
11213 MIB.addOperand(Op);
11215 BuildMI(*BB, MI, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
11216 .addReg(X86::XMM0);
11218 MI->eraseFromParent();
11222 MachineBasicBlock *
11223 X86TargetLowering::EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB) const {
11224 DebugLoc dl = MI->getDebugLoc();
11225 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11227 // Address into RAX/EAX, other two args into ECX, EDX.
11228 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
11229 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
11230 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
11231 for (int i = 0; i < X86::AddrNumOperands; ++i)
11232 MIB.addOperand(MI->getOperand(i));
11234 unsigned ValOps = X86::AddrNumOperands;
11235 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
11236 .addReg(MI->getOperand(ValOps).getReg());
11237 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
11238 .addReg(MI->getOperand(ValOps+1).getReg());
11240 // The instruction doesn't actually take any operands though.
11241 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
11243 MI->eraseFromParent(); // The pseudo is gone now.
11247 MachineBasicBlock *
11248 X86TargetLowering::EmitMwait(MachineInstr *MI, MachineBasicBlock *BB) const {
11249 DebugLoc dl = MI->getDebugLoc();
11250 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11252 // First arg in ECX, the second in EAX.
11253 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
11254 .addReg(MI->getOperand(0).getReg());
11255 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EAX)
11256 .addReg(MI->getOperand(1).getReg());
11258 // The instruction doesn't actually take any operands though.
11259 BuildMI(*BB, MI, dl, TII->get(X86::MWAITrr));
11261 MI->eraseFromParent(); // The pseudo is gone now.
11265 MachineBasicBlock *
11266 X86TargetLowering::EmitVAARG64WithCustomInserter(
11268 MachineBasicBlock *MBB) const {
11269 // Emit va_arg instruction on X86-64.
11271 // Operands to this pseudo-instruction:
11272 // 0 ) Output : destination address (reg)
11273 // 1-5) Input : va_list address (addr, i64mem)
11274 // 6 ) ArgSize : Size (in bytes) of vararg type
11275 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
11276 // 8 ) Align : Alignment of type
11277 // 9 ) EFLAGS (implicit-def)
11279 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
11280 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
11282 unsigned DestReg = MI->getOperand(0).getReg();
11283 MachineOperand &Base = MI->getOperand(1);
11284 MachineOperand &Scale = MI->getOperand(2);
11285 MachineOperand &Index = MI->getOperand(3);
11286 MachineOperand &Disp = MI->getOperand(4);
11287 MachineOperand &Segment = MI->getOperand(5);
11288 unsigned ArgSize = MI->getOperand(6).getImm();
11289 unsigned ArgMode = MI->getOperand(7).getImm();
11290 unsigned Align = MI->getOperand(8).getImm();
11292 // Memory Reference
11293 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
11294 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
11295 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
11297 // Machine Information
11298 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11299 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
11300 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
11301 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
11302 DebugLoc DL = MI->getDebugLoc();
11304 // struct va_list {
11307 // i64 overflow_area (address)
11308 // i64 reg_save_area (address)
11310 // sizeof(va_list) = 24
11311 // alignment(va_list) = 8
11313 unsigned TotalNumIntRegs = 6;
11314 unsigned TotalNumXMMRegs = 8;
11315 bool UseGPOffset = (ArgMode == 1);
11316 bool UseFPOffset = (ArgMode == 2);
11317 unsigned MaxOffset = TotalNumIntRegs * 8 +
11318 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
11320 /* Align ArgSize to a multiple of 8 */
11321 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
11322 bool NeedsAlign = (Align > 8);
11324 MachineBasicBlock *thisMBB = MBB;
11325 MachineBasicBlock *overflowMBB;
11326 MachineBasicBlock *offsetMBB;
11327 MachineBasicBlock *endMBB;
11329 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
11330 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
11331 unsigned OffsetReg = 0;
11333 if (!UseGPOffset && !UseFPOffset) {
11334 // If we only pull from the overflow region, we don't create a branch.
11335 // We don't need to alter control flow.
11336 OffsetDestReg = 0; // unused
11337 OverflowDestReg = DestReg;
11340 overflowMBB = thisMBB;
11343 // First emit code to check if gp_offset (or fp_offset) is below the bound.
11344 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
11345 // If not, pull from overflow_area. (branch to overflowMBB)
11350 // offsetMBB overflowMBB
11355 // Registers for the PHI in endMBB
11356 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
11357 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
11359 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
11360 MachineFunction *MF = MBB->getParent();
11361 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
11362 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
11363 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
11365 MachineFunction::iterator MBBIter = MBB;
11368 // Insert the new basic blocks
11369 MF->insert(MBBIter, offsetMBB);
11370 MF->insert(MBBIter, overflowMBB);
11371 MF->insert(MBBIter, endMBB);
11373 // Transfer the remainder of MBB and its successor edges to endMBB.
11374 endMBB->splice(endMBB->begin(), thisMBB,
11375 llvm::next(MachineBasicBlock::iterator(MI)),
11377 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
11379 // Make offsetMBB and overflowMBB successors of thisMBB
11380 thisMBB->addSuccessor(offsetMBB);
11381 thisMBB->addSuccessor(overflowMBB);
11383 // endMBB is a successor of both offsetMBB and overflowMBB
11384 offsetMBB->addSuccessor(endMBB);
11385 overflowMBB->addSuccessor(endMBB);
11387 // Load the offset value into a register
11388 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
11389 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
11393 .addDisp(Disp, UseFPOffset ? 4 : 0)
11394 .addOperand(Segment)
11395 .setMemRefs(MMOBegin, MMOEnd);
11397 // Check if there is enough room left to pull this argument.
11398 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
11400 .addImm(MaxOffset + 8 - ArgSizeA8);
11402 // Branch to "overflowMBB" if offset >= max
11403 // Fall through to "offsetMBB" otherwise
11404 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
11405 .addMBB(overflowMBB);
11408 // In offsetMBB, emit code to use the reg_save_area.
11410 assert(OffsetReg != 0);
11412 // Read the reg_save_area address.
11413 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
11414 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
11419 .addOperand(Segment)
11420 .setMemRefs(MMOBegin, MMOEnd);
11422 // Zero-extend the offset
11423 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
11424 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
11427 .addImm(X86::sub_32bit);
11429 // Add the offset to the reg_save_area to get the final address.
11430 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
11431 .addReg(OffsetReg64)
11432 .addReg(RegSaveReg);
11434 // Compute the offset for the next argument
11435 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
11436 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
11438 .addImm(UseFPOffset ? 16 : 8);
11440 // Store it back into the va_list.
11441 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
11445 .addDisp(Disp, UseFPOffset ? 4 : 0)
11446 .addOperand(Segment)
11447 .addReg(NextOffsetReg)
11448 .setMemRefs(MMOBegin, MMOEnd);
11451 BuildMI(offsetMBB, DL, TII->get(X86::JMP_4))
11456 // Emit code to use overflow area
11459 // Load the overflow_area address into a register.
11460 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
11461 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
11466 .addOperand(Segment)
11467 .setMemRefs(MMOBegin, MMOEnd);
11469 // If we need to align it, do so. Otherwise, just copy the address
11470 // to OverflowDestReg.
11472 // Align the overflow address
11473 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
11474 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
11476 // aligned_addr = (addr + (align-1)) & ~(align-1)
11477 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
11478 .addReg(OverflowAddrReg)
11481 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
11483 .addImm(~(uint64_t)(Align-1));
11485 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
11486 .addReg(OverflowAddrReg);
11489 // Compute the next overflow address after this argument.
11490 // (the overflow address should be kept 8-byte aligned)
11491 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
11492 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
11493 .addReg(OverflowDestReg)
11494 .addImm(ArgSizeA8);
11496 // Store the new overflow address.
11497 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
11502 .addOperand(Segment)
11503 .addReg(NextAddrReg)
11504 .setMemRefs(MMOBegin, MMOEnd);
11506 // If we branched, emit the PHI to the front of endMBB.
11508 BuildMI(*endMBB, endMBB->begin(), DL,
11509 TII->get(X86::PHI), DestReg)
11510 .addReg(OffsetDestReg).addMBB(offsetMBB)
11511 .addReg(OverflowDestReg).addMBB(overflowMBB);
11514 // Erase the pseudo instruction
11515 MI->eraseFromParent();
11520 MachineBasicBlock *
11521 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
11523 MachineBasicBlock *MBB) const {
11524 // Emit code to save XMM registers to the stack. The ABI says that the
11525 // number of registers to save is given in %al, so it's theoretically
11526 // possible to do an indirect jump trick to avoid saving all of them,
11527 // however this code takes a simpler approach and just executes all
11528 // of the stores if %al is non-zero. It's less code, and it's probably
11529 // easier on the hardware branch predictor, and stores aren't all that
11530 // expensive anyway.
11532 // Create the new basic blocks. One block contains all the XMM stores,
11533 // and one block is the final destination regardless of whether any
11534 // stores were performed.
11535 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
11536 MachineFunction *F = MBB->getParent();
11537 MachineFunction::iterator MBBIter = MBB;
11539 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
11540 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
11541 F->insert(MBBIter, XMMSaveMBB);
11542 F->insert(MBBIter, EndMBB);
11544 // Transfer the remainder of MBB and its successor edges to EndMBB.
11545 EndMBB->splice(EndMBB->begin(), MBB,
11546 llvm::next(MachineBasicBlock::iterator(MI)),
11548 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
11550 // The original block will now fall through to the XMM save block.
11551 MBB->addSuccessor(XMMSaveMBB);
11552 // The XMMSaveMBB will fall through to the end block.
11553 XMMSaveMBB->addSuccessor(EndMBB);
11555 // Now add the instructions.
11556 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11557 DebugLoc DL = MI->getDebugLoc();
11559 unsigned CountReg = MI->getOperand(0).getReg();
11560 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
11561 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
11563 if (!Subtarget->isTargetWin64()) {
11564 // If %al is 0, branch around the XMM save block.
11565 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
11566 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
11567 MBB->addSuccessor(EndMBB);
11570 // In the XMM save block, save all the XMM argument registers.
11571 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
11572 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
11573 MachineMemOperand *MMO =
11574 F->getMachineMemOperand(
11575 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
11576 MachineMemOperand::MOStore,
11577 /*Size=*/16, /*Align=*/16);
11578 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
11579 .addFrameIndex(RegSaveFrameIndex)
11580 .addImm(/*Scale=*/1)
11581 .addReg(/*IndexReg=*/0)
11582 .addImm(/*Disp=*/Offset)
11583 .addReg(/*Segment=*/0)
11584 .addReg(MI->getOperand(i).getReg())
11585 .addMemOperand(MMO);
11588 MI->eraseFromParent(); // The pseudo instruction is gone now.
11593 MachineBasicBlock *
11594 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
11595 MachineBasicBlock *BB) const {
11596 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11597 DebugLoc DL = MI->getDebugLoc();
11599 // To "insert" a SELECT_CC instruction, we actually have to insert the
11600 // diamond control-flow pattern. The incoming instruction knows the
11601 // destination vreg to set, the condition code register to branch on, the
11602 // true/false values to select between, and a branch opcode to use.
11603 const BasicBlock *LLVM_BB = BB->getBasicBlock();
11604 MachineFunction::iterator It = BB;
11610 // cmpTY ccX, r1, r2
11612 // fallthrough --> copy0MBB
11613 MachineBasicBlock *thisMBB = BB;
11614 MachineFunction *F = BB->getParent();
11615 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
11616 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
11617 F->insert(It, copy0MBB);
11618 F->insert(It, sinkMBB);
11620 // If the EFLAGS register isn't dead in the terminator, then claim that it's
11621 // live into the sink and copy blocks.
11622 const MachineFunction *MF = BB->getParent();
11623 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
11624 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
11626 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
11627 const MachineOperand &MO = MI->getOperand(I);
11628 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
11629 unsigned Reg = MO.getReg();
11630 if (Reg != X86::EFLAGS) continue;
11631 copy0MBB->addLiveIn(Reg);
11632 sinkMBB->addLiveIn(Reg);
11635 // Transfer the remainder of BB and its successor edges to sinkMBB.
11636 sinkMBB->splice(sinkMBB->begin(), BB,
11637 llvm::next(MachineBasicBlock::iterator(MI)),
11639 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
11641 // Add the true and fallthrough blocks as its successors.
11642 BB->addSuccessor(copy0MBB);
11643 BB->addSuccessor(sinkMBB);
11645 // Create the conditional branch instruction.
11647 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
11648 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
11651 // %FalseValue = ...
11652 // # fallthrough to sinkMBB
11653 copy0MBB->addSuccessor(sinkMBB);
11656 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
11658 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
11659 TII->get(X86::PHI), MI->getOperand(0).getReg())
11660 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
11661 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
11663 MI->eraseFromParent(); // The pseudo instruction is gone now.
11667 MachineBasicBlock *
11668 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI, MachineBasicBlock *BB,
11669 bool Is64Bit) const {
11670 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11671 DebugLoc DL = MI->getDebugLoc();
11672 MachineFunction *MF = BB->getParent();
11673 const BasicBlock *LLVM_BB = BB->getBasicBlock();
11675 assert(EnableSegmentedStacks);
11677 unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
11678 unsigned TlsOffset = Is64Bit ? 0x70 : 0x30;
11681 // ... [Till the alloca]
11682 // If stacklet is not large enough, jump to mallocMBB
11685 // Allocate by subtracting from RSP
11686 // Jump to continueMBB
11689 // Allocate by call to runtime
11693 // [rest of original BB]
11696 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
11697 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
11698 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
11700 MachineRegisterInfo &MRI = MF->getRegInfo();
11701 const TargetRegisterClass *AddrRegClass =
11702 getRegClassFor(Is64Bit ? MVT::i64:MVT::i32);
11704 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
11705 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
11706 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
11707 sizeVReg = MI->getOperand(1).getReg(),
11708 physSPReg = Is64Bit ? X86::RSP : X86::ESP;
11710 MachineFunction::iterator MBBIter = BB;
11713 MF->insert(MBBIter, bumpMBB);
11714 MF->insert(MBBIter, mallocMBB);
11715 MF->insert(MBBIter, continueMBB);
11717 continueMBB->splice(continueMBB->begin(), BB, llvm::next
11718 (MachineBasicBlock::iterator(MI)), BB->end());
11719 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
11721 // Add code to the main basic block to check if the stack limit has been hit,
11722 // and if so, jump to mallocMBB otherwise to bumpMBB.
11723 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
11724 BuildMI(BB, DL, TII->get(Is64Bit ? X86::SUB64rr:X86::SUB32rr), tmpSPVReg)
11725 .addReg(tmpSPVReg).addReg(sizeVReg);
11726 BuildMI(BB, DL, TII->get(Is64Bit ? X86::CMP64mr:X86::CMP32mr))
11727 .addReg(0).addImm(0).addReg(0).addImm(TlsOffset).addReg(TlsReg)
11728 .addReg(tmpSPVReg);
11729 BuildMI(BB, DL, TII->get(X86::JG_4)).addMBB(mallocMBB);
11731 // bumpMBB simply decreases the stack pointer, since we know the current
11732 // stacklet has enough space.
11733 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
11734 .addReg(tmpSPVReg);
11735 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
11736 .addReg(tmpSPVReg);
11737 BuildMI(bumpMBB, DL, TII->get(X86::JMP_4)).addMBB(continueMBB);
11739 // Calls into a routine in libgcc to allocate more space from the heap.
11741 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
11743 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
11744 .addExternalSymbol("__morestack_allocate_stack_space").addReg(X86::RDI);
11746 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
11748 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
11749 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
11750 .addExternalSymbol("__morestack_allocate_stack_space");
11754 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
11757 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
11758 .addReg(Is64Bit ? X86::RAX : X86::EAX);
11759 BuildMI(mallocMBB, DL, TII->get(X86::JMP_4)).addMBB(continueMBB);
11761 // Set up the CFG correctly.
11762 BB->addSuccessor(bumpMBB);
11763 BB->addSuccessor(mallocMBB);
11764 mallocMBB->addSuccessor(continueMBB);
11765 bumpMBB->addSuccessor(continueMBB);
11767 // Take care of the PHI nodes.
11768 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
11769 MI->getOperand(0).getReg())
11770 .addReg(mallocPtrVReg).addMBB(mallocMBB)
11771 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
11773 // Delete the original pseudo instruction.
11774 MI->eraseFromParent();
11777 return continueMBB;
11780 MachineBasicBlock *
11781 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
11782 MachineBasicBlock *BB) const {
11783 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11784 DebugLoc DL = MI->getDebugLoc();
11786 assert(!Subtarget->isTargetEnvMacho());
11788 // The lowering is pretty easy: we're just emitting the call to _alloca. The
11789 // non-trivial part is impdef of ESP.
11791 if (Subtarget->isTargetWin64()) {
11792 if (Subtarget->isTargetCygMing()) {
11793 // ___chkstk(Mingw64):
11794 // Clobbers R10, R11, RAX and EFLAGS.
11796 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
11797 .addExternalSymbol("___chkstk")
11798 .addReg(X86::RAX, RegState::Implicit)
11799 .addReg(X86::RSP, RegState::Implicit)
11800 .addReg(X86::RAX, RegState::Define | RegState::Implicit)
11801 .addReg(X86::RSP, RegState::Define | RegState::Implicit)
11802 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
11804 // __chkstk(MSVCRT): does not update stack pointer.
11805 // Clobbers R10, R11 and EFLAGS.
11806 // FIXME: RAX(allocated size) might be reused and not killed.
11807 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
11808 .addExternalSymbol("__chkstk")
11809 .addReg(X86::RAX, RegState::Implicit)
11810 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
11811 // RAX has the offset to subtracted from RSP.
11812 BuildMI(*BB, MI, DL, TII->get(X86::SUB64rr), X86::RSP)
11817 const char *StackProbeSymbol =
11818 Subtarget->isTargetWindows() ? "_chkstk" : "_alloca";
11820 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
11821 .addExternalSymbol(StackProbeSymbol)
11822 .addReg(X86::EAX, RegState::Implicit)
11823 .addReg(X86::ESP, RegState::Implicit)
11824 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
11825 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
11826 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
11829 MI->eraseFromParent(); // The pseudo instruction is gone now.
11833 MachineBasicBlock *
11834 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
11835 MachineBasicBlock *BB) const {
11836 // This is pretty easy. We're taking the value that we received from
11837 // our load from the relocation, sticking it in either RDI (x86-64)
11838 // or EAX and doing an indirect call. The return value will then
11839 // be in the normal return register.
11840 const X86InstrInfo *TII
11841 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
11842 DebugLoc DL = MI->getDebugLoc();
11843 MachineFunction *F = BB->getParent();
11845 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
11846 assert(MI->getOperand(3).isGlobal() && "This should be a global");
11848 if (Subtarget->is64Bit()) {
11849 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
11850 TII->get(X86::MOV64rm), X86::RDI)
11852 .addImm(0).addReg(0)
11853 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
11854 MI->getOperand(3).getTargetFlags())
11856 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
11857 addDirectMem(MIB, X86::RDI);
11858 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
11859 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
11860 TII->get(X86::MOV32rm), X86::EAX)
11862 .addImm(0).addReg(0)
11863 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
11864 MI->getOperand(3).getTargetFlags())
11866 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
11867 addDirectMem(MIB, X86::EAX);
11869 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
11870 TII->get(X86::MOV32rm), X86::EAX)
11871 .addReg(TII->getGlobalBaseReg(F))
11872 .addImm(0).addReg(0)
11873 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
11874 MI->getOperand(3).getTargetFlags())
11876 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
11877 addDirectMem(MIB, X86::EAX);
11880 MI->eraseFromParent(); // The pseudo instruction is gone now.
11884 MachineBasicBlock *
11885 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
11886 MachineBasicBlock *BB) const {
11887 switch (MI->getOpcode()) {
11888 default: assert(false && "Unexpected instr type to insert");
11889 case X86::TAILJMPd64:
11890 case X86::TAILJMPr64:
11891 case X86::TAILJMPm64:
11892 assert(!"TAILJMP64 would not be touched here.");
11893 case X86::TCRETURNdi64:
11894 case X86::TCRETURNri64:
11895 case X86::TCRETURNmi64:
11896 // Defs of TCRETURNxx64 has Win64's callee-saved registers, as subset.
11897 // On AMD64, additional defs should be added before register allocation.
11898 if (!Subtarget->isTargetWin64()) {
11899 MI->addRegisterDefined(X86::RSI);
11900 MI->addRegisterDefined(X86::RDI);
11901 MI->addRegisterDefined(X86::XMM6);
11902 MI->addRegisterDefined(X86::XMM7);
11903 MI->addRegisterDefined(X86::XMM8);
11904 MI->addRegisterDefined(X86::XMM9);
11905 MI->addRegisterDefined(X86::XMM10);
11906 MI->addRegisterDefined(X86::XMM11);
11907 MI->addRegisterDefined(X86::XMM12);
11908 MI->addRegisterDefined(X86::XMM13);
11909 MI->addRegisterDefined(X86::XMM14);
11910 MI->addRegisterDefined(X86::XMM15);
11913 case X86::WIN_ALLOCA:
11914 return EmitLoweredWinAlloca(MI, BB);
11915 case X86::SEG_ALLOCA_32:
11916 return EmitLoweredSegAlloca(MI, BB, false);
11917 case X86::SEG_ALLOCA_64:
11918 return EmitLoweredSegAlloca(MI, BB, true);
11919 case X86::TLSCall_32:
11920 case X86::TLSCall_64:
11921 return EmitLoweredTLSCall(MI, BB);
11922 case X86::CMOV_GR8:
11923 case X86::CMOV_FR32:
11924 case X86::CMOV_FR64:
11925 case X86::CMOV_V4F32:
11926 case X86::CMOV_V2F64:
11927 case X86::CMOV_V2I64:
11928 case X86::CMOV_V8F32:
11929 case X86::CMOV_V4F64:
11930 case X86::CMOV_V4I64:
11931 case X86::CMOV_GR16:
11932 case X86::CMOV_GR32:
11933 case X86::CMOV_RFP32:
11934 case X86::CMOV_RFP64:
11935 case X86::CMOV_RFP80:
11936 return EmitLoweredSelect(MI, BB);
11938 case X86::FP32_TO_INT16_IN_MEM:
11939 case X86::FP32_TO_INT32_IN_MEM:
11940 case X86::FP32_TO_INT64_IN_MEM:
11941 case X86::FP64_TO_INT16_IN_MEM:
11942 case X86::FP64_TO_INT32_IN_MEM:
11943 case X86::FP64_TO_INT64_IN_MEM:
11944 case X86::FP80_TO_INT16_IN_MEM:
11945 case X86::FP80_TO_INT32_IN_MEM:
11946 case X86::FP80_TO_INT64_IN_MEM: {
11947 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11948 DebugLoc DL = MI->getDebugLoc();
11950 // Change the floating point control register to use "round towards zero"
11951 // mode when truncating to an integer value.
11952 MachineFunction *F = BB->getParent();
11953 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
11954 addFrameReference(BuildMI(*BB, MI, DL,
11955 TII->get(X86::FNSTCW16m)), CWFrameIdx);
11957 // Load the old value of the high byte of the control word...
11959 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
11960 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
11963 // Set the high part to be round to zero...
11964 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
11967 // Reload the modified control word now...
11968 addFrameReference(BuildMI(*BB, MI, DL,
11969 TII->get(X86::FLDCW16m)), CWFrameIdx);
11971 // Restore the memory image of control word to original value
11972 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
11975 // Get the X86 opcode to use.
11977 switch (MI->getOpcode()) {
11978 default: llvm_unreachable("illegal opcode!");
11979 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
11980 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
11981 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
11982 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
11983 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
11984 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
11985 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
11986 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
11987 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
11991 MachineOperand &Op = MI->getOperand(0);
11993 AM.BaseType = X86AddressMode::RegBase;
11994 AM.Base.Reg = Op.getReg();
11996 AM.BaseType = X86AddressMode::FrameIndexBase;
11997 AM.Base.FrameIndex = Op.getIndex();
11999 Op = MI->getOperand(1);
12001 AM.Scale = Op.getImm();
12002 Op = MI->getOperand(2);
12004 AM.IndexReg = Op.getImm();
12005 Op = MI->getOperand(3);
12006 if (Op.isGlobal()) {
12007 AM.GV = Op.getGlobal();
12009 AM.Disp = Op.getImm();
12011 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
12012 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
12014 // Reload the original control word now.
12015 addFrameReference(BuildMI(*BB, MI, DL,
12016 TII->get(X86::FLDCW16m)), CWFrameIdx);
12018 MI->eraseFromParent(); // The pseudo instruction is gone now.
12021 // String/text processing lowering.
12022 case X86::PCMPISTRM128REG:
12023 case X86::VPCMPISTRM128REG:
12024 return EmitPCMP(MI, BB, 3, false /* in-mem */);
12025 case X86::PCMPISTRM128MEM:
12026 case X86::VPCMPISTRM128MEM:
12027 return EmitPCMP(MI, BB, 3, true /* in-mem */);
12028 case X86::PCMPESTRM128REG:
12029 case X86::VPCMPESTRM128REG:
12030 return EmitPCMP(MI, BB, 5, false /* in mem */);
12031 case X86::PCMPESTRM128MEM:
12032 case X86::VPCMPESTRM128MEM:
12033 return EmitPCMP(MI, BB, 5, true /* in mem */);
12035 // Thread synchronization.
12037 return EmitMonitor(MI, BB);
12039 return EmitMwait(MI, BB);
12041 // Atomic Lowering.
12042 case X86::ATOMAND32:
12043 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
12044 X86::AND32ri, X86::MOV32rm,
12046 X86::NOT32r, X86::EAX,
12047 X86::GR32RegisterClass);
12048 case X86::ATOMOR32:
12049 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
12050 X86::OR32ri, X86::MOV32rm,
12052 X86::NOT32r, X86::EAX,
12053 X86::GR32RegisterClass);
12054 case X86::ATOMXOR32:
12055 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
12056 X86::XOR32ri, X86::MOV32rm,
12058 X86::NOT32r, X86::EAX,
12059 X86::GR32RegisterClass);
12060 case X86::ATOMNAND32:
12061 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
12062 X86::AND32ri, X86::MOV32rm,
12064 X86::NOT32r, X86::EAX,
12065 X86::GR32RegisterClass, true);
12066 case X86::ATOMMIN32:
12067 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
12068 case X86::ATOMMAX32:
12069 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
12070 case X86::ATOMUMIN32:
12071 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
12072 case X86::ATOMUMAX32:
12073 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
12075 case X86::ATOMAND16:
12076 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
12077 X86::AND16ri, X86::MOV16rm,
12079 X86::NOT16r, X86::AX,
12080 X86::GR16RegisterClass);
12081 case X86::ATOMOR16:
12082 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
12083 X86::OR16ri, X86::MOV16rm,
12085 X86::NOT16r, X86::AX,
12086 X86::GR16RegisterClass);
12087 case X86::ATOMXOR16:
12088 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
12089 X86::XOR16ri, X86::MOV16rm,
12091 X86::NOT16r, X86::AX,
12092 X86::GR16RegisterClass);
12093 case X86::ATOMNAND16:
12094 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
12095 X86::AND16ri, X86::MOV16rm,
12097 X86::NOT16r, X86::AX,
12098 X86::GR16RegisterClass, true);
12099 case X86::ATOMMIN16:
12100 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
12101 case X86::ATOMMAX16:
12102 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
12103 case X86::ATOMUMIN16:
12104 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
12105 case X86::ATOMUMAX16:
12106 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
12108 case X86::ATOMAND8:
12109 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
12110 X86::AND8ri, X86::MOV8rm,
12112 X86::NOT8r, X86::AL,
12113 X86::GR8RegisterClass);
12115 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
12116 X86::OR8ri, X86::MOV8rm,
12118 X86::NOT8r, X86::AL,
12119 X86::GR8RegisterClass);
12120 case X86::ATOMXOR8:
12121 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
12122 X86::XOR8ri, X86::MOV8rm,
12124 X86::NOT8r, X86::AL,
12125 X86::GR8RegisterClass);
12126 case X86::ATOMNAND8:
12127 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
12128 X86::AND8ri, X86::MOV8rm,
12130 X86::NOT8r, X86::AL,
12131 X86::GR8RegisterClass, true);
12132 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
12133 // This group is for 64-bit host.
12134 case X86::ATOMAND64:
12135 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
12136 X86::AND64ri32, X86::MOV64rm,
12138 X86::NOT64r, X86::RAX,
12139 X86::GR64RegisterClass);
12140 case X86::ATOMOR64:
12141 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
12142 X86::OR64ri32, X86::MOV64rm,
12144 X86::NOT64r, X86::RAX,
12145 X86::GR64RegisterClass);
12146 case X86::ATOMXOR64:
12147 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
12148 X86::XOR64ri32, X86::MOV64rm,
12150 X86::NOT64r, X86::RAX,
12151 X86::GR64RegisterClass);
12152 case X86::ATOMNAND64:
12153 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
12154 X86::AND64ri32, X86::MOV64rm,
12156 X86::NOT64r, X86::RAX,
12157 X86::GR64RegisterClass, true);
12158 case X86::ATOMMIN64:
12159 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
12160 case X86::ATOMMAX64:
12161 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
12162 case X86::ATOMUMIN64:
12163 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
12164 case X86::ATOMUMAX64:
12165 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
12167 // This group does 64-bit operations on a 32-bit host.
12168 case X86::ATOMAND6432:
12169 return EmitAtomicBit6432WithCustomInserter(MI, BB,
12170 X86::AND32rr, X86::AND32rr,
12171 X86::AND32ri, X86::AND32ri,
12173 case X86::ATOMOR6432:
12174 return EmitAtomicBit6432WithCustomInserter(MI, BB,
12175 X86::OR32rr, X86::OR32rr,
12176 X86::OR32ri, X86::OR32ri,
12178 case X86::ATOMXOR6432:
12179 return EmitAtomicBit6432WithCustomInserter(MI, BB,
12180 X86::XOR32rr, X86::XOR32rr,
12181 X86::XOR32ri, X86::XOR32ri,
12183 case X86::ATOMNAND6432:
12184 return EmitAtomicBit6432WithCustomInserter(MI, BB,
12185 X86::AND32rr, X86::AND32rr,
12186 X86::AND32ri, X86::AND32ri,
12188 case X86::ATOMADD6432:
12189 return EmitAtomicBit6432WithCustomInserter(MI, BB,
12190 X86::ADD32rr, X86::ADC32rr,
12191 X86::ADD32ri, X86::ADC32ri,
12193 case X86::ATOMSUB6432:
12194 return EmitAtomicBit6432WithCustomInserter(MI, BB,
12195 X86::SUB32rr, X86::SBB32rr,
12196 X86::SUB32ri, X86::SBB32ri,
12198 case X86::ATOMSWAP6432:
12199 return EmitAtomicBit6432WithCustomInserter(MI, BB,
12200 X86::MOV32rr, X86::MOV32rr,
12201 X86::MOV32ri, X86::MOV32ri,
12203 case X86::VASTART_SAVE_XMM_REGS:
12204 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
12206 case X86::VAARG_64:
12207 return EmitVAARG64WithCustomInserter(MI, BB);
12211 //===----------------------------------------------------------------------===//
12212 // X86 Optimization Hooks
12213 //===----------------------------------------------------------------------===//
12215 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
12219 const SelectionDAG &DAG,
12220 unsigned Depth) const {
12221 unsigned Opc = Op.getOpcode();
12222 assert((Opc >= ISD::BUILTIN_OP_END ||
12223 Opc == ISD::INTRINSIC_WO_CHAIN ||
12224 Opc == ISD::INTRINSIC_W_CHAIN ||
12225 Opc == ISD::INTRINSIC_VOID) &&
12226 "Should use MaskedValueIsZero if you don't know whether Op"
12227 " is a target node!");
12229 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
12243 // These nodes' second result is a boolean.
12244 if (Op.getResNo() == 0)
12247 case X86ISD::SETCC:
12248 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
12249 Mask.getBitWidth() - 1);
12254 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
12255 unsigned Depth) const {
12256 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
12257 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
12258 return Op.getValueType().getScalarType().getSizeInBits();
12264 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
12265 /// node is a GlobalAddress + offset.
12266 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
12267 const GlobalValue* &GA,
12268 int64_t &Offset) const {
12269 if (N->getOpcode() == X86ISD::Wrapper) {
12270 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
12271 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
12272 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
12276 return TargetLowering::isGAPlusOffset(N, GA, Offset);
12279 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
12280 /// same as extracting the high 128-bit part of 256-bit vector and then
12281 /// inserting the result into the low part of a new 256-bit vector
12282 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
12283 EVT VT = SVOp->getValueType(0);
12284 int NumElems = VT.getVectorNumElements();
12286 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
12287 for (int i = 0, j = NumElems/2; i < NumElems/2; ++i, ++j)
12288 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
12289 SVOp->getMaskElt(j) >= 0)
12295 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
12296 /// same as extracting the low 128-bit part of 256-bit vector and then
12297 /// inserting the result into the high part of a new 256-bit vector
12298 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
12299 EVT VT = SVOp->getValueType(0);
12300 int NumElems = VT.getVectorNumElements();
12302 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
12303 for (int i = NumElems/2, j = 0; i < NumElems; ++i, ++j)
12304 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
12305 SVOp->getMaskElt(j) >= 0)
12311 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
12312 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
12313 TargetLowering::DAGCombinerInfo &DCI) {
12314 DebugLoc dl = N->getDebugLoc();
12315 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
12316 SDValue V1 = SVOp->getOperand(0);
12317 SDValue V2 = SVOp->getOperand(1);
12318 EVT VT = SVOp->getValueType(0);
12319 int NumElems = VT.getVectorNumElements();
12321 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
12322 V2.getOpcode() == ISD::CONCAT_VECTORS) {
12326 // V UNDEF BUILD_VECTOR UNDEF
12328 // CONCAT_VECTOR CONCAT_VECTOR
12331 // RESULT: V + zero extended
12333 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
12334 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
12335 V1.getOperand(1).getOpcode() != ISD::UNDEF)
12338 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
12341 // To match the shuffle mask, the first half of the mask should
12342 // be exactly the first vector, and all the rest a splat with the
12343 // first element of the second one.
12344 for (int i = 0; i < NumElems/2; ++i)
12345 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
12346 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
12349 // Emit a zeroed vector and insert the desired subvector on its
12351 SDValue Zeros = getZeroVector(VT, true /* HasSSE2 */, DAG, dl);
12352 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0),
12353 DAG.getConstant(0, MVT::i32), DAG, dl);
12354 return DCI.CombineTo(N, InsV);
12357 //===--------------------------------------------------------------------===//
12358 // Combine some shuffles into subvector extracts and inserts:
12361 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
12362 if (isShuffleHigh128VectorInsertLow(SVOp)) {
12363 SDValue V = Extract128BitVector(V1, DAG.getConstant(NumElems/2, MVT::i32),
12365 SDValue InsV = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT),
12366 V, DAG.getConstant(0, MVT::i32), DAG, dl);
12367 return DCI.CombineTo(N, InsV);
12370 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
12371 if (isShuffleLow128VectorInsertHigh(SVOp)) {
12372 SDValue V = Extract128BitVector(V1, DAG.getConstant(0, MVT::i32), DAG, dl);
12373 SDValue InsV = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT),
12374 V, DAG.getConstant(NumElems/2, MVT::i32), DAG, dl);
12375 return DCI.CombineTo(N, InsV);
12381 /// PerformShuffleCombine - Performs several different shuffle combines.
12382 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
12383 TargetLowering::DAGCombinerInfo &DCI,
12384 const X86Subtarget *Subtarget) {
12385 DebugLoc dl = N->getDebugLoc();
12386 EVT VT = N->getValueType(0);
12388 // Don't create instructions with illegal types after legalize types has run.
12389 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12390 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
12393 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
12394 if (Subtarget->hasAVX() && VT.getSizeInBits() == 256 &&
12395 N->getOpcode() == ISD::VECTOR_SHUFFLE)
12396 return PerformShuffleCombine256(N, DAG, DCI);
12398 // Only handle 128 wide vector from here on.
12399 if (VT.getSizeInBits() != 128)
12402 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
12403 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
12404 // consecutive, non-overlapping, and in the right order.
12405 SmallVector<SDValue, 16> Elts;
12406 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
12407 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
12409 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
12412 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
12413 /// generation and convert it from being a bunch of shuffles and extracts
12414 /// to a simple store and scalar loads to extract the elements.
12415 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
12416 const TargetLowering &TLI) {
12417 SDValue InputVector = N->getOperand(0);
12419 // Only operate on vectors of 4 elements, where the alternative shuffling
12420 // gets to be more expensive.
12421 if (InputVector.getValueType() != MVT::v4i32)
12424 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
12425 // single use which is a sign-extend or zero-extend, and all elements are
12427 SmallVector<SDNode *, 4> Uses;
12428 unsigned ExtractedElements = 0;
12429 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
12430 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
12431 if (UI.getUse().getResNo() != InputVector.getResNo())
12434 SDNode *Extract = *UI;
12435 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
12438 if (Extract->getValueType(0) != MVT::i32)
12440 if (!Extract->hasOneUse())
12442 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
12443 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
12445 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
12448 // Record which element was extracted.
12449 ExtractedElements |=
12450 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
12452 Uses.push_back(Extract);
12455 // If not all the elements were used, this may not be worthwhile.
12456 if (ExtractedElements != 15)
12459 // Ok, we've now decided to do the transformation.
12460 DebugLoc dl = InputVector.getDebugLoc();
12462 // Store the value to a temporary stack slot.
12463 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
12464 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
12465 MachinePointerInfo(), false, false, 0);
12467 // Replace each use (extract) with a load of the appropriate element.
12468 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
12469 UE = Uses.end(); UI != UE; ++UI) {
12470 SDNode *Extract = *UI;
12472 // cOMpute the element's address.
12473 SDValue Idx = Extract->getOperand(1);
12475 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
12476 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
12477 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
12479 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
12480 StackPtr, OffsetVal);
12482 // Load the scalar.
12483 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
12484 ScalarAddr, MachinePointerInfo(),
12487 // Replace the exact with the load.
12488 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
12491 // The replacement was made in place; don't return anything.
12495 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
12496 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
12497 const X86Subtarget *Subtarget) {
12498 DebugLoc DL = N->getDebugLoc();
12499 SDValue Cond = N->getOperand(0);
12500 // Get the LHS/RHS of the select.
12501 SDValue LHS = N->getOperand(1);
12502 SDValue RHS = N->getOperand(2);
12504 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
12505 // instructions match the semantics of the common C idiom x<y?x:y but not
12506 // x<=y?x:y, because of how they handle negative zero (which can be
12507 // ignored in unsafe-math mode).
12508 if (Subtarget->hasSSE2() &&
12509 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
12510 Cond.getOpcode() == ISD::SETCC) {
12511 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
12513 unsigned Opcode = 0;
12514 // Check for x CC y ? x : y.
12515 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
12516 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
12520 // Converting this to a min would handle NaNs incorrectly, and swapping
12521 // the operands would cause it to handle comparisons between positive
12522 // and negative zero incorrectly.
12523 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
12524 if (!UnsafeFPMath &&
12525 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
12527 std::swap(LHS, RHS);
12529 Opcode = X86ISD::FMIN;
12532 // Converting this to a min would handle comparisons between positive
12533 // and negative zero incorrectly.
12534 if (!UnsafeFPMath &&
12535 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
12537 Opcode = X86ISD::FMIN;
12540 // Converting this to a min would handle both negative zeros and NaNs
12541 // incorrectly, but we can swap the operands to fix both.
12542 std::swap(LHS, RHS);
12546 Opcode = X86ISD::FMIN;
12550 // Converting this to a max would handle comparisons between positive
12551 // and negative zero incorrectly.
12552 if (!UnsafeFPMath &&
12553 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
12555 Opcode = X86ISD::FMAX;
12558 // Converting this to a max would handle NaNs incorrectly, and swapping
12559 // the operands would cause it to handle comparisons between positive
12560 // and negative zero incorrectly.
12561 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
12562 if (!UnsafeFPMath &&
12563 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
12565 std::swap(LHS, RHS);
12567 Opcode = X86ISD::FMAX;
12570 // Converting this to a max would handle both negative zeros and NaNs
12571 // incorrectly, but we can swap the operands to fix both.
12572 std::swap(LHS, RHS);
12576 Opcode = X86ISD::FMAX;
12579 // Check for x CC y ? y : x -- a min/max with reversed arms.
12580 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
12581 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
12585 // Converting this to a min would handle comparisons between positive
12586 // and negative zero incorrectly, and swapping the operands would
12587 // cause it to handle NaNs incorrectly.
12588 if (!UnsafeFPMath &&
12589 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
12590 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
12592 std::swap(LHS, RHS);
12594 Opcode = X86ISD::FMIN;
12597 // Converting this to a min would handle NaNs incorrectly.
12598 if (!UnsafeFPMath &&
12599 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
12601 Opcode = X86ISD::FMIN;
12604 // Converting this to a min would handle both negative zeros and NaNs
12605 // incorrectly, but we can swap the operands to fix both.
12606 std::swap(LHS, RHS);
12610 Opcode = X86ISD::FMIN;
12614 // Converting this to a max would handle NaNs incorrectly.
12615 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
12617 Opcode = X86ISD::FMAX;
12620 // Converting this to a max would handle comparisons between positive
12621 // and negative zero incorrectly, and swapping the operands would
12622 // cause it to handle NaNs incorrectly.
12623 if (!UnsafeFPMath &&
12624 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
12625 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
12627 std::swap(LHS, RHS);
12629 Opcode = X86ISD::FMAX;
12632 // Converting this to a max would handle both negative zeros and NaNs
12633 // incorrectly, but we can swap the operands to fix both.
12634 std::swap(LHS, RHS);
12638 Opcode = X86ISD::FMAX;
12644 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
12647 // If this is a select between two integer constants, try to do some
12649 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
12650 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
12651 // Don't do this for crazy integer types.
12652 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
12653 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
12654 // so that TrueC (the true value) is larger than FalseC.
12655 bool NeedsCondInvert = false;
12657 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
12658 // Efficiently invertible.
12659 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
12660 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
12661 isa<ConstantSDNode>(Cond.getOperand(1))))) {
12662 NeedsCondInvert = true;
12663 std::swap(TrueC, FalseC);
12666 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
12667 if (FalseC->getAPIntValue() == 0 &&
12668 TrueC->getAPIntValue().isPowerOf2()) {
12669 if (NeedsCondInvert) // Invert the condition if needed.
12670 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
12671 DAG.getConstant(1, Cond.getValueType()));
12673 // Zero extend the condition if needed.
12674 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
12676 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
12677 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
12678 DAG.getConstant(ShAmt, MVT::i8));
12681 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
12682 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
12683 if (NeedsCondInvert) // Invert the condition if needed.
12684 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
12685 DAG.getConstant(1, Cond.getValueType()));
12687 // Zero extend the condition if needed.
12688 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
12689 FalseC->getValueType(0), Cond);
12690 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
12691 SDValue(FalseC, 0));
12694 // Optimize cases that will turn into an LEA instruction. This requires
12695 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
12696 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
12697 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
12698 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
12700 bool isFastMultiplier = false;
12702 switch ((unsigned char)Diff) {
12704 case 1: // result = add base, cond
12705 case 2: // result = lea base( , cond*2)
12706 case 3: // result = lea base(cond, cond*2)
12707 case 4: // result = lea base( , cond*4)
12708 case 5: // result = lea base(cond, cond*4)
12709 case 8: // result = lea base( , cond*8)
12710 case 9: // result = lea base(cond, cond*8)
12711 isFastMultiplier = true;
12716 if (isFastMultiplier) {
12717 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
12718 if (NeedsCondInvert) // Invert the condition if needed.
12719 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
12720 DAG.getConstant(1, Cond.getValueType()));
12722 // Zero extend the condition if needed.
12723 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
12725 // Scale the condition by the difference.
12727 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
12728 DAG.getConstant(Diff, Cond.getValueType()));
12730 // Add the base if non-zero.
12731 if (FalseC->getAPIntValue() != 0)
12732 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
12733 SDValue(FalseC, 0));
12743 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
12744 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
12745 TargetLowering::DAGCombinerInfo &DCI) {
12746 DebugLoc DL = N->getDebugLoc();
12748 // If the flag operand isn't dead, don't touch this CMOV.
12749 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
12752 SDValue FalseOp = N->getOperand(0);
12753 SDValue TrueOp = N->getOperand(1);
12754 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
12755 SDValue Cond = N->getOperand(3);
12756 if (CC == X86::COND_E || CC == X86::COND_NE) {
12757 switch (Cond.getOpcode()) {
12761 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
12762 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
12763 return (CC == X86::COND_E) ? FalseOp : TrueOp;
12767 // If this is a select between two integer constants, try to do some
12768 // optimizations. Note that the operands are ordered the opposite of SELECT
12770 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
12771 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
12772 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
12773 // larger than FalseC (the false value).
12774 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
12775 CC = X86::GetOppositeBranchCondition(CC);
12776 std::swap(TrueC, FalseC);
12779 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
12780 // This is efficient for any integer data type (including i8/i16) and
12782 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
12783 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
12784 DAG.getConstant(CC, MVT::i8), Cond);
12786 // Zero extend the condition if needed.
12787 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
12789 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
12790 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
12791 DAG.getConstant(ShAmt, MVT::i8));
12792 if (N->getNumValues() == 2) // Dead flag value?
12793 return DCI.CombineTo(N, Cond, SDValue());
12797 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
12798 // for any integer data type, including i8/i16.
12799 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
12800 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
12801 DAG.getConstant(CC, MVT::i8), Cond);
12803 // Zero extend the condition if needed.
12804 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
12805 FalseC->getValueType(0), Cond);
12806 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
12807 SDValue(FalseC, 0));
12809 if (N->getNumValues() == 2) // Dead flag value?
12810 return DCI.CombineTo(N, Cond, SDValue());
12814 // Optimize cases that will turn into an LEA instruction. This requires
12815 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
12816 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
12817 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
12818 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
12820 bool isFastMultiplier = false;
12822 switch ((unsigned char)Diff) {
12824 case 1: // result = add base, cond
12825 case 2: // result = lea base( , cond*2)
12826 case 3: // result = lea base(cond, cond*2)
12827 case 4: // result = lea base( , cond*4)
12828 case 5: // result = lea base(cond, cond*4)
12829 case 8: // result = lea base( , cond*8)
12830 case 9: // result = lea base(cond, cond*8)
12831 isFastMultiplier = true;
12836 if (isFastMultiplier) {
12837 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
12838 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
12839 DAG.getConstant(CC, MVT::i8), Cond);
12840 // Zero extend the condition if needed.
12841 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
12843 // Scale the condition by the difference.
12845 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
12846 DAG.getConstant(Diff, Cond.getValueType()));
12848 // Add the base if non-zero.
12849 if (FalseC->getAPIntValue() != 0)
12850 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
12851 SDValue(FalseC, 0));
12852 if (N->getNumValues() == 2) // Dead flag value?
12853 return DCI.CombineTo(N, Cond, SDValue());
12863 /// PerformMulCombine - Optimize a single multiply with constant into two
12864 /// in order to implement it with two cheaper instructions, e.g.
12865 /// LEA + SHL, LEA + LEA.
12866 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
12867 TargetLowering::DAGCombinerInfo &DCI) {
12868 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
12871 EVT VT = N->getValueType(0);
12872 if (VT != MVT::i64)
12875 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
12878 uint64_t MulAmt = C->getZExtValue();
12879 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
12882 uint64_t MulAmt1 = 0;
12883 uint64_t MulAmt2 = 0;
12884 if ((MulAmt % 9) == 0) {
12886 MulAmt2 = MulAmt / 9;
12887 } else if ((MulAmt % 5) == 0) {
12889 MulAmt2 = MulAmt / 5;
12890 } else if ((MulAmt % 3) == 0) {
12892 MulAmt2 = MulAmt / 3;
12895 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
12896 DebugLoc DL = N->getDebugLoc();
12898 if (isPowerOf2_64(MulAmt2) &&
12899 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
12900 // If second multiplifer is pow2, issue it first. We want the multiply by
12901 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
12903 std::swap(MulAmt1, MulAmt2);
12906 if (isPowerOf2_64(MulAmt1))
12907 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
12908 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
12910 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
12911 DAG.getConstant(MulAmt1, VT));
12913 if (isPowerOf2_64(MulAmt2))
12914 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
12915 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
12917 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
12918 DAG.getConstant(MulAmt2, VT));
12920 // Do not add new nodes to DAG combiner worklist.
12921 DCI.CombineTo(N, NewMul, false);
12926 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
12927 SDValue N0 = N->getOperand(0);
12928 SDValue N1 = N->getOperand(1);
12929 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
12930 EVT VT = N0.getValueType();
12932 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
12933 // since the result of setcc_c is all zero's or all ones.
12934 if (N1C && N0.getOpcode() == ISD::AND &&
12935 N0.getOperand(1).getOpcode() == ISD::Constant) {
12936 SDValue N00 = N0.getOperand(0);
12937 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
12938 ((N00.getOpcode() == ISD::ANY_EXTEND ||
12939 N00.getOpcode() == ISD::ZERO_EXTEND) &&
12940 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
12941 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
12942 APInt ShAmt = N1C->getAPIntValue();
12943 Mask = Mask.shl(ShAmt);
12945 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
12946 N00, DAG.getConstant(Mask, VT));
12953 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
12955 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
12956 const X86Subtarget *Subtarget) {
12957 EVT VT = N->getValueType(0);
12958 if (!VT.isVector() && VT.isInteger() &&
12959 N->getOpcode() == ISD::SHL)
12960 return PerformSHLCombine(N, DAG);
12962 // On X86 with SSE2 support, we can transform this to a vector shift if
12963 // all elements are shifted by the same amount. We can't do this in legalize
12964 // because the a constant vector is typically transformed to a constant pool
12965 // so we have no knowledge of the shift amount.
12966 if (!(Subtarget->hasSSE2() || Subtarget->hasAVX()))
12969 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
12972 SDValue ShAmtOp = N->getOperand(1);
12973 EVT EltVT = VT.getVectorElementType();
12974 DebugLoc DL = N->getDebugLoc();
12975 SDValue BaseShAmt = SDValue();
12976 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
12977 unsigned NumElts = VT.getVectorNumElements();
12979 for (; i != NumElts; ++i) {
12980 SDValue Arg = ShAmtOp.getOperand(i);
12981 if (Arg.getOpcode() == ISD::UNDEF) continue;
12985 for (; i != NumElts; ++i) {
12986 SDValue Arg = ShAmtOp.getOperand(i);
12987 if (Arg.getOpcode() == ISD::UNDEF) continue;
12988 if (Arg != BaseShAmt) {
12992 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
12993 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
12994 SDValue InVec = ShAmtOp.getOperand(0);
12995 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
12996 unsigned NumElts = InVec.getValueType().getVectorNumElements();
12998 for (; i != NumElts; ++i) {
12999 SDValue Arg = InVec.getOperand(i);
13000 if (Arg.getOpcode() == ISD::UNDEF) continue;
13004 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
13005 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
13006 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
13007 if (C->getZExtValue() == SplatIdx)
13008 BaseShAmt = InVec.getOperand(1);
13011 if (BaseShAmt.getNode() == 0)
13012 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
13013 DAG.getIntPtrConstant(0));
13017 // The shift amount is an i32.
13018 if (EltVT.bitsGT(MVT::i32))
13019 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
13020 else if (EltVT.bitsLT(MVT::i32))
13021 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
13023 // The shift amount is identical so we can do a vector shift.
13024 SDValue ValOp = N->getOperand(0);
13025 switch (N->getOpcode()) {
13027 llvm_unreachable("Unknown shift opcode!");
13030 if (VT == MVT::v2i64)
13031 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
13032 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
13034 if (VT == MVT::v4i32)
13035 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
13036 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
13038 if (VT == MVT::v8i16)
13039 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
13040 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
13044 if (VT == MVT::v4i32)
13045 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
13046 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
13048 if (VT == MVT::v8i16)
13049 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
13050 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
13054 if (VT == MVT::v2i64)
13055 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
13056 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
13058 if (VT == MVT::v4i32)
13059 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
13060 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
13062 if (VT == MVT::v8i16)
13063 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
13064 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
13072 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
13073 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
13074 // and friends. Likewise for OR -> CMPNEQSS.
13075 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
13076 TargetLowering::DAGCombinerInfo &DCI,
13077 const X86Subtarget *Subtarget) {
13080 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
13081 // we're requiring SSE2 for both.
13082 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
13083 SDValue N0 = N->getOperand(0);
13084 SDValue N1 = N->getOperand(1);
13085 SDValue CMP0 = N0->getOperand(1);
13086 SDValue CMP1 = N1->getOperand(1);
13087 DebugLoc DL = N->getDebugLoc();
13089 // The SETCCs should both refer to the same CMP.
13090 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
13093 SDValue CMP00 = CMP0->getOperand(0);
13094 SDValue CMP01 = CMP0->getOperand(1);
13095 EVT VT = CMP00.getValueType();
13097 if (VT == MVT::f32 || VT == MVT::f64) {
13098 bool ExpectingFlags = false;
13099 // Check for any users that want flags:
13100 for (SDNode::use_iterator UI = N->use_begin(),
13102 !ExpectingFlags && UI != UE; ++UI)
13103 switch (UI->getOpcode()) {
13108 ExpectingFlags = true;
13110 case ISD::CopyToReg:
13111 case ISD::SIGN_EXTEND:
13112 case ISD::ZERO_EXTEND:
13113 case ISD::ANY_EXTEND:
13117 if (!ExpectingFlags) {
13118 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
13119 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
13121 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
13122 X86::CondCode tmp = cc0;
13127 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
13128 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
13129 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
13130 X86ISD::NodeType NTOperator = is64BitFP ?
13131 X86ISD::FSETCCsd : X86ISD::FSETCCss;
13132 // FIXME: need symbolic constants for these magic numbers.
13133 // See X86ATTInstPrinter.cpp:printSSECC().
13134 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
13135 SDValue OnesOrZeroesF = DAG.getNode(NTOperator, DL, MVT::f32, CMP00, CMP01,
13136 DAG.getConstant(x86cc, MVT::i8));
13137 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, MVT::i32,
13139 SDValue ANDed = DAG.getNode(ISD::AND, DL, MVT::i32, OnesOrZeroesI,
13140 DAG.getConstant(1, MVT::i32));
13141 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
13142 return OneBitOfTruth;
13150 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
13151 /// so it can be folded inside ANDNP.
13152 static bool CanFoldXORWithAllOnes(const SDNode *N) {
13153 EVT VT = N->getValueType(0);
13155 // Match direct AllOnes for 128 and 256-bit vectors
13156 if (ISD::isBuildVectorAllOnes(N))
13159 // Look through a bit convert.
13160 if (N->getOpcode() == ISD::BITCAST)
13161 N = N->getOperand(0).getNode();
13163 // Sometimes the operand may come from a insert_subvector building a 256-bit
13165 if (VT.getSizeInBits() == 256 &&
13166 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
13167 SDValue V1 = N->getOperand(0);
13168 SDValue V2 = N->getOperand(1);
13170 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
13171 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
13172 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
13173 ISD::isBuildVectorAllOnes(V2.getNode()))
13180 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
13181 TargetLowering::DAGCombinerInfo &DCI,
13182 const X86Subtarget *Subtarget) {
13183 if (DCI.isBeforeLegalizeOps())
13186 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
13190 // Want to form ANDNP nodes:
13191 // 1) In the hopes of then easily combining them with OR and AND nodes
13192 // to form PBLEND/PSIGN.
13193 // 2) To match ANDN packed intrinsics
13194 EVT VT = N->getValueType(0);
13195 if (VT != MVT::v2i64 && VT != MVT::v4i64)
13198 SDValue N0 = N->getOperand(0);
13199 SDValue N1 = N->getOperand(1);
13200 DebugLoc DL = N->getDebugLoc();
13202 // Check LHS for vnot
13203 if (N0.getOpcode() == ISD::XOR &&
13204 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
13205 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
13206 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
13208 // Check RHS for vnot
13209 if (N1.getOpcode() == ISD::XOR &&
13210 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
13211 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
13212 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
13217 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
13218 TargetLowering::DAGCombinerInfo &DCI,
13219 const X86Subtarget *Subtarget) {
13220 if (DCI.isBeforeLegalizeOps())
13223 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
13227 EVT VT = N->getValueType(0);
13228 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64 && VT != MVT::v2i64)
13231 SDValue N0 = N->getOperand(0);
13232 SDValue N1 = N->getOperand(1);
13234 // look for psign/blend
13235 if (Subtarget->hasSSSE3()) {
13236 if (VT == MVT::v2i64) {
13237 // Canonicalize pandn to RHS
13238 if (N0.getOpcode() == X86ISD::ANDNP)
13240 // or (and (m, x), (pandn m, y))
13241 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
13242 SDValue Mask = N1.getOperand(0);
13243 SDValue X = N1.getOperand(1);
13245 if (N0.getOperand(0) == Mask)
13246 Y = N0.getOperand(1);
13247 if (N0.getOperand(1) == Mask)
13248 Y = N0.getOperand(0);
13250 // Check to see if the mask appeared in both the AND and ANDNP and
13254 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
13255 if (Mask.getOpcode() != ISD::BITCAST ||
13256 X.getOpcode() != ISD::BITCAST ||
13257 Y.getOpcode() != ISD::BITCAST)
13260 // Look through mask bitcast.
13261 Mask = Mask.getOperand(0);
13262 EVT MaskVT = Mask.getValueType();
13264 // Validate that the Mask operand is a vector sra node. The sra node
13265 // will be an intrinsic.
13266 if (Mask.getOpcode() != ISD::INTRINSIC_WO_CHAIN)
13269 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
13270 // there is no psrai.b
13271 switch (cast<ConstantSDNode>(Mask.getOperand(0))->getZExtValue()) {
13272 case Intrinsic::x86_sse2_psrai_w:
13273 case Intrinsic::x86_sse2_psrai_d:
13275 default: return SDValue();
13278 // Check that the SRA is all signbits.
13279 SDValue SraC = Mask.getOperand(2);
13280 unsigned SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
13281 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
13282 if ((SraAmt + 1) != EltBits)
13285 DebugLoc DL = N->getDebugLoc();
13287 // Now we know we at least have a plendvb with the mask val. See if
13288 // we can form a psignb/w/d.
13289 // psign = x.type == y.type == mask.type && y = sub(0, x);
13290 X = X.getOperand(0);
13291 Y = Y.getOperand(0);
13292 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
13293 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
13294 X.getValueType() == MaskVT && X.getValueType() == Y.getValueType()){
13297 case 8: Opc = X86ISD::PSIGNB; break;
13298 case 16: Opc = X86ISD::PSIGNW; break;
13299 case 32: Opc = X86ISD::PSIGND; break;
13303 SDValue Sign = DAG.getNode(Opc, DL, MaskVT, X, Mask.getOperand(1));
13304 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Sign);
13307 // PBLENDVB only available on SSE 4.1
13308 if (!Subtarget->hasSSE41())
13311 X = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, X);
13312 Y = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Y);
13313 Mask = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Mask);
13314 Mask = DAG.getNode(X86ISD::PBLENDVB, DL, MVT::v16i8, X, Y, Mask);
13315 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Mask);
13320 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
13321 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
13323 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
13325 if (!N0.hasOneUse() || !N1.hasOneUse())
13328 SDValue ShAmt0 = N0.getOperand(1);
13329 if (ShAmt0.getValueType() != MVT::i8)
13331 SDValue ShAmt1 = N1.getOperand(1);
13332 if (ShAmt1.getValueType() != MVT::i8)
13334 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
13335 ShAmt0 = ShAmt0.getOperand(0);
13336 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
13337 ShAmt1 = ShAmt1.getOperand(0);
13339 DebugLoc DL = N->getDebugLoc();
13340 unsigned Opc = X86ISD::SHLD;
13341 SDValue Op0 = N0.getOperand(0);
13342 SDValue Op1 = N1.getOperand(0);
13343 if (ShAmt0.getOpcode() == ISD::SUB) {
13344 Opc = X86ISD::SHRD;
13345 std::swap(Op0, Op1);
13346 std::swap(ShAmt0, ShAmt1);
13349 unsigned Bits = VT.getSizeInBits();
13350 if (ShAmt1.getOpcode() == ISD::SUB) {
13351 SDValue Sum = ShAmt1.getOperand(0);
13352 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
13353 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
13354 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
13355 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
13356 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
13357 return DAG.getNode(Opc, DL, VT,
13359 DAG.getNode(ISD::TRUNCATE, DL,
13362 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
13363 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
13365 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
13366 return DAG.getNode(Opc, DL, VT,
13367 N0.getOperand(0), N1.getOperand(0),
13368 DAG.getNode(ISD::TRUNCATE, DL,
13375 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
13376 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
13377 const X86Subtarget *Subtarget) {
13378 StoreSDNode *St = cast<StoreSDNode>(N);
13379 EVT VT = St->getValue().getValueType();
13380 EVT StVT = St->getMemoryVT();
13381 DebugLoc dl = St->getDebugLoc();
13382 SDValue StoredVal = St->getOperand(1);
13383 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13385 // If we are saving a concatination of two XMM registers, perform two stores.
13386 // This is better in Sandy Bridge cause one 256-bit mem op is done via two
13387 // 128-bit ones. If in the future the cost becomes only one memory access the
13388 // first version would be better.
13389 if (VT.getSizeInBits() == 256 &&
13390 StoredVal.getNode()->getOpcode() == ISD::CONCAT_VECTORS &&
13391 StoredVal.getNumOperands() == 2) {
13393 SDValue Value0 = StoredVal.getOperand(0);
13394 SDValue Value1 = StoredVal.getOperand(1);
13396 SDValue Stride = DAG.getConstant(16, TLI.getPointerTy());
13397 SDValue Ptr0 = St->getBasePtr();
13398 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
13400 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
13401 St->getPointerInfo(), St->isVolatile(),
13402 St->isNonTemporal(), St->getAlignment());
13403 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
13404 St->getPointerInfo(), St->isVolatile(),
13405 St->isNonTemporal(), St->getAlignment());
13406 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
13409 // Optimize trunc store (of multiple scalars) to shuffle and store.
13410 // First, pack all of the elements in one place. Next, store to memory
13411 // in fewer chunks.
13412 if (St->isTruncatingStore() && VT.isVector()) {
13413 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13414 unsigned NumElems = VT.getVectorNumElements();
13415 assert(StVT != VT && "Cannot truncate to the same type");
13416 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
13417 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
13419 // From, To sizes and ElemCount must be pow of two
13420 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
13421 // We are going to use the original vector elt for storing.
13422 // accumulated smaller vector elements must be a multiple of bigger size.
13423 if (0 != (NumElems * ToSz) % FromSz) return SDValue();
13424 unsigned SizeRatio = FromSz / ToSz;
13426 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
13428 // Create a type on which we perform the shuffle
13429 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
13430 StVT.getScalarType(), NumElems*SizeRatio);
13432 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
13434 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, St->getValue());
13435 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
13436 for (unsigned i = 0; i < NumElems; i++ ) ShuffleVec[i] = i * SizeRatio;
13438 // Can't shuffle using an illegal type
13439 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
13441 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
13442 DAG.getUNDEF(WideVec.getValueType()),
13443 ShuffleVec.data());
13444 // At this point all of the data is stored at the bottom of the
13445 // register. We now need to save it to mem.
13447 // Find the largest store unit
13448 MVT StoreType = MVT::i8;
13449 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
13450 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
13451 MVT Tp = (MVT::SimpleValueType)tp;
13452 if (TLI.isTypeLegal(Tp) && StoreType.getSizeInBits() < NumElems * ToSz)
13456 // Bitcast the original vector into a vector of store-size units
13457 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
13458 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
13459 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
13460 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, dl, StoreVecVT, Shuff);
13461 SmallVector<SDValue, 8> Chains;
13462 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
13463 TLI.getPointerTy());
13464 SDValue Ptr = St->getBasePtr();
13466 // Perform one or more big stores into memory.
13467 for (unsigned i = 0; i < (ToSz*NumElems)/StoreType.getSizeInBits() ; i++) {
13468 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
13469 StoreType, ShuffWide,
13470 DAG.getIntPtrConstant(i));
13471 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
13472 St->getPointerInfo(), St->isVolatile(),
13473 St->isNonTemporal(), St->getAlignment());
13474 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
13475 Chains.push_back(Ch);
13478 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0],
13483 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
13484 // the FP state in cases where an emms may be missing.
13485 // A preferable solution to the general problem is to figure out the right
13486 // places to insert EMMS. This qualifies as a quick hack.
13488 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
13489 if (VT.getSizeInBits() != 64)
13492 const Function *F = DAG.getMachineFunction().getFunction();
13493 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
13494 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
13495 && Subtarget->hasSSE2();
13496 if ((VT.isVector() ||
13497 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
13498 isa<LoadSDNode>(St->getValue()) &&
13499 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
13500 St->getChain().hasOneUse() && !St->isVolatile()) {
13501 SDNode* LdVal = St->getValue().getNode();
13502 LoadSDNode *Ld = 0;
13503 int TokenFactorIndex = -1;
13504 SmallVector<SDValue, 8> Ops;
13505 SDNode* ChainVal = St->getChain().getNode();
13506 // Must be a store of a load. We currently handle two cases: the load
13507 // is a direct child, and it's under an intervening TokenFactor. It is
13508 // possible to dig deeper under nested TokenFactors.
13509 if (ChainVal == LdVal)
13510 Ld = cast<LoadSDNode>(St->getChain());
13511 else if (St->getValue().hasOneUse() &&
13512 ChainVal->getOpcode() == ISD::TokenFactor) {
13513 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
13514 if (ChainVal->getOperand(i).getNode() == LdVal) {
13515 TokenFactorIndex = i;
13516 Ld = cast<LoadSDNode>(St->getValue());
13518 Ops.push_back(ChainVal->getOperand(i));
13522 if (!Ld || !ISD::isNormalLoad(Ld))
13525 // If this is not the MMX case, i.e. we are just turning i64 load/store
13526 // into f64 load/store, avoid the transformation if there are multiple
13527 // uses of the loaded value.
13528 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
13531 DebugLoc LdDL = Ld->getDebugLoc();
13532 DebugLoc StDL = N->getDebugLoc();
13533 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
13534 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
13536 if (Subtarget->is64Bit() || F64IsLegal) {
13537 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
13538 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
13539 Ld->getPointerInfo(), Ld->isVolatile(),
13540 Ld->isNonTemporal(), Ld->getAlignment());
13541 SDValue NewChain = NewLd.getValue(1);
13542 if (TokenFactorIndex != -1) {
13543 Ops.push_back(NewChain);
13544 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
13547 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
13548 St->getPointerInfo(),
13549 St->isVolatile(), St->isNonTemporal(),
13550 St->getAlignment());
13553 // Otherwise, lower to two pairs of 32-bit loads / stores.
13554 SDValue LoAddr = Ld->getBasePtr();
13555 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
13556 DAG.getConstant(4, MVT::i32));
13558 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
13559 Ld->getPointerInfo(),
13560 Ld->isVolatile(), Ld->isNonTemporal(),
13561 Ld->getAlignment());
13562 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
13563 Ld->getPointerInfo().getWithOffset(4),
13564 Ld->isVolatile(), Ld->isNonTemporal(),
13565 MinAlign(Ld->getAlignment(), 4));
13567 SDValue NewChain = LoLd.getValue(1);
13568 if (TokenFactorIndex != -1) {
13569 Ops.push_back(LoLd);
13570 Ops.push_back(HiLd);
13571 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
13575 LoAddr = St->getBasePtr();
13576 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
13577 DAG.getConstant(4, MVT::i32));
13579 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
13580 St->getPointerInfo(),
13581 St->isVolatile(), St->isNonTemporal(),
13582 St->getAlignment());
13583 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
13584 St->getPointerInfo().getWithOffset(4),
13586 St->isNonTemporal(),
13587 MinAlign(St->getAlignment(), 4));
13588 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
13593 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
13594 /// X86ISD::FXOR nodes.
13595 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
13596 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
13597 // F[X]OR(0.0, x) -> x
13598 // F[X]OR(x, 0.0) -> x
13599 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
13600 if (C->getValueAPF().isPosZero())
13601 return N->getOperand(1);
13602 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
13603 if (C->getValueAPF().isPosZero())
13604 return N->getOperand(0);
13608 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
13609 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
13610 // FAND(0.0, x) -> 0.0
13611 // FAND(x, 0.0) -> 0.0
13612 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
13613 if (C->getValueAPF().isPosZero())
13614 return N->getOperand(0);
13615 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
13616 if (C->getValueAPF().isPosZero())
13617 return N->getOperand(1);
13621 static SDValue PerformBTCombine(SDNode *N,
13623 TargetLowering::DAGCombinerInfo &DCI) {
13624 // BT ignores high bits in the bit index operand.
13625 SDValue Op1 = N->getOperand(1);
13626 if (Op1.hasOneUse()) {
13627 unsigned BitWidth = Op1.getValueSizeInBits();
13628 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
13629 APInt KnownZero, KnownOne;
13630 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
13631 !DCI.isBeforeLegalizeOps());
13632 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13633 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
13634 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
13635 DCI.CommitTargetLoweringOpt(TLO);
13640 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
13641 SDValue Op = N->getOperand(0);
13642 if (Op.getOpcode() == ISD::BITCAST)
13643 Op = Op.getOperand(0);
13644 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
13645 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
13646 VT.getVectorElementType().getSizeInBits() ==
13647 OpVT.getVectorElementType().getSizeInBits()) {
13648 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
13653 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
13654 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
13655 // (and (i32 x86isd::setcc_carry), 1)
13656 // This eliminates the zext. This transformation is necessary because
13657 // ISD::SETCC is always legalized to i8.
13658 DebugLoc dl = N->getDebugLoc();
13659 SDValue N0 = N->getOperand(0);
13660 EVT VT = N->getValueType(0);
13661 if (N0.getOpcode() == ISD::AND &&
13663 N0.getOperand(0).hasOneUse()) {
13664 SDValue N00 = N0.getOperand(0);
13665 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
13667 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
13668 if (!C || C->getZExtValue() != 1)
13670 return DAG.getNode(ISD::AND, dl, VT,
13671 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
13672 N00.getOperand(0), N00.getOperand(1)),
13673 DAG.getConstant(1, VT));
13679 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
13680 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG) {
13681 unsigned X86CC = N->getConstantOperandVal(0);
13682 SDValue EFLAG = N->getOperand(1);
13683 DebugLoc DL = N->getDebugLoc();
13685 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
13686 // a zext and produces an all-ones bit which is more useful than 0/1 in some
13688 if (X86CC == X86::COND_B)
13689 return DAG.getNode(ISD::AND, DL, MVT::i8,
13690 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
13691 DAG.getConstant(X86CC, MVT::i8), EFLAG),
13692 DAG.getConstant(1, MVT::i8));
13697 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
13698 const X86TargetLowering *XTLI) {
13699 SDValue Op0 = N->getOperand(0);
13700 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
13701 // a 32-bit target where SSE doesn't support i64->FP operations.
13702 if (Op0.getOpcode() == ISD::LOAD) {
13703 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
13704 EVT VT = Ld->getValueType(0);
13705 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
13706 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
13707 !XTLI->getSubtarget()->is64Bit() &&
13708 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
13709 SDValue FILDChain = XTLI->BuildFILD(SDValue(N, 0), Ld->getValueType(0),
13710 Ld->getChain(), Op0, DAG);
13711 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
13718 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
13719 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
13720 X86TargetLowering::DAGCombinerInfo &DCI) {
13721 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
13722 // the result is either zero or one (depending on the input carry bit).
13723 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
13724 if (X86::isZeroNode(N->getOperand(0)) &&
13725 X86::isZeroNode(N->getOperand(1)) &&
13726 // We don't have a good way to replace an EFLAGS use, so only do this when
13728 SDValue(N, 1).use_empty()) {
13729 DebugLoc DL = N->getDebugLoc();
13730 EVT VT = N->getValueType(0);
13731 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
13732 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
13733 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
13734 DAG.getConstant(X86::COND_B,MVT::i8),
13736 DAG.getConstant(1, VT));
13737 return DCI.CombineTo(N, Res1, CarryOut);
13743 // fold (add Y, (sete X, 0)) -> adc 0, Y
13744 // (add Y, (setne X, 0)) -> sbb -1, Y
13745 // (sub (sete X, 0), Y) -> sbb 0, Y
13746 // (sub (setne X, 0), Y) -> adc -1, Y
13747 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
13748 DebugLoc DL = N->getDebugLoc();
13750 // Look through ZExts.
13751 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
13752 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
13755 SDValue SetCC = Ext.getOperand(0);
13756 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
13759 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
13760 if (CC != X86::COND_E && CC != X86::COND_NE)
13763 SDValue Cmp = SetCC.getOperand(1);
13764 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
13765 !X86::isZeroNode(Cmp.getOperand(1)) ||
13766 !Cmp.getOperand(0).getValueType().isInteger())
13769 SDValue CmpOp0 = Cmp.getOperand(0);
13770 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
13771 DAG.getConstant(1, CmpOp0.getValueType()));
13773 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
13774 if (CC == X86::COND_NE)
13775 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
13776 DL, OtherVal.getValueType(), OtherVal,
13777 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
13778 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
13779 DL, OtherVal.getValueType(), OtherVal,
13780 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
13783 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG) {
13784 SDValue Op0 = N->getOperand(0);
13785 SDValue Op1 = N->getOperand(1);
13787 // X86 can't encode an immediate LHS of a sub. See if we can push the
13788 // negation into a preceding instruction.
13789 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
13790 // If the RHS of the sub is a XOR with one use and a constant, invert the
13791 // immediate. Then add one to the LHS of the sub so we can turn
13792 // X-Y -> X+~Y+1, saving one register.
13793 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
13794 isa<ConstantSDNode>(Op1.getOperand(1))) {
13795 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
13796 EVT VT = Op0.getValueType();
13797 SDValue NewXor = DAG.getNode(ISD::XOR, Op1.getDebugLoc(), VT,
13799 DAG.getConstant(~XorC, VT));
13800 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, NewXor,
13801 DAG.getConstant(C->getAPIntValue()+1, VT));
13805 return OptimizeConditionalInDecrement(N, DAG);
13808 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
13809 DAGCombinerInfo &DCI) const {
13810 SelectionDAG &DAG = DCI.DAG;
13811 switch (N->getOpcode()) {
13813 case ISD::EXTRACT_VECTOR_ELT:
13814 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
13815 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
13816 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
13817 case ISD::ADD: return OptimizeConditionalInDecrement(N, DAG);
13818 case ISD::SUB: return PerformSubCombine(N, DAG);
13819 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
13820 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
13823 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
13824 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
13825 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
13826 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
13827 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, this);
13829 case X86ISD::FOR: return PerformFORCombine(N, DAG);
13830 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
13831 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
13832 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
13833 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
13834 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG);
13835 case X86ISD::SHUFPS: // Handle all target specific shuffles
13836 case X86ISD::SHUFPD:
13837 case X86ISD::PALIGN:
13838 case X86ISD::PUNPCKHBW:
13839 case X86ISD::PUNPCKHWD:
13840 case X86ISD::PUNPCKHDQ:
13841 case X86ISD::PUNPCKHQDQ:
13842 case X86ISD::UNPCKHPS:
13843 case X86ISD::UNPCKHPD:
13844 case X86ISD::VUNPCKHPSY:
13845 case X86ISD::VUNPCKHPDY:
13846 case X86ISD::PUNPCKLBW:
13847 case X86ISD::PUNPCKLWD:
13848 case X86ISD::PUNPCKLDQ:
13849 case X86ISD::PUNPCKLQDQ:
13850 case X86ISD::UNPCKLPS:
13851 case X86ISD::UNPCKLPD:
13852 case X86ISD::VUNPCKLPSY:
13853 case X86ISD::VUNPCKLPDY:
13854 case X86ISD::MOVHLPS:
13855 case X86ISD::MOVLHPS:
13856 case X86ISD::PSHUFD:
13857 case X86ISD::PSHUFHW:
13858 case X86ISD::PSHUFLW:
13859 case X86ISD::MOVSS:
13860 case X86ISD::MOVSD:
13861 case X86ISD::VPERMILPS:
13862 case X86ISD::VPERMILPSY:
13863 case X86ISD::VPERMILPD:
13864 case X86ISD::VPERMILPDY:
13865 case X86ISD::VPERM2F128:
13866 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
13872 /// isTypeDesirableForOp - Return true if the target has native support for
13873 /// the specified value type and it is 'desirable' to use the type for the
13874 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
13875 /// instruction encodings are longer and some i16 instructions are slow.
13876 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
13877 if (!isTypeLegal(VT))
13879 if (VT != MVT::i16)
13886 case ISD::SIGN_EXTEND:
13887 case ISD::ZERO_EXTEND:
13888 case ISD::ANY_EXTEND:
13901 /// IsDesirableToPromoteOp - This method query the target whether it is
13902 /// beneficial for dag combiner to promote the specified node. If true, it
13903 /// should return the desired promotion type by reference.
13904 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
13905 EVT VT = Op.getValueType();
13906 if (VT != MVT::i16)
13909 bool Promote = false;
13910 bool Commute = false;
13911 switch (Op.getOpcode()) {
13914 LoadSDNode *LD = cast<LoadSDNode>(Op);
13915 // If the non-extending load has a single use and it's not live out, then it
13916 // might be folded.
13917 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
13918 Op.hasOneUse()*/) {
13919 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13920 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
13921 // The only case where we'd want to promote LOAD (rather then it being
13922 // promoted as an operand is when it's only use is liveout.
13923 if (UI->getOpcode() != ISD::CopyToReg)
13930 case ISD::SIGN_EXTEND:
13931 case ISD::ZERO_EXTEND:
13932 case ISD::ANY_EXTEND:
13937 SDValue N0 = Op.getOperand(0);
13938 // Look out for (store (shl (load), x)).
13939 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
13952 SDValue N0 = Op.getOperand(0);
13953 SDValue N1 = Op.getOperand(1);
13954 if (!Commute && MayFoldLoad(N1))
13956 // Avoid disabling potential load folding opportunities.
13957 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
13959 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
13969 //===----------------------------------------------------------------------===//
13970 // X86 Inline Assembly Support
13971 //===----------------------------------------------------------------------===//
13973 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
13974 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
13976 std::string AsmStr = IA->getAsmString();
13978 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
13979 SmallVector<StringRef, 4> AsmPieces;
13980 SplitString(AsmStr, AsmPieces, ";\n");
13982 switch (AsmPieces.size()) {
13983 default: return false;
13985 AsmStr = AsmPieces[0];
13987 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
13989 // FIXME: this should verify that we are targeting a 486 or better. If not,
13990 // we will turn this bswap into something that will be lowered to logical ops
13991 // instead of emitting the bswap asm. For now, we don't support 486 or lower
13992 // so don't worry about this.
13994 if (AsmPieces.size() == 2 &&
13995 (AsmPieces[0] == "bswap" ||
13996 AsmPieces[0] == "bswapq" ||
13997 AsmPieces[0] == "bswapl") &&
13998 (AsmPieces[1] == "$0" ||
13999 AsmPieces[1] == "${0:q}")) {
14000 // No need to check constraints, nothing other than the equivalent of
14001 // "=r,0" would be valid here.
14002 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
14003 if (!Ty || Ty->getBitWidth() % 16 != 0)
14005 return IntrinsicLowering::LowerToByteSwap(CI);
14007 // rorw $$8, ${0:w} --> llvm.bswap.i16
14008 if (CI->getType()->isIntegerTy(16) &&
14009 AsmPieces.size() == 3 &&
14010 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
14011 AsmPieces[1] == "$$8," &&
14012 AsmPieces[2] == "${0:w}" &&
14013 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
14015 const std::string &ConstraintsStr = IA->getConstraintString();
14016 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
14017 std::sort(AsmPieces.begin(), AsmPieces.end());
14018 if (AsmPieces.size() == 4 &&
14019 AsmPieces[0] == "~{cc}" &&
14020 AsmPieces[1] == "~{dirflag}" &&
14021 AsmPieces[2] == "~{flags}" &&
14022 AsmPieces[3] == "~{fpsr}") {
14023 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
14024 if (!Ty || Ty->getBitWidth() % 16 != 0)
14026 return IntrinsicLowering::LowerToByteSwap(CI);
14031 if (CI->getType()->isIntegerTy(32) &&
14032 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
14033 SmallVector<StringRef, 4> Words;
14034 SplitString(AsmPieces[0], Words, " \t,");
14035 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
14036 Words[2] == "${0:w}") {
14038 SplitString(AsmPieces[1], Words, " \t,");
14039 if (Words.size() == 3 && Words[0] == "rorl" && Words[1] == "$$16" &&
14040 Words[2] == "$0") {
14042 SplitString(AsmPieces[2], Words, " \t,");
14043 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
14044 Words[2] == "${0:w}") {
14046 const std::string &ConstraintsStr = IA->getConstraintString();
14047 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
14048 std::sort(AsmPieces.begin(), AsmPieces.end());
14049 if (AsmPieces.size() == 4 &&
14050 AsmPieces[0] == "~{cc}" &&
14051 AsmPieces[1] == "~{dirflag}" &&
14052 AsmPieces[2] == "~{flags}" &&
14053 AsmPieces[3] == "~{fpsr}") {
14054 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
14055 if (!Ty || Ty->getBitWidth() % 16 != 0)
14057 return IntrinsicLowering::LowerToByteSwap(CI);
14064 if (CI->getType()->isIntegerTy(64)) {
14065 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
14066 if (Constraints.size() >= 2 &&
14067 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
14068 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
14069 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
14070 SmallVector<StringRef, 4> Words;
14071 SplitString(AsmPieces[0], Words, " \t");
14072 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
14074 SplitString(AsmPieces[1], Words, " \t");
14075 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
14077 SplitString(AsmPieces[2], Words, " \t,");
14078 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
14079 Words[2] == "%edx") {
14080 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
14081 if (!Ty || Ty->getBitWidth() % 16 != 0)
14083 return IntrinsicLowering::LowerToByteSwap(CI);
14096 /// getConstraintType - Given a constraint letter, return the type of
14097 /// constraint it is for this target.
14098 X86TargetLowering::ConstraintType
14099 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
14100 if (Constraint.size() == 1) {
14101 switch (Constraint[0]) {
14112 return C_RegisterClass;
14136 return TargetLowering::getConstraintType(Constraint);
14139 /// Examine constraint type and operand type and determine a weight value.
14140 /// This object must already have been set up with the operand type
14141 /// and the current alternative constraint selected.
14142 TargetLowering::ConstraintWeight
14143 X86TargetLowering::getSingleConstraintMatchWeight(
14144 AsmOperandInfo &info, const char *constraint) const {
14145 ConstraintWeight weight = CW_Invalid;
14146 Value *CallOperandVal = info.CallOperandVal;
14147 // If we don't have a value, we can't do a match,
14148 // but allow it at the lowest weight.
14149 if (CallOperandVal == NULL)
14151 Type *type = CallOperandVal->getType();
14152 // Look at the constraint type.
14153 switch (*constraint) {
14155 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
14166 if (CallOperandVal->getType()->isIntegerTy())
14167 weight = CW_SpecificReg;
14172 if (type->isFloatingPointTy())
14173 weight = CW_SpecificReg;
14176 if (type->isX86_MMXTy() && Subtarget->hasMMX())
14177 weight = CW_SpecificReg;
14181 if ((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasXMM())
14182 weight = CW_Register;
14185 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
14186 if (C->getZExtValue() <= 31)
14187 weight = CW_Constant;
14191 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
14192 if (C->getZExtValue() <= 63)
14193 weight = CW_Constant;
14197 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
14198 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
14199 weight = CW_Constant;
14203 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
14204 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
14205 weight = CW_Constant;
14209 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
14210 if (C->getZExtValue() <= 3)
14211 weight = CW_Constant;
14215 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
14216 if (C->getZExtValue() <= 0xff)
14217 weight = CW_Constant;
14222 if (dyn_cast<ConstantFP>(CallOperandVal)) {
14223 weight = CW_Constant;
14227 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
14228 if ((C->getSExtValue() >= -0x80000000LL) &&
14229 (C->getSExtValue() <= 0x7fffffffLL))
14230 weight = CW_Constant;
14234 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
14235 if (C->getZExtValue() <= 0xffffffff)
14236 weight = CW_Constant;
14243 /// LowerXConstraint - try to replace an X constraint, which matches anything,
14244 /// with another that has more specific requirements based on the type of the
14245 /// corresponding operand.
14246 const char *X86TargetLowering::
14247 LowerXConstraint(EVT ConstraintVT) const {
14248 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
14249 // 'f' like normal targets.
14250 if (ConstraintVT.isFloatingPoint()) {
14251 if (Subtarget->hasXMMInt())
14253 if (Subtarget->hasXMM())
14257 return TargetLowering::LowerXConstraint(ConstraintVT);
14260 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
14261 /// vector. If it is invalid, don't add anything to Ops.
14262 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
14263 std::string &Constraint,
14264 std::vector<SDValue>&Ops,
14265 SelectionDAG &DAG) const {
14266 SDValue Result(0, 0);
14268 // Only support length 1 constraints for now.
14269 if (Constraint.length() > 1) return;
14271 char ConstraintLetter = Constraint[0];
14272 switch (ConstraintLetter) {
14275 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
14276 if (C->getZExtValue() <= 31) {
14277 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
14283 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
14284 if (C->getZExtValue() <= 63) {
14285 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
14291 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
14292 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
14293 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
14299 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
14300 if (C->getZExtValue() <= 255) {
14301 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
14307 // 32-bit signed value
14308 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
14309 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
14310 C->getSExtValue())) {
14311 // Widen to 64 bits here to get it sign extended.
14312 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
14315 // FIXME gcc accepts some relocatable values here too, but only in certain
14316 // memory models; it's complicated.
14321 // 32-bit unsigned value
14322 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
14323 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
14324 C->getZExtValue())) {
14325 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
14329 // FIXME gcc accepts some relocatable values here too, but only in certain
14330 // memory models; it's complicated.
14334 // Literal immediates are always ok.
14335 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
14336 // Widen to 64 bits here to get it sign extended.
14337 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
14341 // In any sort of PIC mode addresses need to be computed at runtime by
14342 // adding in a register or some sort of table lookup. These can't
14343 // be used as immediates.
14344 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
14347 // If we are in non-pic codegen mode, we allow the address of a global (with
14348 // an optional displacement) to be used with 'i'.
14349 GlobalAddressSDNode *GA = 0;
14350 int64_t Offset = 0;
14352 // Match either (GA), (GA+C), (GA+C1+C2), etc.
14354 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
14355 Offset += GA->getOffset();
14357 } else if (Op.getOpcode() == ISD::ADD) {
14358 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
14359 Offset += C->getZExtValue();
14360 Op = Op.getOperand(0);
14363 } else if (Op.getOpcode() == ISD::SUB) {
14364 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
14365 Offset += -C->getZExtValue();
14366 Op = Op.getOperand(0);
14371 // Otherwise, this isn't something we can handle, reject it.
14375 const GlobalValue *GV = GA->getGlobal();
14376 // If we require an extra load to get this address, as in PIC mode, we
14377 // can't accept it.
14378 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
14379 getTargetMachine())))
14382 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
14383 GA->getValueType(0), Offset);
14388 if (Result.getNode()) {
14389 Ops.push_back(Result);
14392 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
14395 std::pair<unsigned, const TargetRegisterClass*>
14396 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
14398 // First, see if this is a constraint that directly corresponds to an LLVM
14400 if (Constraint.size() == 1) {
14401 // GCC Constraint Letters
14402 switch (Constraint[0]) {
14404 // TODO: Slight differences here in allocation order and leaving
14405 // RIP in the class. Do they matter any more here than they do
14406 // in the normal allocation?
14407 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
14408 if (Subtarget->is64Bit()) {
14409 if (VT == MVT::i32 || VT == MVT::f32)
14410 return std::make_pair(0U, X86::GR32RegisterClass);
14411 else if (VT == MVT::i16)
14412 return std::make_pair(0U, X86::GR16RegisterClass);
14413 else if (VT == MVT::i8 || VT == MVT::i1)
14414 return std::make_pair(0U, X86::GR8RegisterClass);
14415 else if (VT == MVT::i64 || VT == MVT::f64)
14416 return std::make_pair(0U, X86::GR64RegisterClass);
14419 // 32-bit fallthrough
14420 case 'Q': // Q_REGS
14421 if (VT == MVT::i32 || VT == MVT::f32)
14422 return std::make_pair(0U, X86::GR32_ABCDRegisterClass);
14423 else if (VT == MVT::i16)
14424 return std::make_pair(0U, X86::GR16_ABCDRegisterClass);
14425 else if (VT == MVT::i8 || VT == MVT::i1)
14426 return std::make_pair(0U, X86::GR8_ABCD_LRegisterClass);
14427 else if (VT == MVT::i64)
14428 return std::make_pair(0U, X86::GR64_ABCDRegisterClass);
14430 case 'r': // GENERAL_REGS
14431 case 'l': // INDEX_REGS
14432 if (VT == MVT::i8 || VT == MVT::i1)
14433 return std::make_pair(0U, X86::GR8RegisterClass);
14434 if (VT == MVT::i16)
14435 return std::make_pair(0U, X86::GR16RegisterClass);
14436 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
14437 return std::make_pair(0U, X86::GR32RegisterClass);
14438 return std::make_pair(0U, X86::GR64RegisterClass);
14439 case 'R': // LEGACY_REGS
14440 if (VT == MVT::i8 || VT == MVT::i1)
14441 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
14442 if (VT == MVT::i16)
14443 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
14444 if (VT == MVT::i32 || !Subtarget->is64Bit())
14445 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
14446 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
14447 case 'f': // FP Stack registers.
14448 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
14449 // value to the correct fpstack register class.
14450 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
14451 return std::make_pair(0U, X86::RFP32RegisterClass);
14452 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
14453 return std::make_pair(0U, X86::RFP64RegisterClass);
14454 return std::make_pair(0U, X86::RFP80RegisterClass);
14455 case 'y': // MMX_REGS if MMX allowed.
14456 if (!Subtarget->hasMMX()) break;
14457 return std::make_pair(0U, X86::VR64RegisterClass);
14458 case 'Y': // SSE_REGS if SSE2 allowed
14459 if (!Subtarget->hasXMMInt()) break;
14461 case 'x': // SSE_REGS if SSE1 allowed
14462 if (!Subtarget->hasXMM()) break;
14464 switch (VT.getSimpleVT().SimpleTy) {
14466 // Scalar SSE types.
14469 return std::make_pair(0U, X86::FR32RegisterClass);
14472 return std::make_pair(0U, X86::FR64RegisterClass);
14480 return std::make_pair(0U, X86::VR128RegisterClass);
14486 // Use the default implementation in TargetLowering to convert the register
14487 // constraint into a member of a register class.
14488 std::pair<unsigned, const TargetRegisterClass*> Res;
14489 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
14491 // Not found as a standard register?
14492 if (Res.second == 0) {
14493 // Map st(0) -> st(7) -> ST0
14494 if (Constraint.size() == 7 && Constraint[0] == '{' &&
14495 tolower(Constraint[1]) == 's' &&
14496 tolower(Constraint[2]) == 't' &&
14497 Constraint[3] == '(' &&
14498 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
14499 Constraint[5] == ')' &&
14500 Constraint[6] == '}') {
14502 Res.first = X86::ST0+Constraint[4]-'0';
14503 Res.second = X86::RFP80RegisterClass;
14507 // GCC allows "st(0)" to be called just plain "st".
14508 if (StringRef("{st}").equals_lower(Constraint)) {
14509 Res.first = X86::ST0;
14510 Res.second = X86::RFP80RegisterClass;
14515 if (StringRef("{flags}").equals_lower(Constraint)) {
14516 Res.first = X86::EFLAGS;
14517 Res.second = X86::CCRRegisterClass;
14521 // 'A' means EAX + EDX.
14522 if (Constraint == "A") {
14523 Res.first = X86::EAX;
14524 Res.second = X86::GR32_ADRegisterClass;
14530 // Otherwise, check to see if this is a register class of the wrong value
14531 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
14532 // turn into {ax},{dx}.
14533 if (Res.second->hasType(VT))
14534 return Res; // Correct type already, nothing to do.
14536 // All of the single-register GCC register classes map their values onto
14537 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
14538 // really want an 8-bit or 32-bit register, map to the appropriate register
14539 // class and return the appropriate register.
14540 if (Res.second == X86::GR16RegisterClass) {
14541 if (VT == MVT::i8) {
14542 unsigned DestReg = 0;
14543 switch (Res.first) {
14545 case X86::AX: DestReg = X86::AL; break;
14546 case X86::DX: DestReg = X86::DL; break;
14547 case X86::CX: DestReg = X86::CL; break;
14548 case X86::BX: DestReg = X86::BL; break;
14551 Res.first = DestReg;
14552 Res.second = X86::GR8RegisterClass;
14554 } else if (VT == MVT::i32) {
14555 unsigned DestReg = 0;
14556 switch (Res.first) {
14558 case X86::AX: DestReg = X86::EAX; break;
14559 case X86::DX: DestReg = X86::EDX; break;
14560 case X86::CX: DestReg = X86::ECX; break;
14561 case X86::BX: DestReg = X86::EBX; break;
14562 case X86::SI: DestReg = X86::ESI; break;
14563 case X86::DI: DestReg = X86::EDI; break;
14564 case X86::BP: DestReg = X86::EBP; break;
14565 case X86::SP: DestReg = X86::ESP; break;
14568 Res.first = DestReg;
14569 Res.second = X86::GR32RegisterClass;
14571 } else if (VT == MVT::i64) {
14572 unsigned DestReg = 0;
14573 switch (Res.first) {
14575 case X86::AX: DestReg = X86::RAX; break;
14576 case X86::DX: DestReg = X86::RDX; break;
14577 case X86::CX: DestReg = X86::RCX; break;
14578 case X86::BX: DestReg = X86::RBX; break;
14579 case X86::SI: DestReg = X86::RSI; break;
14580 case X86::DI: DestReg = X86::RDI; break;
14581 case X86::BP: DestReg = X86::RBP; break;
14582 case X86::SP: DestReg = X86::RSP; break;
14585 Res.first = DestReg;
14586 Res.second = X86::GR64RegisterClass;
14589 } else if (Res.second == X86::FR32RegisterClass ||
14590 Res.second == X86::FR64RegisterClass ||
14591 Res.second == X86::VR128RegisterClass) {
14592 // Handle references to XMM physical registers that got mapped into the
14593 // wrong class. This can happen with constraints like {xmm0} where the
14594 // target independent register mapper will just pick the first match it can
14595 // find, ignoring the required type.
14596 if (VT == MVT::f32)
14597 Res.second = X86::FR32RegisterClass;
14598 else if (VT == MVT::f64)
14599 Res.second = X86::FR64RegisterClass;
14600 else if (X86::VR128RegisterClass->hasType(VT))
14601 Res.second = X86::VR128RegisterClass;