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 "X86MCTargetExpr.h"
20 #include "X86TargetMachine.h"
21 #include "X86TargetObjectFile.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/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineJumpTableInfo.h"
35 #include "llvm/CodeGen/MachineModuleInfo.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/PseudoSourceValue.h"
38 #include "llvm/MC/MCAsmInfo.h"
39 #include "llvm/MC/MCContext.h"
40 #include "llvm/MC/MCSymbol.h"
41 #include "llvm/ADT/BitVector.h"
42 #include "llvm/ADT/SmallSet.h"
43 #include "llvm/ADT/Statistic.h"
44 #include "llvm/ADT/StringExtras.h"
45 #include "llvm/ADT/VectorExtras.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/ErrorHandling.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Support/raw_ostream.h"
53 STATISTIC(NumTailCalls, "Number of tail calls");
56 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
58 // Disable16Bit - 16-bit operations typically have a larger encoding than
59 // corresponding 32-bit instructions, and 16-bit code is slow on some
60 // processors. This is an experimental flag to disable 16-bit operations
61 // (which forces them to be Legalized to 32-bit operations).
63 Disable16Bit("disable-16bit", cl::Hidden,
64 cl::desc("Disable use of 16-bit instructions"));
66 // Forward declarations.
67 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
70 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
71 switch (TM.getSubtarget<X86Subtarget>().TargetType) {
72 default: llvm_unreachable("unknown subtarget type");
73 case X86Subtarget::isDarwin:
74 if (TM.getSubtarget<X86Subtarget>().is64Bit())
75 return new X8664_MachoTargetObjectFile();
76 return new TargetLoweringObjectFileMachO();
77 case X86Subtarget::isELF:
78 if (TM.getSubtarget<X86Subtarget>().is64Bit())
79 return new X8664_ELFTargetObjectFile(TM);
80 return new X8632_ELFTargetObjectFile(TM);
81 case X86Subtarget::isMingw:
82 case X86Subtarget::isCygwin:
83 case X86Subtarget::isWindows:
84 return new TargetLoweringObjectFileCOFF();
88 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
89 : TargetLowering(TM, createTLOF(TM)) {
90 Subtarget = &TM.getSubtarget<X86Subtarget>();
91 X86ScalarSSEf64 = Subtarget->hasSSE2();
92 X86ScalarSSEf32 = Subtarget->hasSSE1();
93 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
95 RegInfo = TM.getRegisterInfo();
98 // Set up the TargetLowering object.
100 // X86 is weird, it always uses i8 for shift amounts and setcc results.
101 setShiftAmountType(MVT::i8);
102 setBooleanContents(ZeroOrOneBooleanContent);
103 setSchedulingPreference(SchedulingForRegPressure);
104 setStackPointerRegisterToSaveRestore(X86StackPtr);
106 if (Subtarget->isTargetDarwin()) {
107 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
108 setUseUnderscoreSetJmp(false);
109 setUseUnderscoreLongJmp(false);
110 } else if (Subtarget->isTargetMingw()) {
111 // MS runtime is weird: it exports _setjmp, but longjmp!
112 setUseUnderscoreSetJmp(true);
113 setUseUnderscoreLongJmp(false);
115 setUseUnderscoreSetJmp(true);
116 setUseUnderscoreLongJmp(true);
119 // Set up the register classes.
120 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
122 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
123 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
124 if (Subtarget->is64Bit())
125 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
127 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
129 // We don't accept any truncstore of integer registers.
130 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
132 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
133 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
135 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
136 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
137 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
139 // SETOEQ and SETUNE require checking two conditions.
140 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
141 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
142 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
143 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
144 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
145 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
147 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
149 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
150 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
151 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
153 if (Subtarget->is64Bit()) {
154 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
155 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
156 } else if (!UseSoftFloat) {
157 if (X86ScalarSSEf64) {
158 // We have an impenetrably clever algorithm for ui64->double only.
159 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
161 // We have an algorithm for SSE2, and we turn this into a 64-bit
162 // FILD for other targets.
163 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
166 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
168 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
169 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
172 // SSE has no i16 to fp conversion, only i32
173 if (X86ScalarSSEf32) {
174 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
175 // f32 and f64 cases are Legal, f80 case is not
176 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
178 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
179 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
182 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
183 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
186 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
187 // are Legal, f80 is custom lowered.
188 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
189 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
191 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
193 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
194 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
196 if (X86ScalarSSEf32) {
197 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
198 // f32 and f64 cases are Legal, f80 case is not
199 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
201 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
202 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
205 // Handle FP_TO_UINT by promoting the destination to a larger signed
207 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
208 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
209 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
211 if (Subtarget->is64Bit()) {
212 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
213 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
214 } else if (!UseSoftFloat) {
215 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
216 // Expand FP_TO_UINT into a select.
217 // FIXME: We would like to use a Custom expander here eventually to do
218 // the optimal thing for SSE vs. the default expansion in the legalizer.
219 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
221 // With SSE3 we can use fisttpll to convert to a signed i64; without
222 // SSE, we're stuck with a fistpll.
223 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
226 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
227 if (!X86ScalarSSEf64) {
228 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
229 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
232 // Scalar integer divide and remainder are lowered to use operations that
233 // produce two results, to match the available instructions. This exposes
234 // the two-result form to trivial CSE, which is able to combine x/y and x%y
235 // into a single instruction.
237 // Scalar integer multiply-high is also lowered to use two-result
238 // operations, to match the available instructions. However, plain multiply
239 // (low) operations are left as Legal, as there are single-result
240 // instructions for this in x86. Using the two-result multiply instructions
241 // when both high and low results are needed must be arranged by dagcombine.
242 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
243 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
244 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
245 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
246 setOperationAction(ISD::SREM , MVT::i8 , Expand);
247 setOperationAction(ISD::UREM , MVT::i8 , Expand);
248 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
249 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
250 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
251 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
252 setOperationAction(ISD::SREM , MVT::i16 , Expand);
253 setOperationAction(ISD::UREM , MVT::i16 , Expand);
254 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
255 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
256 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
257 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
258 setOperationAction(ISD::SREM , MVT::i32 , Expand);
259 setOperationAction(ISD::UREM , MVT::i32 , Expand);
260 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
261 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
262 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
263 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
264 setOperationAction(ISD::SREM , MVT::i64 , Expand);
265 setOperationAction(ISD::UREM , MVT::i64 , Expand);
267 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
268 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
269 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
270 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
271 if (Subtarget->is64Bit())
272 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
273 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
274 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
275 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
276 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
277 setOperationAction(ISD::FREM , MVT::f32 , Expand);
278 setOperationAction(ISD::FREM , MVT::f64 , Expand);
279 setOperationAction(ISD::FREM , MVT::f80 , Expand);
280 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
282 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
283 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
284 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
285 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
287 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
288 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
290 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
291 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
293 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
294 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
295 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
296 if (Subtarget->is64Bit()) {
297 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
298 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
299 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
302 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
303 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
305 // These should be promoted to a larger select which is supported.
306 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
307 // X86 wants to expand cmov itself.
308 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
310 setOperationAction(ISD::SELECT , MVT::i16 , Expand);
312 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
313 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
314 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
315 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
316 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
317 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
319 setOperationAction(ISD::SETCC , MVT::i16 , Expand);
321 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
322 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
323 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
324 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
325 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
326 if (Subtarget->is64Bit()) {
327 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
328 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
330 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
333 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
334 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
335 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
336 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
337 if (Subtarget->is64Bit())
338 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
339 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
340 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
341 if (Subtarget->is64Bit()) {
342 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
343 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
344 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
345 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
346 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
348 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
349 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
350 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
351 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
352 if (Subtarget->is64Bit()) {
353 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
354 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
355 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
358 if (Subtarget->hasSSE1())
359 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
361 if (!Subtarget->hasSSE2())
362 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
364 // Expand certain atomics
365 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
366 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
367 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
368 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
370 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
371 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
372 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
373 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
375 if (!Subtarget->is64Bit()) {
376 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
377 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
378 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
379 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
380 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
381 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
382 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
385 // FIXME - use subtarget debug flags
386 if (!Subtarget->isTargetDarwin() &&
387 !Subtarget->isTargetELF() &&
388 !Subtarget->isTargetCygMing()) {
389 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
392 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
393 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
394 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
395 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
396 if (Subtarget->is64Bit()) {
397 setExceptionPointerRegister(X86::RAX);
398 setExceptionSelectorRegister(X86::RDX);
400 setExceptionPointerRegister(X86::EAX);
401 setExceptionSelectorRegister(X86::EDX);
403 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
404 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
406 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
408 setOperationAction(ISD::TRAP, MVT::Other, Legal);
410 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
411 setOperationAction(ISD::VASTART , MVT::Other, Custom);
412 setOperationAction(ISD::VAEND , MVT::Other, Expand);
413 if (Subtarget->is64Bit()) {
414 setOperationAction(ISD::VAARG , MVT::Other, Custom);
415 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
417 setOperationAction(ISD::VAARG , MVT::Other, Expand);
418 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
421 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
422 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
423 if (Subtarget->is64Bit())
424 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
425 if (Subtarget->isTargetCygMing())
426 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
428 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
430 if (!UseSoftFloat && X86ScalarSSEf64) {
431 // f32 and f64 use SSE.
432 // Set up the FP register classes.
433 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
434 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
436 // Use ANDPD to simulate FABS.
437 setOperationAction(ISD::FABS , MVT::f64, Custom);
438 setOperationAction(ISD::FABS , MVT::f32, Custom);
440 // Use XORP to simulate FNEG.
441 setOperationAction(ISD::FNEG , MVT::f64, Custom);
442 setOperationAction(ISD::FNEG , MVT::f32, Custom);
444 // Use ANDPD and ORPD to simulate FCOPYSIGN.
445 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
446 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
448 // We don't support sin/cos/fmod
449 setOperationAction(ISD::FSIN , MVT::f64, Expand);
450 setOperationAction(ISD::FCOS , MVT::f64, Expand);
451 setOperationAction(ISD::FSIN , MVT::f32, Expand);
452 setOperationAction(ISD::FCOS , MVT::f32, Expand);
454 // Expand FP immediates into loads from the stack, except for the special
456 addLegalFPImmediate(APFloat(+0.0)); // xorpd
457 addLegalFPImmediate(APFloat(+0.0f)); // xorps
458 } else if (!UseSoftFloat && X86ScalarSSEf32) {
459 // Use SSE for f32, x87 for f64.
460 // Set up the FP register classes.
461 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
462 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
464 // Use ANDPS to simulate FABS.
465 setOperationAction(ISD::FABS , MVT::f32, Custom);
467 // Use XORP to simulate FNEG.
468 setOperationAction(ISD::FNEG , MVT::f32, Custom);
470 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
472 // Use ANDPS and ORPS to simulate FCOPYSIGN.
473 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
474 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
476 // We don't support sin/cos/fmod
477 setOperationAction(ISD::FSIN , MVT::f32, Expand);
478 setOperationAction(ISD::FCOS , MVT::f32, Expand);
480 // Special cases we handle for FP constants.
481 addLegalFPImmediate(APFloat(+0.0f)); // xorps
482 addLegalFPImmediate(APFloat(+0.0)); // FLD0
483 addLegalFPImmediate(APFloat(+1.0)); // FLD1
484 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
485 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
488 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
489 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
491 } else if (!UseSoftFloat) {
492 // f32 and f64 in x87.
493 // Set up the FP register classes.
494 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
495 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
497 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
498 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
499 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
500 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
503 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
504 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
506 addLegalFPImmediate(APFloat(+0.0)); // FLD0
507 addLegalFPImmediate(APFloat(+1.0)); // FLD1
508 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
509 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
510 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
511 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
512 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
513 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
516 // Long double always uses X87.
518 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
519 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
520 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
523 APFloat TmpFlt(+0.0);
524 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
526 addLegalFPImmediate(TmpFlt); // FLD0
528 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
529 APFloat TmpFlt2(+1.0);
530 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
532 addLegalFPImmediate(TmpFlt2); // FLD1
533 TmpFlt2.changeSign();
534 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
538 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
539 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
543 // Always use a library call for pow.
544 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
545 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
546 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
548 setOperationAction(ISD::FLOG, MVT::f80, Expand);
549 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
550 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
551 setOperationAction(ISD::FEXP, MVT::f80, Expand);
552 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
554 // First set operation action for all vector types to either promote
555 // (for widening) or expand (for scalarization). Then we will selectively
556 // turn on ones that can be effectively codegen'd.
557 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
558 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
559 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
574 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
575 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
599 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
601 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
603 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
604 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
605 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
606 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
607 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
608 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
609 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
610 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
611 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
612 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
613 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
614 setTruncStoreAction((MVT::SimpleValueType)VT,
615 (MVT::SimpleValueType)InnerVT, Expand);
616 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
617 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
618 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
621 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
622 // with -msoft-float, disable use of MMX as well.
623 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
624 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
625 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
626 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
627 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
628 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
630 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
631 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
632 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
633 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
635 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
636 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
637 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
638 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
640 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
641 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
643 setOperationAction(ISD::AND, MVT::v8i8, Promote);
644 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
645 setOperationAction(ISD::AND, MVT::v4i16, Promote);
646 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
647 setOperationAction(ISD::AND, MVT::v2i32, Promote);
648 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
649 setOperationAction(ISD::AND, MVT::v1i64, Legal);
651 setOperationAction(ISD::OR, MVT::v8i8, Promote);
652 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
653 setOperationAction(ISD::OR, MVT::v4i16, Promote);
654 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
655 setOperationAction(ISD::OR, MVT::v2i32, Promote);
656 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
657 setOperationAction(ISD::OR, MVT::v1i64, Legal);
659 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
660 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
661 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
662 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
663 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
664 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
665 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
667 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
668 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
669 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
670 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
671 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
672 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
673 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
674 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
675 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
677 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
678 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
679 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
680 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
681 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
683 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
684 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
685 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
686 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
688 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
689 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
690 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
691 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
693 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
695 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
696 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
697 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
698 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
699 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
700 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
701 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
704 if (!UseSoftFloat && Subtarget->hasSSE1()) {
705 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
707 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
708 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
709 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
710 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
711 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
712 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
713 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
714 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
715 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
716 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
717 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
718 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
721 if (!UseSoftFloat && Subtarget->hasSSE2()) {
722 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
724 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
725 // registers cannot be used even for integer operations.
726 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
727 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
728 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
729 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
731 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
732 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
733 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
734 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
735 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
736 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
737 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
738 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
739 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
740 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
741 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
742 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
743 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
744 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
745 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
746 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
748 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
749 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
750 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
751 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
753 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
754 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
755 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
756 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
757 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
759 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
760 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
761 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
762 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
763 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
765 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
766 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
767 EVT VT = (MVT::SimpleValueType)i;
768 // Do not attempt to custom lower non-power-of-2 vectors
769 if (!isPowerOf2_32(VT.getVectorNumElements()))
771 // Do not attempt to custom lower non-128-bit vectors
772 if (!VT.is128BitVector())
774 setOperationAction(ISD::BUILD_VECTOR,
775 VT.getSimpleVT().SimpleTy, Custom);
776 setOperationAction(ISD::VECTOR_SHUFFLE,
777 VT.getSimpleVT().SimpleTy, Custom);
778 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
779 VT.getSimpleVT().SimpleTy, Custom);
782 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
783 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
784 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
785 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
786 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
787 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
789 if (Subtarget->is64Bit()) {
790 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
791 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
794 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
795 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
796 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
799 // Do not attempt to promote non-128-bit vectors
800 if (!VT.is128BitVector()) {
803 setOperationAction(ISD::AND, SVT, Promote);
804 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
805 setOperationAction(ISD::OR, SVT, Promote);
806 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
807 setOperationAction(ISD::XOR, SVT, Promote);
808 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
809 setOperationAction(ISD::LOAD, SVT, Promote);
810 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
811 setOperationAction(ISD::SELECT, SVT, Promote);
812 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
815 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
817 // Custom lower v2i64 and v2f64 selects.
818 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
819 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
820 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
821 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
823 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
824 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
825 if (!DisableMMX && Subtarget->hasMMX()) {
826 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
827 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
831 if (Subtarget->hasSSE41()) {
832 // FIXME: Do we need to handle scalar-to-vector here?
833 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
835 // i8 and i16 vectors are custom , because the source register and source
836 // source memory operand types are not the same width. f32 vectors are
837 // custom since the immediate controlling the insert encodes additional
839 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
840 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
841 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
842 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
844 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
845 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
846 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
847 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
849 if (Subtarget->is64Bit()) {
850 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
851 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
855 if (Subtarget->hasSSE42()) {
856 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
859 if (!UseSoftFloat && Subtarget->hasAVX()) {
860 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
861 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
862 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
863 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
865 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
866 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
867 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
868 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
869 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
870 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
871 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
872 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
873 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
874 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
875 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
876 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
877 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
878 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
879 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
881 // Operations to consider commented out -v16i16 v32i8
882 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
883 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
884 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
885 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
886 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
887 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
888 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
889 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
890 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
891 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
892 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
893 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
894 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
895 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
897 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
898 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
899 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
900 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
902 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
903 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
904 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
905 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
906 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
908 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
909 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
910 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
911 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
912 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
913 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
916 // Not sure we want to do this since there are no 256-bit integer
919 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
920 // This includes 256-bit vectors
921 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
922 EVT VT = (MVT::SimpleValueType)i;
924 // Do not attempt to custom lower non-power-of-2 vectors
925 if (!isPowerOf2_32(VT.getVectorNumElements()))
928 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
929 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
930 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
933 if (Subtarget->is64Bit()) {
934 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
935 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
940 // Not sure we want to do this since there are no 256-bit integer
943 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
944 // Including 256-bit vectors
945 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
946 EVT VT = (MVT::SimpleValueType)i;
948 if (!VT.is256BitVector()) {
951 setOperationAction(ISD::AND, VT, Promote);
952 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
953 setOperationAction(ISD::OR, VT, Promote);
954 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
955 setOperationAction(ISD::XOR, VT, Promote);
956 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
957 setOperationAction(ISD::LOAD, VT, Promote);
958 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
959 setOperationAction(ISD::SELECT, VT, Promote);
960 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
963 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
967 // We want to custom lower some of our intrinsics.
968 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
970 // Add/Sub/Mul with overflow operations are custom lowered.
971 setOperationAction(ISD::SADDO, MVT::i32, Custom);
972 setOperationAction(ISD::SADDO, MVT::i64, Custom);
973 setOperationAction(ISD::UADDO, MVT::i32, Custom);
974 setOperationAction(ISD::UADDO, MVT::i64, Custom);
975 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
976 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
977 setOperationAction(ISD::USUBO, MVT::i32, Custom);
978 setOperationAction(ISD::USUBO, MVT::i64, Custom);
979 setOperationAction(ISD::SMULO, MVT::i32, Custom);
980 setOperationAction(ISD::SMULO, MVT::i64, Custom);
982 if (!Subtarget->is64Bit()) {
983 // These libcalls are not available in 32-bit.
984 setLibcallName(RTLIB::SHL_I128, 0);
985 setLibcallName(RTLIB::SRL_I128, 0);
986 setLibcallName(RTLIB::SRA_I128, 0);
989 // We have target-specific dag combine patterns for the following nodes:
990 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
991 setTargetDAGCombine(ISD::BUILD_VECTOR);
992 setTargetDAGCombine(ISD::SELECT);
993 setTargetDAGCombine(ISD::AND);
994 setTargetDAGCombine(ISD::SHL);
995 setTargetDAGCombine(ISD::SRA);
996 setTargetDAGCombine(ISD::SRL);
997 setTargetDAGCombine(ISD::OR);
998 setTargetDAGCombine(ISD::STORE);
999 setTargetDAGCombine(ISD::MEMBARRIER);
1000 setTargetDAGCombine(ISD::ZERO_EXTEND);
1001 if (Subtarget->is64Bit())
1002 setTargetDAGCombine(ISD::MUL);
1004 computeRegisterProperties();
1006 // FIXME: These should be based on subtarget info. Plus, the values should
1007 // be smaller when we are in optimizing for size mode.
1008 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1009 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
1010 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1011 setPrefLoopAlignment(16);
1012 benefitFromCodePlacementOpt = true;
1016 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1021 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1022 /// the desired ByVal argument alignment.
1023 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1026 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1027 if (VTy->getBitWidth() == 128)
1029 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1030 unsigned EltAlign = 0;
1031 getMaxByValAlign(ATy->getElementType(), EltAlign);
1032 if (EltAlign > MaxAlign)
1033 MaxAlign = EltAlign;
1034 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1035 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1036 unsigned EltAlign = 0;
1037 getMaxByValAlign(STy->getElementType(i), EltAlign);
1038 if (EltAlign > MaxAlign)
1039 MaxAlign = EltAlign;
1047 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1048 /// function arguments in the caller parameter area. For X86, aggregates
1049 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1050 /// are at 4-byte boundaries.
1051 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1052 if (Subtarget->is64Bit()) {
1053 // Max of 8 and alignment of type.
1054 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1061 if (Subtarget->hasSSE1())
1062 getMaxByValAlign(Ty, Align);
1066 /// getOptimalMemOpType - Returns the target specific optimal type for load
1067 /// and store operations as a result of memset, memcpy, and memmove
1068 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
1071 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1072 bool isSrcConst, bool isSrcStr,
1073 SelectionDAG &DAG) const {
1074 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1075 // linux. This is because the stack realignment code can't handle certain
1076 // cases like PR2962. This should be removed when PR2962 is fixed.
1077 const Function *F = DAG.getMachineFunction().getFunction();
1078 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1079 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1080 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1082 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1085 if (Subtarget->is64Bit() && Size >= 8)
1090 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1091 /// current function. The returned value is a member of the
1092 /// MachineJumpTableInfo::JTEntryKind enum.
1093 unsigned X86TargetLowering::getJumpTableEncoding() const {
1094 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1096 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1097 Subtarget->isPICStyleGOT())
1098 return MachineJumpTableInfo::EK_Custom32;
1100 // Otherwise, use the normal jump table encoding heuristics.
1101 return TargetLowering::getJumpTableEncoding();
1104 /// getPICBaseSymbol - Return the X86-32 PIC base.
1106 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1107 MCContext &Ctx) const {
1108 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1109 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1110 Twine(MF->getFunctionNumber())+"$pb");
1115 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1116 const MachineBasicBlock *MBB,
1117 unsigned uid,MCContext &Ctx) const{
1118 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1119 Subtarget->isPICStyleGOT());
1120 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1122 return X86MCTargetExpr::Create(MBB->getSymbol(Ctx),
1123 X86MCTargetExpr::GOTOFF, Ctx);
1126 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1128 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1129 SelectionDAG &DAG) const {
1130 if (!Subtarget->is64Bit())
1131 // This doesn't have DebugLoc associated with it, but is not really the
1132 // same as a Register.
1133 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1138 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1139 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1141 const MCExpr *X86TargetLowering::
1142 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1143 MCContext &Ctx) const {
1144 // X86-64 uses RIP relative addressing based on the jump table label.
1145 if (Subtarget->isPICStyleRIPRel())
1146 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1148 // Otherwise, the reference is relative to the PIC base.
1149 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1152 /// getFunctionAlignment - Return the Log2 alignment of this function.
1153 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1154 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1157 //===----------------------------------------------------------------------===//
1158 // Return Value Calling Convention Implementation
1159 //===----------------------------------------------------------------------===//
1161 #include "X86GenCallingConv.inc"
1164 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1165 const SmallVectorImpl<EVT> &OutTys,
1166 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1167 SelectionDAG &DAG) {
1168 SmallVector<CCValAssign, 16> RVLocs;
1169 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1170 RVLocs, *DAG.getContext());
1171 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1175 X86TargetLowering::LowerReturn(SDValue Chain,
1176 CallingConv::ID CallConv, bool isVarArg,
1177 const SmallVectorImpl<ISD::OutputArg> &Outs,
1178 DebugLoc dl, SelectionDAG &DAG) {
1180 SmallVector<CCValAssign, 16> RVLocs;
1181 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1182 RVLocs, *DAG.getContext());
1183 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1185 // Add the regs to the liveout set for the function.
1186 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1187 for (unsigned i = 0; i != RVLocs.size(); ++i)
1188 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1189 MRI.addLiveOut(RVLocs[i].getLocReg());
1193 SmallVector<SDValue, 6> RetOps;
1194 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1195 // Operand #1 = Bytes To Pop
1196 RetOps.push_back(DAG.getTargetConstant(getBytesToPopOnReturn(), MVT::i16));
1198 // Copy the result values into the output registers.
1199 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1200 CCValAssign &VA = RVLocs[i];
1201 assert(VA.isRegLoc() && "Can only return in registers!");
1202 SDValue ValToCopy = Outs[i].Val;
1204 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1205 // the RET instruction and handled by the FP Stackifier.
1206 if (VA.getLocReg() == X86::ST0 ||
1207 VA.getLocReg() == X86::ST1) {
1208 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1209 // change the value to the FP stack register class.
1210 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1211 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1212 RetOps.push_back(ValToCopy);
1213 // Don't emit a copytoreg.
1217 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1218 // which is returned in RAX / RDX.
1219 if (Subtarget->is64Bit()) {
1220 EVT ValVT = ValToCopy.getValueType();
1221 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1222 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1223 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1224 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1228 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1229 Flag = Chain.getValue(1);
1232 // The x86-64 ABI for returning structs by value requires that we copy
1233 // the sret argument into %rax for the return. We saved the argument into
1234 // a virtual register in the entry block, so now we copy the value out
1236 if (Subtarget->is64Bit() &&
1237 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1238 MachineFunction &MF = DAG.getMachineFunction();
1239 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1240 unsigned Reg = FuncInfo->getSRetReturnReg();
1242 Reg = MRI.createVirtualRegister(getRegClassFor(MVT::i64));
1243 FuncInfo->setSRetReturnReg(Reg);
1245 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1247 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1248 Flag = Chain.getValue(1);
1250 // RAX now acts like a return value.
1251 MRI.addLiveOut(X86::RAX);
1254 RetOps[0] = Chain; // Update chain.
1256 // Add the flag if we have it.
1258 RetOps.push_back(Flag);
1260 return DAG.getNode(X86ISD::RET_FLAG, dl,
1261 MVT::Other, &RetOps[0], RetOps.size());
1264 /// LowerCallResult - Lower the result values of a call into the
1265 /// appropriate copies out of appropriate physical registers.
1268 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1269 CallingConv::ID CallConv, bool isVarArg,
1270 const SmallVectorImpl<ISD::InputArg> &Ins,
1271 DebugLoc dl, SelectionDAG &DAG,
1272 SmallVectorImpl<SDValue> &InVals) {
1274 // Assign locations to each value returned by this call.
1275 SmallVector<CCValAssign, 16> RVLocs;
1276 bool Is64Bit = Subtarget->is64Bit();
1277 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1278 RVLocs, *DAG.getContext());
1279 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1281 // Copy all of the result registers out of their specified physreg.
1282 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1283 CCValAssign &VA = RVLocs[i];
1284 EVT CopyVT = VA.getValVT();
1286 // If this is x86-64, and we disabled SSE, we can't return FP values
1287 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1288 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1289 llvm_report_error("SSE register return with SSE disabled");
1292 // If this is a call to a function that returns an fp value on the floating
1293 // point stack, but where we prefer to use the value in xmm registers, copy
1294 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1295 if ((VA.getLocReg() == X86::ST0 ||
1296 VA.getLocReg() == X86::ST1) &&
1297 isScalarFPTypeInSSEReg(VA.getValVT())) {
1302 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1303 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1304 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1305 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1306 MVT::v2i64, InFlag).getValue(1);
1307 Val = Chain.getValue(0);
1308 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1309 Val, DAG.getConstant(0, MVT::i64));
1311 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1312 MVT::i64, InFlag).getValue(1);
1313 Val = Chain.getValue(0);
1315 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1317 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1318 CopyVT, InFlag).getValue(1);
1319 Val = Chain.getValue(0);
1321 InFlag = Chain.getValue(2);
1323 if (CopyVT != VA.getValVT()) {
1324 // Round the F80 the right size, which also moves to the appropriate xmm
1326 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1327 // This truncation won't change the value.
1328 DAG.getIntPtrConstant(1));
1331 InVals.push_back(Val);
1338 //===----------------------------------------------------------------------===//
1339 // C & StdCall & Fast Calling Convention implementation
1340 //===----------------------------------------------------------------------===//
1341 // StdCall calling convention seems to be standard for many Windows' API
1342 // routines and around. It differs from C calling convention just a little:
1343 // callee should clean up the stack, not caller. Symbols should be also
1344 // decorated in some fancy way :) It doesn't support any vector arguments.
1345 // For info on fast calling convention see Fast Calling Convention (tail call)
1346 // implementation LowerX86_32FastCCCallTo.
1348 /// CallIsStructReturn - Determines whether a call uses struct return
1350 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1354 return Outs[0].Flags.isSRet();
1357 /// ArgsAreStructReturn - Determines whether a function uses struct
1358 /// return semantics.
1360 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1364 return Ins[0].Flags.isSRet();
1367 /// IsCalleePop - Determines whether the callee is required to pop its
1368 /// own arguments. Callee pop is necessary to support tail calls.
1369 bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
1373 switch (CallingConv) {
1376 case CallingConv::X86_StdCall:
1377 return !Subtarget->is64Bit();
1378 case CallingConv::X86_FastCall:
1379 return !Subtarget->is64Bit();
1380 case CallingConv::Fast:
1381 return GuaranteedTailCallOpt;
1385 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1386 /// given CallingConvention value.
1387 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1388 if (Subtarget->is64Bit()) {
1389 if (Subtarget->isTargetWin64())
1390 return CC_X86_Win64_C;
1395 if (CC == CallingConv::X86_FastCall)
1396 return CC_X86_32_FastCall;
1397 else if (CC == CallingConv::Fast)
1398 return CC_X86_32_FastCC;
1403 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1404 /// by "Src" to address "Dst" with size and alignment information specified by
1405 /// the specific parameter attribute. The copy will be passed as a byval
1406 /// function parameter.
1408 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1409 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1411 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1412 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1413 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1416 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1417 /// a tailcall target by changing its ABI.
1418 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1419 return GuaranteedTailCallOpt && CC == CallingConv::Fast;
1423 X86TargetLowering::LowerMemArgument(SDValue Chain,
1424 CallingConv::ID CallConv,
1425 const SmallVectorImpl<ISD::InputArg> &Ins,
1426 DebugLoc dl, SelectionDAG &DAG,
1427 const CCValAssign &VA,
1428 MachineFrameInfo *MFI,
1430 // Create the nodes corresponding to a load from this parameter slot.
1431 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1432 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1433 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1436 // If value is passed by pointer we have address passed instead of the value
1438 if (VA.getLocInfo() == CCValAssign::Indirect)
1439 ValVT = VA.getLocVT();
1441 ValVT = VA.getValVT();
1443 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1444 // changed with more analysis.
1445 // In case of tail call optimization mark all arguments mutable. Since they
1446 // could be overwritten by lowering of arguments in case of a tail call.
1447 if (Flags.isByVal()) {
1448 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1449 VA.getLocMemOffset(), isImmutable, false);
1450 return DAG.getFrameIndex(FI, getPointerTy());
1452 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1453 VA.getLocMemOffset(), isImmutable, false);
1454 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1455 return DAG.getLoad(ValVT, dl, Chain, FIN,
1456 PseudoSourceValue::getFixedStack(FI), 0,
1462 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1463 CallingConv::ID CallConv,
1465 const SmallVectorImpl<ISD::InputArg> &Ins,
1468 SmallVectorImpl<SDValue> &InVals) {
1470 MachineFunction &MF = DAG.getMachineFunction();
1471 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1473 const Function* Fn = MF.getFunction();
1474 if (Fn->hasExternalLinkage() &&
1475 Subtarget->isTargetCygMing() &&
1476 Fn->getName() == "main")
1477 FuncInfo->setForceFramePointer(true);
1479 MachineFrameInfo *MFI = MF.getFrameInfo();
1480 bool Is64Bit = Subtarget->is64Bit();
1481 bool IsWin64 = Subtarget->isTargetWin64();
1483 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1484 "Var args not supported with calling convention fastcc");
1486 // Assign locations to all of the incoming arguments.
1487 SmallVector<CCValAssign, 16> ArgLocs;
1488 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1489 ArgLocs, *DAG.getContext());
1490 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1492 unsigned LastVal = ~0U;
1494 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1495 CCValAssign &VA = ArgLocs[i];
1496 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1498 assert(VA.getValNo() != LastVal &&
1499 "Don't support value assigned to multiple locs yet");
1500 LastVal = VA.getValNo();
1502 if (VA.isRegLoc()) {
1503 EVT RegVT = VA.getLocVT();
1504 TargetRegisterClass *RC = NULL;
1505 if (RegVT == MVT::i32)
1506 RC = X86::GR32RegisterClass;
1507 else if (Is64Bit && RegVT == MVT::i64)
1508 RC = X86::GR64RegisterClass;
1509 else if (RegVT == MVT::f32)
1510 RC = X86::FR32RegisterClass;
1511 else if (RegVT == MVT::f64)
1512 RC = X86::FR64RegisterClass;
1513 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1514 RC = X86::VR128RegisterClass;
1515 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1516 RC = X86::VR64RegisterClass;
1518 llvm_unreachable("Unknown argument type!");
1520 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1521 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1523 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1524 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1526 if (VA.getLocInfo() == CCValAssign::SExt)
1527 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1528 DAG.getValueType(VA.getValVT()));
1529 else if (VA.getLocInfo() == CCValAssign::ZExt)
1530 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1531 DAG.getValueType(VA.getValVT()));
1532 else if (VA.getLocInfo() == CCValAssign::BCvt)
1533 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1535 if (VA.isExtInLoc()) {
1536 // Handle MMX values passed in XMM regs.
1537 if (RegVT.isVector()) {
1538 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1539 ArgValue, DAG.getConstant(0, MVT::i64));
1540 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1542 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1545 assert(VA.isMemLoc());
1546 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1549 // If value is passed via pointer - do a load.
1550 if (VA.getLocInfo() == CCValAssign::Indirect)
1551 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1554 InVals.push_back(ArgValue);
1557 // The x86-64 ABI for returning structs by value requires that we copy
1558 // the sret argument into %rax for the return. Save the argument into
1559 // a virtual register so that we can access it from the return points.
1560 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1561 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1562 unsigned Reg = FuncInfo->getSRetReturnReg();
1564 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1565 FuncInfo->setSRetReturnReg(Reg);
1567 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1568 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1571 unsigned StackSize = CCInfo.getNextStackOffset();
1572 // Align stack specially for tail calls.
1573 if (FuncIsMadeTailCallSafe(CallConv))
1574 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1576 // If the function takes variable number of arguments, make a frame index for
1577 // the start of the first vararg value... for expansion of llvm.va_start.
1579 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1580 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize, true, false);
1583 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1585 // FIXME: We should really autogenerate these arrays
1586 static const unsigned GPR64ArgRegsWin64[] = {
1587 X86::RCX, X86::RDX, X86::R8, X86::R9
1589 static const unsigned XMMArgRegsWin64[] = {
1590 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1592 static const unsigned GPR64ArgRegs64Bit[] = {
1593 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1595 static const unsigned XMMArgRegs64Bit[] = {
1596 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1597 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1599 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1602 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1603 GPR64ArgRegs = GPR64ArgRegsWin64;
1604 XMMArgRegs = XMMArgRegsWin64;
1606 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1607 GPR64ArgRegs = GPR64ArgRegs64Bit;
1608 XMMArgRegs = XMMArgRegs64Bit;
1610 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1612 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1615 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1616 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1617 "SSE register cannot be used when SSE is disabled!");
1618 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1619 "SSE register cannot be used when SSE is disabled!");
1620 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1621 // Kernel mode asks for SSE to be disabled, so don't push them
1623 TotalNumXMMRegs = 0;
1625 // For X86-64, if there are vararg parameters that are passed via
1626 // registers, then we must store them to their spots on the stack so they
1627 // may be loaded by deferencing the result of va_next.
1628 VarArgsGPOffset = NumIntRegs * 8;
1629 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1630 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1631 TotalNumXMMRegs * 16, 16,
1634 // Store the integer parameter registers.
1635 SmallVector<SDValue, 8> MemOps;
1636 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1637 unsigned Offset = VarArgsGPOffset;
1638 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1639 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1640 DAG.getIntPtrConstant(Offset));
1641 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1642 X86::GR64RegisterClass);
1643 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1645 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1646 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1647 Offset, false, false, 0);
1648 MemOps.push_back(Store);
1652 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1653 // Now store the XMM (fp + vector) parameter registers.
1654 SmallVector<SDValue, 11> SaveXMMOps;
1655 SaveXMMOps.push_back(Chain);
1657 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1658 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1659 SaveXMMOps.push_back(ALVal);
1661 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1662 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1664 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1665 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1666 X86::VR128RegisterClass);
1667 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1668 SaveXMMOps.push_back(Val);
1670 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1672 &SaveXMMOps[0], SaveXMMOps.size()));
1675 if (!MemOps.empty())
1676 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1677 &MemOps[0], MemOps.size());
1681 // Some CCs need callee pop.
1682 if (IsCalleePop(isVarArg, CallConv)) {
1683 BytesToPopOnReturn = StackSize; // Callee pops everything.
1685 BytesToPopOnReturn = 0; // Callee pops nothing.
1686 // If this is an sret function, the return should pop the hidden pointer.
1687 if (!Is64Bit && CallConv != CallingConv::Fast && ArgsAreStructReturn(Ins))
1688 BytesToPopOnReturn = 4;
1692 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1693 if (CallConv == CallingConv::X86_FastCall)
1694 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1697 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1703 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1704 SDValue StackPtr, SDValue Arg,
1705 DebugLoc dl, SelectionDAG &DAG,
1706 const CCValAssign &VA,
1707 ISD::ArgFlagsTy Flags) {
1708 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1709 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1710 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1711 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1712 if (Flags.isByVal()) {
1713 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1715 return DAG.getStore(Chain, dl, Arg, PtrOff,
1716 PseudoSourceValue::getStack(), LocMemOffset,
1720 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1721 /// optimization is performed and it is required.
1723 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1724 SDValue &OutRetAddr, SDValue Chain,
1725 bool IsTailCall, bool Is64Bit,
1726 int FPDiff, DebugLoc dl) {
1727 // Adjust the Return address stack slot.
1728 EVT VT = getPointerTy();
1729 OutRetAddr = getReturnAddressFrameIndex(DAG);
1731 // Load the "old" Return address.
1732 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1733 return SDValue(OutRetAddr.getNode(), 1);
1736 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1737 /// optimization is performed and it is required (FPDiff!=0).
1739 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1740 SDValue Chain, SDValue RetAddrFrIdx,
1741 bool Is64Bit, int FPDiff, DebugLoc dl) {
1742 // Store the return address to the appropriate stack slot.
1743 if (!FPDiff) return Chain;
1744 // Calculate the new stack slot for the return address.
1745 int SlotSize = Is64Bit ? 8 : 4;
1746 int NewReturnAddrFI =
1747 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false, false);
1748 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1749 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1750 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1751 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1757 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1758 CallingConv::ID CallConv, bool isVarArg,
1760 const SmallVectorImpl<ISD::OutputArg> &Outs,
1761 const SmallVectorImpl<ISD::InputArg> &Ins,
1762 DebugLoc dl, SelectionDAG &DAG,
1763 SmallVectorImpl<SDValue> &InVals) {
1764 MachineFunction &MF = DAG.getMachineFunction();
1765 bool Is64Bit = Subtarget->is64Bit();
1766 bool IsStructRet = CallIsStructReturn(Outs);
1767 bool IsSibcall = false;
1770 // Check if it's really possible to do a tail call.
1771 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
1774 // Sibcalls are automatically detected tailcalls which do not require
1776 if (!GuaranteedTailCallOpt && isTailCall)
1783 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1784 "Var args not supported with calling convention fastcc");
1786 // Analyze operands of the call, assigning locations to each operand.
1787 SmallVector<CCValAssign, 16> ArgLocs;
1788 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1789 ArgLocs, *DAG.getContext());
1790 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1792 // Get a count of how many bytes are to be pushed on the stack.
1793 unsigned NumBytes = CCInfo.getNextStackOffset();
1795 // This is a sibcall. The memory operands are available in caller's
1796 // own caller's stack.
1798 else if (GuaranteedTailCallOpt && CallConv == CallingConv::Fast)
1799 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1802 if (isTailCall && !IsSibcall) {
1803 // Lower arguments at fp - stackoffset + fpdiff.
1804 unsigned NumBytesCallerPushed =
1805 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1806 FPDiff = NumBytesCallerPushed - NumBytes;
1808 // Set the delta of movement of the returnaddr stackslot.
1809 // But only set if delta is greater than previous delta.
1810 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1811 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1815 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1817 SDValue RetAddrFrIdx;
1818 // Load return adress for tail calls.
1819 if (isTailCall && FPDiff)
1820 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1821 Is64Bit, FPDiff, dl);
1823 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1824 SmallVector<SDValue, 8> MemOpChains;
1827 // Walk the register/memloc assignments, inserting copies/loads. In the case
1828 // of tail call optimization arguments are handle later.
1829 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1830 CCValAssign &VA = ArgLocs[i];
1831 EVT RegVT = VA.getLocVT();
1832 SDValue Arg = Outs[i].Val;
1833 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1834 bool isByVal = Flags.isByVal();
1836 // Promote the value if needed.
1837 switch (VA.getLocInfo()) {
1838 default: llvm_unreachable("Unknown loc info!");
1839 case CCValAssign::Full: break;
1840 case CCValAssign::SExt:
1841 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1843 case CCValAssign::ZExt:
1844 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1846 case CCValAssign::AExt:
1847 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1848 // Special case: passing MMX values in XMM registers.
1849 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1850 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1851 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1853 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1855 case CCValAssign::BCvt:
1856 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1858 case CCValAssign::Indirect: {
1859 // Store the argument.
1860 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1861 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1862 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1863 PseudoSourceValue::getFixedStack(FI), 0,
1870 if (VA.isRegLoc()) {
1871 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1872 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1873 assert(VA.isMemLoc());
1874 if (StackPtr.getNode() == 0)
1875 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1876 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1877 dl, DAG, VA, Flags));
1881 if (!MemOpChains.empty())
1882 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1883 &MemOpChains[0], MemOpChains.size());
1885 // Build a sequence of copy-to-reg nodes chained together with token chain
1886 // and flag operands which copy the outgoing args into registers.
1888 // Tail call byval lowering might overwrite argument registers so in case of
1889 // tail call optimization the copies to registers are lowered later.
1891 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1892 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1893 RegsToPass[i].second, InFlag);
1894 InFlag = Chain.getValue(1);
1897 if (Subtarget->isPICStyleGOT()) {
1898 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1901 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1902 DAG.getNode(X86ISD::GlobalBaseReg,
1903 DebugLoc::getUnknownLoc(),
1906 InFlag = Chain.getValue(1);
1908 // If we are tail calling and generating PIC/GOT style code load the
1909 // address of the callee into ECX. The value in ecx is used as target of
1910 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1911 // for tail calls on PIC/GOT architectures. Normally we would just put the
1912 // address of GOT into ebx and then call target@PLT. But for tail calls
1913 // ebx would be restored (since ebx is callee saved) before jumping to the
1916 // Note: The actual moving to ECX is done further down.
1917 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1918 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1919 !G->getGlobal()->hasProtectedVisibility())
1920 Callee = LowerGlobalAddress(Callee, DAG);
1921 else if (isa<ExternalSymbolSDNode>(Callee))
1922 Callee = LowerExternalSymbol(Callee, DAG);
1926 if (Is64Bit && isVarArg) {
1927 // From AMD64 ABI document:
1928 // For calls that may call functions that use varargs or stdargs
1929 // (prototype-less calls or calls to functions containing ellipsis (...) in
1930 // the declaration) %al is used as hidden argument to specify the number
1931 // of SSE registers used. The contents of %al do not need to match exactly
1932 // the number of registers, but must be an ubound on the number of SSE
1933 // registers used and is in the range 0 - 8 inclusive.
1935 // FIXME: Verify this on Win64
1936 // Count the number of XMM registers allocated.
1937 static const unsigned XMMArgRegs[] = {
1938 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1939 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1941 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1942 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1943 && "SSE registers cannot be used when SSE is disabled");
1945 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1946 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1947 InFlag = Chain.getValue(1);
1951 // For tail calls lower the arguments to the 'real' stack slot.
1953 // Force all the incoming stack arguments to be loaded from the stack
1954 // before any new outgoing arguments are stored to the stack, because the
1955 // outgoing stack slots may alias the incoming argument stack slots, and
1956 // the alias isn't otherwise explicit. This is slightly more conservative
1957 // than necessary, because it means that each store effectively depends
1958 // on every argument instead of just those arguments it would clobber.
1959 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1961 SmallVector<SDValue, 8> MemOpChains2;
1964 // Do not flag preceeding copytoreg stuff together with the following stuff.
1966 if (GuaranteedTailCallOpt) {
1967 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1968 CCValAssign &VA = ArgLocs[i];
1971 assert(VA.isMemLoc());
1972 SDValue Arg = Outs[i].Val;
1973 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1974 // Create frame index.
1975 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1976 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1977 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
1978 FIN = DAG.getFrameIndex(FI, getPointerTy());
1980 if (Flags.isByVal()) {
1981 // Copy relative to framepointer.
1982 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1983 if (StackPtr.getNode() == 0)
1984 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1986 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1988 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
1992 // Store relative to framepointer.
1993 MemOpChains2.push_back(
1994 DAG.getStore(ArgChain, dl, Arg, FIN,
1995 PseudoSourceValue::getFixedStack(FI), 0,
2001 if (!MemOpChains2.empty())
2002 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2003 &MemOpChains2[0], MemOpChains2.size());
2005 // Copy arguments to their registers.
2006 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2007 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2008 RegsToPass[i].second, InFlag);
2009 InFlag = Chain.getValue(1);
2013 // Store the return address to the appropriate stack slot.
2014 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2018 bool WasGlobalOrExternal = false;
2019 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2020 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2021 // In the 64-bit large code model, we have to make all calls
2022 // through a register, since the call instruction's 32-bit
2023 // pc-relative offset may not be large enough to hold the whole
2025 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2026 WasGlobalOrExternal = true;
2027 // If the callee is a GlobalAddress node (quite common, every direct call
2028 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2031 // We should use extra load for direct calls to dllimported functions in
2033 GlobalValue *GV = G->getGlobal();
2034 if (!GV->hasDLLImportLinkage()) {
2035 unsigned char OpFlags = 0;
2037 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2038 // external symbols most go through the PLT in PIC mode. If the symbol
2039 // has hidden or protected visibility, or if it is static or local, then
2040 // we don't need to use the PLT - we can directly call it.
2041 if (Subtarget->isTargetELF() &&
2042 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2043 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2044 OpFlags = X86II::MO_PLT;
2045 } else if (Subtarget->isPICStyleStubAny() &&
2046 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2047 Subtarget->getDarwinVers() < 9) {
2048 // PC-relative references to external symbols should go through $stub,
2049 // unless we're building with the leopard linker or later, which
2050 // automatically synthesizes these stubs.
2051 OpFlags = X86II::MO_DARWIN_STUB;
2054 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2055 G->getOffset(), OpFlags);
2057 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2058 WasGlobalOrExternal = true;
2059 unsigned char OpFlags = 0;
2061 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2062 // symbols should go through the PLT.
2063 if (Subtarget->isTargetELF() &&
2064 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2065 OpFlags = X86II::MO_PLT;
2066 } else if (Subtarget->isPICStyleStubAny() &&
2067 Subtarget->getDarwinVers() < 9) {
2068 // PC-relative references to external symbols should go through $stub,
2069 // unless we're building with the leopard linker or later, which
2070 // automatically synthesizes these stubs.
2071 OpFlags = X86II::MO_DARWIN_STUB;
2074 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2078 if (isTailCall && !WasGlobalOrExternal) {
2079 // Force the address into a (call preserved) caller-saved register since
2080 // tailcall must happen after callee-saved registers are poped.
2081 // FIXME: Give it a special register class that contains caller-saved
2082 // register instead?
2083 unsigned TCReg = Is64Bit ? X86::R11 : X86::EAX;
2084 Chain = DAG.getCopyToReg(Chain, dl,
2085 DAG.getRegister(TCReg, getPointerTy()),
2087 Callee = DAG.getRegister(TCReg, getPointerTy());
2090 // Returns a chain & a flag for retval copy to use.
2091 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2092 SmallVector<SDValue, 8> Ops;
2094 if (!IsSibcall && isTailCall) {
2095 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2096 DAG.getIntPtrConstant(0, true), InFlag);
2097 InFlag = Chain.getValue(1);
2100 Ops.push_back(Chain);
2101 Ops.push_back(Callee);
2104 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2106 // Add argument registers to the end of the list so that they are known live
2108 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2109 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2110 RegsToPass[i].second.getValueType()));
2112 // Add an implicit use GOT pointer in EBX.
2113 if (!isTailCall && Subtarget->isPICStyleGOT())
2114 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2116 // Add an implicit use of AL for x86 vararg functions.
2117 if (Is64Bit && isVarArg)
2118 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2120 if (InFlag.getNode())
2121 Ops.push_back(InFlag);
2124 // If this is the first return lowered for this function, add the regs
2125 // to the liveout set for the function.
2126 if (MF.getRegInfo().liveout_empty()) {
2127 SmallVector<CCValAssign, 16> RVLocs;
2128 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2130 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2131 for (unsigned i = 0; i != RVLocs.size(); ++i)
2132 if (RVLocs[i].isRegLoc())
2133 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2136 assert(((Callee.getOpcode() == ISD::Register &&
2137 (cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
2138 cast<RegisterSDNode>(Callee)->getReg() == X86::R11)) ||
2139 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2140 Callee.getOpcode() == ISD::TargetGlobalAddress) &&
2141 "Expecting a global address, external symbol, or scratch register");
2143 return DAG.getNode(X86ISD::TC_RETURN, dl,
2144 NodeTys, &Ops[0], Ops.size());
2147 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2148 InFlag = Chain.getValue(1);
2150 // Create the CALLSEQ_END node.
2151 unsigned NumBytesForCalleeToPush;
2152 if (IsCalleePop(isVarArg, CallConv))
2153 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2154 else if (!Is64Bit && CallConv != CallingConv::Fast && IsStructRet)
2155 // If this is a call to a struct-return function, the callee
2156 // pops the hidden struct pointer, so we have to push it back.
2157 // This is common for Darwin/X86, Linux & Mingw32 targets.
2158 NumBytesForCalleeToPush = 4;
2160 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2162 // Returns a flag for retval copy to use.
2164 Chain = DAG.getCALLSEQ_END(Chain,
2165 DAG.getIntPtrConstant(NumBytes, true),
2166 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2169 InFlag = Chain.getValue(1);
2172 // Handle result values, copying them out of physregs into vregs that we
2174 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2175 Ins, dl, DAG, InVals);
2179 //===----------------------------------------------------------------------===//
2180 // Fast Calling Convention (tail call) implementation
2181 //===----------------------------------------------------------------------===//
2183 // Like std call, callee cleans arguments, convention except that ECX is
2184 // reserved for storing the tail called function address. Only 2 registers are
2185 // free for argument passing (inreg). Tail call optimization is performed
2187 // * tailcallopt is enabled
2188 // * caller/callee are fastcc
2189 // On X86_64 architecture with GOT-style position independent code only local
2190 // (within module) calls are supported at the moment.
2191 // To keep the stack aligned according to platform abi the function
2192 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2193 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2194 // If a tail called function callee has more arguments than the caller the
2195 // caller needs to make sure that there is room to move the RETADDR to. This is
2196 // achieved by reserving an area the size of the argument delta right after the
2197 // original REtADDR, but before the saved framepointer or the spilled registers
2198 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2210 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2211 /// for a 16 byte align requirement.
2212 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2213 SelectionDAG& DAG) {
2214 MachineFunction &MF = DAG.getMachineFunction();
2215 const TargetMachine &TM = MF.getTarget();
2216 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2217 unsigned StackAlignment = TFI.getStackAlignment();
2218 uint64_t AlignMask = StackAlignment - 1;
2219 int64_t Offset = StackSize;
2220 uint64_t SlotSize = TD->getPointerSize();
2221 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2222 // Number smaller than 12 so just add the difference.
2223 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2225 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2226 Offset = ((~AlignMask) & Offset) + StackAlignment +
2227 (StackAlignment-SlotSize);
2232 /// MatchingStackOffset - Return true if the given stack call argument is
2233 /// already available in the same position (relatively) of the caller's
2234 /// incoming argument stack.
2236 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2237 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2238 const X86InstrInfo *TII) {
2240 if (Arg.getOpcode() == ISD::CopyFromReg) {
2241 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2242 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2244 MachineInstr *Def = MRI->getVRegDef(VR);
2247 if (!Flags.isByVal()) {
2248 if (!TII->isLoadFromStackSlot(Def, FI))
2251 unsigned Opcode = Def->getOpcode();
2252 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2253 Def->getOperand(1).isFI()) {
2254 FI = Def->getOperand(1).getIndex();
2255 if (MFI->getObjectSize(FI) != Flags.getByValSize())
2261 LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg);
2264 SDValue Ptr = Ld->getBasePtr();
2265 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2268 FI = FINode->getIndex();
2271 if (!MFI->isFixedObjectIndex(FI))
2273 return Offset == MFI->getObjectOffset(FI);
2276 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2277 /// for tail call optimization. Targets which want to do tail call
2278 /// optimization should implement this function.
2280 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2281 CallingConv::ID CalleeCC,
2283 const SmallVectorImpl<ISD::OutputArg> &Outs,
2284 const SmallVectorImpl<ISD::InputArg> &Ins,
2285 SelectionDAG& DAG) const {
2286 if (CalleeCC != CallingConv::Fast &&
2287 CalleeCC != CallingConv::C)
2290 // If -tailcallopt is specified, make fastcc functions tail-callable.
2291 const Function *CallerF = DAG.getMachineFunction().getFunction();
2292 if (GuaranteedTailCallOpt) {
2293 if (CalleeCC == CallingConv::Fast &&
2294 CallerF->getCallingConv() == CalleeCC)
2299 // Look for obvious safe cases to perform tail call optimization that does not
2300 // requite ABI changes. This is what gcc calls sibcall.
2302 // Do not tail call optimize vararg calls for now.
2306 // If the callee takes no arguments then go on to check the results of the
2308 if (!Outs.empty()) {
2309 // Check if stack adjustment is needed. For now, do not do this if any
2310 // argument is passed on the stack.
2311 SmallVector<CCValAssign, 16> ArgLocs;
2312 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2313 ArgLocs, *DAG.getContext());
2314 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2315 if (CCInfo.getNextStackOffset()) {
2316 MachineFunction &MF = DAG.getMachineFunction();
2317 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2319 if (Subtarget->isTargetWin64())
2320 // Win64 ABI has additional complications.
2323 // Check if the arguments are already laid out in the right way as
2324 // the caller's fixed stack objects.
2325 MachineFrameInfo *MFI = MF.getFrameInfo();
2326 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2327 const X86InstrInfo *TII =
2328 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2329 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2330 CCValAssign &VA = ArgLocs[i];
2331 EVT RegVT = VA.getLocVT();
2332 SDValue Arg = Outs[i].Val;
2333 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2334 if (VA.getLocInfo() == CCValAssign::Indirect)
2336 if (!VA.isRegLoc()) {
2337 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2349 X86TargetLowering::createFastISel(MachineFunction &mf, MachineModuleInfo *mmo,
2351 DenseMap<const Value *, unsigned> &vm,
2352 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2353 DenseMap<const AllocaInst *, int> &am
2355 , SmallSet<Instruction*, 8> &cil
2358 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2366 //===----------------------------------------------------------------------===//
2367 // Other Lowering Hooks
2368 //===----------------------------------------------------------------------===//
2371 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2372 MachineFunction &MF = DAG.getMachineFunction();
2373 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2374 int ReturnAddrIndex = FuncInfo->getRAIndex();
2376 if (ReturnAddrIndex == 0) {
2377 // Set up a frame object for the return address.
2378 uint64_t SlotSize = TD->getPointerSize();
2379 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2381 FuncInfo->setRAIndex(ReturnAddrIndex);
2384 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2388 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2389 bool hasSymbolicDisplacement) {
2390 // Offset should fit into 32 bit immediate field.
2391 if (!isInt32(Offset))
2394 // If we don't have a symbolic displacement - we don't have any extra
2396 if (!hasSymbolicDisplacement)
2399 // FIXME: Some tweaks might be needed for medium code model.
2400 if (M != CodeModel::Small && M != CodeModel::Kernel)
2403 // For small code model we assume that latest object is 16MB before end of 31
2404 // bits boundary. We may also accept pretty large negative constants knowing
2405 // that all objects are in the positive half of address space.
2406 if (M == CodeModel::Small && Offset < 16*1024*1024)
2409 // For kernel code model we know that all object resist in the negative half
2410 // of 32bits address space. We may not accept negative offsets, since they may
2411 // be just off and we may accept pretty large positive ones.
2412 if (M == CodeModel::Kernel && Offset > 0)
2418 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2419 /// specific condition code, returning the condition code and the LHS/RHS of the
2420 /// comparison to make.
2421 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2422 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2424 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2425 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2426 // X > -1 -> X == 0, jump !sign.
2427 RHS = DAG.getConstant(0, RHS.getValueType());
2428 return X86::COND_NS;
2429 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2430 // X < 0 -> X == 0, jump on sign.
2432 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2434 RHS = DAG.getConstant(0, RHS.getValueType());
2435 return X86::COND_LE;
2439 switch (SetCCOpcode) {
2440 default: llvm_unreachable("Invalid integer condition!");
2441 case ISD::SETEQ: return X86::COND_E;
2442 case ISD::SETGT: return X86::COND_G;
2443 case ISD::SETGE: return X86::COND_GE;
2444 case ISD::SETLT: return X86::COND_L;
2445 case ISD::SETLE: return X86::COND_LE;
2446 case ISD::SETNE: return X86::COND_NE;
2447 case ISD::SETULT: return X86::COND_B;
2448 case ISD::SETUGT: return X86::COND_A;
2449 case ISD::SETULE: return X86::COND_BE;
2450 case ISD::SETUGE: return X86::COND_AE;
2454 // First determine if it is required or is profitable to flip the operands.
2456 // If LHS is a foldable load, but RHS is not, flip the condition.
2457 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2458 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2459 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2460 std::swap(LHS, RHS);
2463 switch (SetCCOpcode) {
2469 std::swap(LHS, RHS);
2473 // On a floating point condition, the flags are set as follows:
2475 // 0 | 0 | 0 | X > Y
2476 // 0 | 0 | 1 | X < Y
2477 // 1 | 0 | 0 | X == Y
2478 // 1 | 1 | 1 | unordered
2479 switch (SetCCOpcode) {
2480 default: llvm_unreachable("Condcode should be pre-legalized away");
2482 case ISD::SETEQ: return X86::COND_E;
2483 case ISD::SETOLT: // flipped
2485 case ISD::SETGT: return X86::COND_A;
2486 case ISD::SETOLE: // flipped
2488 case ISD::SETGE: return X86::COND_AE;
2489 case ISD::SETUGT: // flipped
2491 case ISD::SETLT: return X86::COND_B;
2492 case ISD::SETUGE: // flipped
2494 case ISD::SETLE: return X86::COND_BE;
2496 case ISD::SETNE: return X86::COND_NE;
2497 case ISD::SETUO: return X86::COND_P;
2498 case ISD::SETO: return X86::COND_NP;
2500 case ISD::SETUNE: return X86::COND_INVALID;
2504 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2505 /// code. Current x86 isa includes the following FP cmov instructions:
2506 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2507 static bool hasFPCMov(unsigned X86CC) {
2523 /// isFPImmLegal - Returns true if the target can instruction select the
2524 /// specified FP immediate natively. If false, the legalizer will
2525 /// materialize the FP immediate as a load from a constant pool.
2526 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2527 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2528 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2534 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2535 /// the specified range (L, H].
2536 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2537 return (Val < 0) || (Val >= Low && Val < Hi);
2540 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2541 /// specified value.
2542 static bool isUndefOrEqual(int Val, int CmpVal) {
2543 if (Val < 0 || Val == CmpVal)
2548 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2549 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2550 /// the second operand.
2551 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2552 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2553 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2554 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2555 return (Mask[0] < 2 && Mask[1] < 2);
2559 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2560 SmallVector<int, 8> M;
2562 return ::isPSHUFDMask(M, N->getValueType(0));
2565 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2566 /// is suitable for input to PSHUFHW.
2567 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2568 if (VT != MVT::v8i16)
2571 // Lower quadword copied in order or undef.
2572 for (int i = 0; i != 4; ++i)
2573 if (Mask[i] >= 0 && Mask[i] != i)
2576 // Upper quadword shuffled.
2577 for (int i = 4; i != 8; ++i)
2578 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2584 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2585 SmallVector<int, 8> M;
2587 return ::isPSHUFHWMask(M, N->getValueType(0));
2590 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2591 /// is suitable for input to PSHUFLW.
2592 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2593 if (VT != MVT::v8i16)
2596 // Upper quadword copied in order.
2597 for (int i = 4; i != 8; ++i)
2598 if (Mask[i] >= 0 && Mask[i] != i)
2601 // Lower quadword shuffled.
2602 for (int i = 0; i != 4; ++i)
2609 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2610 SmallVector<int, 8> M;
2612 return ::isPSHUFLWMask(M, N->getValueType(0));
2615 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2616 /// is suitable for input to PALIGNR.
2617 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2619 int i, e = VT.getVectorNumElements();
2621 // Do not handle v2i64 / v2f64 shuffles with palignr.
2622 if (e < 4 || !hasSSSE3)
2625 for (i = 0; i != e; ++i)
2629 // All undef, not a palignr.
2633 // Determine if it's ok to perform a palignr with only the LHS, since we
2634 // don't have access to the actual shuffle elements to see if RHS is undef.
2635 bool Unary = Mask[i] < (int)e;
2636 bool NeedsUnary = false;
2638 int s = Mask[i] - i;
2640 // Check the rest of the elements to see if they are consecutive.
2641 for (++i; i != e; ++i) {
2646 Unary = Unary && (m < (int)e);
2647 NeedsUnary = NeedsUnary || (m < s);
2649 if (NeedsUnary && !Unary)
2651 if (Unary && m != ((s+i) & (e-1)))
2653 if (!Unary && m != (s+i))
2659 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2660 SmallVector<int, 8> M;
2662 return ::isPALIGNRMask(M, N->getValueType(0), true);
2665 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2666 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2667 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2668 int NumElems = VT.getVectorNumElements();
2669 if (NumElems != 2 && NumElems != 4)
2672 int Half = NumElems / 2;
2673 for (int i = 0; i < Half; ++i)
2674 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2676 for (int i = Half; i < NumElems; ++i)
2677 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2683 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2684 SmallVector<int, 8> M;
2686 return ::isSHUFPMask(M, N->getValueType(0));
2689 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2690 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2691 /// half elements to come from vector 1 (which would equal the dest.) and
2692 /// the upper half to come from vector 2.
2693 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2694 int NumElems = VT.getVectorNumElements();
2696 if (NumElems != 2 && NumElems != 4)
2699 int Half = NumElems / 2;
2700 for (int i = 0; i < Half; ++i)
2701 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2703 for (int i = Half; i < NumElems; ++i)
2704 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2709 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2710 SmallVector<int, 8> M;
2712 return isCommutedSHUFPMask(M, N->getValueType(0));
2715 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2716 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2717 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2718 if (N->getValueType(0).getVectorNumElements() != 4)
2721 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2722 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2723 isUndefOrEqual(N->getMaskElt(1), 7) &&
2724 isUndefOrEqual(N->getMaskElt(2), 2) &&
2725 isUndefOrEqual(N->getMaskElt(3), 3);
2728 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2729 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2731 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2732 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2737 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2738 isUndefOrEqual(N->getMaskElt(1), 3) &&
2739 isUndefOrEqual(N->getMaskElt(2), 2) &&
2740 isUndefOrEqual(N->getMaskElt(3), 3);
2743 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2744 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2745 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2746 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2748 if (NumElems != 2 && NumElems != 4)
2751 for (unsigned i = 0; i < NumElems/2; ++i)
2752 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2755 for (unsigned i = NumElems/2; i < NumElems; ++i)
2756 if (!isUndefOrEqual(N->getMaskElt(i), i))
2762 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2763 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2764 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2765 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2767 if (NumElems != 2 && NumElems != 4)
2770 for (unsigned i = 0; i < NumElems/2; ++i)
2771 if (!isUndefOrEqual(N->getMaskElt(i), i))
2774 for (unsigned i = 0; i < NumElems/2; ++i)
2775 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2781 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2782 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2783 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2784 bool V2IsSplat = false) {
2785 int NumElts = VT.getVectorNumElements();
2786 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2789 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2791 int BitI1 = Mask[i+1];
2792 if (!isUndefOrEqual(BitI, j))
2795 if (!isUndefOrEqual(BitI1, NumElts))
2798 if (!isUndefOrEqual(BitI1, j + NumElts))
2805 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2806 SmallVector<int, 8> M;
2808 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2811 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2812 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2813 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2814 bool V2IsSplat = false) {
2815 int NumElts = VT.getVectorNumElements();
2816 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2819 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2821 int BitI1 = Mask[i+1];
2822 if (!isUndefOrEqual(BitI, j + NumElts/2))
2825 if (isUndefOrEqual(BitI1, NumElts))
2828 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2835 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2836 SmallVector<int, 8> M;
2838 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2841 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2842 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2844 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2845 int NumElems = VT.getVectorNumElements();
2846 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2849 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2851 int BitI1 = Mask[i+1];
2852 if (!isUndefOrEqual(BitI, j))
2854 if (!isUndefOrEqual(BitI1, j))
2860 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2861 SmallVector<int, 8> M;
2863 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2866 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2867 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2869 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2870 int NumElems = VT.getVectorNumElements();
2871 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2874 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2876 int BitI1 = Mask[i+1];
2877 if (!isUndefOrEqual(BitI, j))
2879 if (!isUndefOrEqual(BitI1, j))
2885 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2886 SmallVector<int, 8> M;
2888 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2891 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2892 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2893 /// MOVSD, and MOVD, i.e. setting the lowest element.
2894 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2895 if (VT.getVectorElementType().getSizeInBits() < 32)
2898 int NumElts = VT.getVectorNumElements();
2900 if (!isUndefOrEqual(Mask[0], NumElts))
2903 for (int i = 1; i < NumElts; ++i)
2904 if (!isUndefOrEqual(Mask[i], i))
2910 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2911 SmallVector<int, 8> M;
2913 return ::isMOVLMask(M, N->getValueType(0));
2916 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2917 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2918 /// element of vector 2 and the other elements to come from vector 1 in order.
2919 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2920 bool V2IsSplat = false, bool V2IsUndef = false) {
2921 int NumOps = VT.getVectorNumElements();
2922 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2925 if (!isUndefOrEqual(Mask[0], 0))
2928 for (int i = 1; i < NumOps; ++i)
2929 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2930 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2931 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2937 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2938 bool V2IsUndef = false) {
2939 SmallVector<int, 8> M;
2941 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2944 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2945 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2946 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2947 if (N->getValueType(0).getVectorNumElements() != 4)
2950 // Expect 1, 1, 3, 3
2951 for (unsigned i = 0; i < 2; ++i) {
2952 int Elt = N->getMaskElt(i);
2953 if (Elt >= 0 && Elt != 1)
2958 for (unsigned i = 2; i < 4; ++i) {
2959 int Elt = N->getMaskElt(i);
2960 if (Elt >= 0 && Elt != 3)
2965 // Don't use movshdup if it can be done with a shufps.
2966 // FIXME: verify that matching u, u, 3, 3 is what we want.
2970 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2971 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2972 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2973 if (N->getValueType(0).getVectorNumElements() != 4)
2976 // Expect 0, 0, 2, 2
2977 for (unsigned i = 0; i < 2; ++i)
2978 if (N->getMaskElt(i) > 0)
2982 for (unsigned i = 2; i < 4; ++i) {
2983 int Elt = N->getMaskElt(i);
2984 if (Elt >= 0 && Elt != 2)
2989 // Don't use movsldup if it can be done with a shufps.
2993 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2994 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2995 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2996 int e = N->getValueType(0).getVectorNumElements() / 2;
2998 for (int i = 0; i < e; ++i)
2999 if (!isUndefOrEqual(N->getMaskElt(i), i))
3001 for (int i = 0; i < e; ++i)
3002 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3007 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3008 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3009 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3010 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3011 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3013 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3015 for (int i = 0; i < NumOperands; ++i) {
3016 int Val = SVOp->getMaskElt(NumOperands-i-1);
3017 if (Val < 0) Val = 0;
3018 if (Val >= NumOperands) Val -= NumOperands;
3020 if (i != NumOperands - 1)
3026 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3027 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3028 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3029 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3031 // 8 nodes, but we only care about the last 4.
3032 for (unsigned i = 7; i >= 4; --i) {
3033 int Val = SVOp->getMaskElt(i);
3042 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3043 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3044 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3045 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3047 // 8 nodes, but we only care about the first 4.
3048 for (int i = 3; i >= 0; --i) {
3049 int Val = SVOp->getMaskElt(i);
3058 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3059 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3060 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3061 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3062 EVT VVT = N->getValueType(0);
3063 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3067 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3068 Val = SVOp->getMaskElt(i);
3072 return (Val - i) * EltSize;
3075 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3077 bool X86::isZeroNode(SDValue Elt) {
3078 return ((isa<ConstantSDNode>(Elt) &&
3079 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
3080 (isa<ConstantFPSDNode>(Elt) &&
3081 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3084 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3085 /// their permute mask.
3086 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3087 SelectionDAG &DAG) {
3088 EVT VT = SVOp->getValueType(0);
3089 unsigned NumElems = VT.getVectorNumElements();
3090 SmallVector<int, 8> MaskVec;
3092 for (unsigned i = 0; i != NumElems; ++i) {
3093 int idx = SVOp->getMaskElt(i);
3095 MaskVec.push_back(idx);
3096 else if (idx < (int)NumElems)
3097 MaskVec.push_back(idx + NumElems);
3099 MaskVec.push_back(idx - NumElems);
3101 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3102 SVOp->getOperand(0), &MaskVec[0]);
3105 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3106 /// the two vector operands have swapped position.
3107 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3108 unsigned NumElems = VT.getVectorNumElements();
3109 for (unsigned i = 0; i != NumElems; ++i) {
3113 else if (idx < (int)NumElems)
3114 Mask[i] = idx + NumElems;
3116 Mask[i] = idx - NumElems;
3120 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3121 /// match movhlps. The lower half elements should come from upper half of
3122 /// V1 (and in order), and the upper half elements should come from the upper
3123 /// half of V2 (and in order).
3124 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3125 if (Op->getValueType(0).getVectorNumElements() != 4)
3127 for (unsigned i = 0, e = 2; i != e; ++i)
3128 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3130 for (unsigned i = 2; i != 4; ++i)
3131 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3136 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3137 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3139 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3140 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3142 N = N->getOperand(0).getNode();
3143 if (!ISD::isNON_EXTLoad(N))
3146 *LD = cast<LoadSDNode>(N);
3150 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3151 /// match movlp{s|d}. The lower half elements should come from lower half of
3152 /// V1 (and in order), and the upper half elements should come from the upper
3153 /// half of V2 (and in order). And since V1 will become the source of the
3154 /// MOVLP, it must be either a vector load or a scalar load to vector.
3155 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3156 ShuffleVectorSDNode *Op) {
3157 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3159 // Is V2 is a vector load, don't do this transformation. We will try to use
3160 // load folding shufps op.
3161 if (ISD::isNON_EXTLoad(V2))
3164 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3166 if (NumElems != 2 && NumElems != 4)
3168 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3169 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3171 for (unsigned i = NumElems/2; i != NumElems; ++i)
3172 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3177 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3179 static bool isSplatVector(SDNode *N) {
3180 if (N->getOpcode() != ISD::BUILD_VECTOR)
3183 SDValue SplatValue = N->getOperand(0);
3184 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3185 if (N->getOperand(i) != SplatValue)
3190 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3191 /// to an zero vector.
3192 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3193 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3194 SDValue V1 = N->getOperand(0);
3195 SDValue V2 = N->getOperand(1);
3196 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3197 for (unsigned i = 0; i != NumElems; ++i) {
3198 int Idx = N->getMaskElt(i);
3199 if (Idx >= (int)NumElems) {
3200 unsigned Opc = V2.getOpcode();
3201 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3203 if (Opc != ISD::BUILD_VECTOR ||
3204 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3206 } else if (Idx >= 0) {
3207 unsigned Opc = V1.getOpcode();
3208 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3210 if (Opc != ISD::BUILD_VECTOR ||
3211 !X86::isZeroNode(V1.getOperand(Idx)))
3218 /// getZeroVector - Returns a vector of specified type with all zero elements.
3220 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3222 assert(VT.isVector() && "Expected a vector type");
3224 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3225 // type. This ensures they get CSE'd.
3227 if (VT.getSizeInBits() == 64) { // MMX
3228 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3229 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3230 } else if (HasSSE2) { // SSE2
3231 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3232 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3234 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3235 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3237 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3240 /// getOnesVector - Returns a vector of specified type with all bits set.
3242 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3243 assert(VT.isVector() && "Expected a vector type");
3245 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3246 // type. This ensures they get CSE'd.
3247 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3249 if (VT.getSizeInBits() == 64) // MMX
3250 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3252 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3253 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3257 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3258 /// that point to V2 points to its first element.
3259 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3260 EVT VT = SVOp->getValueType(0);
3261 unsigned NumElems = VT.getVectorNumElements();
3263 bool Changed = false;
3264 SmallVector<int, 8> MaskVec;
3265 SVOp->getMask(MaskVec);
3267 for (unsigned i = 0; i != NumElems; ++i) {
3268 if (MaskVec[i] > (int)NumElems) {
3269 MaskVec[i] = NumElems;
3274 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3275 SVOp->getOperand(1), &MaskVec[0]);
3276 return SDValue(SVOp, 0);
3279 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3280 /// operation of specified width.
3281 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3283 unsigned NumElems = VT.getVectorNumElements();
3284 SmallVector<int, 8> Mask;
3285 Mask.push_back(NumElems);
3286 for (unsigned i = 1; i != NumElems; ++i)
3288 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3291 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3292 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3294 unsigned NumElems = VT.getVectorNumElements();
3295 SmallVector<int, 8> Mask;
3296 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3298 Mask.push_back(i + NumElems);
3300 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3303 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3304 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3306 unsigned NumElems = VT.getVectorNumElements();
3307 unsigned Half = NumElems/2;
3308 SmallVector<int, 8> Mask;
3309 for (unsigned i = 0; i != Half; ++i) {
3310 Mask.push_back(i + Half);
3311 Mask.push_back(i + NumElems + Half);
3313 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3316 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3317 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3319 if (SV->getValueType(0).getVectorNumElements() <= 4)
3320 return SDValue(SV, 0);
3322 EVT PVT = MVT::v4f32;
3323 EVT VT = SV->getValueType(0);
3324 DebugLoc dl = SV->getDebugLoc();
3325 SDValue V1 = SV->getOperand(0);
3326 int NumElems = VT.getVectorNumElements();
3327 int EltNo = SV->getSplatIndex();
3329 // unpack elements to the correct location
3330 while (NumElems > 4) {
3331 if (EltNo < NumElems/2) {
3332 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3334 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3335 EltNo -= NumElems/2;
3340 // Perform the splat.
3341 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3342 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3343 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3344 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3347 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3348 /// vector of zero or undef vector. This produces a shuffle where the low
3349 /// element of V2 is swizzled into the zero/undef vector, landing at element
3350 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3351 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3352 bool isZero, bool HasSSE2,
3353 SelectionDAG &DAG) {
3354 EVT VT = V2.getValueType();
3356 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3357 unsigned NumElems = VT.getVectorNumElements();
3358 SmallVector<int, 16> MaskVec;
3359 for (unsigned i = 0; i != NumElems; ++i)
3360 // If this is the insertion idx, put the low elt of V2 here.
3361 MaskVec.push_back(i == Idx ? NumElems : i);
3362 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3365 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3366 /// a shuffle that is zero.
3368 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3369 bool Low, SelectionDAG &DAG) {
3370 unsigned NumZeros = 0;
3371 for (int i = 0; i < NumElems; ++i) {
3372 unsigned Index = Low ? i : NumElems-i-1;
3373 int Idx = SVOp->getMaskElt(Index);
3378 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3379 if (Elt.getNode() && X86::isZeroNode(Elt))
3387 /// isVectorShift - Returns true if the shuffle can be implemented as a
3388 /// logical left or right shift of a vector.
3389 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3390 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3391 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3392 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3395 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3398 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3402 bool SeenV1 = false;
3403 bool SeenV2 = false;
3404 for (int i = NumZeros; i < NumElems; ++i) {
3405 int Val = isLeft ? (i - NumZeros) : i;
3406 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3418 if (SeenV1 && SeenV2)
3421 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3427 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3429 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3430 unsigned NumNonZero, unsigned NumZero,
3431 SelectionDAG &DAG, TargetLowering &TLI) {
3435 DebugLoc dl = Op.getDebugLoc();
3438 for (unsigned i = 0; i < 16; ++i) {
3439 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3440 if (ThisIsNonZero && First) {
3442 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3444 V = DAG.getUNDEF(MVT::v8i16);
3449 SDValue ThisElt(0, 0), LastElt(0, 0);
3450 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3451 if (LastIsNonZero) {
3452 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3453 MVT::i16, Op.getOperand(i-1));
3455 if (ThisIsNonZero) {
3456 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3457 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3458 ThisElt, DAG.getConstant(8, MVT::i8));
3460 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3464 if (ThisElt.getNode())
3465 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3466 DAG.getIntPtrConstant(i/2));
3470 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3473 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3475 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3476 unsigned NumNonZero, unsigned NumZero,
3477 SelectionDAG &DAG, TargetLowering &TLI) {
3481 DebugLoc dl = Op.getDebugLoc();
3484 for (unsigned i = 0; i < 8; ++i) {
3485 bool isNonZero = (NonZeros & (1 << i)) != 0;
3489 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3491 V = DAG.getUNDEF(MVT::v8i16);
3494 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3495 MVT::v8i16, V, Op.getOperand(i),
3496 DAG.getIntPtrConstant(i));
3503 /// getVShift - Return a vector logical shift node.
3505 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3506 unsigned NumBits, SelectionDAG &DAG,
3507 const TargetLowering &TLI, DebugLoc dl) {
3508 bool isMMX = VT.getSizeInBits() == 64;
3509 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3510 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3511 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3512 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3513 DAG.getNode(Opc, dl, ShVT, SrcOp,
3514 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3518 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3519 SelectionDAG &DAG) {
3521 // Check if the scalar load can be widened into a vector load. And if
3522 // the address is "base + cst" see if the cst can be "absorbed" into
3523 // the shuffle mask.
3524 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3525 SDValue Ptr = LD->getBasePtr();
3526 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3528 EVT PVT = LD->getValueType(0);
3529 if (PVT != MVT::i32 && PVT != MVT::f32)
3534 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3535 FI = FINode->getIndex();
3537 } else if (Ptr.getOpcode() == ISD::ADD &&
3538 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3539 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3540 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3541 Offset = Ptr.getConstantOperandVal(1);
3542 Ptr = Ptr.getOperand(0);
3547 SDValue Chain = LD->getChain();
3548 // Make sure the stack object alignment is at least 16.
3549 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3550 if (DAG.InferPtrAlignment(Ptr) < 16) {
3551 if (MFI->isFixedObjectIndex(FI)) {
3552 // Can't change the alignment. FIXME: It's possible to compute
3553 // the exact stack offset and reference FI + adjust offset instead.
3554 // If someone *really* cares about this. That's the way to implement it.
3557 MFI->setObjectAlignment(FI, 16);
3561 // (Offset % 16) must be multiple of 4. Then address is then
3562 // Ptr + (Offset & ~15).
3565 if ((Offset % 16) & 3)
3567 int64_t StartOffset = Offset & ~15;
3569 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3570 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3572 int EltNo = (Offset - StartOffset) >> 2;
3573 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3574 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3575 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3577 // Canonicalize it to a v4i32 shuffle.
3578 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3579 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3580 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3581 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3588 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3589 DebugLoc dl = Op.getDebugLoc();
3590 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3591 if (ISD::isBuildVectorAllZeros(Op.getNode())
3592 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3593 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3594 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3595 // eliminated on x86-32 hosts.
3596 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3599 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3600 return getOnesVector(Op.getValueType(), DAG, dl);
3601 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3604 EVT VT = Op.getValueType();
3605 EVT ExtVT = VT.getVectorElementType();
3606 unsigned EVTBits = ExtVT.getSizeInBits();
3608 unsigned NumElems = Op.getNumOperands();
3609 unsigned NumZero = 0;
3610 unsigned NumNonZero = 0;
3611 unsigned NonZeros = 0;
3612 bool IsAllConstants = true;
3613 SmallSet<SDValue, 8> Values;
3614 for (unsigned i = 0; i < NumElems; ++i) {
3615 SDValue Elt = Op.getOperand(i);
3616 if (Elt.getOpcode() == ISD::UNDEF)
3619 if (Elt.getOpcode() != ISD::Constant &&
3620 Elt.getOpcode() != ISD::ConstantFP)
3621 IsAllConstants = false;
3622 if (X86::isZeroNode(Elt))
3625 NonZeros |= (1 << i);
3630 if (NumNonZero == 0) {
3631 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3632 return DAG.getUNDEF(VT);
3635 // Special case for single non-zero, non-undef, element.
3636 if (NumNonZero == 1) {
3637 unsigned Idx = CountTrailingZeros_32(NonZeros);
3638 SDValue Item = Op.getOperand(Idx);
3640 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3641 // the value are obviously zero, truncate the value to i32 and do the
3642 // insertion that way. Only do this if the value is non-constant or if the
3643 // value is a constant being inserted into element 0. It is cheaper to do
3644 // a constant pool load than it is to do a movd + shuffle.
3645 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3646 (!IsAllConstants || Idx == 0)) {
3647 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3648 // Handle MMX and SSE both.
3649 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3650 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3652 // Truncate the value (which may itself be a constant) to i32, and
3653 // convert it to a vector with movd (S2V+shuffle to zero extend).
3654 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3655 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3656 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3657 Subtarget->hasSSE2(), DAG);
3659 // Now we have our 32-bit value zero extended in the low element of
3660 // a vector. If Idx != 0, swizzle it into place.
3662 SmallVector<int, 4> Mask;
3663 Mask.push_back(Idx);
3664 for (unsigned i = 1; i != VecElts; ++i)
3666 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3667 DAG.getUNDEF(Item.getValueType()),
3670 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3674 // If we have a constant or non-constant insertion into the low element of
3675 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3676 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3677 // depending on what the source datatype is.
3680 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3681 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3682 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3683 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3684 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3685 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3687 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3688 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3689 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3690 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3691 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3692 Subtarget->hasSSE2(), DAG);
3693 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3697 // Is it a vector logical left shift?
3698 if (NumElems == 2 && Idx == 1 &&
3699 X86::isZeroNode(Op.getOperand(0)) &&
3700 !X86::isZeroNode(Op.getOperand(1))) {
3701 unsigned NumBits = VT.getSizeInBits();
3702 return getVShift(true, VT,
3703 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3704 VT, Op.getOperand(1)),
3705 NumBits/2, DAG, *this, dl);
3708 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3711 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3712 // is a non-constant being inserted into an element other than the low one,
3713 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3714 // movd/movss) to move this into the low element, then shuffle it into
3716 if (EVTBits == 32) {
3717 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3719 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3720 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3721 Subtarget->hasSSE2(), DAG);
3722 SmallVector<int, 8> MaskVec;
3723 for (unsigned i = 0; i < NumElems; i++)
3724 MaskVec.push_back(i == Idx ? 0 : 1);
3725 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3729 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3730 if (Values.size() == 1) {
3731 if (EVTBits == 32) {
3732 // Instead of a shuffle like this:
3733 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3734 // Check if it's possible to issue this instead.
3735 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3736 unsigned Idx = CountTrailingZeros_32(NonZeros);
3737 SDValue Item = Op.getOperand(Idx);
3738 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3739 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3744 // A vector full of immediates; various special cases are already
3745 // handled, so this is best done with a single constant-pool load.
3749 // Let legalizer expand 2-wide build_vectors.
3750 if (EVTBits == 64) {
3751 if (NumNonZero == 1) {
3752 // One half is zero or undef.
3753 unsigned Idx = CountTrailingZeros_32(NonZeros);
3754 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3755 Op.getOperand(Idx));
3756 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3757 Subtarget->hasSSE2(), DAG);
3762 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3763 if (EVTBits == 8 && NumElems == 16) {
3764 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3766 if (V.getNode()) return V;
3769 if (EVTBits == 16 && NumElems == 8) {
3770 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3772 if (V.getNode()) return V;
3775 // If element VT is == 32 bits, turn it into a number of shuffles.
3776 SmallVector<SDValue, 8> V;
3778 if (NumElems == 4 && NumZero > 0) {
3779 for (unsigned i = 0; i < 4; ++i) {
3780 bool isZero = !(NonZeros & (1 << i));
3782 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3784 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3787 for (unsigned i = 0; i < 2; ++i) {
3788 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3791 V[i] = V[i*2]; // Must be a zero vector.
3794 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3797 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3800 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3805 SmallVector<int, 8> MaskVec;
3806 bool Reverse = (NonZeros & 0x3) == 2;
3807 for (unsigned i = 0; i < 2; ++i)
3808 MaskVec.push_back(Reverse ? 1-i : i);
3809 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3810 for (unsigned i = 0; i < 2; ++i)
3811 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3812 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3815 if (Values.size() > 2) {
3816 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3817 // values to be inserted is equal to the number of elements, in which case
3818 // use the unpack code below in the hopes of matching the consecutive elts
3819 // load merge pattern for shuffles.
3820 // FIXME: We could probably just check that here directly.
3821 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3822 getSubtarget()->hasSSE41()) {
3823 V[0] = DAG.getUNDEF(VT);
3824 for (unsigned i = 0; i < NumElems; ++i)
3825 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3826 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3827 Op.getOperand(i), DAG.getIntPtrConstant(i));
3830 // Expand into a number of unpckl*.
3832 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3833 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3834 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3835 for (unsigned i = 0; i < NumElems; ++i)
3836 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3838 while (NumElems != 0) {
3839 for (unsigned i = 0; i < NumElems; ++i)
3840 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3850 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3851 // We support concatenate two MMX registers and place them in a MMX
3852 // register. This is better than doing a stack convert.
3853 DebugLoc dl = Op.getDebugLoc();
3854 EVT ResVT = Op.getValueType();
3855 assert(Op.getNumOperands() == 2);
3856 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
3857 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
3859 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
3860 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3861 InVec = Op.getOperand(1);
3862 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3863 unsigned NumElts = ResVT.getVectorNumElements();
3864 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3865 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
3866 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
3868 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
3869 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3870 Mask[0] = 0; Mask[1] = 2;
3871 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
3873 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3876 // v8i16 shuffles - Prefer shuffles in the following order:
3877 // 1. [all] pshuflw, pshufhw, optional move
3878 // 2. [ssse3] 1 x pshufb
3879 // 3. [ssse3] 2 x pshufb + 1 x por
3880 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3882 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3883 SelectionDAG &DAG, X86TargetLowering &TLI) {
3884 SDValue V1 = SVOp->getOperand(0);
3885 SDValue V2 = SVOp->getOperand(1);
3886 DebugLoc dl = SVOp->getDebugLoc();
3887 SmallVector<int, 8> MaskVals;
3889 // Determine if more than 1 of the words in each of the low and high quadwords
3890 // of the result come from the same quadword of one of the two inputs. Undef
3891 // mask values count as coming from any quadword, for better codegen.
3892 SmallVector<unsigned, 4> LoQuad(4);
3893 SmallVector<unsigned, 4> HiQuad(4);
3894 BitVector InputQuads(4);
3895 for (unsigned i = 0; i < 8; ++i) {
3896 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3897 int EltIdx = SVOp->getMaskElt(i);
3898 MaskVals.push_back(EltIdx);
3907 InputQuads.set(EltIdx / 4);
3910 int BestLoQuad = -1;
3911 unsigned MaxQuad = 1;
3912 for (unsigned i = 0; i < 4; ++i) {
3913 if (LoQuad[i] > MaxQuad) {
3915 MaxQuad = LoQuad[i];
3919 int BestHiQuad = -1;
3921 for (unsigned i = 0; i < 4; ++i) {
3922 if (HiQuad[i] > MaxQuad) {
3924 MaxQuad = HiQuad[i];
3928 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3929 // of the two input vectors, shuffle them into one input vector so only a
3930 // single pshufb instruction is necessary. If There are more than 2 input
3931 // quads, disable the next transformation since it does not help SSSE3.
3932 bool V1Used = InputQuads[0] || InputQuads[1];
3933 bool V2Used = InputQuads[2] || InputQuads[3];
3934 if (TLI.getSubtarget()->hasSSSE3()) {
3935 if (InputQuads.count() == 2 && V1Used && V2Used) {
3936 BestLoQuad = InputQuads.find_first();
3937 BestHiQuad = InputQuads.find_next(BestLoQuad);
3939 if (InputQuads.count() > 2) {
3945 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3946 // the shuffle mask. If a quad is scored as -1, that means that it contains
3947 // words from all 4 input quadwords.
3949 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3950 SmallVector<int, 8> MaskV;
3951 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3952 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3953 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3954 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3955 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3956 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3958 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3959 // source words for the shuffle, to aid later transformations.
3960 bool AllWordsInNewV = true;
3961 bool InOrder[2] = { true, true };
3962 for (unsigned i = 0; i != 8; ++i) {
3963 int idx = MaskVals[i];
3965 InOrder[i/4] = false;
3966 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3968 AllWordsInNewV = false;
3972 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3973 if (AllWordsInNewV) {
3974 for (int i = 0; i != 8; ++i) {
3975 int idx = MaskVals[i];
3978 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3979 if ((idx != i) && idx < 4)
3981 if ((idx != i) && idx > 3)
3990 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3991 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3992 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3993 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3994 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3998 // If we have SSSE3, and all words of the result are from 1 input vector,
3999 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4000 // is present, fall back to case 4.
4001 if (TLI.getSubtarget()->hasSSSE3()) {
4002 SmallVector<SDValue,16> pshufbMask;
4004 // If we have elements from both input vectors, set the high bit of the
4005 // shuffle mask element to zero out elements that come from V2 in the V1
4006 // mask, and elements that come from V1 in the V2 mask, so that the two
4007 // results can be OR'd together.
4008 bool TwoInputs = V1Used && V2Used;
4009 for (unsigned i = 0; i != 8; ++i) {
4010 int EltIdx = MaskVals[i] * 2;
4011 if (TwoInputs && (EltIdx >= 16)) {
4012 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4013 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4016 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4017 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4019 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4020 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4021 DAG.getNode(ISD::BUILD_VECTOR, dl,
4022 MVT::v16i8, &pshufbMask[0], 16));
4024 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4026 // Calculate the shuffle mask for the second input, shuffle it, and
4027 // OR it with the first shuffled input.
4029 for (unsigned i = 0; i != 8; ++i) {
4030 int EltIdx = MaskVals[i] * 2;
4032 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4033 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4036 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4037 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4039 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4040 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4041 DAG.getNode(ISD::BUILD_VECTOR, dl,
4042 MVT::v16i8, &pshufbMask[0], 16));
4043 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4044 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4047 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4048 // and update MaskVals with new element order.
4049 BitVector InOrder(8);
4050 if (BestLoQuad >= 0) {
4051 SmallVector<int, 8> MaskV;
4052 for (int i = 0; i != 4; ++i) {
4053 int idx = MaskVals[i];
4055 MaskV.push_back(-1);
4057 } else if ((idx / 4) == BestLoQuad) {
4058 MaskV.push_back(idx & 3);
4061 MaskV.push_back(-1);
4064 for (unsigned i = 4; i != 8; ++i)
4066 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4070 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4071 // and update MaskVals with the new element order.
4072 if (BestHiQuad >= 0) {
4073 SmallVector<int, 8> MaskV;
4074 for (unsigned i = 0; i != 4; ++i)
4076 for (unsigned i = 4; i != 8; ++i) {
4077 int idx = MaskVals[i];
4079 MaskV.push_back(-1);
4081 } else if ((idx / 4) == BestHiQuad) {
4082 MaskV.push_back((idx & 3) + 4);
4085 MaskV.push_back(-1);
4088 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4092 // In case BestHi & BestLo were both -1, which means each quadword has a word
4093 // from each of the four input quadwords, calculate the InOrder bitvector now
4094 // before falling through to the insert/extract cleanup.
4095 if (BestLoQuad == -1 && BestHiQuad == -1) {
4097 for (int i = 0; i != 8; ++i)
4098 if (MaskVals[i] < 0 || MaskVals[i] == i)
4102 // The other elements are put in the right place using pextrw and pinsrw.
4103 for (unsigned i = 0; i != 8; ++i) {
4106 int EltIdx = MaskVals[i];
4109 SDValue ExtOp = (EltIdx < 8)
4110 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4111 DAG.getIntPtrConstant(EltIdx))
4112 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4113 DAG.getIntPtrConstant(EltIdx - 8));
4114 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4115 DAG.getIntPtrConstant(i));
4120 // v16i8 shuffles - Prefer shuffles in the following order:
4121 // 1. [ssse3] 1 x pshufb
4122 // 2. [ssse3] 2 x pshufb + 1 x por
4123 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4125 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4126 SelectionDAG &DAG, X86TargetLowering &TLI) {
4127 SDValue V1 = SVOp->getOperand(0);
4128 SDValue V2 = SVOp->getOperand(1);
4129 DebugLoc dl = SVOp->getDebugLoc();
4130 SmallVector<int, 16> MaskVals;
4131 SVOp->getMask(MaskVals);
4133 // If we have SSSE3, case 1 is generated when all result bytes come from
4134 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4135 // present, fall back to case 3.
4136 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4139 for (unsigned i = 0; i < 16; ++i) {
4140 int EltIdx = MaskVals[i];
4149 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4150 if (TLI.getSubtarget()->hasSSSE3()) {
4151 SmallVector<SDValue,16> pshufbMask;
4153 // If all result elements are from one input vector, then only translate
4154 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4156 // Otherwise, we have elements from both input vectors, and must zero out
4157 // elements that come from V2 in the first mask, and V1 in the second mask
4158 // so that we can OR them together.
4159 bool TwoInputs = !(V1Only || V2Only);
4160 for (unsigned i = 0; i != 16; ++i) {
4161 int EltIdx = MaskVals[i];
4162 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4163 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4166 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4168 // If all the elements are from V2, assign it to V1 and return after
4169 // building the first pshufb.
4172 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4173 DAG.getNode(ISD::BUILD_VECTOR, dl,
4174 MVT::v16i8, &pshufbMask[0], 16));
4178 // Calculate the shuffle mask for the second input, shuffle it, and
4179 // OR it with the first shuffled input.
4181 for (unsigned i = 0; i != 16; ++i) {
4182 int EltIdx = MaskVals[i];
4184 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4187 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4189 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4190 DAG.getNode(ISD::BUILD_VECTOR, dl,
4191 MVT::v16i8, &pshufbMask[0], 16));
4192 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4195 // No SSSE3 - Calculate in place words and then fix all out of place words
4196 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4197 // the 16 different words that comprise the two doublequadword input vectors.
4198 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4199 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4200 SDValue NewV = V2Only ? V2 : V1;
4201 for (int i = 0; i != 8; ++i) {
4202 int Elt0 = MaskVals[i*2];
4203 int Elt1 = MaskVals[i*2+1];
4205 // This word of the result is all undef, skip it.
4206 if (Elt0 < 0 && Elt1 < 0)
4209 // This word of the result is already in the correct place, skip it.
4210 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4212 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4215 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4216 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4219 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4220 // using a single extract together, load it and store it.
4221 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4222 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4223 DAG.getIntPtrConstant(Elt1 / 2));
4224 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4225 DAG.getIntPtrConstant(i));
4229 // If Elt1 is defined, extract it from the appropriate source. If the
4230 // source byte is not also odd, shift the extracted word left 8 bits
4231 // otherwise clear the bottom 8 bits if we need to do an or.
4233 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4234 DAG.getIntPtrConstant(Elt1 / 2));
4235 if ((Elt1 & 1) == 0)
4236 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4237 DAG.getConstant(8, TLI.getShiftAmountTy()));
4239 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4240 DAG.getConstant(0xFF00, MVT::i16));
4242 // If Elt0 is defined, extract it from the appropriate source. If the
4243 // source byte is not also even, shift the extracted word right 8 bits. If
4244 // Elt1 was also defined, OR the extracted values together before
4245 // inserting them in the result.
4247 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4248 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4249 if ((Elt0 & 1) != 0)
4250 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4251 DAG.getConstant(8, TLI.getShiftAmountTy()));
4253 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4254 DAG.getConstant(0x00FF, MVT::i16));
4255 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4258 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4259 DAG.getIntPtrConstant(i));
4261 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4264 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4265 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4266 /// done when every pair / quad of shuffle mask elements point to elements in
4267 /// the right sequence. e.g.
4268 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4270 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4272 TargetLowering &TLI, DebugLoc dl) {
4273 EVT VT = SVOp->getValueType(0);
4274 SDValue V1 = SVOp->getOperand(0);
4275 SDValue V2 = SVOp->getOperand(1);
4276 unsigned NumElems = VT.getVectorNumElements();
4277 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4278 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4279 EVT MaskEltVT = MaskVT.getVectorElementType();
4281 switch (VT.getSimpleVT().SimpleTy) {
4282 default: assert(false && "Unexpected!");
4283 case MVT::v4f32: NewVT = MVT::v2f64; break;
4284 case MVT::v4i32: NewVT = MVT::v2i64; break;
4285 case MVT::v8i16: NewVT = MVT::v4i32; break;
4286 case MVT::v16i8: NewVT = MVT::v4i32; break;
4289 if (NewWidth == 2) {
4295 int Scale = NumElems / NewWidth;
4296 SmallVector<int, 8> MaskVec;
4297 for (unsigned i = 0; i < NumElems; i += Scale) {
4299 for (int j = 0; j < Scale; ++j) {
4300 int EltIdx = SVOp->getMaskElt(i+j);
4304 StartIdx = EltIdx - (EltIdx % Scale);
4305 if (EltIdx != StartIdx + j)
4309 MaskVec.push_back(-1);
4311 MaskVec.push_back(StartIdx / Scale);
4314 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4315 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4316 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4319 /// getVZextMovL - Return a zero-extending vector move low node.
4321 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4322 SDValue SrcOp, SelectionDAG &DAG,
4323 const X86Subtarget *Subtarget, DebugLoc dl) {
4324 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4325 LoadSDNode *LD = NULL;
4326 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4327 LD = dyn_cast<LoadSDNode>(SrcOp);
4329 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4331 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4332 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4333 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4334 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4335 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4337 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4338 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4339 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4340 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4348 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4349 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4350 DAG.getNode(ISD::BIT_CONVERT, dl,
4354 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4357 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4358 SDValue V1 = SVOp->getOperand(0);
4359 SDValue V2 = SVOp->getOperand(1);
4360 DebugLoc dl = SVOp->getDebugLoc();
4361 EVT VT = SVOp->getValueType(0);
4363 SmallVector<std::pair<int, int>, 8> Locs;
4365 SmallVector<int, 8> Mask1(4U, -1);
4366 SmallVector<int, 8> PermMask;
4367 SVOp->getMask(PermMask);
4371 for (unsigned i = 0; i != 4; ++i) {
4372 int Idx = PermMask[i];
4374 Locs[i] = std::make_pair(-1, -1);
4376 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4378 Locs[i] = std::make_pair(0, NumLo);
4382 Locs[i] = std::make_pair(1, NumHi);
4384 Mask1[2+NumHi] = Idx;
4390 if (NumLo <= 2 && NumHi <= 2) {
4391 // If no more than two elements come from either vector. This can be
4392 // implemented with two shuffles. First shuffle gather the elements.
4393 // The second shuffle, which takes the first shuffle as both of its
4394 // vector operands, put the elements into the right order.
4395 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4397 SmallVector<int, 8> Mask2(4U, -1);
4399 for (unsigned i = 0; i != 4; ++i) {
4400 if (Locs[i].first == -1)
4403 unsigned Idx = (i < 2) ? 0 : 4;
4404 Idx += Locs[i].first * 2 + Locs[i].second;
4409 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4410 } else if (NumLo == 3 || NumHi == 3) {
4411 // Otherwise, we must have three elements from one vector, call it X, and
4412 // one element from the other, call it Y. First, use a shufps to build an
4413 // intermediate vector with the one element from Y and the element from X
4414 // that will be in the same half in the final destination (the indexes don't
4415 // matter). Then, use a shufps to build the final vector, taking the half
4416 // containing the element from Y from the intermediate, and the other half
4419 // Normalize it so the 3 elements come from V1.
4420 CommuteVectorShuffleMask(PermMask, VT);
4424 // Find the element from V2.
4426 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4427 int Val = PermMask[HiIndex];
4434 Mask1[0] = PermMask[HiIndex];
4436 Mask1[2] = PermMask[HiIndex^1];
4438 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4441 Mask1[0] = PermMask[0];
4442 Mask1[1] = PermMask[1];
4443 Mask1[2] = HiIndex & 1 ? 6 : 4;
4444 Mask1[3] = HiIndex & 1 ? 4 : 6;
4445 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4447 Mask1[0] = HiIndex & 1 ? 2 : 0;
4448 Mask1[1] = HiIndex & 1 ? 0 : 2;
4449 Mask1[2] = PermMask[2];
4450 Mask1[3] = PermMask[3];
4455 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4459 // Break it into (shuffle shuffle_hi, shuffle_lo).
4461 SmallVector<int,8> LoMask(4U, -1);
4462 SmallVector<int,8> HiMask(4U, -1);
4464 SmallVector<int,8> *MaskPtr = &LoMask;
4465 unsigned MaskIdx = 0;
4468 for (unsigned i = 0; i != 4; ++i) {
4475 int Idx = PermMask[i];
4477 Locs[i] = std::make_pair(-1, -1);
4478 } else if (Idx < 4) {
4479 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4480 (*MaskPtr)[LoIdx] = Idx;
4483 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4484 (*MaskPtr)[HiIdx] = Idx;
4489 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4490 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4491 SmallVector<int, 8> MaskOps;
4492 for (unsigned i = 0; i != 4; ++i) {
4493 if (Locs[i].first == -1) {
4494 MaskOps.push_back(-1);
4496 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4497 MaskOps.push_back(Idx);
4500 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4504 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4505 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4506 SDValue V1 = Op.getOperand(0);
4507 SDValue V2 = Op.getOperand(1);
4508 EVT VT = Op.getValueType();
4509 DebugLoc dl = Op.getDebugLoc();
4510 unsigned NumElems = VT.getVectorNumElements();
4511 bool isMMX = VT.getSizeInBits() == 64;
4512 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4513 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4514 bool V1IsSplat = false;
4515 bool V2IsSplat = false;
4517 if (isZeroShuffle(SVOp))
4518 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4520 // Promote splats to v4f32.
4521 if (SVOp->isSplat()) {
4522 if (isMMX || NumElems < 4)
4524 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4527 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4529 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4530 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4531 if (NewOp.getNode())
4532 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4533 LowerVECTOR_SHUFFLE(NewOp, DAG));
4534 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4535 // FIXME: Figure out a cleaner way to do this.
4536 // Try to make use of movq to zero out the top part.
4537 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4538 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4539 if (NewOp.getNode()) {
4540 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4541 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4542 DAG, Subtarget, dl);
4544 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4545 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4546 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4547 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4548 DAG, Subtarget, dl);
4552 if (X86::isPSHUFDMask(SVOp))
4555 // Check if this can be converted into a logical shift.
4556 bool isLeft = false;
4559 bool isShift = getSubtarget()->hasSSE2() &&
4560 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4561 if (isShift && ShVal.hasOneUse()) {
4562 // If the shifted value has multiple uses, it may be cheaper to use
4563 // v_set0 + movlhps or movhlps, etc.
4564 EVT EltVT = VT.getVectorElementType();
4565 ShAmt *= EltVT.getSizeInBits();
4566 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4569 if (X86::isMOVLMask(SVOp)) {
4572 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4573 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4578 // FIXME: fold these into legal mask.
4579 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4580 X86::isMOVSLDUPMask(SVOp) ||
4581 X86::isMOVHLPSMask(SVOp) ||
4582 X86::isMOVLHPSMask(SVOp) ||
4583 X86::isMOVLPMask(SVOp)))
4586 if (ShouldXformToMOVHLPS(SVOp) ||
4587 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4588 return CommuteVectorShuffle(SVOp, DAG);
4591 // No better options. Use a vshl / vsrl.
4592 EVT EltVT = VT.getVectorElementType();
4593 ShAmt *= EltVT.getSizeInBits();
4594 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4597 bool Commuted = false;
4598 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4599 // 1,1,1,1 -> v8i16 though.
4600 V1IsSplat = isSplatVector(V1.getNode());
4601 V2IsSplat = isSplatVector(V2.getNode());
4603 // Canonicalize the splat or undef, if present, to be on the RHS.
4604 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4605 Op = CommuteVectorShuffle(SVOp, DAG);
4606 SVOp = cast<ShuffleVectorSDNode>(Op);
4607 V1 = SVOp->getOperand(0);
4608 V2 = SVOp->getOperand(1);
4609 std::swap(V1IsSplat, V2IsSplat);
4610 std::swap(V1IsUndef, V2IsUndef);
4614 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4615 // Shuffling low element of v1 into undef, just return v1.
4618 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4619 // the instruction selector will not match, so get a canonical MOVL with
4620 // swapped operands to undo the commute.
4621 return getMOVL(DAG, dl, VT, V2, V1);
4624 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4625 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4626 X86::isUNPCKLMask(SVOp) ||
4627 X86::isUNPCKHMask(SVOp))
4631 // Normalize mask so all entries that point to V2 points to its first
4632 // element then try to match unpck{h|l} again. If match, return a
4633 // new vector_shuffle with the corrected mask.
4634 SDValue NewMask = NormalizeMask(SVOp, DAG);
4635 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4636 if (NSVOp != SVOp) {
4637 if (X86::isUNPCKLMask(NSVOp, true)) {
4639 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4646 // Commute is back and try unpck* again.
4647 // FIXME: this seems wrong.
4648 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4649 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4650 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4651 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4652 X86::isUNPCKLMask(NewSVOp) ||
4653 X86::isUNPCKHMask(NewSVOp))
4657 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4659 // Normalize the node to match x86 shuffle ops if needed
4660 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4661 return CommuteVectorShuffle(SVOp, DAG);
4663 // Check for legal shuffle and return?
4664 SmallVector<int, 16> PermMask;
4665 SVOp->getMask(PermMask);
4666 if (isShuffleMaskLegal(PermMask, VT))
4669 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4670 if (VT == MVT::v8i16) {
4671 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4672 if (NewOp.getNode())
4676 if (VT == MVT::v16i8) {
4677 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4678 if (NewOp.getNode())
4682 // Handle all 4 wide cases with a number of shuffles except for MMX.
4683 if (NumElems == 4 && !isMMX)
4684 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4690 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4691 SelectionDAG &DAG) {
4692 EVT VT = Op.getValueType();
4693 DebugLoc dl = Op.getDebugLoc();
4694 if (VT.getSizeInBits() == 8) {
4695 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4696 Op.getOperand(0), Op.getOperand(1));
4697 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4698 DAG.getValueType(VT));
4699 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4700 } else if (VT.getSizeInBits() == 16) {
4701 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4702 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4704 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4705 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4706 DAG.getNode(ISD::BIT_CONVERT, dl,
4710 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4711 Op.getOperand(0), Op.getOperand(1));
4712 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4713 DAG.getValueType(VT));
4714 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4715 } else if (VT == MVT::f32) {
4716 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4717 // the result back to FR32 register. It's only worth matching if the
4718 // result has a single use which is a store or a bitcast to i32. And in
4719 // the case of a store, it's not worth it if the index is a constant 0,
4720 // because a MOVSSmr can be used instead, which is smaller and faster.
4721 if (!Op.hasOneUse())
4723 SDNode *User = *Op.getNode()->use_begin();
4724 if ((User->getOpcode() != ISD::STORE ||
4725 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4726 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4727 (User->getOpcode() != ISD::BIT_CONVERT ||
4728 User->getValueType(0) != MVT::i32))
4730 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4731 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4734 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4735 } else if (VT == MVT::i32) {
4736 // ExtractPS works with constant index.
4737 if (isa<ConstantSDNode>(Op.getOperand(1)))
4745 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4746 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4749 if (Subtarget->hasSSE41()) {
4750 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4755 EVT VT = Op.getValueType();
4756 DebugLoc dl = Op.getDebugLoc();
4757 // TODO: handle v16i8.
4758 if (VT.getSizeInBits() == 16) {
4759 SDValue Vec = Op.getOperand(0);
4760 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4762 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4763 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4764 DAG.getNode(ISD::BIT_CONVERT, dl,
4767 // Transform it so it match pextrw which produces a 32-bit result.
4768 EVT EltVT = MVT::i32;
4769 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4770 Op.getOperand(0), Op.getOperand(1));
4771 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4772 DAG.getValueType(VT));
4773 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4774 } else if (VT.getSizeInBits() == 32) {
4775 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4779 // SHUFPS the element to the lowest double word, then movss.
4780 int Mask[4] = { Idx, -1, -1, -1 };
4781 EVT VVT = Op.getOperand(0).getValueType();
4782 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4783 DAG.getUNDEF(VVT), Mask);
4784 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4785 DAG.getIntPtrConstant(0));
4786 } else if (VT.getSizeInBits() == 64) {
4787 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4788 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4789 // to match extract_elt for f64.
4790 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4794 // UNPCKHPD the element to the lowest double word, then movsd.
4795 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4796 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4797 int Mask[2] = { 1, -1 };
4798 EVT VVT = Op.getOperand(0).getValueType();
4799 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4800 DAG.getUNDEF(VVT), Mask);
4801 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4802 DAG.getIntPtrConstant(0));
4809 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4810 EVT VT = Op.getValueType();
4811 EVT EltVT = VT.getVectorElementType();
4812 DebugLoc dl = Op.getDebugLoc();
4814 SDValue N0 = Op.getOperand(0);
4815 SDValue N1 = Op.getOperand(1);
4816 SDValue N2 = Op.getOperand(2);
4818 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4819 isa<ConstantSDNode>(N2)) {
4821 if (VT == MVT::v8i16)
4822 Opc = X86ISD::PINSRW;
4823 else if (VT == MVT::v4i16)
4824 Opc = X86ISD::MMX_PINSRW;
4825 else if (VT == MVT::v16i8)
4826 Opc = X86ISD::PINSRB;
4828 Opc = X86ISD::PINSRB;
4830 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4832 if (N1.getValueType() != MVT::i32)
4833 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4834 if (N2.getValueType() != MVT::i32)
4835 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4836 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4837 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4838 // Bits [7:6] of the constant are the source select. This will always be
4839 // zero here. The DAG Combiner may combine an extract_elt index into these
4840 // bits. For example (insert (extract, 3), 2) could be matched by putting
4841 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4842 // Bits [5:4] of the constant are the destination select. This is the
4843 // value of the incoming immediate.
4844 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4845 // combine either bitwise AND or insert of float 0.0 to set these bits.
4846 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4847 // Create this as a scalar to vector..
4848 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4849 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4850 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4851 // PINSR* works with constant index.
4858 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4859 EVT VT = Op.getValueType();
4860 EVT EltVT = VT.getVectorElementType();
4862 if (Subtarget->hasSSE41())
4863 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4865 if (EltVT == MVT::i8)
4868 DebugLoc dl = Op.getDebugLoc();
4869 SDValue N0 = Op.getOperand(0);
4870 SDValue N1 = Op.getOperand(1);
4871 SDValue N2 = Op.getOperand(2);
4873 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4874 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4875 // as its second argument.
4876 if (N1.getValueType() != MVT::i32)
4877 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4878 if (N2.getValueType() != MVT::i32)
4879 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4880 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
4881 dl, VT, N0, N1, N2);
4887 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4888 DebugLoc dl = Op.getDebugLoc();
4889 if (Op.getValueType() == MVT::v2f32)
4890 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4891 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4892 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4893 Op.getOperand(0))));
4895 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4896 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4898 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4899 EVT VT = MVT::v2i32;
4900 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4907 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4908 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4911 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4912 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4913 // one of the above mentioned nodes. It has to be wrapped because otherwise
4914 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4915 // be used to form addressing mode. These wrapped nodes will be selected
4918 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4919 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4921 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4923 unsigned char OpFlag = 0;
4924 unsigned WrapperKind = X86ISD::Wrapper;
4925 CodeModel::Model M = getTargetMachine().getCodeModel();
4927 if (Subtarget->isPICStyleRIPRel() &&
4928 (M == CodeModel::Small || M == CodeModel::Kernel))
4929 WrapperKind = X86ISD::WrapperRIP;
4930 else if (Subtarget->isPICStyleGOT())
4931 OpFlag = X86II::MO_GOTOFF;
4932 else if (Subtarget->isPICStyleStubPIC())
4933 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4935 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4937 CP->getOffset(), OpFlag);
4938 DebugLoc DL = CP->getDebugLoc();
4939 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4940 // With PIC, the address is actually $g + Offset.
4942 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4943 DAG.getNode(X86ISD::GlobalBaseReg,
4944 DebugLoc::getUnknownLoc(), getPointerTy()),
4951 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4952 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4954 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4956 unsigned char OpFlag = 0;
4957 unsigned WrapperKind = X86ISD::Wrapper;
4958 CodeModel::Model M = getTargetMachine().getCodeModel();
4960 if (Subtarget->isPICStyleRIPRel() &&
4961 (M == CodeModel::Small || M == CodeModel::Kernel))
4962 WrapperKind = X86ISD::WrapperRIP;
4963 else if (Subtarget->isPICStyleGOT())
4964 OpFlag = X86II::MO_GOTOFF;
4965 else if (Subtarget->isPICStyleStubPIC())
4966 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4968 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4970 DebugLoc DL = JT->getDebugLoc();
4971 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4973 // With PIC, the address is actually $g + Offset.
4975 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4976 DAG.getNode(X86ISD::GlobalBaseReg,
4977 DebugLoc::getUnknownLoc(), getPointerTy()),
4985 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4986 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4988 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4990 unsigned char OpFlag = 0;
4991 unsigned WrapperKind = X86ISD::Wrapper;
4992 CodeModel::Model M = getTargetMachine().getCodeModel();
4994 if (Subtarget->isPICStyleRIPRel() &&
4995 (M == CodeModel::Small || M == CodeModel::Kernel))
4996 WrapperKind = X86ISD::WrapperRIP;
4997 else if (Subtarget->isPICStyleGOT())
4998 OpFlag = X86II::MO_GOTOFF;
4999 else if (Subtarget->isPICStyleStubPIC())
5000 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5002 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5004 DebugLoc DL = Op.getDebugLoc();
5005 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5008 // With PIC, the address is actually $g + Offset.
5009 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5010 !Subtarget->is64Bit()) {
5011 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5012 DAG.getNode(X86ISD::GlobalBaseReg,
5013 DebugLoc::getUnknownLoc(),
5022 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
5023 // Create the TargetBlockAddressAddress node.
5024 unsigned char OpFlags =
5025 Subtarget->ClassifyBlockAddressReference();
5026 CodeModel::Model M = getTargetMachine().getCodeModel();
5027 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5028 DebugLoc dl = Op.getDebugLoc();
5029 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5030 /*isTarget=*/true, OpFlags);
5032 if (Subtarget->isPICStyleRIPRel() &&
5033 (M == CodeModel::Small || M == CodeModel::Kernel))
5034 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5036 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5038 // With PIC, the address is actually $g + Offset.
5039 if (isGlobalRelativeToPICBase(OpFlags)) {
5040 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5041 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5049 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5051 SelectionDAG &DAG) const {
5052 // Create the TargetGlobalAddress node, folding in the constant
5053 // offset if it is legal.
5054 unsigned char OpFlags =
5055 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5056 CodeModel::Model M = getTargetMachine().getCodeModel();
5058 if (OpFlags == X86II::MO_NO_FLAG &&
5059 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5060 // A direct static reference to a global.
5061 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5064 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5067 if (Subtarget->isPICStyleRIPRel() &&
5068 (M == CodeModel::Small || M == CodeModel::Kernel))
5069 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5071 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5073 // With PIC, the address is actually $g + Offset.
5074 if (isGlobalRelativeToPICBase(OpFlags)) {
5075 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5076 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5080 // For globals that require a load from a stub to get the address, emit the
5082 if (isGlobalStubReference(OpFlags))
5083 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5084 PseudoSourceValue::getGOT(), 0, false, false, 0);
5086 // If there was a non-zero offset that we didn't fold, create an explicit
5089 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5090 DAG.getConstant(Offset, getPointerTy()));
5096 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
5097 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5098 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5099 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5103 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5104 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5105 unsigned char OperandFlags) {
5106 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5107 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5108 DebugLoc dl = GA->getDebugLoc();
5109 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5110 GA->getValueType(0),
5114 SDValue Ops[] = { Chain, TGA, *InFlag };
5115 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5117 SDValue Ops[] = { Chain, TGA };
5118 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5121 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5122 MFI->setHasCalls(true);
5124 SDValue Flag = Chain.getValue(1);
5125 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5128 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5130 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5133 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5134 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5135 DAG.getNode(X86ISD::GlobalBaseReg,
5136 DebugLoc::getUnknownLoc(),
5138 InFlag = Chain.getValue(1);
5140 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5143 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5145 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5147 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5148 X86::RAX, X86II::MO_TLSGD);
5151 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5152 // "local exec" model.
5153 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5154 const EVT PtrVT, TLSModel::Model model,
5156 DebugLoc dl = GA->getDebugLoc();
5157 // Get the Thread Pointer
5158 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5159 DebugLoc::getUnknownLoc(), PtrVT,
5160 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5163 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5164 NULL, 0, false, false, 0);
5166 unsigned char OperandFlags = 0;
5167 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5169 unsigned WrapperKind = X86ISD::Wrapper;
5170 if (model == TLSModel::LocalExec) {
5171 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5172 } else if (is64Bit) {
5173 assert(model == TLSModel::InitialExec);
5174 OperandFlags = X86II::MO_GOTTPOFF;
5175 WrapperKind = X86ISD::WrapperRIP;
5177 assert(model == TLSModel::InitialExec);
5178 OperandFlags = X86II::MO_INDNTPOFF;
5181 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5183 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5184 GA->getOffset(), OperandFlags);
5185 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5187 if (model == TLSModel::InitialExec)
5188 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5189 PseudoSourceValue::getGOT(), 0, false, false, 0);
5191 // The address of the thread local variable is the add of the thread
5192 // pointer with the offset of the variable.
5193 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5197 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
5198 // TODO: implement the "local dynamic" model
5199 // TODO: implement the "initial exec"model for pic executables
5200 assert(Subtarget->isTargetELF() &&
5201 "TLS not implemented for non-ELF targets");
5202 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5203 const GlobalValue *GV = GA->getGlobal();
5205 // If GV is an alias then use the aliasee for determining
5206 // thread-localness.
5207 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5208 GV = GA->resolveAliasedGlobal(false);
5210 TLSModel::Model model = getTLSModel(GV,
5211 getTargetMachine().getRelocationModel());
5214 case TLSModel::GeneralDynamic:
5215 case TLSModel::LocalDynamic: // not implemented
5216 if (Subtarget->is64Bit())
5217 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5218 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5220 case TLSModel::InitialExec:
5221 case TLSModel::LocalExec:
5222 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5223 Subtarget->is64Bit());
5226 llvm_unreachable("Unreachable");
5231 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5232 /// take a 2 x i32 value to shift plus a shift amount.
5233 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
5234 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5235 EVT VT = Op.getValueType();
5236 unsigned VTBits = VT.getSizeInBits();
5237 DebugLoc dl = Op.getDebugLoc();
5238 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5239 SDValue ShOpLo = Op.getOperand(0);
5240 SDValue ShOpHi = Op.getOperand(1);
5241 SDValue ShAmt = Op.getOperand(2);
5242 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5243 DAG.getConstant(VTBits - 1, MVT::i8))
5244 : DAG.getConstant(0, VT);
5247 if (Op.getOpcode() == ISD::SHL_PARTS) {
5248 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5249 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5251 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5252 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5255 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5256 DAG.getConstant(VTBits, MVT::i8));
5257 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5258 AndNode, DAG.getConstant(0, MVT::i8));
5261 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5262 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5263 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5265 if (Op.getOpcode() == ISD::SHL_PARTS) {
5266 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5267 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5269 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5270 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5273 SDValue Ops[2] = { Lo, Hi };
5274 return DAG.getMergeValues(Ops, 2, dl);
5277 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5278 EVT SrcVT = Op.getOperand(0).getValueType();
5280 if (SrcVT.isVector()) {
5281 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5287 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5288 "Unknown SINT_TO_FP to lower!");
5290 // These are really Legal; return the operand so the caller accepts it as
5292 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5294 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5295 Subtarget->is64Bit()) {
5299 DebugLoc dl = Op.getDebugLoc();
5300 unsigned Size = SrcVT.getSizeInBits()/8;
5301 MachineFunction &MF = DAG.getMachineFunction();
5302 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5303 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5304 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5306 PseudoSourceValue::getFixedStack(SSFI), 0,
5308 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5311 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5313 SelectionDAG &DAG) {
5315 DebugLoc dl = Op.getDebugLoc();
5317 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5319 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5321 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5322 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5323 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5324 Tys, Ops, array_lengthof(Ops));
5327 Chain = Result.getValue(1);
5328 SDValue InFlag = Result.getValue(2);
5330 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5331 // shouldn't be necessary except that RFP cannot be live across
5332 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5333 MachineFunction &MF = DAG.getMachineFunction();
5334 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5335 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5336 Tys = DAG.getVTList(MVT::Other);
5338 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5340 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5341 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5342 PseudoSourceValue::getFixedStack(SSFI), 0,
5349 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5350 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5351 // This algorithm is not obvious. Here it is in C code, more or less:
5353 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5354 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5355 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5357 // Copy ints to xmm registers.
5358 __m128i xh = _mm_cvtsi32_si128( hi );
5359 __m128i xl = _mm_cvtsi32_si128( lo );
5361 // Combine into low half of a single xmm register.
5362 __m128i x = _mm_unpacklo_epi32( xh, xl );
5366 // Merge in appropriate exponents to give the integer bits the right
5368 x = _mm_unpacklo_epi32( x, exp );
5370 // Subtract away the biases to deal with the IEEE-754 double precision
5372 d = _mm_sub_pd( (__m128d) x, bias );
5374 // All conversions up to here are exact. The correctly rounded result is
5375 // calculated using the current rounding mode using the following
5377 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5378 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5379 // store doesn't really need to be here (except
5380 // maybe to zero the other double)
5385 DebugLoc dl = Op.getDebugLoc();
5386 LLVMContext *Context = DAG.getContext();
5388 // Build some magic constants.
5389 std::vector<Constant*> CV0;
5390 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5391 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5392 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5393 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5394 Constant *C0 = ConstantVector::get(CV0);
5395 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5397 std::vector<Constant*> CV1;
5399 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5401 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5402 Constant *C1 = ConstantVector::get(CV1);
5403 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5405 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5406 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5408 DAG.getIntPtrConstant(1)));
5409 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5410 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5412 DAG.getIntPtrConstant(0)));
5413 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5414 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5415 PseudoSourceValue::getConstantPool(), 0,
5417 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5418 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5419 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5420 PseudoSourceValue::getConstantPool(), 0,
5422 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5424 // Add the halves; easiest way is to swap them into another reg first.
5425 int ShufMask[2] = { 1, -1 };
5426 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5427 DAG.getUNDEF(MVT::v2f64), ShufMask);
5428 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5429 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5430 DAG.getIntPtrConstant(0));
5433 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5434 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5435 DebugLoc dl = Op.getDebugLoc();
5436 // FP constant to bias correct the final result.
5437 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5440 // Load the 32-bit value into an XMM register.
5441 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5442 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5444 DAG.getIntPtrConstant(0)));
5446 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5447 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5448 DAG.getIntPtrConstant(0));
5450 // Or the load with the bias.
5451 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5452 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5453 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5455 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5456 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5457 MVT::v2f64, Bias)));
5458 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5459 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5460 DAG.getIntPtrConstant(0));
5462 // Subtract the bias.
5463 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5465 // Handle final rounding.
5466 EVT DestVT = Op.getValueType();
5468 if (DestVT.bitsLT(MVT::f64)) {
5469 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5470 DAG.getIntPtrConstant(0));
5471 } else if (DestVT.bitsGT(MVT::f64)) {
5472 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5475 // Handle final rounding.
5479 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5480 SDValue N0 = Op.getOperand(0);
5481 DebugLoc dl = Op.getDebugLoc();
5483 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5484 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5485 // the optimization here.
5486 if (DAG.SignBitIsZero(N0))
5487 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5489 EVT SrcVT = N0.getValueType();
5490 if (SrcVT == MVT::i64) {
5491 // We only handle SSE2 f64 target here; caller can expand the rest.
5492 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5495 return LowerUINT_TO_FP_i64(Op, DAG);
5496 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5497 return LowerUINT_TO_FP_i32(Op, DAG);
5500 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5502 // Make a 64-bit buffer, and use it to build an FILD.
5503 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5504 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5505 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5506 getPointerTy(), StackSlot, WordOff);
5507 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5508 StackSlot, NULL, 0, false, false, 0);
5509 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5510 OffsetSlot, NULL, 0, false, false, 0);
5511 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5514 std::pair<SDValue,SDValue> X86TargetLowering::
5515 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5516 DebugLoc dl = Op.getDebugLoc();
5518 EVT DstTy = Op.getValueType();
5521 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5525 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5526 DstTy.getSimpleVT() >= MVT::i16 &&
5527 "Unknown FP_TO_SINT to lower!");
5529 // These are really Legal.
5530 if (DstTy == MVT::i32 &&
5531 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5532 return std::make_pair(SDValue(), SDValue());
5533 if (Subtarget->is64Bit() &&
5534 DstTy == MVT::i64 &&
5535 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5536 return std::make_pair(SDValue(), SDValue());
5538 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5540 MachineFunction &MF = DAG.getMachineFunction();
5541 unsigned MemSize = DstTy.getSizeInBits()/8;
5542 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5543 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5546 switch (DstTy.getSimpleVT().SimpleTy) {
5547 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5548 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5549 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5550 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5553 SDValue Chain = DAG.getEntryNode();
5554 SDValue Value = Op.getOperand(0);
5555 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5556 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5557 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5558 PseudoSourceValue::getFixedStack(SSFI), 0,
5560 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5562 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5564 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5565 Chain = Value.getValue(1);
5566 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5567 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5570 // Build the FP_TO_INT*_IN_MEM
5571 SDValue Ops[] = { Chain, Value, StackSlot };
5572 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5574 return std::make_pair(FIST, StackSlot);
5577 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5578 if (Op.getValueType().isVector()) {
5579 if (Op.getValueType() == MVT::v2i32 &&
5580 Op.getOperand(0).getValueType() == MVT::v2f64) {
5586 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5587 SDValue FIST = Vals.first, StackSlot = Vals.second;
5588 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5589 if (FIST.getNode() == 0) return Op;
5592 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5593 FIST, StackSlot, NULL, 0, false, false, 0);
5596 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5597 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5598 SDValue FIST = Vals.first, StackSlot = Vals.second;
5599 assert(FIST.getNode() && "Unexpected failure");
5602 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5603 FIST, StackSlot, NULL, 0, false, false, 0);
5606 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5607 LLVMContext *Context = DAG.getContext();
5608 DebugLoc dl = Op.getDebugLoc();
5609 EVT VT = Op.getValueType();
5612 EltVT = VT.getVectorElementType();
5613 std::vector<Constant*> CV;
5614 if (EltVT == MVT::f64) {
5615 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5619 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5625 Constant *C = ConstantVector::get(CV);
5626 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5627 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5628 PseudoSourceValue::getConstantPool(), 0,
5630 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5633 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5634 LLVMContext *Context = DAG.getContext();
5635 DebugLoc dl = Op.getDebugLoc();
5636 EVT VT = Op.getValueType();
5639 EltVT = VT.getVectorElementType();
5640 std::vector<Constant*> CV;
5641 if (EltVT == MVT::f64) {
5642 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5646 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5652 Constant *C = ConstantVector::get(CV);
5653 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5654 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5655 PseudoSourceValue::getConstantPool(), 0,
5657 if (VT.isVector()) {
5658 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5659 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5660 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5662 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5664 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5668 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5669 LLVMContext *Context = DAG.getContext();
5670 SDValue Op0 = Op.getOperand(0);
5671 SDValue Op1 = Op.getOperand(1);
5672 DebugLoc dl = Op.getDebugLoc();
5673 EVT VT = Op.getValueType();
5674 EVT SrcVT = Op1.getValueType();
5676 // If second operand is smaller, extend it first.
5677 if (SrcVT.bitsLT(VT)) {
5678 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5681 // And if it is bigger, shrink it first.
5682 if (SrcVT.bitsGT(VT)) {
5683 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5687 // At this point the operands and the result should have the same
5688 // type, and that won't be f80 since that is not custom lowered.
5690 // First get the sign bit of second operand.
5691 std::vector<Constant*> CV;
5692 if (SrcVT == MVT::f64) {
5693 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5694 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5696 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5697 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5698 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5699 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5701 Constant *C = ConstantVector::get(CV);
5702 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5703 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5704 PseudoSourceValue::getConstantPool(), 0,
5706 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5708 // Shift sign bit right or left if the two operands have different types.
5709 if (SrcVT.bitsGT(VT)) {
5710 // Op0 is MVT::f32, Op1 is MVT::f64.
5711 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5712 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5713 DAG.getConstant(32, MVT::i32));
5714 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5715 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5716 DAG.getIntPtrConstant(0));
5719 // Clear first operand sign bit.
5721 if (VT == MVT::f64) {
5722 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5723 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5725 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5726 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5727 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5728 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5730 C = ConstantVector::get(CV);
5731 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5732 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5733 PseudoSourceValue::getConstantPool(), 0,
5735 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5737 // Or the value with the sign bit.
5738 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5741 /// Emit nodes that will be selected as "test Op0,Op0", or something
5743 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5744 SelectionDAG &DAG) {
5745 DebugLoc dl = Op.getDebugLoc();
5747 // CF and OF aren't always set the way we want. Determine which
5748 // of these we need.
5749 bool NeedCF = false;
5750 bool NeedOF = false;
5752 case X86::COND_A: case X86::COND_AE:
5753 case X86::COND_B: case X86::COND_BE:
5756 case X86::COND_G: case X86::COND_GE:
5757 case X86::COND_L: case X86::COND_LE:
5758 case X86::COND_O: case X86::COND_NO:
5764 // See if we can use the EFLAGS value from the operand instead of
5765 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5766 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5767 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5768 unsigned Opcode = 0;
5769 unsigned NumOperands = 0;
5770 switch (Op.getNode()->getOpcode()) {
5772 // Due to an isel shortcoming, be conservative if this add is likely to
5773 // be selected as part of a load-modify-store instruction. When the root
5774 // node in a match is a store, isel doesn't know how to remap non-chain
5775 // non-flag uses of other nodes in the match, such as the ADD in this
5776 // case. This leads to the ADD being left around and reselected, with
5777 // the result being two adds in the output.
5778 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5779 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5780 if (UI->getOpcode() == ISD::STORE)
5782 if (ConstantSDNode *C =
5783 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5784 // An add of one will be selected as an INC.
5785 if (C->getAPIntValue() == 1) {
5786 Opcode = X86ISD::INC;
5790 // An add of negative one (subtract of one) will be selected as a DEC.
5791 if (C->getAPIntValue().isAllOnesValue()) {
5792 Opcode = X86ISD::DEC;
5797 // Otherwise use a regular EFLAGS-setting add.
5798 Opcode = X86ISD::ADD;
5802 // If the primary and result isn't used, don't bother using X86ISD::AND,
5803 // because a TEST instruction will be better.
5804 bool NonFlagUse = false;
5805 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5806 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
5808 unsigned UOpNo = UI.getOperandNo();
5809 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
5810 // Look pass truncate.
5811 UOpNo = User->use_begin().getOperandNo();
5812 User = *User->use_begin();
5814 if (User->getOpcode() != ISD::BRCOND &&
5815 User->getOpcode() != ISD::SETCC &&
5816 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
5828 // Due to the ISEL shortcoming noted above, be conservative if this op is
5829 // likely to be selected as part of a load-modify-store instruction.
5830 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5831 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5832 if (UI->getOpcode() == ISD::STORE)
5834 // Otherwise use a regular EFLAGS-setting instruction.
5835 switch (Op.getNode()->getOpcode()) {
5836 case ISD::SUB: Opcode = X86ISD::SUB; break;
5837 case ISD::OR: Opcode = X86ISD::OR; break;
5838 case ISD::XOR: Opcode = X86ISD::XOR; break;
5839 case ISD::AND: Opcode = X86ISD::AND; break;
5840 default: llvm_unreachable("unexpected operator!");
5851 return SDValue(Op.getNode(), 1);
5857 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5858 SmallVector<SDValue, 4> Ops;
5859 for (unsigned i = 0; i != NumOperands; ++i)
5860 Ops.push_back(Op.getOperand(i));
5861 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5862 DAG.ReplaceAllUsesWith(Op, New);
5863 return SDValue(New.getNode(), 1);
5867 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5868 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5869 DAG.getConstant(0, Op.getValueType()));
5872 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5874 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5875 SelectionDAG &DAG) {
5876 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5877 if (C->getAPIntValue() == 0)
5878 return EmitTest(Op0, X86CC, DAG);
5880 DebugLoc dl = Op0.getDebugLoc();
5881 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5884 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
5885 /// if it's possible.
5886 static SDValue LowerToBT(SDValue Op0, ISD::CondCode CC,
5887 DebugLoc dl, SelectionDAG &DAG) {
5889 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5890 if (ConstantSDNode *Op010C =
5891 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5892 if (Op010C->getZExtValue() == 1) {
5893 LHS = Op0.getOperand(0);
5894 RHS = Op0.getOperand(1).getOperand(1);
5896 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5897 if (ConstantSDNode *Op000C =
5898 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5899 if (Op000C->getZExtValue() == 1) {
5900 LHS = Op0.getOperand(1);
5901 RHS = Op0.getOperand(0).getOperand(1);
5903 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5904 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5905 SDValue AndLHS = Op0.getOperand(0);
5906 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5907 LHS = AndLHS.getOperand(0);
5908 RHS = AndLHS.getOperand(1);
5912 if (LHS.getNode()) {
5913 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5914 // instruction. Since the shift amount is in-range-or-undefined, we know
5915 // that doing a bittest on the i16 value is ok. We extend to i32 because
5916 // the encoding for the i16 version is larger than the i32 version.
5917 if (LHS.getValueType() == MVT::i8)
5918 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5920 // If the operand types disagree, extend the shift amount to match. Since
5921 // BT ignores high bits (like shifts) we can use anyextend.
5922 if (LHS.getValueType() != RHS.getValueType())
5923 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5925 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5926 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5927 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5928 DAG.getConstant(Cond, MVT::i8), BT);
5934 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5935 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5936 SDValue Op0 = Op.getOperand(0);
5937 SDValue Op1 = Op.getOperand(1);
5938 DebugLoc dl = Op.getDebugLoc();
5939 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5941 // Optimize to BT if possible.
5942 // Lower (X & (1 << N)) == 0 to BT(X, N).
5943 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5944 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5945 if (Op0.getOpcode() == ISD::AND &&
5947 Op1.getOpcode() == ISD::Constant &&
5948 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5949 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5950 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
5951 if (NewSetCC.getNode())
5955 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5956 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5957 if (X86CC == X86::COND_INVALID)
5960 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5962 // Use sbb x, x to materialize carry bit into a GPR.
5963 if (X86CC == X86::COND_B)
5964 return DAG.getNode(ISD::AND, dl, MVT::i8,
5965 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
5966 DAG.getConstant(X86CC, MVT::i8), Cond),
5967 DAG.getConstant(1, MVT::i8));
5969 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5970 DAG.getConstant(X86CC, MVT::i8), Cond);
5973 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5975 SDValue Op0 = Op.getOperand(0);
5976 SDValue Op1 = Op.getOperand(1);
5977 SDValue CC = Op.getOperand(2);
5978 EVT VT = Op.getValueType();
5979 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5980 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5981 DebugLoc dl = Op.getDebugLoc();
5985 EVT VT0 = Op0.getValueType();
5986 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5987 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5990 switch (SetCCOpcode) {
5993 case ISD::SETEQ: SSECC = 0; break;
5995 case ISD::SETGT: Swap = true; // Fallthrough
5997 case ISD::SETOLT: SSECC = 1; break;
5999 case ISD::SETGE: Swap = true; // Fallthrough
6001 case ISD::SETOLE: SSECC = 2; break;
6002 case ISD::SETUO: SSECC = 3; break;
6004 case ISD::SETNE: SSECC = 4; break;
6005 case ISD::SETULE: Swap = true;
6006 case ISD::SETUGE: SSECC = 5; break;
6007 case ISD::SETULT: Swap = true;
6008 case ISD::SETUGT: SSECC = 6; break;
6009 case ISD::SETO: SSECC = 7; break;
6012 std::swap(Op0, Op1);
6014 // In the two special cases we can't handle, emit two comparisons.
6016 if (SetCCOpcode == ISD::SETUEQ) {
6018 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6019 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6020 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6022 else if (SetCCOpcode == ISD::SETONE) {
6024 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6025 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6026 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6028 llvm_unreachable("Illegal FP comparison");
6030 // Handle all other FP comparisons here.
6031 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6034 // We are handling one of the integer comparisons here. Since SSE only has
6035 // GT and EQ comparisons for integer, swapping operands and multiple
6036 // operations may be required for some comparisons.
6037 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6038 bool Swap = false, Invert = false, FlipSigns = false;
6040 switch (VT.getSimpleVT().SimpleTy) {
6043 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6045 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6047 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6048 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6051 switch (SetCCOpcode) {
6053 case ISD::SETNE: Invert = true;
6054 case ISD::SETEQ: Opc = EQOpc; break;
6055 case ISD::SETLT: Swap = true;
6056 case ISD::SETGT: Opc = GTOpc; break;
6057 case ISD::SETGE: Swap = true;
6058 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6059 case ISD::SETULT: Swap = true;
6060 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6061 case ISD::SETUGE: Swap = true;
6062 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6065 std::swap(Op0, Op1);
6067 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6068 // bits of the inputs before performing those operations.
6070 EVT EltVT = VT.getVectorElementType();
6071 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6073 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6074 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6076 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6077 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6080 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6082 // If the logical-not of the result is required, perform that now.
6084 Result = DAG.getNOT(dl, Result, VT);
6089 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6090 static bool isX86LogicalCmp(SDValue Op) {
6091 unsigned Opc = Op.getNode()->getOpcode();
6092 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6094 if (Op.getResNo() == 1 &&
6095 (Opc == X86ISD::ADD ||
6096 Opc == X86ISD::SUB ||
6097 Opc == X86ISD::SMUL ||
6098 Opc == X86ISD::UMUL ||
6099 Opc == X86ISD::INC ||
6100 Opc == X86ISD::DEC ||
6101 Opc == X86ISD::OR ||
6102 Opc == X86ISD::XOR ||
6103 Opc == X86ISD::AND))
6109 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
6110 bool addTest = true;
6111 SDValue Cond = Op.getOperand(0);
6112 DebugLoc dl = Op.getDebugLoc();
6115 if (Cond.getOpcode() == ISD::SETCC) {
6116 SDValue NewCond = LowerSETCC(Cond, DAG);
6117 if (NewCond.getNode())
6121 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6122 SDValue Op1 = Op.getOperand(1);
6123 SDValue Op2 = Op.getOperand(2);
6124 if (Cond.getOpcode() == X86ISD::SETCC &&
6125 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6126 SDValue Cmp = Cond.getOperand(1);
6127 if (Cmp.getOpcode() == X86ISD::CMP) {
6128 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6129 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6130 ConstantSDNode *RHSC =
6131 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6132 if (N1C && N1C->isAllOnesValue() &&
6133 N2C && N2C->isNullValue() &&
6134 RHSC && RHSC->isNullValue()) {
6135 SDValue CmpOp0 = Cmp.getOperand(0);
6136 Cmp = DAG.getNode(X86ISD::CMP, dl, CmpOp0.getValueType(),
6137 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6138 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6139 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6144 // Look pass (and (setcc_carry (cmp ...)), 1).
6145 if (Cond.getOpcode() == ISD::AND &&
6146 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6147 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6148 if (C && C->getAPIntValue() == 1)
6149 Cond = Cond.getOperand(0);
6152 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6153 // setting operand in place of the X86ISD::SETCC.
6154 if (Cond.getOpcode() == X86ISD::SETCC ||
6155 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6156 CC = Cond.getOperand(0);
6158 SDValue Cmp = Cond.getOperand(1);
6159 unsigned Opc = Cmp.getOpcode();
6160 EVT VT = Op.getValueType();
6162 bool IllegalFPCMov = false;
6163 if (VT.isFloatingPoint() && !VT.isVector() &&
6164 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6165 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6167 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6168 Opc == X86ISD::BT) { // FIXME
6175 // Look pass the truncate.
6176 if (Cond.getOpcode() == ISD::TRUNCATE)
6177 Cond = Cond.getOperand(0);
6179 // We know the result of AND is compared against zero. Try to match
6181 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6182 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6183 if (NewSetCC.getNode()) {
6184 CC = NewSetCC.getOperand(0);
6185 Cond = NewSetCC.getOperand(1);
6192 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6193 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6196 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6197 // condition is true.
6198 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6199 SDValue Ops[] = { Op2, Op1, CC, Cond };
6200 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6203 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6204 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6205 // from the AND / OR.
6206 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6207 Opc = Op.getOpcode();
6208 if (Opc != ISD::OR && Opc != ISD::AND)
6210 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6211 Op.getOperand(0).hasOneUse() &&
6212 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6213 Op.getOperand(1).hasOneUse());
6216 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6217 // 1 and that the SETCC node has a single use.
6218 static bool isXor1OfSetCC(SDValue Op) {
6219 if (Op.getOpcode() != ISD::XOR)
6221 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6222 if (N1C && N1C->getAPIntValue() == 1) {
6223 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6224 Op.getOperand(0).hasOneUse();
6229 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
6230 bool addTest = true;
6231 SDValue Chain = Op.getOperand(0);
6232 SDValue Cond = Op.getOperand(1);
6233 SDValue Dest = Op.getOperand(2);
6234 DebugLoc dl = Op.getDebugLoc();
6237 if (Cond.getOpcode() == ISD::SETCC) {
6238 SDValue NewCond = LowerSETCC(Cond, DAG);
6239 if (NewCond.getNode())
6243 // FIXME: LowerXALUO doesn't handle these!!
6244 else if (Cond.getOpcode() == X86ISD::ADD ||
6245 Cond.getOpcode() == X86ISD::SUB ||
6246 Cond.getOpcode() == X86ISD::SMUL ||
6247 Cond.getOpcode() == X86ISD::UMUL)
6248 Cond = LowerXALUO(Cond, DAG);
6251 // Look pass (and (setcc_carry (cmp ...)), 1).
6252 if (Cond.getOpcode() == ISD::AND &&
6253 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6254 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6255 if (C && C->getAPIntValue() == 1)
6256 Cond = Cond.getOperand(0);
6259 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6260 // setting operand in place of the X86ISD::SETCC.
6261 if (Cond.getOpcode() == X86ISD::SETCC ||
6262 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6263 CC = Cond.getOperand(0);
6265 SDValue Cmp = Cond.getOperand(1);
6266 unsigned Opc = Cmp.getOpcode();
6267 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6268 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6272 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6276 // These can only come from an arithmetic instruction with overflow,
6277 // e.g. SADDO, UADDO.
6278 Cond = Cond.getNode()->getOperand(1);
6285 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6286 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6287 if (CondOpc == ISD::OR) {
6288 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6289 // two branches instead of an explicit OR instruction with a
6291 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6292 isX86LogicalCmp(Cmp)) {
6293 CC = Cond.getOperand(0).getOperand(0);
6294 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6295 Chain, Dest, CC, Cmp);
6296 CC = Cond.getOperand(1).getOperand(0);
6300 } else { // ISD::AND
6301 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6302 // two branches instead of an explicit AND instruction with a
6303 // separate test. However, we only do this if this block doesn't
6304 // have a fall-through edge, because this requires an explicit
6305 // jmp when the condition is false.
6306 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6307 isX86LogicalCmp(Cmp) &&
6308 Op.getNode()->hasOneUse()) {
6309 X86::CondCode CCode =
6310 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6311 CCode = X86::GetOppositeBranchCondition(CCode);
6312 CC = DAG.getConstant(CCode, MVT::i8);
6313 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6314 // Look for an unconditional branch following this conditional branch.
6315 // We need this because we need to reverse the successors in order
6316 // to implement FCMP_OEQ.
6317 if (User.getOpcode() == ISD::BR) {
6318 SDValue FalseBB = User.getOperand(1);
6320 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6321 assert(NewBR == User);
6324 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6325 Chain, Dest, CC, Cmp);
6326 X86::CondCode CCode =
6327 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6328 CCode = X86::GetOppositeBranchCondition(CCode);
6329 CC = DAG.getConstant(CCode, MVT::i8);
6335 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6336 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6337 // It should be transformed during dag combiner except when the condition
6338 // is set by a arithmetics with overflow node.
6339 X86::CondCode CCode =
6340 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6341 CCode = X86::GetOppositeBranchCondition(CCode);
6342 CC = DAG.getConstant(CCode, MVT::i8);
6343 Cond = Cond.getOperand(0).getOperand(1);
6349 // Look pass the truncate.
6350 if (Cond.getOpcode() == ISD::TRUNCATE)
6351 Cond = Cond.getOperand(0);
6353 // We know the result of AND is compared against zero. Try to match
6355 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6356 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6357 if (NewSetCC.getNode()) {
6358 CC = NewSetCC.getOperand(0);
6359 Cond = NewSetCC.getOperand(1);
6366 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6367 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6369 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6370 Chain, Dest, CC, Cond);
6374 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6375 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6376 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6377 // that the guard pages used by the OS virtual memory manager are allocated in
6378 // correct sequence.
6380 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6381 SelectionDAG &DAG) {
6382 assert(Subtarget->isTargetCygMing() &&
6383 "This should be used only on Cygwin/Mingw targets");
6384 DebugLoc dl = Op.getDebugLoc();
6387 SDValue Chain = Op.getOperand(0);
6388 SDValue Size = Op.getOperand(1);
6389 // FIXME: Ensure alignment here
6393 EVT IntPtr = getPointerTy();
6394 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6396 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
6398 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6399 Flag = Chain.getValue(1);
6401 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6402 SDValue Ops[] = { Chain,
6403 DAG.getTargetExternalSymbol("_alloca", IntPtr),
6404 DAG.getRegister(X86::EAX, IntPtr),
6405 DAG.getRegister(X86StackPtr, SPTy),
6407 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
6408 Flag = Chain.getValue(1);
6410 Chain = DAG.getCALLSEQ_END(Chain,
6411 DAG.getIntPtrConstant(0, true),
6412 DAG.getIntPtrConstant(0, true),
6415 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6417 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6418 return DAG.getMergeValues(Ops1, 2, dl);
6422 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6424 SDValue Dst, SDValue Src,
6425 SDValue Size, unsigned Align,
6427 uint64_t DstSVOff) {
6428 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6430 // If not DWORD aligned or size is more than the threshold, call the library.
6431 // The libc version is likely to be faster for these cases. It can use the
6432 // address value and run time information about the CPU.
6433 if ((Align & 3) != 0 ||
6435 ConstantSize->getZExtValue() >
6436 getSubtarget()->getMaxInlineSizeThreshold()) {
6437 SDValue InFlag(0, 0);
6439 // Check to see if there is a specialized entry-point for memory zeroing.
6440 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6442 if (const char *bzeroEntry = V &&
6443 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6444 EVT IntPtr = getPointerTy();
6445 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6446 TargetLowering::ArgListTy Args;
6447 TargetLowering::ArgListEntry Entry;
6449 Entry.Ty = IntPtrTy;
6450 Args.push_back(Entry);
6452 Args.push_back(Entry);
6453 std::pair<SDValue,SDValue> CallResult =
6454 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6455 false, false, false, false,
6456 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6457 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl,
6458 DAG.GetOrdering(Chain.getNode()));
6459 return CallResult.second;
6462 // Otherwise have the target-independent code call memset.
6466 uint64_t SizeVal = ConstantSize->getZExtValue();
6467 SDValue InFlag(0, 0);
6470 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6471 unsigned BytesLeft = 0;
6472 bool TwoRepStos = false;
6475 uint64_t Val = ValC->getZExtValue() & 255;
6477 // If the value is a constant, then we can potentially use larger sets.
6478 switch (Align & 3) {
6479 case 2: // WORD aligned
6482 Val = (Val << 8) | Val;
6484 case 0: // DWORD aligned
6487 Val = (Val << 8) | Val;
6488 Val = (Val << 16) | Val;
6489 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6492 Val = (Val << 32) | Val;
6495 default: // Byte aligned
6498 Count = DAG.getIntPtrConstant(SizeVal);
6502 if (AVT.bitsGT(MVT::i8)) {
6503 unsigned UBytes = AVT.getSizeInBits() / 8;
6504 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6505 BytesLeft = SizeVal % UBytes;
6508 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6510 InFlag = Chain.getValue(1);
6513 Count = DAG.getIntPtrConstant(SizeVal);
6514 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6515 InFlag = Chain.getValue(1);
6518 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6521 InFlag = Chain.getValue(1);
6522 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6525 InFlag = Chain.getValue(1);
6527 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6528 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6529 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6532 InFlag = Chain.getValue(1);
6534 EVT CVT = Count.getValueType();
6535 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6536 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6537 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6540 InFlag = Chain.getValue(1);
6541 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6542 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
6543 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6544 } else if (BytesLeft) {
6545 // Handle the last 1 - 7 bytes.
6546 unsigned Offset = SizeVal - BytesLeft;
6547 EVT AddrVT = Dst.getValueType();
6548 EVT SizeVT = Size.getValueType();
6550 Chain = DAG.getMemset(Chain, dl,
6551 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6552 DAG.getConstant(Offset, AddrVT)),
6554 DAG.getConstant(BytesLeft, SizeVT),
6555 Align, DstSV, DstSVOff + Offset);
6558 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6563 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6564 SDValue Chain, SDValue Dst, SDValue Src,
6565 SDValue Size, unsigned Align,
6567 const Value *DstSV, uint64_t DstSVOff,
6568 const Value *SrcSV, uint64_t SrcSVOff) {
6569 // This requires the copy size to be a constant, preferrably
6570 // within a subtarget-specific limit.
6571 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6574 uint64_t SizeVal = ConstantSize->getZExtValue();
6575 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6578 /// If not DWORD aligned, call the library.
6579 if ((Align & 3) != 0)
6584 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6587 unsigned UBytes = AVT.getSizeInBits() / 8;
6588 unsigned CountVal = SizeVal / UBytes;
6589 SDValue Count = DAG.getIntPtrConstant(CountVal);
6590 unsigned BytesLeft = SizeVal % UBytes;
6592 SDValue InFlag(0, 0);
6593 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6596 InFlag = Chain.getValue(1);
6597 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6600 InFlag = Chain.getValue(1);
6601 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6604 InFlag = Chain.getValue(1);
6606 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6607 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6608 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
6609 array_lengthof(Ops));
6611 SmallVector<SDValue, 4> Results;
6612 Results.push_back(RepMovs);
6614 // Handle the last 1 - 7 bytes.
6615 unsigned Offset = SizeVal - BytesLeft;
6616 EVT DstVT = Dst.getValueType();
6617 EVT SrcVT = Src.getValueType();
6618 EVT SizeVT = Size.getValueType();
6619 Results.push_back(DAG.getMemcpy(Chain, dl,
6620 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6621 DAG.getConstant(Offset, DstVT)),
6622 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6623 DAG.getConstant(Offset, SrcVT)),
6624 DAG.getConstant(BytesLeft, SizeVT),
6625 Align, AlwaysInline,
6626 DstSV, DstSVOff + Offset,
6627 SrcSV, SrcSVOff + Offset));
6630 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6631 &Results[0], Results.size());
6634 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6635 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6636 DebugLoc dl = Op.getDebugLoc();
6638 if (!Subtarget->is64Bit()) {
6639 // vastart just stores the address of the VarArgsFrameIndex slot into the
6640 // memory location argument.
6641 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6642 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6647 // gp_offset (0 - 6 * 8)
6648 // fp_offset (48 - 48 + 8 * 16)
6649 // overflow_arg_area (point to parameters coming in memory).
6651 SmallVector<SDValue, 8> MemOps;
6652 SDValue FIN = Op.getOperand(1);
6654 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6655 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6656 FIN, SV, 0, false, false, 0);
6657 MemOps.push_back(Store);
6660 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6661 FIN, DAG.getIntPtrConstant(4));
6662 Store = DAG.getStore(Op.getOperand(0), dl,
6663 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6664 FIN, SV, 0, false, false, 0);
6665 MemOps.push_back(Store);
6667 // Store ptr to overflow_arg_area
6668 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6669 FIN, DAG.getIntPtrConstant(4));
6670 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6671 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0,
6673 MemOps.push_back(Store);
6675 // Store ptr to reg_save_area.
6676 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6677 FIN, DAG.getIntPtrConstant(8));
6678 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6679 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0,
6681 MemOps.push_back(Store);
6682 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6683 &MemOps[0], MemOps.size());
6686 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6687 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6688 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6689 SDValue Chain = Op.getOperand(0);
6690 SDValue SrcPtr = Op.getOperand(1);
6691 SDValue SrcSV = Op.getOperand(2);
6693 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6697 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6698 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6699 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6700 SDValue Chain = Op.getOperand(0);
6701 SDValue DstPtr = Op.getOperand(1);
6702 SDValue SrcPtr = Op.getOperand(2);
6703 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6704 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6705 DebugLoc dl = Op.getDebugLoc();
6707 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6708 DAG.getIntPtrConstant(24), 8, false,
6709 DstSV, 0, SrcSV, 0);
6713 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6714 DebugLoc dl = Op.getDebugLoc();
6715 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6717 default: return SDValue(); // Don't custom lower most intrinsics.
6718 // Comparison intrinsics.
6719 case Intrinsic::x86_sse_comieq_ss:
6720 case Intrinsic::x86_sse_comilt_ss:
6721 case Intrinsic::x86_sse_comile_ss:
6722 case Intrinsic::x86_sse_comigt_ss:
6723 case Intrinsic::x86_sse_comige_ss:
6724 case Intrinsic::x86_sse_comineq_ss:
6725 case Intrinsic::x86_sse_ucomieq_ss:
6726 case Intrinsic::x86_sse_ucomilt_ss:
6727 case Intrinsic::x86_sse_ucomile_ss:
6728 case Intrinsic::x86_sse_ucomigt_ss:
6729 case Intrinsic::x86_sse_ucomige_ss:
6730 case Intrinsic::x86_sse_ucomineq_ss:
6731 case Intrinsic::x86_sse2_comieq_sd:
6732 case Intrinsic::x86_sse2_comilt_sd:
6733 case Intrinsic::x86_sse2_comile_sd:
6734 case Intrinsic::x86_sse2_comigt_sd:
6735 case Intrinsic::x86_sse2_comige_sd:
6736 case Intrinsic::x86_sse2_comineq_sd:
6737 case Intrinsic::x86_sse2_ucomieq_sd:
6738 case Intrinsic::x86_sse2_ucomilt_sd:
6739 case Intrinsic::x86_sse2_ucomile_sd:
6740 case Intrinsic::x86_sse2_ucomigt_sd:
6741 case Intrinsic::x86_sse2_ucomige_sd:
6742 case Intrinsic::x86_sse2_ucomineq_sd: {
6744 ISD::CondCode CC = ISD::SETCC_INVALID;
6747 case Intrinsic::x86_sse_comieq_ss:
6748 case Intrinsic::x86_sse2_comieq_sd:
6752 case Intrinsic::x86_sse_comilt_ss:
6753 case Intrinsic::x86_sse2_comilt_sd:
6757 case Intrinsic::x86_sse_comile_ss:
6758 case Intrinsic::x86_sse2_comile_sd:
6762 case Intrinsic::x86_sse_comigt_ss:
6763 case Intrinsic::x86_sse2_comigt_sd:
6767 case Intrinsic::x86_sse_comige_ss:
6768 case Intrinsic::x86_sse2_comige_sd:
6772 case Intrinsic::x86_sse_comineq_ss:
6773 case Intrinsic::x86_sse2_comineq_sd:
6777 case Intrinsic::x86_sse_ucomieq_ss:
6778 case Intrinsic::x86_sse2_ucomieq_sd:
6779 Opc = X86ISD::UCOMI;
6782 case Intrinsic::x86_sse_ucomilt_ss:
6783 case Intrinsic::x86_sse2_ucomilt_sd:
6784 Opc = X86ISD::UCOMI;
6787 case Intrinsic::x86_sse_ucomile_ss:
6788 case Intrinsic::x86_sse2_ucomile_sd:
6789 Opc = X86ISD::UCOMI;
6792 case Intrinsic::x86_sse_ucomigt_ss:
6793 case Intrinsic::x86_sse2_ucomigt_sd:
6794 Opc = X86ISD::UCOMI;
6797 case Intrinsic::x86_sse_ucomige_ss:
6798 case Intrinsic::x86_sse2_ucomige_sd:
6799 Opc = X86ISD::UCOMI;
6802 case Intrinsic::x86_sse_ucomineq_ss:
6803 case Intrinsic::x86_sse2_ucomineq_sd:
6804 Opc = X86ISD::UCOMI;
6809 SDValue LHS = Op.getOperand(1);
6810 SDValue RHS = Op.getOperand(2);
6811 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6812 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6813 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6814 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6815 DAG.getConstant(X86CC, MVT::i8), Cond);
6816 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6818 // ptest intrinsics. The intrinsic these come from are designed to return
6819 // an integer value, not just an instruction so lower it to the ptest
6820 // pattern and a setcc for the result.
6821 case Intrinsic::x86_sse41_ptestz:
6822 case Intrinsic::x86_sse41_ptestc:
6823 case Intrinsic::x86_sse41_ptestnzc:{
6826 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6827 case Intrinsic::x86_sse41_ptestz:
6829 X86CC = X86::COND_E;
6831 case Intrinsic::x86_sse41_ptestc:
6833 X86CC = X86::COND_B;
6835 case Intrinsic::x86_sse41_ptestnzc:
6837 X86CC = X86::COND_A;
6841 SDValue LHS = Op.getOperand(1);
6842 SDValue RHS = Op.getOperand(2);
6843 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6844 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6845 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6846 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6849 // Fix vector shift instructions where the last operand is a non-immediate
6851 case Intrinsic::x86_sse2_pslli_w:
6852 case Intrinsic::x86_sse2_pslli_d:
6853 case Intrinsic::x86_sse2_pslli_q:
6854 case Intrinsic::x86_sse2_psrli_w:
6855 case Intrinsic::x86_sse2_psrli_d:
6856 case Intrinsic::x86_sse2_psrli_q:
6857 case Intrinsic::x86_sse2_psrai_w:
6858 case Intrinsic::x86_sse2_psrai_d:
6859 case Intrinsic::x86_mmx_pslli_w:
6860 case Intrinsic::x86_mmx_pslli_d:
6861 case Intrinsic::x86_mmx_pslli_q:
6862 case Intrinsic::x86_mmx_psrli_w:
6863 case Intrinsic::x86_mmx_psrli_d:
6864 case Intrinsic::x86_mmx_psrli_q:
6865 case Intrinsic::x86_mmx_psrai_w:
6866 case Intrinsic::x86_mmx_psrai_d: {
6867 SDValue ShAmt = Op.getOperand(2);
6868 if (isa<ConstantSDNode>(ShAmt))
6871 unsigned NewIntNo = 0;
6872 EVT ShAmtVT = MVT::v4i32;
6874 case Intrinsic::x86_sse2_pslli_w:
6875 NewIntNo = Intrinsic::x86_sse2_psll_w;
6877 case Intrinsic::x86_sse2_pslli_d:
6878 NewIntNo = Intrinsic::x86_sse2_psll_d;
6880 case Intrinsic::x86_sse2_pslli_q:
6881 NewIntNo = Intrinsic::x86_sse2_psll_q;
6883 case Intrinsic::x86_sse2_psrli_w:
6884 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6886 case Intrinsic::x86_sse2_psrli_d:
6887 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6889 case Intrinsic::x86_sse2_psrli_q:
6890 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6892 case Intrinsic::x86_sse2_psrai_w:
6893 NewIntNo = Intrinsic::x86_sse2_psra_w;
6895 case Intrinsic::x86_sse2_psrai_d:
6896 NewIntNo = Intrinsic::x86_sse2_psra_d;
6899 ShAmtVT = MVT::v2i32;
6901 case Intrinsic::x86_mmx_pslli_w:
6902 NewIntNo = Intrinsic::x86_mmx_psll_w;
6904 case Intrinsic::x86_mmx_pslli_d:
6905 NewIntNo = Intrinsic::x86_mmx_psll_d;
6907 case Intrinsic::x86_mmx_pslli_q:
6908 NewIntNo = Intrinsic::x86_mmx_psll_q;
6910 case Intrinsic::x86_mmx_psrli_w:
6911 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6913 case Intrinsic::x86_mmx_psrli_d:
6914 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6916 case Intrinsic::x86_mmx_psrli_q:
6917 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6919 case Intrinsic::x86_mmx_psrai_w:
6920 NewIntNo = Intrinsic::x86_mmx_psra_w;
6922 case Intrinsic::x86_mmx_psrai_d:
6923 NewIntNo = Intrinsic::x86_mmx_psra_d;
6925 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6931 // The vector shift intrinsics with scalars uses 32b shift amounts but
6932 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
6936 ShOps[1] = DAG.getConstant(0, MVT::i32);
6937 if (ShAmtVT == MVT::v4i32) {
6938 ShOps[2] = DAG.getUNDEF(MVT::i32);
6939 ShOps[3] = DAG.getUNDEF(MVT::i32);
6940 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
6942 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
6945 EVT VT = Op.getValueType();
6946 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
6947 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6948 DAG.getConstant(NewIntNo, MVT::i32),
6949 Op.getOperand(1), ShAmt);
6954 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6955 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6956 DebugLoc dl = Op.getDebugLoc();
6959 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6961 DAG.getConstant(TD->getPointerSize(),
6962 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6963 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6964 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6966 NULL, 0, false, false, 0);
6969 // Just load the return address.
6970 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6971 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6972 RetAddrFI, NULL, 0, false, false, 0);
6975 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6976 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6977 MFI->setFrameAddressIsTaken(true);
6978 EVT VT = Op.getValueType();
6979 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6980 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6981 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6982 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6984 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
6989 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6990 SelectionDAG &DAG) {
6991 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6994 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6996 MachineFunction &MF = DAG.getMachineFunction();
6997 SDValue Chain = Op.getOperand(0);
6998 SDValue Offset = Op.getOperand(1);
6999 SDValue Handler = Op.getOperand(2);
7000 DebugLoc dl = Op.getDebugLoc();
7002 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7004 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7006 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7007 DAG.getIntPtrConstant(-TD->getPointerSize()));
7008 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7009 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7010 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7011 MF.getRegInfo().addLiveOut(StoreAddrReg);
7013 return DAG.getNode(X86ISD::EH_RETURN, dl,
7015 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7018 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7019 SelectionDAG &DAG) {
7020 SDValue Root = Op.getOperand(0);
7021 SDValue Trmp = Op.getOperand(1); // trampoline
7022 SDValue FPtr = Op.getOperand(2); // nested function
7023 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7024 DebugLoc dl = Op.getDebugLoc();
7026 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7028 if (Subtarget->is64Bit()) {
7029 SDValue OutChains[6];
7031 // Large code-model.
7032 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7033 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7035 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7036 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7038 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7040 // Load the pointer to the nested function into R11.
7041 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7042 SDValue Addr = Trmp;
7043 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7044 Addr, TrmpAddr, 0, false, false, 0);
7046 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7047 DAG.getConstant(2, MVT::i64));
7048 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7051 // Load the 'nest' parameter value into R10.
7052 // R10 is specified in X86CallingConv.td
7053 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7054 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7055 DAG.getConstant(10, MVT::i64));
7056 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7057 Addr, TrmpAddr, 10, false, false, 0);
7059 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7060 DAG.getConstant(12, MVT::i64));
7061 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7064 // Jump to the nested function.
7065 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7066 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7067 DAG.getConstant(20, MVT::i64));
7068 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7069 Addr, TrmpAddr, 20, false, false, 0);
7071 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7072 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7073 DAG.getConstant(22, MVT::i64));
7074 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7075 TrmpAddr, 22, false, false, 0);
7078 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7079 return DAG.getMergeValues(Ops, 2, dl);
7081 const Function *Func =
7082 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7083 CallingConv::ID CC = Func->getCallingConv();
7088 llvm_unreachable("Unsupported calling convention");
7089 case CallingConv::C:
7090 case CallingConv::X86_StdCall: {
7091 // Pass 'nest' parameter in ECX.
7092 // Must be kept in sync with X86CallingConv.td
7095 // Check that ECX wasn't needed by an 'inreg' parameter.
7096 const FunctionType *FTy = Func->getFunctionType();
7097 const AttrListPtr &Attrs = Func->getAttributes();
7099 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7100 unsigned InRegCount = 0;
7103 for (FunctionType::param_iterator I = FTy->param_begin(),
7104 E = FTy->param_end(); I != E; ++I, ++Idx)
7105 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7106 // FIXME: should only count parameters that are lowered to integers.
7107 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7109 if (InRegCount > 2) {
7110 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
7115 case CallingConv::X86_FastCall:
7116 case CallingConv::Fast:
7117 // Pass 'nest' parameter in EAX.
7118 // Must be kept in sync with X86CallingConv.td
7123 SDValue OutChains[4];
7126 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7127 DAG.getConstant(10, MVT::i32));
7128 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7130 // This is storing the opcode for MOV32ri.
7131 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7132 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7133 OutChains[0] = DAG.getStore(Root, dl,
7134 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7135 Trmp, TrmpAddr, 0, false, false, 0);
7137 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7138 DAG.getConstant(1, MVT::i32));
7139 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7142 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7143 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7144 DAG.getConstant(5, MVT::i32));
7145 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7146 TrmpAddr, 5, false, false, 1);
7148 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7149 DAG.getConstant(6, MVT::i32));
7150 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7154 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7155 return DAG.getMergeValues(Ops, 2, dl);
7159 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
7161 The rounding mode is in bits 11:10 of FPSR, and has the following
7168 FLT_ROUNDS, on the other hand, expects the following:
7175 To perform the conversion, we do:
7176 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7179 MachineFunction &MF = DAG.getMachineFunction();
7180 const TargetMachine &TM = MF.getTarget();
7181 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7182 unsigned StackAlignment = TFI.getStackAlignment();
7183 EVT VT = Op.getValueType();
7184 DebugLoc dl = Op.getDebugLoc();
7186 // Save FP Control Word to stack slot
7187 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7188 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7190 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7191 DAG.getEntryNode(), StackSlot);
7193 // Load FP Control Word from stack slot
7194 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7197 // Transform as necessary
7199 DAG.getNode(ISD::SRL, dl, MVT::i16,
7200 DAG.getNode(ISD::AND, dl, MVT::i16,
7201 CWD, DAG.getConstant(0x800, MVT::i16)),
7202 DAG.getConstant(11, MVT::i8));
7204 DAG.getNode(ISD::SRL, dl, MVT::i16,
7205 DAG.getNode(ISD::AND, dl, MVT::i16,
7206 CWD, DAG.getConstant(0x400, MVT::i16)),
7207 DAG.getConstant(9, MVT::i8));
7210 DAG.getNode(ISD::AND, dl, MVT::i16,
7211 DAG.getNode(ISD::ADD, dl, MVT::i16,
7212 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7213 DAG.getConstant(1, MVT::i16)),
7214 DAG.getConstant(3, MVT::i16));
7217 return DAG.getNode((VT.getSizeInBits() < 16 ?
7218 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7221 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
7222 EVT VT = Op.getValueType();
7224 unsigned NumBits = VT.getSizeInBits();
7225 DebugLoc dl = Op.getDebugLoc();
7227 Op = Op.getOperand(0);
7228 if (VT == MVT::i8) {
7229 // Zero extend to i32 since there is not an i8 bsr.
7231 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7234 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7235 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7236 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7238 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7241 DAG.getConstant(NumBits+NumBits-1, OpVT),
7242 DAG.getConstant(X86::COND_E, MVT::i8),
7245 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7247 // Finally xor with NumBits-1.
7248 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7251 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7255 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
7256 EVT VT = Op.getValueType();
7258 unsigned NumBits = VT.getSizeInBits();
7259 DebugLoc dl = Op.getDebugLoc();
7261 Op = Op.getOperand(0);
7262 if (VT == MVT::i8) {
7264 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7267 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7268 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7269 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7271 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7274 DAG.getConstant(NumBits, OpVT),
7275 DAG.getConstant(X86::COND_E, MVT::i8),
7278 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7281 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7285 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
7286 EVT VT = Op.getValueType();
7287 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7288 DebugLoc dl = Op.getDebugLoc();
7290 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7291 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7292 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7293 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7294 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7296 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7297 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7298 // return AloBlo + AloBhi + AhiBlo;
7300 SDValue A = Op.getOperand(0);
7301 SDValue B = Op.getOperand(1);
7303 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7304 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7305 A, DAG.getConstant(32, MVT::i32));
7306 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7307 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7308 B, DAG.getConstant(32, MVT::i32));
7309 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7310 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7312 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7313 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7315 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7316 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7318 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7319 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7320 AloBhi, DAG.getConstant(32, MVT::i32));
7321 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7322 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7323 AhiBlo, DAG.getConstant(32, MVT::i32));
7324 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7325 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7330 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
7331 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7332 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7333 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7334 // has only one use.
7335 SDNode *N = Op.getNode();
7336 SDValue LHS = N->getOperand(0);
7337 SDValue RHS = N->getOperand(1);
7338 unsigned BaseOp = 0;
7340 DebugLoc dl = Op.getDebugLoc();
7342 switch (Op.getOpcode()) {
7343 default: llvm_unreachable("Unknown ovf instruction!");
7345 // A subtract of one will be selected as a INC. Note that INC doesn't
7346 // set CF, so we can't do this for UADDO.
7347 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7348 if (C->getAPIntValue() == 1) {
7349 BaseOp = X86ISD::INC;
7353 BaseOp = X86ISD::ADD;
7357 BaseOp = X86ISD::ADD;
7361 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7362 // set CF, so we can't do this for USUBO.
7363 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7364 if (C->getAPIntValue() == 1) {
7365 BaseOp = X86ISD::DEC;
7369 BaseOp = X86ISD::SUB;
7373 BaseOp = X86ISD::SUB;
7377 BaseOp = X86ISD::SMUL;
7381 BaseOp = X86ISD::UMUL;
7386 // Also sets EFLAGS.
7387 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7388 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7391 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7392 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7394 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7398 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
7399 EVT T = Op.getValueType();
7400 DebugLoc dl = Op.getDebugLoc();
7403 switch(T.getSimpleVT().SimpleTy) {
7405 assert(false && "Invalid value type!");
7406 case MVT::i8: Reg = X86::AL; size = 1; break;
7407 case MVT::i16: Reg = X86::AX; size = 2; break;
7408 case MVT::i32: Reg = X86::EAX; size = 4; break;
7410 assert(Subtarget->is64Bit() && "Node not type legal!");
7411 Reg = X86::RAX; size = 8;
7414 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7415 Op.getOperand(2), SDValue());
7416 SDValue Ops[] = { cpIn.getValue(0),
7419 DAG.getTargetConstant(size, MVT::i8),
7421 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7422 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7424 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7428 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7429 SelectionDAG &DAG) {
7430 assert(Subtarget->is64Bit() && "Result not type legalized?");
7431 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7432 SDValue TheChain = Op.getOperand(0);
7433 DebugLoc dl = Op.getDebugLoc();
7434 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7435 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7436 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7438 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7439 DAG.getConstant(32, MVT::i8));
7441 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7444 return DAG.getMergeValues(Ops, 2, dl);
7447 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7448 SDNode *Node = Op.getNode();
7449 DebugLoc dl = Node->getDebugLoc();
7450 EVT T = Node->getValueType(0);
7451 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7452 DAG.getConstant(0, T), Node->getOperand(2));
7453 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7454 cast<AtomicSDNode>(Node)->getMemoryVT(),
7455 Node->getOperand(0),
7456 Node->getOperand(1), negOp,
7457 cast<AtomicSDNode>(Node)->getSrcValue(),
7458 cast<AtomicSDNode>(Node)->getAlignment());
7461 /// LowerOperation - Provide custom lowering hooks for some operations.
7463 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7464 switch (Op.getOpcode()) {
7465 default: llvm_unreachable("Should not custom lower this!");
7466 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7467 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7468 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7469 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7470 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7471 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7472 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7473 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7474 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7475 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7476 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7477 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7478 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7479 case ISD::SHL_PARTS:
7480 case ISD::SRA_PARTS:
7481 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7482 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7483 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7484 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7485 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7486 case ISD::FABS: return LowerFABS(Op, DAG);
7487 case ISD::FNEG: return LowerFNEG(Op, DAG);
7488 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7489 case ISD::SETCC: return LowerSETCC(Op, DAG);
7490 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7491 case ISD::SELECT: return LowerSELECT(Op, DAG);
7492 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7493 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7494 case ISD::VASTART: return LowerVASTART(Op, DAG);
7495 case ISD::VAARG: return LowerVAARG(Op, DAG);
7496 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7497 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7498 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7499 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7500 case ISD::FRAME_TO_ARGS_OFFSET:
7501 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7502 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7503 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7504 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7505 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7506 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7507 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7508 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7514 case ISD::UMULO: return LowerXALUO(Op, DAG);
7515 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7519 void X86TargetLowering::
7520 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7521 SelectionDAG &DAG, unsigned NewOp) {
7522 EVT T = Node->getValueType(0);
7523 DebugLoc dl = Node->getDebugLoc();
7524 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7526 SDValue Chain = Node->getOperand(0);
7527 SDValue In1 = Node->getOperand(1);
7528 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7529 Node->getOperand(2), DAG.getIntPtrConstant(0));
7530 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7531 Node->getOperand(2), DAG.getIntPtrConstant(1));
7532 SDValue Ops[] = { Chain, In1, In2L, In2H };
7533 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7535 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7536 cast<MemSDNode>(Node)->getMemOperand());
7537 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7538 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7539 Results.push_back(Result.getValue(2));
7542 /// ReplaceNodeResults - Replace a node with an illegal result type
7543 /// with a new node built out of custom code.
7544 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7545 SmallVectorImpl<SDValue>&Results,
7546 SelectionDAG &DAG) {
7547 DebugLoc dl = N->getDebugLoc();
7548 switch (N->getOpcode()) {
7550 assert(false && "Do not know how to custom type legalize this operation!");
7552 case ISD::FP_TO_SINT: {
7553 std::pair<SDValue,SDValue> Vals =
7554 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7555 SDValue FIST = Vals.first, StackSlot = Vals.second;
7556 if (FIST.getNode() != 0) {
7557 EVT VT = N->getValueType(0);
7558 // Return a load from the stack slot.
7559 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7564 case ISD::READCYCLECOUNTER: {
7565 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7566 SDValue TheChain = N->getOperand(0);
7567 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7568 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7570 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7572 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7573 SDValue Ops[] = { eax, edx };
7574 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7575 Results.push_back(edx.getValue(1));
7578 case ISD::ATOMIC_CMP_SWAP: {
7579 EVT T = N->getValueType(0);
7580 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7581 SDValue cpInL, cpInH;
7582 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7583 DAG.getConstant(0, MVT::i32));
7584 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7585 DAG.getConstant(1, MVT::i32));
7586 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7587 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7589 SDValue swapInL, swapInH;
7590 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7591 DAG.getConstant(0, MVT::i32));
7592 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7593 DAG.getConstant(1, MVT::i32));
7594 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7596 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7597 swapInL.getValue(1));
7598 SDValue Ops[] = { swapInH.getValue(0),
7600 swapInH.getValue(1) };
7601 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7602 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7603 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7604 MVT::i32, Result.getValue(1));
7605 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7606 MVT::i32, cpOutL.getValue(2));
7607 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7608 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7609 Results.push_back(cpOutH.getValue(1));
7612 case ISD::ATOMIC_LOAD_ADD:
7613 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7615 case ISD::ATOMIC_LOAD_AND:
7616 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7618 case ISD::ATOMIC_LOAD_NAND:
7619 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7621 case ISD::ATOMIC_LOAD_OR:
7622 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7624 case ISD::ATOMIC_LOAD_SUB:
7625 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7627 case ISD::ATOMIC_LOAD_XOR:
7628 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7630 case ISD::ATOMIC_SWAP:
7631 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7636 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7638 default: return NULL;
7639 case X86ISD::BSF: return "X86ISD::BSF";
7640 case X86ISD::BSR: return "X86ISD::BSR";
7641 case X86ISD::SHLD: return "X86ISD::SHLD";
7642 case X86ISD::SHRD: return "X86ISD::SHRD";
7643 case X86ISD::FAND: return "X86ISD::FAND";
7644 case X86ISD::FOR: return "X86ISD::FOR";
7645 case X86ISD::FXOR: return "X86ISD::FXOR";
7646 case X86ISD::FSRL: return "X86ISD::FSRL";
7647 case X86ISD::FILD: return "X86ISD::FILD";
7648 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7649 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7650 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7651 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7652 case X86ISD::FLD: return "X86ISD::FLD";
7653 case X86ISD::FST: return "X86ISD::FST";
7654 case X86ISD::CALL: return "X86ISD::CALL";
7655 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7656 case X86ISD::BT: return "X86ISD::BT";
7657 case X86ISD::CMP: return "X86ISD::CMP";
7658 case X86ISD::COMI: return "X86ISD::COMI";
7659 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7660 case X86ISD::SETCC: return "X86ISD::SETCC";
7661 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7662 case X86ISD::CMOV: return "X86ISD::CMOV";
7663 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7664 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7665 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7666 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7667 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7668 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7669 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7670 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7671 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7672 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7673 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7674 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7675 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
7676 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7677 case X86ISD::FMAX: return "X86ISD::FMAX";
7678 case X86ISD::FMIN: return "X86ISD::FMIN";
7679 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7680 case X86ISD::FRCP: return "X86ISD::FRCP";
7681 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7682 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7683 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7684 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7685 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7686 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7687 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7688 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7689 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7690 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7691 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7692 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7693 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7694 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7695 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7696 case X86ISD::VSHL: return "X86ISD::VSHL";
7697 case X86ISD::VSRL: return "X86ISD::VSRL";
7698 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7699 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7700 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7701 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7702 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7703 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7704 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7705 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7706 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7707 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7708 case X86ISD::ADD: return "X86ISD::ADD";
7709 case X86ISD::SUB: return "X86ISD::SUB";
7710 case X86ISD::SMUL: return "X86ISD::SMUL";
7711 case X86ISD::UMUL: return "X86ISD::UMUL";
7712 case X86ISD::INC: return "X86ISD::INC";
7713 case X86ISD::DEC: return "X86ISD::DEC";
7714 case X86ISD::OR: return "X86ISD::OR";
7715 case X86ISD::XOR: return "X86ISD::XOR";
7716 case X86ISD::AND: return "X86ISD::AND";
7717 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7718 case X86ISD::PTEST: return "X86ISD::PTEST";
7719 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7723 // isLegalAddressingMode - Return true if the addressing mode represented
7724 // by AM is legal for this target, for a load/store of the specified type.
7725 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7726 const Type *Ty) const {
7727 // X86 supports extremely general addressing modes.
7728 CodeModel::Model M = getTargetMachine().getCodeModel();
7730 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7731 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7736 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7738 // If a reference to this global requires an extra load, we can't fold it.
7739 if (isGlobalStubReference(GVFlags))
7742 // If BaseGV requires a register for the PIC base, we cannot also have a
7743 // BaseReg specified.
7744 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7747 // If lower 4G is not available, then we must use rip-relative addressing.
7748 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7758 // These scales always work.
7763 // These scales are formed with basereg+scalereg. Only accept if there is
7768 default: // Other stuff never works.
7776 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7777 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7779 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7780 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7781 if (NumBits1 <= NumBits2)
7786 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7787 if (!VT1.isInteger() || !VT2.isInteger())
7789 unsigned NumBits1 = VT1.getSizeInBits();
7790 unsigned NumBits2 = VT2.getSizeInBits();
7791 if (NumBits1 <= NumBits2)
7796 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7797 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7798 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7801 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7802 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7803 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7806 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7807 // i16 instructions are longer (0x66 prefix) and potentially slower.
7808 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7811 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7812 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7813 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7814 /// are assumed to be legal.
7816 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7818 // Only do shuffles on 128-bit vector types for now.
7819 if (VT.getSizeInBits() == 64)
7822 // FIXME: pshufb, blends, shifts.
7823 return (VT.getVectorNumElements() == 2 ||
7824 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7825 isMOVLMask(M, VT) ||
7826 isSHUFPMask(M, VT) ||
7827 isPSHUFDMask(M, VT) ||
7828 isPSHUFHWMask(M, VT) ||
7829 isPSHUFLWMask(M, VT) ||
7830 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7831 isUNPCKLMask(M, VT) ||
7832 isUNPCKHMask(M, VT) ||
7833 isUNPCKL_v_undef_Mask(M, VT) ||
7834 isUNPCKH_v_undef_Mask(M, VT));
7838 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7840 unsigned NumElts = VT.getVectorNumElements();
7841 // FIXME: This collection of masks seems suspect.
7844 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7845 return (isMOVLMask(Mask, VT) ||
7846 isCommutedMOVLMask(Mask, VT, true) ||
7847 isSHUFPMask(Mask, VT) ||
7848 isCommutedSHUFPMask(Mask, VT));
7853 //===----------------------------------------------------------------------===//
7854 // X86 Scheduler Hooks
7855 //===----------------------------------------------------------------------===//
7857 // private utility function
7859 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7860 MachineBasicBlock *MBB,
7868 TargetRegisterClass *RC,
7869 bool invSrc) const {
7870 // For the atomic bitwise operator, we generate
7873 // ld t1 = [bitinstr.addr]
7874 // op t2 = t1, [bitinstr.val]
7876 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7878 // fallthrough -->nextMBB
7879 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7880 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7881 MachineFunction::iterator MBBIter = MBB;
7884 /// First build the CFG
7885 MachineFunction *F = MBB->getParent();
7886 MachineBasicBlock *thisMBB = MBB;
7887 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7888 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7889 F->insert(MBBIter, newMBB);
7890 F->insert(MBBIter, nextMBB);
7892 // Move all successors to thisMBB to nextMBB
7893 nextMBB->transferSuccessors(thisMBB);
7895 // Update thisMBB to fall through to newMBB
7896 thisMBB->addSuccessor(newMBB);
7898 // newMBB jumps to itself and fall through to nextMBB
7899 newMBB->addSuccessor(nextMBB);
7900 newMBB->addSuccessor(newMBB);
7902 // Insert instructions into newMBB based on incoming instruction
7903 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7904 "unexpected number of operands");
7905 DebugLoc dl = bInstr->getDebugLoc();
7906 MachineOperand& destOper = bInstr->getOperand(0);
7907 MachineOperand* argOpers[2 + X86AddrNumOperands];
7908 int numArgs = bInstr->getNumOperands() - 1;
7909 for (int i=0; i < numArgs; ++i)
7910 argOpers[i] = &bInstr->getOperand(i+1);
7912 // x86 address has 4 operands: base, index, scale, and displacement
7913 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7914 int valArgIndx = lastAddrIndx + 1;
7916 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7917 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7918 for (int i=0; i <= lastAddrIndx; ++i)
7919 (*MIB).addOperand(*argOpers[i]);
7921 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7923 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7928 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7929 assert((argOpers[valArgIndx]->isReg() ||
7930 argOpers[valArgIndx]->isImm()) &&
7932 if (argOpers[valArgIndx]->isReg())
7933 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7935 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7937 (*MIB).addOperand(*argOpers[valArgIndx]);
7939 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7942 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7943 for (int i=0; i <= lastAddrIndx; ++i)
7944 (*MIB).addOperand(*argOpers[i]);
7946 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7947 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7948 bInstr->memoperands_end());
7950 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7954 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
7956 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7960 // private utility function: 64 bit atomics on 32 bit host.
7962 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7963 MachineBasicBlock *MBB,
7968 bool invSrc) const {
7969 // For the atomic bitwise operator, we generate
7970 // thisMBB (instructions are in pairs, except cmpxchg8b)
7971 // ld t1,t2 = [bitinstr.addr]
7973 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7974 // op t5, t6 <- out1, out2, [bitinstr.val]
7975 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7976 // mov ECX, EBX <- t5, t6
7977 // mov EAX, EDX <- t1, t2
7978 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7979 // mov t3, t4 <- EAX, EDX
7981 // result in out1, out2
7982 // fallthrough -->nextMBB
7984 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7985 const unsigned LoadOpc = X86::MOV32rm;
7986 const unsigned copyOpc = X86::MOV32rr;
7987 const unsigned NotOpc = X86::NOT32r;
7988 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7989 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7990 MachineFunction::iterator MBBIter = MBB;
7993 /// First build the CFG
7994 MachineFunction *F = MBB->getParent();
7995 MachineBasicBlock *thisMBB = MBB;
7996 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7997 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7998 F->insert(MBBIter, newMBB);
7999 F->insert(MBBIter, nextMBB);
8001 // Move all successors to thisMBB to nextMBB
8002 nextMBB->transferSuccessors(thisMBB);
8004 // Update thisMBB to fall through to newMBB
8005 thisMBB->addSuccessor(newMBB);
8007 // newMBB jumps to itself and fall through to nextMBB
8008 newMBB->addSuccessor(nextMBB);
8009 newMBB->addSuccessor(newMBB);
8011 DebugLoc dl = bInstr->getDebugLoc();
8012 // Insert instructions into newMBB based on incoming instruction
8013 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8014 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
8015 "unexpected number of operands");
8016 MachineOperand& dest1Oper = bInstr->getOperand(0);
8017 MachineOperand& dest2Oper = bInstr->getOperand(1);
8018 MachineOperand* argOpers[2 + X86AddrNumOperands];
8019 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
8020 argOpers[i] = &bInstr->getOperand(i+2);
8022 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8023 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8025 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8026 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8027 for (int i=0; i <= lastAddrIndx; ++i)
8028 (*MIB).addOperand(*argOpers[i]);
8029 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8030 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8031 // add 4 to displacement.
8032 for (int i=0; i <= lastAddrIndx-2; ++i)
8033 (*MIB).addOperand(*argOpers[i]);
8034 MachineOperand newOp3 = *(argOpers[3]);
8036 newOp3.setImm(newOp3.getImm()+4);
8038 newOp3.setOffset(newOp3.getOffset()+4);
8039 (*MIB).addOperand(newOp3);
8040 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8042 // t3/4 are defined later, at the bottom of the loop
8043 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8044 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8045 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8046 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8047 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8048 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8050 // The subsequent operations should be using the destination registers of
8051 //the PHI instructions.
8053 t1 = F->getRegInfo().createVirtualRegister(RC);
8054 t2 = F->getRegInfo().createVirtualRegister(RC);
8055 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8056 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8058 t1 = dest1Oper.getReg();
8059 t2 = dest2Oper.getReg();
8062 int valArgIndx = lastAddrIndx + 1;
8063 assert((argOpers[valArgIndx]->isReg() ||
8064 argOpers[valArgIndx]->isImm()) &&
8066 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8067 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8068 if (argOpers[valArgIndx]->isReg())
8069 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8071 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8072 if (regOpcL != X86::MOV32rr)
8074 (*MIB).addOperand(*argOpers[valArgIndx]);
8075 assert(argOpers[valArgIndx + 1]->isReg() ==
8076 argOpers[valArgIndx]->isReg());
8077 assert(argOpers[valArgIndx + 1]->isImm() ==
8078 argOpers[valArgIndx]->isImm());
8079 if (argOpers[valArgIndx + 1]->isReg())
8080 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8082 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8083 if (regOpcH != X86::MOV32rr)
8085 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8087 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8089 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8092 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8094 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8097 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8098 for (int i=0; i <= lastAddrIndx; ++i)
8099 (*MIB).addOperand(*argOpers[i]);
8101 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8102 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8103 bInstr->memoperands_end());
8105 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8106 MIB.addReg(X86::EAX);
8107 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8108 MIB.addReg(X86::EDX);
8111 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8113 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8117 // private utility function
8119 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8120 MachineBasicBlock *MBB,
8121 unsigned cmovOpc) const {
8122 // For the atomic min/max operator, we generate
8125 // ld t1 = [min/max.addr]
8126 // mov t2 = [min/max.val]
8128 // cmov[cond] t2 = t1
8130 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8132 // fallthrough -->nextMBB
8134 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8135 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8136 MachineFunction::iterator MBBIter = MBB;
8139 /// First build the CFG
8140 MachineFunction *F = MBB->getParent();
8141 MachineBasicBlock *thisMBB = MBB;
8142 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8143 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8144 F->insert(MBBIter, newMBB);
8145 F->insert(MBBIter, nextMBB);
8147 // Move all successors of thisMBB to nextMBB
8148 nextMBB->transferSuccessors(thisMBB);
8150 // Update thisMBB to fall through to newMBB
8151 thisMBB->addSuccessor(newMBB);
8153 // newMBB jumps to newMBB and fall through to nextMBB
8154 newMBB->addSuccessor(nextMBB);
8155 newMBB->addSuccessor(newMBB);
8157 DebugLoc dl = mInstr->getDebugLoc();
8158 // Insert instructions into newMBB based on incoming instruction
8159 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8160 "unexpected number of operands");
8161 MachineOperand& destOper = mInstr->getOperand(0);
8162 MachineOperand* argOpers[2 + X86AddrNumOperands];
8163 int numArgs = mInstr->getNumOperands() - 1;
8164 for (int i=0; i < numArgs; ++i)
8165 argOpers[i] = &mInstr->getOperand(i+1);
8167 // x86 address has 4 operands: base, index, scale, and displacement
8168 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8169 int valArgIndx = lastAddrIndx + 1;
8171 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8172 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8173 for (int i=0; i <= lastAddrIndx; ++i)
8174 (*MIB).addOperand(*argOpers[i]);
8176 // We only support register and immediate values
8177 assert((argOpers[valArgIndx]->isReg() ||
8178 argOpers[valArgIndx]->isImm()) &&
8181 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8182 if (argOpers[valArgIndx]->isReg())
8183 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8185 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8186 (*MIB).addOperand(*argOpers[valArgIndx]);
8188 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8191 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8196 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8197 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8201 // Cmp and exchange if none has modified the memory location
8202 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8203 for (int i=0; i <= lastAddrIndx; ++i)
8204 (*MIB).addOperand(*argOpers[i]);
8206 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8207 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8208 mInstr->memoperands_end());
8210 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8211 MIB.addReg(X86::EAX);
8214 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8216 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8220 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8221 // all of this code can be replaced with that in the .td file.
8223 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8224 unsigned numArgs, bool memArg) const {
8226 MachineFunction *F = BB->getParent();
8227 DebugLoc dl = MI->getDebugLoc();
8228 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8232 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8234 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8236 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8238 for (unsigned i = 0; i < numArgs; ++i) {
8239 MachineOperand &Op = MI->getOperand(i+1);
8241 if (!(Op.isReg() && Op.isImplicit()))
8245 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8248 F->DeleteMachineInstr(MI);
8254 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8256 MachineBasicBlock *MBB) const {
8257 // Emit code to save XMM registers to the stack. The ABI says that the
8258 // number of registers to save is given in %al, so it's theoretically
8259 // possible to do an indirect jump trick to avoid saving all of them,
8260 // however this code takes a simpler approach and just executes all
8261 // of the stores if %al is non-zero. It's less code, and it's probably
8262 // easier on the hardware branch predictor, and stores aren't all that
8263 // expensive anyway.
8265 // Create the new basic blocks. One block contains all the XMM stores,
8266 // and one block is the final destination regardless of whether any
8267 // stores were performed.
8268 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8269 MachineFunction *F = MBB->getParent();
8270 MachineFunction::iterator MBBIter = MBB;
8272 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8273 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8274 F->insert(MBBIter, XMMSaveMBB);
8275 F->insert(MBBIter, EndMBB);
8278 // Move any original successors of MBB to the end block.
8279 EndMBB->transferSuccessors(MBB);
8280 // The original block will now fall through to the XMM save block.
8281 MBB->addSuccessor(XMMSaveMBB);
8282 // The XMMSaveMBB will fall through to the end block.
8283 XMMSaveMBB->addSuccessor(EndMBB);
8285 // Now add the instructions.
8286 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8287 DebugLoc DL = MI->getDebugLoc();
8289 unsigned CountReg = MI->getOperand(0).getReg();
8290 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8291 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8293 if (!Subtarget->isTargetWin64()) {
8294 // If %al is 0, branch around the XMM save block.
8295 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8296 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8297 MBB->addSuccessor(EndMBB);
8300 // In the XMM save block, save all the XMM argument registers.
8301 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8302 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8303 MachineMemOperand *MMO =
8304 F->getMachineMemOperand(
8305 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8306 MachineMemOperand::MOStore, Offset,
8307 /*Size=*/16, /*Align=*/16);
8308 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8309 .addFrameIndex(RegSaveFrameIndex)
8310 .addImm(/*Scale=*/1)
8311 .addReg(/*IndexReg=*/0)
8312 .addImm(/*Disp=*/Offset)
8313 .addReg(/*Segment=*/0)
8314 .addReg(MI->getOperand(i).getReg())
8315 .addMemOperand(MMO);
8318 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8324 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8325 MachineBasicBlock *BB,
8326 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8327 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8328 DebugLoc DL = MI->getDebugLoc();
8330 // To "insert" a SELECT_CC instruction, we actually have to insert the
8331 // diamond control-flow pattern. The incoming instruction knows the
8332 // destination vreg to set, the condition code register to branch on, the
8333 // true/false values to select between, and a branch opcode to use.
8334 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8335 MachineFunction::iterator It = BB;
8341 // cmpTY ccX, r1, r2
8343 // fallthrough --> copy0MBB
8344 MachineBasicBlock *thisMBB = BB;
8345 MachineFunction *F = BB->getParent();
8346 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8347 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8349 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8350 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8351 F->insert(It, copy0MBB);
8352 F->insert(It, sinkMBB);
8353 // Update machine-CFG edges by first adding all successors of the current
8354 // block to the new block which will contain the Phi node for the select.
8355 // Also inform sdisel of the edge changes.
8356 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8357 E = BB->succ_end(); I != E; ++I) {
8358 EM->insert(std::make_pair(*I, sinkMBB));
8359 sinkMBB->addSuccessor(*I);
8361 // Next, remove all successors of the current block, and add the true
8362 // and fallthrough blocks as its successors.
8363 while (!BB->succ_empty())
8364 BB->removeSuccessor(BB->succ_begin());
8365 // Add the true and fallthrough blocks as its successors.
8366 BB->addSuccessor(copy0MBB);
8367 BB->addSuccessor(sinkMBB);
8370 // %FalseValue = ...
8371 // # fallthrough to sinkMBB
8374 // Update machine-CFG edges
8375 BB->addSuccessor(sinkMBB);
8378 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8381 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8382 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8383 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8385 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8391 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8392 MachineBasicBlock *BB,
8393 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8394 switch (MI->getOpcode()) {
8395 default: assert(false && "Unexpected instr type to insert");
8397 case X86::CMOV_V1I64:
8398 case X86::CMOV_FR32:
8399 case X86::CMOV_FR64:
8400 case X86::CMOV_V4F32:
8401 case X86::CMOV_V2F64:
8402 case X86::CMOV_V2I64:
8403 return EmitLoweredSelect(MI, BB, EM);
8405 case X86::FP32_TO_INT16_IN_MEM:
8406 case X86::FP32_TO_INT32_IN_MEM:
8407 case X86::FP32_TO_INT64_IN_MEM:
8408 case X86::FP64_TO_INT16_IN_MEM:
8409 case X86::FP64_TO_INT32_IN_MEM:
8410 case X86::FP64_TO_INT64_IN_MEM:
8411 case X86::FP80_TO_INT16_IN_MEM:
8412 case X86::FP80_TO_INT32_IN_MEM:
8413 case X86::FP80_TO_INT64_IN_MEM: {
8414 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8415 DebugLoc DL = MI->getDebugLoc();
8417 // Change the floating point control register to use "round towards zero"
8418 // mode when truncating to an integer value.
8419 MachineFunction *F = BB->getParent();
8420 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8421 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8423 // Load the old value of the high byte of the control word...
8425 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8426 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8429 // Set the high part to be round to zero...
8430 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8433 // Reload the modified control word now...
8434 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8436 // Restore the memory image of control word to original value
8437 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8440 // Get the X86 opcode to use.
8442 switch (MI->getOpcode()) {
8443 default: llvm_unreachable("illegal opcode!");
8444 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8445 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8446 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8447 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8448 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8449 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8450 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8451 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8452 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8456 MachineOperand &Op = MI->getOperand(0);
8458 AM.BaseType = X86AddressMode::RegBase;
8459 AM.Base.Reg = Op.getReg();
8461 AM.BaseType = X86AddressMode::FrameIndexBase;
8462 AM.Base.FrameIndex = Op.getIndex();
8464 Op = MI->getOperand(1);
8466 AM.Scale = Op.getImm();
8467 Op = MI->getOperand(2);
8469 AM.IndexReg = Op.getImm();
8470 Op = MI->getOperand(3);
8471 if (Op.isGlobal()) {
8472 AM.GV = Op.getGlobal();
8474 AM.Disp = Op.getImm();
8476 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8477 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8479 // Reload the original control word now.
8480 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8482 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8485 // String/text processing lowering.
8486 case X86::PCMPISTRM128REG:
8487 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8488 case X86::PCMPISTRM128MEM:
8489 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8490 case X86::PCMPESTRM128REG:
8491 return EmitPCMP(MI, BB, 5, false /* in mem */);
8492 case X86::PCMPESTRM128MEM:
8493 return EmitPCMP(MI, BB, 5, true /* in mem */);
8496 case X86::ATOMAND32:
8497 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8498 X86::AND32ri, X86::MOV32rm,
8499 X86::LCMPXCHG32, X86::MOV32rr,
8500 X86::NOT32r, X86::EAX,
8501 X86::GR32RegisterClass);
8503 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8504 X86::OR32ri, X86::MOV32rm,
8505 X86::LCMPXCHG32, X86::MOV32rr,
8506 X86::NOT32r, X86::EAX,
8507 X86::GR32RegisterClass);
8508 case X86::ATOMXOR32:
8509 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8510 X86::XOR32ri, X86::MOV32rm,
8511 X86::LCMPXCHG32, X86::MOV32rr,
8512 X86::NOT32r, X86::EAX,
8513 X86::GR32RegisterClass);
8514 case X86::ATOMNAND32:
8515 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8516 X86::AND32ri, X86::MOV32rm,
8517 X86::LCMPXCHG32, X86::MOV32rr,
8518 X86::NOT32r, X86::EAX,
8519 X86::GR32RegisterClass, true);
8520 case X86::ATOMMIN32:
8521 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8522 case X86::ATOMMAX32:
8523 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8524 case X86::ATOMUMIN32:
8525 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8526 case X86::ATOMUMAX32:
8527 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8529 case X86::ATOMAND16:
8530 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8531 X86::AND16ri, X86::MOV16rm,
8532 X86::LCMPXCHG16, X86::MOV16rr,
8533 X86::NOT16r, X86::AX,
8534 X86::GR16RegisterClass);
8536 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8537 X86::OR16ri, X86::MOV16rm,
8538 X86::LCMPXCHG16, X86::MOV16rr,
8539 X86::NOT16r, X86::AX,
8540 X86::GR16RegisterClass);
8541 case X86::ATOMXOR16:
8542 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8543 X86::XOR16ri, X86::MOV16rm,
8544 X86::LCMPXCHG16, X86::MOV16rr,
8545 X86::NOT16r, X86::AX,
8546 X86::GR16RegisterClass);
8547 case X86::ATOMNAND16:
8548 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8549 X86::AND16ri, X86::MOV16rm,
8550 X86::LCMPXCHG16, X86::MOV16rr,
8551 X86::NOT16r, X86::AX,
8552 X86::GR16RegisterClass, true);
8553 case X86::ATOMMIN16:
8554 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8555 case X86::ATOMMAX16:
8556 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8557 case X86::ATOMUMIN16:
8558 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8559 case X86::ATOMUMAX16:
8560 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8563 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8564 X86::AND8ri, X86::MOV8rm,
8565 X86::LCMPXCHG8, X86::MOV8rr,
8566 X86::NOT8r, X86::AL,
8567 X86::GR8RegisterClass);
8569 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8570 X86::OR8ri, X86::MOV8rm,
8571 X86::LCMPXCHG8, X86::MOV8rr,
8572 X86::NOT8r, X86::AL,
8573 X86::GR8RegisterClass);
8575 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8576 X86::XOR8ri, X86::MOV8rm,
8577 X86::LCMPXCHG8, X86::MOV8rr,
8578 X86::NOT8r, X86::AL,
8579 X86::GR8RegisterClass);
8580 case X86::ATOMNAND8:
8581 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8582 X86::AND8ri, X86::MOV8rm,
8583 X86::LCMPXCHG8, X86::MOV8rr,
8584 X86::NOT8r, X86::AL,
8585 X86::GR8RegisterClass, true);
8586 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8587 // This group is for 64-bit host.
8588 case X86::ATOMAND64:
8589 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8590 X86::AND64ri32, X86::MOV64rm,
8591 X86::LCMPXCHG64, X86::MOV64rr,
8592 X86::NOT64r, X86::RAX,
8593 X86::GR64RegisterClass);
8595 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8596 X86::OR64ri32, X86::MOV64rm,
8597 X86::LCMPXCHG64, X86::MOV64rr,
8598 X86::NOT64r, X86::RAX,
8599 X86::GR64RegisterClass);
8600 case X86::ATOMXOR64:
8601 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8602 X86::XOR64ri32, X86::MOV64rm,
8603 X86::LCMPXCHG64, X86::MOV64rr,
8604 X86::NOT64r, X86::RAX,
8605 X86::GR64RegisterClass);
8606 case X86::ATOMNAND64:
8607 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8608 X86::AND64ri32, X86::MOV64rm,
8609 X86::LCMPXCHG64, X86::MOV64rr,
8610 X86::NOT64r, X86::RAX,
8611 X86::GR64RegisterClass, true);
8612 case X86::ATOMMIN64:
8613 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8614 case X86::ATOMMAX64:
8615 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8616 case X86::ATOMUMIN64:
8617 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8618 case X86::ATOMUMAX64:
8619 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8621 // This group does 64-bit operations on a 32-bit host.
8622 case X86::ATOMAND6432:
8623 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8624 X86::AND32rr, X86::AND32rr,
8625 X86::AND32ri, X86::AND32ri,
8627 case X86::ATOMOR6432:
8628 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8629 X86::OR32rr, X86::OR32rr,
8630 X86::OR32ri, X86::OR32ri,
8632 case X86::ATOMXOR6432:
8633 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8634 X86::XOR32rr, X86::XOR32rr,
8635 X86::XOR32ri, X86::XOR32ri,
8637 case X86::ATOMNAND6432:
8638 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8639 X86::AND32rr, X86::AND32rr,
8640 X86::AND32ri, X86::AND32ri,
8642 case X86::ATOMADD6432:
8643 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8644 X86::ADD32rr, X86::ADC32rr,
8645 X86::ADD32ri, X86::ADC32ri,
8647 case X86::ATOMSUB6432:
8648 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8649 X86::SUB32rr, X86::SBB32rr,
8650 X86::SUB32ri, X86::SBB32ri,
8652 case X86::ATOMSWAP6432:
8653 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8654 X86::MOV32rr, X86::MOV32rr,
8655 X86::MOV32ri, X86::MOV32ri,
8657 case X86::VASTART_SAVE_XMM_REGS:
8658 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8662 //===----------------------------------------------------------------------===//
8663 // X86 Optimization Hooks
8664 //===----------------------------------------------------------------------===//
8666 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8670 const SelectionDAG &DAG,
8671 unsigned Depth) const {
8672 unsigned Opc = Op.getOpcode();
8673 assert((Opc >= ISD::BUILTIN_OP_END ||
8674 Opc == ISD::INTRINSIC_WO_CHAIN ||
8675 Opc == ISD::INTRINSIC_W_CHAIN ||
8676 Opc == ISD::INTRINSIC_VOID) &&
8677 "Should use MaskedValueIsZero if you don't know whether Op"
8678 " is a target node!");
8680 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8692 // These nodes' second result is a boolean.
8693 if (Op.getResNo() == 0)
8697 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8698 Mask.getBitWidth() - 1);
8703 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8704 /// node is a GlobalAddress + offset.
8705 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8706 GlobalValue* &GA, int64_t &Offset) const{
8707 if (N->getOpcode() == X86ISD::Wrapper) {
8708 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8709 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8710 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8714 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8717 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8718 EVT EltVT, LoadSDNode *&LDBase,
8719 unsigned &LastLoadedElt,
8720 SelectionDAG &DAG, MachineFrameInfo *MFI,
8721 const TargetLowering &TLI) {
8723 LastLoadedElt = -1U;
8724 for (unsigned i = 0; i < NumElems; ++i) {
8725 if (N->getMaskElt(i) < 0) {
8731 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8732 if (!Elt.getNode() ||
8733 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8736 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8738 LDBase = cast<LoadSDNode>(Elt.getNode());
8742 if (Elt.getOpcode() == ISD::UNDEF)
8745 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8746 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
8753 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8754 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8755 /// if the load addresses are consecutive, non-overlapping, and in the right
8756 /// order. In the case of v2i64, it will see if it can rewrite the
8757 /// shuffle to be an appropriate build vector so it can take advantage of
8758 // performBuildVectorCombine.
8759 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8760 const TargetLowering &TLI) {
8761 DebugLoc dl = N->getDebugLoc();
8762 EVT VT = N->getValueType(0);
8763 EVT EltVT = VT.getVectorElementType();
8764 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8765 unsigned NumElems = VT.getVectorNumElements();
8767 if (VT.getSizeInBits() != 128)
8770 // Try to combine a vector_shuffle into a 128-bit load.
8771 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8772 LoadSDNode *LD = NULL;
8773 unsigned LastLoadedElt;
8774 if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
8778 if (LastLoadedElt == NumElems - 1) {
8779 if (DAG.InferPtrAlignment(LD->getBasePtr()) >= 16)
8780 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8781 LD->getSrcValue(), LD->getSrcValueOffset(),
8782 LD->isVolatile(), LD->isNonTemporal(), 0);
8783 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8784 LD->getSrcValue(), LD->getSrcValueOffset(),
8785 LD->isVolatile(), LD->isNonTemporal(),
8786 LD->getAlignment());
8787 } else if (NumElems == 4 && LastLoadedElt == 1) {
8788 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8789 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8790 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8791 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8796 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8797 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8798 const X86Subtarget *Subtarget) {
8799 DebugLoc DL = N->getDebugLoc();
8800 SDValue Cond = N->getOperand(0);
8801 // Get the LHS/RHS of the select.
8802 SDValue LHS = N->getOperand(1);
8803 SDValue RHS = N->getOperand(2);
8805 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8806 // instructions match the semantics of the common C idiom x<y?x:y but not
8807 // x<=y?x:y, because of how they handle negative zero (which can be
8808 // ignored in unsafe-math mode).
8809 if (Subtarget->hasSSE2() &&
8810 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8811 Cond.getOpcode() == ISD::SETCC) {
8812 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8814 unsigned Opcode = 0;
8815 // Check for x CC y ? x : y.
8816 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
8817 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
8821 // Converting this to a min would handle NaNs incorrectly, and swapping
8822 // the operands would cause it to handle comparisons between positive
8823 // and negative zero incorrectly.
8824 if (!FiniteOnlyFPMath() &&
8825 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
8826 if (!UnsafeFPMath &&
8827 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8829 std::swap(LHS, RHS);
8831 Opcode = X86ISD::FMIN;
8834 // Converting this to a min would handle comparisons between positive
8835 // and negative zero incorrectly.
8836 if (!UnsafeFPMath &&
8837 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
8839 Opcode = X86ISD::FMIN;
8842 // Converting this to a min would handle both negative zeros and NaNs
8843 // incorrectly, but we can swap the operands to fix both.
8844 std::swap(LHS, RHS);
8848 Opcode = X86ISD::FMIN;
8852 // Converting this to a max would handle comparisons between positive
8853 // and negative zero incorrectly.
8854 if (!UnsafeFPMath &&
8855 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
8857 Opcode = X86ISD::FMAX;
8860 // Converting this to a max would handle NaNs incorrectly, and swapping
8861 // the operands would cause it to handle comparisons between positive
8862 // and negative zero incorrectly.
8863 if (!FiniteOnlyFPMath() &&
8864 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
8865 if (!UnsafeFPMath &&
8866 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8868 std::swap(LHS, RHS);
8870 Opcode = X86ISD::FMAX;
8873 // Converting this to a max would handle both negative zeros and NaNs
8874 // incorrectly, but we can swap the operands to fix both.
8875 std::swap(LHS, RHS);
8879 Opcode = X86ISD::FMAX;
8882 // Check for x CC y ? y : x -- a min/max with reversed arms.
8883 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
8884 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
8888 // Converting this to a min would handle comparisons between positive
8889 // and negative zero incorrectly, and swapping the operands would
8890 // cause it to handle NaNs incorrectly.
8891 if (!UnsafeFPMath &&
8892 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
8893 if (!FiniteOnlyFPMath() &&
8894 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
8896 std::swap(LHS, RHS);
8898 Opcode = X86ISD::FMIN;
8901 // Converting this to a min would handle NaNs incorrectly.
8902 if (!UnsafeFPMath &&
8903 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
8905 Opcode = X86ISD::FMIN;
8908 // Converting this to a min would handle both negative zeros and NaNs
8909 // incorrectly, but we can swap the operands to fix both.
8910 std::swap(LHS, RHS);
8914 Opcode = X86ISD::FMIN;
8918 // Converting this to a max would handle NaNs incorrectly.
8919 if (!FiniteOnlyFPMath() &&
8920 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
8922 Opcode = X86ISD::FMAX;
8925 // Converting this to a max would handle comparisons between positive
8926 // and negative zero incorrectly, and swapping the operands would
8927 // cause it to handle NaNs incorrectly.
8928 if (!UnsafeFPMath &&
8929 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
8930 if (!FiniteOnlyFPMath() &&
8931 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
8933 std::swap(LHS, RHS);
8935 Opcode = X86ISD::FMAX;
8938 // Converting this to a max would handle both negative zeros and NaNs
8939 // incorrectly, but we can swap the operands to fix both.
8940 std::swap(LHS, RHS);
8944 Opcode = X86ISD::FMAX;
8950 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8953 // If this is a select between two integer constants, try to do some
8955 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8956 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8957 // Don't do this for crazy integer types.
8958 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8959 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8960 // so that TrueC (the true value) is larger than FalseC.
8961 bool NeedsCondInvert = false;
8963 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8964 // Efficiently invertible.
8965 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8966 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8967 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8968 NeedsCondInvert = true;
8969 std::swap(TrueC, FalseC);
8972 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8973 if (FalseC->getAPIntValue() == 0 &&
8974 TrueC->getAPIntValue().isPowerOf2()) {
8975 if (NeedsCondInvert) // Invert the condition if needed.
8976 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8977 DAG.getConstant(1, Cond.getValueType()));
8979 // Zero extend the condition if needed.
8980 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8982 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8983 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8984 DAG.getConstant(ShAmt, MVT::i8));
8987 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8988 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8989 if (NeedsCondInvert) // Invert the condition if needed.
8990 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8991 DAG.getConstant(1, Cond.getValueType()));
8993 // Zero extend the condition if needed.
8994 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8995 FalseC->getValueType(0), Cond);
8996 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8997 SDValue(FalseC, 0));
9000 // Optimize cases that will turn into an LEA instruction. This requires
9001 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9002 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9003 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9004 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9006 bool isFastMultiplier = false;
9008 switch ((unsigned char)Diff) {
9010 case 1: // result = add base, cond
9011 case 2: // result = lea base( , cond*2)
9012 case 3: // result = lea base(cond, cond*2)
9013 case 4: // result = lea base( , cond*4)
9014 case 5: // result = lea base(cond, cond*4)
9015 case 8: // result = lea base( , cond*8)
9016 case 9: // result = lea base(cond, cond*8)
9017 isFastMultiplier = true;
9022 if (isFastMultiplier) {
9023 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9024 if (NeedsCondInvert) // Invert the condition if needed.
9025 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9026 DAG.getConstant(1, Cond.getValueType()));
9028 // Zero extend the condition if needed.
9029 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9031 // Scale the condition by the difference.
9033 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9034 DAG.getConstant(Diff, Cond.getValueType()));
9036 // Add the base if non-zero.
9037 if (FalseC->getAPIntValue() != 0)
9038 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9039 SDValue(FalseC, 0));
9049 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9050 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9051 TargetLowering::DAGCombinerInfo &DCI) {
9052 DebugLoc DL = N->getDebugLoc();
9054 // If the flag operand isn't dead, don't touch this CMOV.
9055 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9058 // If this is a select between two integer constants, try to do some
9059 // optimizations. Note that the operands are ordered the opposite of SELECT
9061 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9062 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9063 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9064 // larger than FalseC (the false value).
9065 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9067 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9068 CC = X86::GetOppositeBranchCondition(CC);
9069 std::swap(TrueC, FalseC);
9072 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9073 // This is efficient for any integer data type (including i8/i16) and
9075 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9076 SDValue Cond = N->getOperand(3);
9077 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9078 DAG.getConstant(CC, MVT::i8), Cond);
9080 // Zero extend the condition if needed.
9081 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9083 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9084 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9085 DAG.getConstant(ShAmt, MVT::i8));
9086 if (N->getNumValues() == 2) // Dead flag value?
9087 return DCI.CombineTo(N, Cond, SDValue());
9091 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9092 // for any integer data type, including i8/i16.
9093 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9094 SDValue Cond = N->getOperand(3);
9095 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9096 DAG.getConstant(CC, MVT::i8), Cond);
9098 // Zero extend the condition if needed.
9099 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9100 FalseC->getValueType(0), Cond);
9101 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9102 SDValue(FalseC, 0));
9104 if (N->getNumValues() == 2) // Dead flag value?
9105 return DCI.CombineTo(N, Cond, SDValue());
9109 // Optimize cases that will turn into an LEA instruction. This requires
9110 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9111 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9112 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9113 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9115 bool isFastMultiplier = false;
9117 switch ((unsigned char)Diff) {
9119 case 1: // result = add base, cond
9120 case 2: // result = lea base( , cond*2)
9121 case 3: // result = lea base(cond, cond*2)
9122 case 4: // result = lea base( , cond*4)
9123 case 5: // result = lea base(cond, cond*4)
9124 case 8: // result = lea base( , cond*8)
9125 case 9: // result = lea base(cond, cond*8)
9126 isFastMultiplier = true;
9131 if (isFastMultiplier) {
9132 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9133 SDValue Cond = N->getOperand(3);
9134 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9135 DAG.getConstant(CC, MVT::i8), Cond);
9136 // Zero extend the condition if needed.
9137 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9139 // Scale the condition by the difference.
9141 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9142 DAG.getConstant(Diff, Cond.getValueType()));
9144 // Add the base if non-zero.
9145 if (FalseC->getAPIntValue() != 0)
9146 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9147 SDValue(FalseC, 0));
9148 if (N->getNumValues() == 2) // Dead flag value?
9149 return DCI.CombineTo(N, Cond, SDValue());
9158 /// PerformANDCombine - Look for SSE and instructions of this form:
9159 /// (and x, (build_vector signbit,signbit,signbit,signbit)). If there
9160 /// exists a use of a build_vector that's the bitwise complement of the mask,
9161 /// then transform the node to
9162 /// (and (xor x, (build_vector -1,-1,-1,-1)), (build_vector ~sb,~sb,~sb,~sb)).
9163 static SDValue PerformANDCombine(SDNode *N, SelectionDAG &DAG,
9164 TargetLowering::DAGCombinerInfo &DCI) {
9165 EVT VT = N->getValueType(0);
9166 if (!VT.isVector() || !VT.isInteger())
9169 SDValue N0 = N->getOperand(0);
9170 SDValue N1 = N->getOperand(1);
9171 if (N0.getOpcode() == ISD::XOR || !N1.hasOneUse())
9174 if (N1.getOpcode() == ISD::BUILD_VECTOR) {
9175 unsigned NumElts = VT.getVectorNumElements();
9176 EVT EltVT = VT.getVectorElementType();
9177 SmallVector<SDValue, 8> Mask;
9178 Mask.reserve(NumElts);
9179 for (unsigned i = 0; i != NumElts; ++i) {
9180 SDValue Arg = N1.getOperand(i);
9181 if (Arg.getOpcode() == ISD::UNDEF) {
9182 Mask.push_back(Arg);
9185 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Arg);
9188 if (!C->getAPIntValue().isSignBit() &&
9189 !C->getAPIntValue().isMaxSignedValue())
9191 Mask.push_back(DAG.getConstant(~C->getAPIntValue(), EltVT));
9193 N1 = DAG.getNode(ISD::BUILD_VECTOR, N1.getDebugLoc(), VT,
9195 if (!N1.use_empty()) {
9196 unsigned Bits = EltVT.getSizeInBits();
9198 for (unsigned i = 0; i != NumElts; ++i)
9199 Mask.push_back(DAG.getConstant(APInt::getAllOnesValue(Bits), EltVT));
9200 SDValue NewMask = DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
9201 VT, &Mask[0], NumElts);
9202 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9203 DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
9211 /// PerformMulCombine - Optimize a single multiply with constant into two
9212 /// in order to implement it with two cheaper instructions, e.g.
9213 /// LEA + SHL, LEA + LEA.
9214 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9215 TargetLowering::DAGCombinerInfo &DCI) {
9216 if (DAG.getMachineFunction().
9217 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
9220 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9223 EVT VT = N->getValueType(0);
9227 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9230 uint64_t MulAmt = C->getZExtValue();
9231 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9234 uint64_t MulAmt1 = 0;
9235 uint64_t MulAmt2 = 0;
9236 if ((MulAmt % 9) == 0) {
9238 MulAmt2 = MulAmt / 9;
9239 } else if ((MulAmt % 5) == 0) {
9241 MulAmt2 = MulAmt / 5;
9242 } else if ((MulAmt % 3) == 0) {
9244 MulAmt2 = MulAmt / 3;
9247 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9248 DebugLoc DL = N->getDebugLoc();
9250 if (isPowerOf2_64(MulAmt2) &&
9251 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9252 // If second multiplifer is pow2, issue it first. We want the multiply by
9253 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9255 std::swap(MulAmt1, MulAmt2);
9258 if (isPowerOf2_64(MulAmt1))
9259 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9260 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9262 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9263 DAG.getConstant(MulAmt1, VT));
9265 if (isPowerOf2_64(MulAmt2))
9266 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9267 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9269 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9270 DAG.getConstant(MulAmt2, VT));
9272 // Do not add new nodes to DAG combiner worklist.
9273 DCI.CombineTo(N, NewMul, false);
9278 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9279 SDValue N0 = N->getOperand(0);
9280 SDValue N1 = N->getOperand(1);
9281 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9282 EVT VT = N0.getValueType();
9284 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9285 // since the result of setcc_c is all zero's or all ones.
9286 if (N1C && N0.getOpcode() == ISD::AND &&
9287 N0.getOperand(1).getOpcode() == ISD::Constant) {
9288 SDValue N00 = N0.getOperand(0);
9289 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9290 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9291 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9292 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9293 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9294 APInt ShAmt = N1C->getAPIntValue();
9295 Mask = Mask.shl(ShAmt);
9297 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9298 N00, DAG.getConstant(Mask, VT));
9305 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9307 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9308 const X86Subtarget *Subtarget) {
9309 EVT VT = N->getValueType(0);
9310 if (!VT.isVector() && VT.isInteger() &&
9311 N->getOpcode() == ISD::SHL)
9312 return PerformSHLCombine(N, DAG);
9314 // On X86 with SSE2 support, we can transform this to a vector shift if
9315 // all elements are shifted by the same amount. We can't do this in legalize
9316 // because the a constant vector is typically transformed to a constant pool
9317 // so we have no knowledge of the shift amount.
9318 if (!Subtarget->hasSSE2())
9321 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9324 SDValue ShAmtOp = N->getOperand(1);
9325 EVT EltVT = VT.getVectorElementType();
9326 DebugLoc DL = N->getDebugLoc();
9327 SDValue BaseShAmt = SDValue();
9328 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9329 unsigned NumElts = VT.getVectorNumElements();
9331 for (; i != NumElts; ++i) {
9332 SDValue Arg = ShAmtOp.getOperand(i);
9333 if (Arg.getOpcode() == ISD::UNDEF) continue;
9337 for (; i != NumElts; ++i) {
9338 SDValue Arg = ShAmtOp.getOperand(i);
9339 if (Arg.getOpcode() == ISD::UNDEF) continue;
9340 if (Arg != BaseShAmt) {
9344 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9345 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9346 SDValue InVec = ShAmtOp.getOperand(0);
9347 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9348 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9350 for (; i != NumElts; ++i) {
9351 SDValue Arg = InVec.getOperand(i);
9352 if (Arg.getOpcode() == ISD::UNDEF) continue;
9356 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9357 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9358 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9359 if (C->getZExtValue() == SplatIdx)
9360 BaseShAmt = InVec.getOperand(1);
9363 if (BaseShAmt.getNode() == 0)
9364 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9365 DAG.getIntPtrConstant(0));
9369 // The shift amount is an i32.
9370 if (EltVT.bitsGT(MVT::i32))
9371 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9372 else if (EltVT.bitsLT(MVT::i32))
9373 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9375 // The shift amount is identical so we can do a vector shift.
9376 SDValue ValOp = N->getOperand(0);
9377 switch (N->getOpcode()) {
9379 llvm_unreachable("Unknown shift opcode!");
9382 if (VT == MVT::v2i64)
9383 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9384 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9386 if (VT == MVT::v4i32)
9387 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9388 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9390 if (VT == MVT::v8i16)
9391 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9392 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9396 if (VT == MVT::v4i32)
9397 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9398 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9400 if (VT == MVT::v8i16)
9401 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9402 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9406 if (VT == MVT::v2i64)
9407 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9408 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9410 if (VT == MVT::v4i32)
9411 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9412 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9414 if (VT == MVT::v8i16)
9415 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9416 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9423 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9424 const X86Subtarget *Subtarget) {
9425 EVT VT = N->getValueType(0);
9426 if (VT != MVT::i64 || !Subtarget->is64Bit())
9429 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9430 SDValue N0 = N->getOperand(0);
9431 SDValue N1 = N->getOperand(1);
9432 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9434 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9437 SDValue ShAmt0 = N0.getOperand(1);
9438 if (ShAmt0.getValueType() != MVT::i8)
9440 SDValue ShAmt1 = N1.getOperand(1);
9441 if (ShAmt1.getValueType() != MVT::i8)
9443 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9444 ShAmt0 = ShAmt0.getOperand(0);
9445 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9446 ShAmt1 = ShAmt1.getOperand(0);
9448 DebugLoc DL = N->getDebugLoc();
9449 unsigned Opc = X86ISD::SHLD;
9450 SDValue Op0 = N0.getOperand(0);
9451 SDValue Op1 = N1.getOperand(0);
9452 if (ShAmt0.getOpcode() == ISD::SUB) {
9454 std::swap(Op0, Op1);
9455 std::swap(ShAmt0, ShAmt1);
9458 if (ShAmt1.getOpcode() == ISD::SUB) {
9459 SDValue Sum = ShAmt1.getOperand(0);
9460 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9461 if (SumC->getSExtValue() == 64 &&
9462 ShAmt1.getOperand(1) == ShAmt0)
9463 return DAG.getNode(Opc, DL, VT,
9465 DAG.getNode(ISD::TRUNCATE, DL,
9468 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9469 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9471 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
9472 return DAG.getNode(Opc, DL, VT,
9473 N0.getOperand(0), N1.getOperand(0),
9474 DAG.getNode(ISD::TRUNCATE, DL,
9481 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9482 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9483 const X86Subtarget *Subtarget) {
9484 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9485 // the FP state in cases where an emms may be missing.
9486 // A preferable solution to the general problem is to figure out the right
9487 // places to insert EMMS. This qualifies as a quick hack.
9489 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9490 StoreSDNode *St = cast<StoreSDNode>(N);
9491 EVT VT = St->getValue().getValueType();
9492 if (VT.getSizeInBits() != 64)
9495 const Function *F = DAG.getMachineFunction().getFunction();
9496 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9497 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9498 && Subtarget->hasSSE2();
9499 if ((VT.isVector() ||
9500 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9501 isa<LoadSDNode>(St->getValue()) &&
9502 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9503 St->getChain().hasOneUse() && !St->isVolatile()) {
9504 SDNode* LdVal = St->getValue().getNode();
9506 int TokenFactorIndex = -1;
9507 SmallVector<SDValue, 8> Ops;
9508 SDNode* ChainVal = St->getChain().getNode();
9509 // Must be a store of a load. We currently handle two cases: the load
9510 // is a direct child, and it's under an intervening TokenFactor. It is
9511 // possible to dig deeper under nested TokenFactors.
9512 if (ChainVal == LdVal)
9513 Ld = cast<LoadSDNode>(St->getChain());
9514 else if (St->getValue().hasOneUse() &&
9515 ChainVal->getOpcode() == ISD::TokenFactor) {
9516 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9517 if (ChainVal->getOperand(i).getNode() == LdVal) {
9518 TokenFactorIndex = i;
9519 Ld = cast<LoadSDNode>(St->getValue());
9521 Ops.push_back(ChainVal->getOperand(i));
9525 if (!Ld || !ISD::isNormalLoad(Ld))
9528 // If this is not the MMX case, i.e. we are just turning i64 load/store
9529 // into f64 load/store, avoid the transformation if there are multiple
9530 // uses of the loaded value.
9531 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9534 DebugLoc LdDL = Ld->getDebugLoc();
9535 DebugLoc StDL = N->getDebugLoc();
9536 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9537 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9539 if (Subtarget->is64Bit() || F64IsLegal) {
9540 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9541 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9542 Ld->getBasePtr(), Ld->getSrcValue(),
9543 Ld->getSrcValueOffset(), Ld->isVolatile(),
9544 Ld->isNonTemporal(), Ld->getAlignment());
9545 SDValue NewChain = NewLd.getValue(1);
9546 if (TokenFactorIndex != -1) {
9547 Ops.push_back(NewChain);
9548 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9551 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9552 St->getSrcValue(), St->getSrcValueOffset(),
9553 St->isVolatile(), St->isNonTemporal(),
9554 St->getAlignment());
9557 // Otherwise, lower to two pairs of 32-bit loads / stores.
9558 SDValue LoAddr = Ld->getBasePtr();
9559 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9560 DAG.getConstant(4, MVT::i32));
9562 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9563 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9564 Ld->isVolatile(), Ld->isNonTemporal(),
9565 Ld->getAlignment());
9566 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9567 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9568 Ld->isVolatile(), Ld->isNonTemporal(),
9569 MinAlign(Ld->getAlignment(), 4));
9571 SDValue NewChain = LoLd.getValue(1);
9572 if (TokenFactorIndex != -1) {
9573 Ops.push_back(LoLd);
9574 Ops.push_back(HiLd);
9575 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9579 LoAddr = St->getBasePtr();
9580 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9581 DAG.getConstant(4, MVT::i32));
9583 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9584 St->getSrcValue(), St->getSrcValueOffset(),
9585 St->isVolatile(), St->isNonTemporal(),
9586 St->getAlignment());
9587 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9589 St->getSrcValueOffset() + 4,
9591 St->isNonTemporal(),
9592 MinAlign(St->getAlignment(), 4));
9593 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9598 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9599 /// X86ISD::FXOR nodes.
9600 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9601 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9602 // F[X]OR(0.0, x) -> x
9603 // F[X]OR(x, 0.0) -> x
9604 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9605 if (C->getValueAPF().isPosZero())
9606 return N->getOperand(1);
9607 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9608 if (C->getValueAPF().isPosZero())
9609 return N->getOperand(0);
9613 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9614 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9615 // FAND(0.0, x) -> 0.0
9616 // FAND(x, 0.0) -> 0.0
9617 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9618 if (C->getValueAPF().isPosZero())
9619 return N->getOperand(0);
9620 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9621 if (C->getValueAPF().isPosZero())
9622 return N->getOperand(1);
9626 static SDValue PerformBTCombine(SDNode *N,
9628 TargetLowering::DAGCombinerInfo &DCI) {
9629 // BT ignores high bits in the bit index operand.
9630 SDValue Op1 = N->getOperand(1);
9631 if (Op1.hasOneUse()) {
9632 unsigned BitWidth = Op1.getValueSizeInBits();
9633 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9634 APInt KnownZero, KnownOne;
9635 TargetLowering::TargetLoweringOpt TLO(DAG);
9636 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9637 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9638 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9639 DCI.CommitTargetLoweringOpt(TLO);
9644 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9645 SDValue Op = N->getOperand(0);
9646 if (Op.getOpcode() == ISD::BIT_CONVERT)
9647 Op = Op.getOperand(0);
9648 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9649 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9650 VT.getVectorElementType().getSizeInBits() ==
9651 OpVT.getVectorElementType().getSizeInBits()) {
9652 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9657 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9658 // Locked instructions, in turn, have implicit fence semantics (all memory
9659 // operations are flushed before issuing the locked instruction, and the
9660 // are not buffered), so we can fold away the common pattern of
9661 // fence-atomic-fence.
9662 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9663 SDValue atomic = N->getOperand(0);
9664 switch (atomic.getOpcode()) {
9665 case ISD::ATOMIC_CMP_SWAP:
9666 case ISD::ATOMIC_SWAP:
9667 case ISD::ATOMIC_LOAD_ADD:
9668 case ISD::ATOMIC_LOAD_SUB:
9669 case ISD::ATOMIC_LOAD_AND:
9670 case ISD::ATOMIC_LOAD_OR:
9671 case ISD::ATOMIC_LOAD_XOR:
9672 case ISD::ATOMIC_LOAD_NAND:
9673 case ISD::ATOMIC_LOAD_MIN:
9674 case ISD::ATOMIC_LOAD_MAX:
9675 case ISD::ATOMIC_LOAD_UMIN:
9676 case ISD::ATOMIC_LOAD_UMAX:
9682 SDValue fence = atomic.getOperand(0);
9683 if (fence.getOpcode() != ISD::MEMBARRIER)
9686 switch (atomic.getOpcode()) {
9687 case ISD::ATOMIC_CMP_SWAP:
9688 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9689 atomic.getOperand(1), atomic.getOperand(2),
9690 atomic.getOperand(3));
9691 case ISD::ATOMIC_SWAP:
9692 case ISD::ATOMIC_LOAD_ADD:
9693 case ISD::ATOMIC_LOAD_SUB:
9694 case ISD::ATOMIC_LOAD_AND:
9695 case ISD::ATOMIC_LOAD_OR:
9696 case ISD::ATOMIC_LOAD_XOR:
9697 case ISD::ATOMIC_LOAD_NAND:
9698 case ISD::ATOMIC_LOAD_MIN:
9699 case ISD::ATOMIC_LOAD_MAX:
9700 case ISD::ATOMIC_LOAD_UMIN:
9701 case ISD::ATOMIC_LOAD_UMAX:
9702 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9703 atomic.getOperand(1), atomic.getOperand(2));
9709 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9710 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9711 // (and (i32 x86isd::setcc_carry), 1)
9712 // This eliminates the zext. This transformation is necessary because
9713 // ISD::SETCC is always legalized to i8.
9714 DebugLoc dl = N->getDebugLoc();
9715 SDValue N0 = N->getOperand(0);
9716 EVT VT = N->getValueType(0);
9717 if (N0.getOpcode() == ISD::AND &&
9719 N0.getOperand(0).hasOneUse()) {
9720 SDValue N00 = N0.getOperand(0);
9721 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9723 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9724 if (!C || C->getZExtValue() != 1)
9726 return DAG.getNode(ISD::AND, dl, VT,
9727 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9728 N00.getOperand(0), N00.getOperand(1)),
9729 DAG.getConstant(1, VT));
9735 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9736 DAGCombinerInfo &DCI) const {
9737 SelectionDAG &DAG = DCI.DAG;
9738 switch (N->getOpcode()) {
9740 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9741 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9742 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9743 case ISD::AND: return PerformANDCombine(N, DAG, DCI);
9744 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9747 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9748 case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
9749 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9751 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9752 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9753 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9754 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9755 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9756 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9762 //===----------------------------------------------------------------------===//
9763 // X86 Inline Assembly Support
9764 //===----------------------------------------------------------------------===//
9766 static bool LowerToBSwap(CallInst *CI) {
9767 // FIXME: this should verify that we are targetting a 486 or better. If not,
9768 // we will turn this bswap into something that will be lowered to logical ops
9769 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9770 // so don't worry about this.
9772 // Verify this is a simple bswap.
9773 if (CI->getNumOperands() != 2 ||
9774 CI->getType() != CI->getOperand(1)->getType() ||
9775 !CI->getType()->isIntegerTy())
9778 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9779 if (!Ty || Ty->getBitWidth() % 16 != 0)
9782 // Okay, we can do this xform, do so now.
9783 const Type *Tys[] = { Ty };
9784 Module *M = CI->getParent()->getParent()->getParent();
9785 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9787 Value *Op = CI->getOperand(1);
9788 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9790 CI->replaceAllUsesWith(Op);
9791 CI->eraseFromParent();
9795 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9796 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9797 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9799 std::string AsmStr = IA->getAsmString();
9801 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9802 SmallVector<StringRef, 4> AsmPieces;
9803 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9805 switch (AsmPieces.size()) {
9806 default: return false;
9808 AsmStr = AsmPieces[0];
9810 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9813 if (AsmPieces.size() == 2 &&
9814 (AsmPieces[0] == "bswap" ||
9815 AsmPieces[0] == "bswapq" ||
9816 AsmPieces[0] == "bswapl") &&
9817 (AsmPieces[1] == "$0" ||
9818 AsmPieces[1] == "${0:q}")) {
9819 // No need to check constraints, nothing other than the equivalent of
9820 // "=r,0" would be valid here.
9821 return LowerToBSwap(CI);
9823 // rorw $$8, ${0:w} --> llvm.bswap.i16
9824 if (CI->getType()->isIntegerTy(16) &&
9825 AsmPieces.size() == 3 &&
9826 AsmPieces[0] == "rorw" &&
9827 AsmPieces[1] == "$$8," &&
9828 AsmPieces[2] == "${0:w}" &&
9829 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
9830 return LowerToBSwap(CI);
9834 if (CI->getType()->isIntegerTy(64) &&
9835 Constraints.size() >= 2 &&
9836 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9837 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9838 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9839 SmallVector<StringRef, 4> Words;
9840 SplitString(AsmPieces[0], Words, " \t");
9841 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9843 SplitString(AsmPieces[1], Words, " \t");
9844 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9846 SplitString(AsmPieces[2], Words, " \t,");
9847 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9848 Words[2] == "%edx") {
9849 return LowerToBSwap(CI);
9861 /// getConstraintType - Given a constraint letter, return the type of
9862 /// constraint it is for this target.
9863 X86TargetLowering::ConstraintType
9864 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
9865 if (Constraint.size() == 1) {
9866 switch (Constraint[0]) {
9878 return C_RegisterClass;
9886 return TargetLowering::getConstraintType(Constraint);
9889 /// LowerXConstraint - try to replace an X constraint, which matches anything,
9890 /// with another that has more specific requirements based on the type of the
9891 /// corresponding operand.
9892 const char *X86TargetLowering::
9893 LowerXConstraint(EVT ConstraintVT) const {
9894 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
9895 // 'f' like normal targets.
9896 if (ConstraintVT.isFloatingPoint()) {
9897 if (Subtarget->hasSSE2())
9899 if (Subtarget->hasSSE1())
9903 return TargetLowering::LowerXConstraint(ConstraintVT);
9906 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9907 /// vector. If it is invalid, don't add anything to Ops.
9908 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9911 std::vector<SDValue>&Ops,
9912 SelectionDAG &DAG) const {
9913 SDValue Result(0, 0);
9915 switch (Constraint) {
9918 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9919 if (C->getZExtValue() <= 31) {
9920 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9926 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9927 if (C->getZExtValue() <= 63) {
9928 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9934 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9935 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9936 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9942 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9943 if (C->getZExtValue() <= 255) {
9944 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9950 // 32-bit signed value
9951 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9952 const ConstantInt *CI = C->getConstantIntValue();
9953 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9954 C->getSExtValue())) {
9955 // Widen to 64 bits here to get it sign extended.
9956 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9959 // FIXME gcc accepts some relocatable values here too, but only in certain
9960 // memory models; it's complicated.
9965 // 32-bit unsigned value
9966 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9967 const ConstantInt *CI = C->getConstantIntValue();
9968 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9969 C->getZExtValue())) {
9970 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9974 // FIXME gcc accepts some relocatable values here too, but only in certain
9975 // memory models; it's complicated.
9979 // Literal immediates are always ok.
9980 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
9981 // Widen to 64 bits here to get it sign extended.
9982 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
9986 // If we are in non-pic codegen mode, we allow the address of a global (with
9987 // an optional displacement) to be used with 'i'.
9988 GlobalAddressSDNode *GA = 0;
9991 // Match either (GA), (GA+C), (GA+C1+C2), etc.
9993 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
9994 Offset += GA->getOffset();
9996 } else if (Op.getOpcode() == ISD::ADD) {
9997 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9998 Offset += C->getZExtValue();
9999 Op = Op.getOperand(0);
10002 } else if (Op.getOpcode() == ISD::SUB) {
10003 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10004 Offset += -C->getZExtValue();
10005 Op = Op.getOperand(0);
10010 // Otherwise, this isn't something we can handle, reject it.
10014 GlobalValue *GV = GA->getGlobal();
10015 // If we require an extra load to get this address, as in PIC mode, we
10016 // can't accept it.
10017 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10018 getTargetMachine())))
10022 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
10024 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
10030 if (Result.getNode()) {
10031 Ops.push_back(Result);
10034 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
10038 std::vector<unsigned> X86TargetLowering::
10039 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10041 if (Constraint.size() == 1) {
10042 // FIXME: not handling fp-stack yet!
10043 switch (Constraint[0]) { // GCC X86 Constraint Letters
10044 default: break; // Unknown constraint letter
10045 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10046 if (Subtarget->is64Bit()) {
10047 if (VT == MVT::i32)
10048 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10049 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10050 X86::R10D,X86::R11D,X86::R12D,
10051 X86::R13D,X86::R14D,X86::R15D,
10052 X86::EBP, X86::ESP, 0);
10053 else if (VT == MVT::i16)
10054 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10055 X86::SI, X86::DI, X86::R8W,X86::R9W,
10056 X86::R10W,X86::R11W,X86::R12W,
10057 X86::R13W,X86::R14W,X86::R15W,
10058 X86::BP, X86::SP, 0);
10059 else if (VT == MVT::i8)
10060 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10061 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10062 X86::R10B,X86::R11B,X86::R12B,
10063 X86::R13B,X86::R14B,X86::R15B,
10064 X86::BPL, X86::SPL, 0);
10066 else if (VT == MVT::i64)
10067 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10068 X86::RSI, X86::RDI, X86::R8, X86::R9,
10069 X86::R10, X86::R11, X86::R12,
10070 X86::R13, X86::R14, X86::R15,
10071 X86::RBP, X86::RSP, 0);
10075 // 32-bit fallthrough
10076 case 'Q': // Q_REGS
10077 if (VT == MVT::i32)
10078 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10079 else if (VT == MVT::i16)
10080 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10081 else if (VT == MVT::i8)
10082 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10083 else if (VT == MVT::i64)
10084 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10089 return std::vector<unsigned>();
10092 std::pair<unsigned, const TargetRegisterClass*>
10093 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10095 // First, see if this is a constraint that directly corresponds to an LLVM
10097 if (Constraint.size() == 1) {
10098 // GCC Constraint Letters
10099 switch (Constraint[0]) {
10101 case 'r': // GENERAL_REGS
10102 case 'l': // INDEX_REGS
10104 return std::make_pair(0U, X86::GR8RegisterClass);
10105 if (VT == MVT::i16)
10106 return std::make_pair(0U, X86::GR16RegisterClass);
10107 if (VT == MVT::i32 || !Subtarget->is64Bit())
10108 return std::make_pair(0U, X86::GR32RegisterClass);
10109 return std::make_pair(0U, X86::GR64RegisterClass);
10110 case 'R': // LEGACY_REGS
10112 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10113 if (VT == MVT::i16)
10114 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10115 if (VT == MVT::i32 || !Subtarget->is64Bit())
10116 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10117 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10118 case 'f': // FP Stack registers.
10119 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10120 // value to the correct fpstack register class.
10121 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10122 return std::make_pair(0U, X86::RFP32RegisterClass);
10123 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10124 return std::make_pair(0U, X86::RFP64RegisterClass);
10125 return std::make_pair(0U, X86::RFP80RegisterClass);
10126 case 'y': // MMX_REGS if MMX allowed.
10127 if (!Subtarget->hasMMX()) break;
10128 return std::make_pair(0U, X86::VR64RegisterClass);
10129 case 'Y': // SSE_REGS if SSE2 allowed
10130 if (!Subtarget->hasSSE2()) break;
10132 case 'x': // SSE_REGS if SSE1 allowed
10133 if (!Subtarget->hasSSE1()) break;
10135 switch (VT.getSimpleVT().SimpleTy) {
10137 // Scalar SSE types.
10140 return std::make_pair(0U, X86::FR32RegisterClass);
10143 return std::make_pair(0U, X86::FR64RegisterClass);
10151 return std::make_pair(0U, X86::VR128RegisterClass);
10157 // Use the default implementation in TargetLowering to convert the register
10158 // constraint into a member of a register class.
10159 std::pair<unsigned, const TargetRegisterClass*> Res;
10160 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10162 // Not found as a standard register?
10163 if (Res.second == 0) {
10164 // Map st(0) -> st(7) -> ST0
10165 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10166 tolower(Constraint[1]) == 's' &&
10167 tolower(Constraint[2]) == 't' &&
10168 Constraint[3] == '(' &&
10169 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10170 Constraint[5] == ')' &&
10171 Constraint[6] == '}') {
10173 Res.first = X86::ST0+Constraint[4]-'0';
10174 Res.second = X86::RFP80RegisterClass;
10178 // GCC allows "st(0)" to be called just plain "st".
10179 if (StringRef("{st}").equals_lower(Constraint)) {
10180 Res.first = X86::ST0;
10181 Res.second = X86::RFP80RegisterClass;
10186 if (StringRef("{flags}").equals_lower(Constraint)) {
10187 Res.first = X86::EFLAGS;
10188 Res.second = X86::CCRRegisterClass;
10192 // 'A' means EAX + EDX.
10193 if (Constraint == "A") {
10194 Res.first = X86::EAX;
10195 Res.second = X86::GR32_ADRegisterClass;
10201 // Otherwise, check to see if this is a register class of the wrong value
10202 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10203 // turn into {ax},{dx}.
10204 if (Res.second->hasType(VT))
10205 return Res; // Correct type already, nothing to do.
10207 // All of the single-register GCC register classes map their values onto
10208 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10209 // really want an 8-bit or 32-bit register, map to the appropriate register
10210 // class and return the appropriate register.
10211 if (Res.second == X86::GR16RegisterClass) {
10212 if (VT == MVT::i8) {
10213 unsigned DestReg = 0;
10214 switch (Res.first) {
10216 case X86::AX: DestReg = X86::AL; break;
10217 case X86::DX: DestReg = X86::DL; break;
10218 case X86::CX: DestReg = X86::CL; break;
10219 case X86::BX: DestReg = X86::BL; break;
10222 Res.first = DestReg;
10223 Res.second = X86::GR8RegisterClass;
10225 } else if (VT == MVT::i32) {
10226 unsigned DestReg = 0;
10227 switch (Res.first) {
10229 case X86::AX: DestReg = X86::EAX; break;
10230 case X86::DX: DestReg = X86::EDX; break;
10231 case X86::CX: DestReg = X86::ECX; break;
10232 case X86::BX: DestReg = X86::EBX; break;
10233 case X86::SI: DestReg = X86::ESI; break;
10234 case X86::DI: DestReg = X86::EDI; break;
10235 case X86::BP: DestReg = X86::EBP; break;
10236 case X86::SP: DestReg = X86::ESP; break;
10239 Res.first = DestReg;
10240 Res.second = X86::GR32RegisterClass;
10242 } else if (VT == MVT::i64) {
10243 unsigned DestReg = 0;
10244 switch (Res.first) {
10246 case X86::AX: DestReg = X86::RAX; break;
10247 case X86::DX: DestReg = X86::RDX; break;
10248 case X86::CX: DestReg = X86::RCX; break;
10249 case X86::BX: DestReg = X86::RBX; break;
10250 case X86::SI: DestReg = X86::RSI; break;
10251 case X86::DI: DestReg = X86::RDI; break;
10252 case X86::BP: DestReg = X86::RBP; break;
10253 case X86::SP: DestReg = X86::RSP; break;
10256 Res.first = DestReg;
10257 Res.second = X86::GR64RegisterClass;
10260 } else if (Res.second == X86::FR32RegisterClass ||
10261 Res.second == X86::FR64RegisterClass ||
10262 Res.second == X86::VR128RegisterClass) {
10263 // Handle references to XMM physical registers that got mapped into the
10264 // wrong class. This can happen with constraints like {xmm0} where the
10265 // target independent register mapper will just pick the first match it can
10266 // find, ignoring the required type.
10267 if (VT == MVT::f32)
10268 Res.second = X86::FR32RegisterClass;
10269 else if (VT == MVT::f64)
10270 Res.second = X86::FR64RegisterClass;
10271 else if (X86::VR128RegisterClass->hasType(VT))
10272 Res.second = X86::VR128RegisterClass;
10278 //===----------------------------------------------------------------------===//
10279 // X86 Widen vector type
10280 //===----------------------------------------------------------------------===//
10282 /// getWidenVectorType: given a vector type, returns the type to widen
10283 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
10284 /// If there is no vector type that we want to widen to, returns MVT::Other
10285 /// When and where to widen is target dependent based on the cost of
10286 /// scalarizing vs using the wider vector type.
10288 EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
10289 assert(VT.isVector());
10290 if (isTypeLegal(VT))
10293 // TODO: In computeRegisterProperty, we can compute the list of legal vector
10294 // type based on element type. This would speed up our search (though
10295 // it may not be worth it since the size of the list is relatively
10297 EVT EltVT = VT.getVectorElementType();
10298 unsigned NElts = VT.getVectorNumElements();
10300 // On X86, it make sense to widen any vector wider than 1
10304 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
10305 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
10306 EVT SVT = (MVT::SimpleValueType)nVT;
10308 if (isTypeLegal(SVT) &&
10309 SVT.getVectorElementType() == EltVT &&
10310 SVT.getVectorNumElements() > NElts)