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::SHL);
994 setTargetDAGCombine(ISD::SRA);
995 setTargetDAGCombine(ISD::SRL);
996 setTargetDAGCombine(ISD::OR);
997 setTargetDAGCombine(ISD::STORE);
998 setTargetDAGCombine(ISD::MEMBARRIER);
999 setTargetDAGCombine(ISD::ZERO_EXTEND);
1000 if (Subtarget->is64Bit())
1001 setTargetDAGCombine(ISD::MUL);
1003 computeRegisterProperties();
1005 // FIXME: These should be based on subtarget info. Plus, the values should
1006 // be smaller when we are in optimizing for size mode.
1007 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1008 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
1009 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1010 setPrefLoopAlignment(16);
1011 benefitFromCodePlacementOpt = true;
1015 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1020 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1021 /// the desired ByVal argument alignment.
1022 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1025 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1026 if (VTy->getBitWidth() == 128)
1028 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1029 unsigned EltAlign = 0;
1030 getMaxByValAlign(ATy->getElementType(), EltAlign);
1031 if (EltAlign > MaxAlign)
1032 MaxAlign = EltAlign;
1033 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1034 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1035 unsigned EltAlign = 0;
1036 getMaxByValAlign(STy->getElementType(i), EltAlign);
1037 if (EltAlign > MaxAlign)
1038 MaxAlign = EltAlign;
1046 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1047 /// function arguments in the caller parameter area. For X86, aggregates
1048 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1049 /// are at 4-byte boundaries.
1050 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1051 if (Subtarget->is64Bit()) {
1052 // Max of 8 and alignment of type.
1053 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1060 if (Subtarget->hasSSE1())
1061 getMaxByValAlign(Ty, Align);
1065 /// getOptimalMemOpType - Returns the target specific optimal type for load
1066 /// and store operations as a result of memset, memcpy, and memmove
1067 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
1070 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1071 bool isSrcConst, bool isSrcStr,
1072 SelectionDAG &DAG) const {
1073 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1074 // linux. This is because the stack realignment code can't handle certain
1075 // cases like PR2962. This should be removed when PR2962 is fixed.
1076 const Function *F = DAG.getMachineFunction().getFunction();
1077 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1078 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1079 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1081 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1084 if (Subtarget->is64Bit() && Size >= 8)
1089 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1090 /// current function. The returned value is a member of the
1091 /// MachineJumpTableInfo::JTEntryKind enum.
1092 unsigned X86TargetLowering::getJumpTableEncoding() const {
1093 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1095 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1096 Subtarget->isPICStyleGOT())
1097 return MachineJumpTableInfo::EK_Custom32;
1099 // Otherwise, use the normal jump table encoding heuristics.
1100 return TargetLowering::getJumpTableEncoding();
1103 /// getPICBaseSymbol - Return the X86-32 PIC base.
1105 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1106 MCContext &Ctx) const {
1107 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1108 return Ctx.GetOrCreateTemporarySymbol(Twine(MAI.getPrivateGlobalPrefix())+
1109 Twine(MF->getFunctionNumber())+"$pb");
1114 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1115 const MachineBasicBlock *MBB,
1116 unsigned uid,MCContext &Ctx) const{
1117 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1118 Subtarget->isPICStyleGOT());
1119 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1121 return X86MCTargetExpr::Create(MBB->getSymbol(Ctx),
1122 X86MCTargetExpr::GOTOFF, Ctx);
1125 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1127 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1128 SelectionDAG &DAG) const {
1129 if (!Subtarget->is64Bit())
1130 // This doesn't have DebugLoc associated with it, but is not really the
1131 // same as a Register.
1132 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1137 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1138 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1140 const MCExpr *X86TargetLowering::
1141 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1142 MCContext &Ctx) const {
1143 // X86-64 uses RIP relative addressing based on the jump table label.
1144 if (Subtarget->isPICStyleRIPRel())
1145 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1147 // Otherwise, the reference is relative to the PIC base.
1148 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1151 /// getFunctionAlignment - Return the Log2 alignment of this function.
1152 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1153 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1156 //===----------------------------------------------------------------------===//
1157 // Return Value Calling Convention Implementation
1158 //===----------------------------------------------------------------------===//
1160 #include "X86GenCallingConv.inc"
1163 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1164 const SmallVectorImpl<EVT> &OutTys,
1165 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1166 SelectionDAG &DAG) {
1167 SmallVector<CCValAssign, 16> RVLocs;
1168 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1169 RVLocs, *DAG.getContext());
1170 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1174 X86TargetLowering::LowerReturn(SDValue Chain,
1175 CallingConv::ID CallConv, bool isVarArg,
1176 const SmallVectorImpl<ISD::OutputArg> &Outs,
1177 DebugLoc dl, SelectionDAG &DAG) {
1179 SmallVector<CCValAssign, 16> RVLocs;
1180 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1181 RVLocs, *DAG.getContext());
1182 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1184 // Add the regs to the liveout set for the function.
1185 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1186 for (unsigned i = 0; i != RVLocs.size(); ++i)
1187 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1188 MRI.addLiveOut(RVLocs[i].getLocReg());
1192 SmallVector<SDValue, 6> RetOps;
1193 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1194 // Operand #1 = Bytes To Pop
1195 RetOps.push_back(DAG.getTargetConstant(getBytesToPopOnReturn(), MVT::i16));
1197 // Copy the result values into the output registers.
1198 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1199 CCValAssign &VA = RVLocs[i];
1200 assert(VA.isRegLoc() && "Can only return in registers!");
1201 SDValue ValToCopy = Outs[i].Val;
1203 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1204 // the RET instruction and handled by the FP Stackifier.
1205 if (VA.getLocReg() == X86::ST0 ||
1206 VA.getLocReg() == X86::ST1) {
1207 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1208 // change the value to the FP stack register class.
1209 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1210 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1211 RetOps.push_back(ValToCopy);
1212 // Don't emit a copytoreg.
1216 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1217 // which is returned in RAX / RDX.
1218 if (Subtarget->is64Bit()) {
1219 EVT ValVT = ValToCopy.getValueType();
1220 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1221 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1222 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1223 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1227 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1228 Flag = Chain.getValue(1);
1231 // The x86-64 ABI for returning structs by value requires that we copy
1232 // the sret argument into %rax for the return. We saved the argument into
1233 // a virtual register in the entry block, so now we copy the value out
1235 if (Subtarget->is64Bit() &&
1236 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1237 MachineFunction &MF = DAG.getMachineFunction();
1238 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1239 unsigned Reg = FuncInfo->getSRetReturnReg();
1241 Reg = MRI.createVirtualRegister(getRegClassFor(MVT::i64));
1242 FuncInfo->setSRetReturnReg(Reg);
1244 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1246 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1247 Flag = Chain.getValue(1);
1249 // RAX now acts like a return value.
1250 MRI.addLiveOut(X86::RAX);
1253 RetOps[0] = Chain; // Update chain.
1255 // Add the flag if we have it.
1257 RetOps.push_back(Flag);
1259 return DAG.getNode(X86ISD::RET_FLAG, dl,
1260 MVT::Other, &RetOps[0], RetOps.size());
1263 /// LowerCallResult - Lower the result values of a call into the
1264 /// appropriate copies out of appropriate physical registers.
1267 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1268 CallingConv::ID CallConv, bool isVarArg,
1269 const SmallVectorImpl<ISD::InputArg> &Ins,
1270 DebugLoc dl, SelectionDAG &DAG,
1271 SmallVectorImpl<SDValue> &InVals) {
1273 // Assign locations to each value returned by this call.
1274 SmallVector<CCValAssign, 16> RVLocs;
1275 bool Is64Bit = Subtarget->is64Bit();
1276 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1277 RVLocs, *DAG.getContext());
1278 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1280 // Copy all of the result registers out of their specified physreg.
1281 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1282 CCValAssign &VA = RVLocs[i];
1283 EVT CopyVT = VA.getValVT();
1285 // If this is x86-64, and we disabled SSE, we can't return FP values
1286 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1287 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1288 llvm_report_error("SSE register return with SSE disabled");
1291 // If this is a call to a function that returns an fp value on the floating
1292 // point stack, but where we prefer to use the value in xmm registers, copy
1293 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1294 if ((VA.getLocReg() == X86::ST0 ||
1295 VA.getLocReg() == X86::ST1) &&
1296 isScalarFPTypeInSSEReg(VA.getValVT())) {
1301 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1302 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1303 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1304 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1305 MVT::v2i64, InFlag).getValue(1);
1306 Val = Chain.getValue(0);
1307 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1308 Val, DAG.getConstant(0, MVT::i64));
1310 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1311 MVT::i64, InFlag).getValue(1);
1312 Val = Chain.getValue(0);
1314 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1316 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1317 CopyVT, InFlag).getValue(1);
1318 Val = Chain.getValue(0);
1320 InFlag = Chain.getValue(2);
1322 if (CopyVT != VA.getValVT()) {
1323 // Round the F80 the right size, which also moves to the appropriate xmm
1325 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1326 // This truncation won't change the value.
1327 DAG.getIntPtrConstant(1));
1330 InVals.push_back(Val);
1337 //===----------------------------------------------------------------------===//
1338 // C & StdCall & Fast Calling Convention implementation
1339 //===----------------------------------------------------------------------===//
1340 // StdCall calling convention seems to be standard for many Windows' API
1341 // routines and around. It differs from C calling convention just a little:
1342 // callee should clean up the stack, not caller. Symbols should be also
1343 // decorated in some fancy way :) It doesn't support any vector arguments.
1344 // For info on fast calling convention see Fast Calling Convention (tail call)
1345 // implementation LowerX86_32FastCCCallTo.
1347 /// CallIsStructReturn - Determines whether a call uses struct return
1349 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1353 return Outs[0].Flags.isSRet();
1356 /// ArgsAreStructReturn - Determines whether a function uses struct
1357 /// return semantics.
1359 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1363 return Ins[0].Flags.isSRet();
1366 /// IsCalleePop - Determines whether the callee is required to pop its
1367 /// own arguments. Callee pop is necessary to support tail calls.
1368 bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
1372 switch (CallingConv) {
1375 case CallingConv::X86_StdCall:
1376 return !Subtarget->is64Bit();
1377 case CallingConv::X86_FastCall:
1378 return !Subtarget->is64Bit();
1379 case CallingConv::Fast:
1380 return GuaranteedTailCallOpt;
1384 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1385 /// given CallingConvention value.
1386 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1387 if (Subtarget->is64Bit()) {
1388 if (Subtarget->isTargetWin64())
1389 return CC_X86_Win64_C;
1394 if (CC == CallingConv::X86_FastCall)
1395 return CC_X86_32_FastCall;
1396 else if (CC == CallingConv::Fast)
1397 return CC_X86_32_FastCC;
1402 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1403 /// by "Src" to address "Dst" with size and alignment information specified by
1404 /// the specific parameter attribute. The copy will be passed as a byval
1405 /// function parameter.
1407 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1408 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1410 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1411 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1412 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1415 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1416 /// a tailcall target by changing its ABI.
1417 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1418 return GuaranteedTailCallOpt && CC == CallingConv::Fast;
1422 X86TargetLowering::LowerMemArgument(SDValue Chain,
1423 CallingConv::ID CallConv,
1424 const SmallVectorImpl<ISD::InputArg> &Ins,
1425 DebugLoc dl, SelectionDAG &DAG,
1426 const CCValAssign &VA,
1427 MachineFrameInfo *MFI,
1429 // Create the nodes corresponding to a load from this parameter slot.
1430 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1431 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1432 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1435 // If value is passed by pointer we have address passed instead of the value
1437 if (VA.getLocInfo() == CCValAssign::Indirect)
1438 ValVT = VA.getLocVT();
1440 ValVT = VA.getValVT();
1442 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1443 // changed with more analysis.
1444 // In case of tail call optimization mark all arguments mutable. Since they
1445 // could be overwritten by lowering of arguments in case of a tail call.
1446 if (Flags.isByVal()) {
1447 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1448 VA.getLocMemOffset(), isImmutable, false);
1449 return DAG.getFrameIndex(FI, getPointerTy());
1451 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1452 VA.getLocMemOffset(), isImmutable, false);
1453 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1454 return DAG.getLoad(ValVT, dl, Chain, FIN,
1455 PseudoSourceValue::getFixedStack(FI), 0,
1461 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1462 CallingConv::ID CallConv,
1464 const SmallVectorImpl<ISD::InputArg> &Ins,
1467 SmallVectorImpl<SDValue> &InVals) {
1469 MachineFunction &MF = DAG.getMachineFunction();
1470 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1472 const Function* Fn = MF.getFunction();
1473 if (Fn->hasExternalLinkage() &&
1474 Subtarget->isTargetCygMing() &&
1475 Fn->getName() == "main")
1476 FuncInfo->setForceFramePointer(true);
1478 MachineFrameInfo *MFI = MF.getFrameInfo();
1479 bool Is64Bit = Subtarget->is64Bit();
1480 bool IsWin64 = Subtarget->isTargetWin64();
1482 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1483 "Var args not supported with calling convention fastcc");
1485 // Assign locations to all of the incoming arguments.
1486 SmallVector<CCValAssign, 16> ArgLocs;
1487 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1488 ArgLocs, *DAG.getContext());
1489 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1491 unsigned LastVal = ~0U;
1493 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1494 CCValAssign &VA = ArgLocs[i];
1495 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1497 assert(VA.getValNo() != LastVal &&
1498 "Don't support value assigned to multiple locs yet");
1499 LastVal = VA.getValNo();
1501 if (VA.isRegLoc()) {
1502 EVT RegVT = VA.getLocVT();
1503 TargetRegisterClass *RC = NULL;
1504 if (RegVT == MVT::i32)
1505 RC = X86::GR32RegisterClass;
1506 else if (Is64Bit && RegVT == MVT::i64)
1507 RC = X86::GR64RegisterClass;
1508 else if (RegVT == MVT::f32)
1509 RC = X86::FR32RegisterClass;
1510 else if (RegVT == MVT::f64)
1511 RC = X86::FR64RegisterClass;
1512 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1513 RC = X86::VR128RegisterClass;
1514 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1515 RC = X86::VR64RegisterClass;
1517 llvm_unreachable("Unknown argument type!");
1519 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1520 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1522 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1523 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1525 if (VA.getLocInfo() == CCValAssign::SExt)
1526 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1527 DAG.getValueType(VA.getValVT()));
1528 else if (VA.getLocInfo() == CCValAssign::ZExt)
1529 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1530 DAG.getValueType(VA.getValVT()));
1531 else if (VA.getLocInfo() == CCValAssign::BCvt)
1532 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1534 if (VA.isExtInLoc()) {
1535 // Handle MMX values passed in XMM regs.
1536 if (RegVT.isVector()) {
1537 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1538 ArgValue, DAG.getConstant(0, MVT::i64));
1539 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1541 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1544 assert(VA.isMemLoc());
1545 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1548 // If value is passed via pointer - do a load.
1549 if (VA.getLocInfo() == CCValAssign::Indirect)
1550 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1553 InVals.push_back(ArgValue);
1556 // The x86-64 ABI for returning structs by value requires that we copy
1557 // the sret argument into %rax for the return. Save the argument into
1558 // a virtual register so that we can access it from the return points.
1559 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1560 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1561 unsigned Reg = FuncInfo->getSRetReturnReg();
1563 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1564 FuncInfo->setSRetReturnReg(Reg);
1566 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1567 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1570 unsigned StackSize = CCInfo.getNextStackOffset();
1571 // Align stack specially for tail calls.
1572 if (FuncIsMadeTailCallSafe(CallConv))
1573 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1575 // If the function takes variable number of arguments, make a frame index for
1576 // the start of the first vararg value... for expansion of llvm.va_start.
1578 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1579 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize, true, false);
1582 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1584 // FIXME: We should really autogenerate these arrays
1585 static const unsigned GPR64ArgRegsWin64[] = {
1586 X86::RCX, X86::RDX, X86::R8, X86::R9
1588 static const unsigned XMMArgRegsWin64[] = {
1589 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1591 static const unsigned GPR64ArgRegs64Bit[] = {
1592 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1594 static const unsigned XMMArgRegs64Bit[] = {
1595 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1596 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1598 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1601 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1602 GPR64ArgRegs = GPR64ArgRegsWin64;
1603 XMMArgRegs = XMMArgRegsWin64;
1605 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1606 GPR64ArgRegs = GPR64ArgRegs64Bit;
1607 XMMArgRegs = XMMArgRegs64Bit;
1609 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1611 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1614 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1615 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1616 "SSE register cannot be used when SSE is disabled!");
1617 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1618 "SSE register cannot be used when SSE is disabled!");
1619 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1620 // Kernel mode asks for SSE to be disabled, so don't push them
1622 TotalNumXMMRegs = 0;
1624 // For X86-64, if there are vararg parameters that are passed via
1625 // registers, then we must store them to their spots on the stack so they
1626 // may be loaded by deferencing the result of va_next.
1627 VarArgsGPOffset = NumIntRegs * 8;
1628 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1629 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1630 TotalNumXMMRegs * 16, 16,
1633 // Store the integer parameter registers.
1634 SmallVector<SDValue, 8> MemOps;
1635 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1636 unsigned Offset = VarArgsGPOffset;
1637 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1638 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1639 DAG.getIntPtrConstant(Offset));
1640 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1641 X86::GR64RegisterClass);
1642 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1644 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1645 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1646 Offset, false, false, 0);
1647 MemOps.push_back(Store);
1651 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1652 // Now store the XMM (fp + vector) parameter registers.
1653 SmallVector<SDValue, 11> SaveXMMOps;
1654 SaveXMMOps.push_back(Chain);
1656 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1657 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1658 SaveXMMOps.push_back(ALVal);
1660 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1661 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1663 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1664 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1665 X86::VR128RegisterClass);
1666 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1667 SaveXMMOps.push_back(Val);
1669 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1671 &SaveXMMOps[0], SaveXMMOps.size()));
1674 if (!MemOps.empty())
1675 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1676 &MemOps[0], MemOps.size());
1680 // Some CCs need callee pop.
1681 if (IsCalleePop(isVarArg, CallConv)) {
1682 BytesToPopOnReturn = StackSize; // Callee pops everything.
1684 BytesToPopOnReturn = 0; // Callee pops nothing.
1685 // If this is an sret function, the return should pop the hidden pointer.
1686 if (!Is64Bit && CallConv != CallingConv::Fast && ArgsAreStructReturn(Ins))
1687 BytesToPopOnReturn = 4;
1691 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1692 if (CallConv == CallingConv::X86_FastCall)
1693 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1696 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1702 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1703 SDValue StackPtr, SDValue Arg,
1704 DebugLoc dl, SelectionDAG &DAG,
1705 const CCValAssign &VA,
1706 ISD::ArgFlagsTy Flags) {
1707 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1708 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1709 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1710 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1711 if (Flags.isByVal()) {
1712 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1714 return DAG.getStore(Chain, dl, Arg, PtrOff,
1715 PseudoSourceValue::getStack(), LocMemOffset,
1719 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1720 /// optimization is performed and it is required.
1722 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1723 SDValue &OutRetAddr, SDValue Chain,
1724 bool IsTailCall, bool Is64Bit,
1725 int FPDiff, DebugLoc dl) {
1726 // Adjust the Return address stack slot.
1727 EVT VT = getPointerTy();
1728 OutRetAddr = getReturnAddressFrameIndex(DAG);
1730 // Load the "old" Return address.
1731 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1732 return SDValue(OutRetAddr.getNode(), 1);
1735 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1736 /// optimization is performed and it is required (FPDiff!=0).
1738 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1739 SDValue Chain, SDValue RetAddrFrIdx,
1740 bool Is64Bit, int FPDiff, DebugLoc dl) {
1741 // Store the return address to the appropriate stack slot.
1742 if (!FPDiff) return Chain;
1743 // Calculate the new stack slot for the return address.
1744 int SlotSize = Is64Bit ? 8 : 4;
1745 int NewReturnAddrFI =
1746 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false, false);
1747 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1748 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1749 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1750 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1756 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1757 CallingConv::ID CallConv, bool isVarArg,
1759 const SmallVectorImpl<ISD::OutputArg> &Outs,
1760 const SmallVectorImpl<ISD::InputArg> &Ins,
1761 DebugLoc dl, SelectionDAG &DAG,
1762 SmallVectorImpl<SDValue> &InVals) {
1763 MachineFunction &MF = DAG.getMachineFunction();
1764 bool Is64Bit = Subtarget->is64Bit();
1765 bool IsStructRet = CallIsStructReturn(Outs);
1766 bool IsSibcall = false;
1769 // Check if it's really possible to do a tail call.
1770 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
1773 // Sibcalls are automatically detected tailcalls which do not require
1775 if (!GuaranteedTailCallOpt && isTailCall)
1782 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1783 "Var args not supported with calling convention fastcc");
1785 // Analyze operands of the call, assigning locations to each operand.
1786 SmallVector<CCValAssign, 16> ArgLocs;
1787 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1788 ArgLocs, *DAG.getContext());
1789 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1791 // Get a count of how many bytes are to be pushed on the stack.
1792 unsigned NumBytes = CCInfo.getNextStackOffset();
1794 // This is a sibcall. The memory operands are available in caller's
1795 // own caller's stack.
1797 else if (GuaranteedTailCallOpt && CallConv == CallingConv::Fast)
1798 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1801 if (isTailCall && !IsSibcall) {
1802 // Lower arguments at fp - stackoffset + fpdiff.
1803 unsigned NumBytesCallerPushed =
1804 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1805 FPDiff = NumBytesCallerPushed - NumBytes;
1807 // Set the delta of movement of the returnaddr stackslot.
1808 // But only set if delta is greater than previous delta.
1809 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1810 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1814 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1816 SDValue RetAddrFrIdx;
1817 // Load return adress for tail calls.
1818 if (isTailCall && FPDiff)
1819 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1820 Is64Bit, FPDiff, dl);
1822 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1823 SmallVector<SDValue, 8> MemOpChains;
1826 // Walk the register/memloc assignments, inserting copies/loads. In the case
1827 // of tail call optimization arguments are handle later.
1828 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1829 CCValAssign &VA = ArgLocs[i];
1830 EVT RegVT = VA.getLocVT();
1831 SDValue Arg = Outs[i].Val;
1832 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1833 bool isByVal = Flags.isByVal();
1835 // Promote the value if needed.
1836 switch (VA.getLocInfo()) {
1837 default: llvm_unreachable("Unknown loc info!");
1838 case CCValAssign::Full: break;
1839 case CCValAssign::SExt:
1840 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1842 case CCValAssign::ZExt:
1843 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1845 case CCValAssign::AExt:
1846 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1847 // Special case: passing MMX values in XMM registers.
1848 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1849 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1850 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1852 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1854 case CCValAssign::BCvt:
1855 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1857 case CCValAssign::Indirect: {
1858 // Store the argument.
1859 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1860 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1861 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1862 PseudoSourceValue::getFixedStack(FI), 0,
1869 if (VA.isRegLoc()) {
1870 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1871 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1872 assert(VA.isMemLoc());
1873 if (StackPtr.getNode() == 0)
1874 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1875 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1876 dl, DAG, VA, Flags));
1880 if (!MemOpChains.empty())
1881 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1882 &MemOpChains[0], MemOpChains.size());
1884 // Build a sequence of copy-to-reg nodes chained together with token chain
1885 // and flag operands which copy the outgoing args into registers.
1887 // Tail call byval lowering might overwrite argument registers so in case of
1888 // tail call optimization the copies to registers are lowered later.
1890 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1891 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1892 RegsToPass[i].second, InFlag);
1893 InFlag = Chain.getValue(1);
1896 if (Subtarget->isPICStyleGOT()) {
1897 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1900 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1901 DAG.getNode(X86ISD::GlobalBaseReg,
1902 DebugLoc::getUnknownLoc(),
1905 InFlag = Chain.getValue(1);
1907 // If we are tail calling and generating PIC/GOT style code load the
1908 // address of the callee into ECX. The value in ecx is used as target of
1909 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1910 // for tail calls on PIC/GOT architectures. Normally we would just put the
1911 // address of GOT into ebx and then call target@PLT. But for tail calls
1912 // ebx would be restored (since ebx is callee saved) before jumping to the
1915 // Note: The actual moving to ECX is done further down.
1916 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1917 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1918 !G->getGlobal()->hasProtectedVisibility())
1919 Callee = LowerGlobalAddress(Callee, DAG);
1920 else if (isa<ExternalSymbolSDNode>(Callee))
1921 Callee = LowerExternalSymbol(Callee, DAG);
1925 if (Is64Bit && isVarArg) {
1926 // From AMD64 ABI document:
1927 // For calls that may call functions that use varargs or stdargs
1928 // (prototype-less calls or calls to functions containing ellipsis (...) in
1929 // the declaration) %al is used as hidden argument to specify the number
1930 // of SSE registers used. The contents of %al do not need to match exactly
1931 // the number of registers, but must be an ubound on the number of SSE
1932 // registers used and is in the range 0 - 8 inclusive.
1934 // FIXME: Verify this on Win64
1935 // Count the number of XMM registers allocated.
1936 static const unsigned XMMArgRegs[] = {
1937 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1938 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1940 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1941 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1942 && "SSE registers cannot be used when SSE is disabled");
1944 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1945 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1946 InFlag = Chain.getValue(1);
1950 // For tail calls lower the arguments to the 'real' stack slot.
1952 // Force all the incoming stack arguments to be loaded from the stack
1953 // before any new outgoing arguments are stored to the stack, because the
1954 // outgoing stack slots may alias the incoming argument stack slots, and
1955 // the alias isn't otherwise explicit. This is slightly more conservative
1956 // than necessary, because it means that each store effectively depends
1957 // on every argument instead of just those arguments it would clobber.
1958 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1960 SmallVector<SDValue, 8> MemOpChains2;
1963 // Do not flag preceeding copytoreg stuff together with the following stuff.
1965 if (GuaranteedTailCallOpt) {
1966 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1967 CCValAssign &VA = ArgLocs[i];
1970 assert(VA.isMemLoc());
1971 SDValue Arg = Outs[i].Val;
1972 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1973 // Create frame index.
1974 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1975 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1976 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
1977 FIN = DAG.getFrameIndex(FI, getPointerTy());
1979 if (Flags.isByVal()) {
1980 // Copy relative to framepointer.
1981 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1982 if (StackPtr.getNode() == 0)
1983 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1985 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1987 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
1991 // Store relative to framepointer.
1992 MemOpChains2.push_back(
1993 DAG.getStore(ArgChain, dl, Arg, FIN,
1994 PseudoSourceValue::getFixedStack(FI), 0,
2000 if (!MemOpChains2.empty())
2001 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2002 &MemOpChains2[0], MemOpChains2.size());
2004 // Copy arguments to their registers.
2005 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2006 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2007 RegsToPass[i].second, InFlag);
2008 InFlag = Chain.getValue(1);
2012 // Store the return address to the appropriate stack slot.
2013 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2017 bool WasGlobalOrExternal = false;
2018 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2019 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2020 // In the 64-bit large code model, we have to make all calls
2021 // through a register, since the call instruction's 32-bit
2022 // pc-relative offset may not be large enough to hold the whole
2024 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2025 WasGlobalOrExternal = true;
2026 // If the callee is a GlobalAddress node (quite common, every direct call
2027 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2030 // We should use extra load for direct calls to dllimported functions in
2032 GlobalValue *GV = G->getGlobal();
2033 if (!GV->hasDLLImportLinkage()) {
2034 unsigned char OpFlags = 0;
2036 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2037 // external symbols most go through the PLT in PIC mode. If the symbol
2038 // has hidden or protected visibility, or if it is static or local, then
2039 // we don't need to use the PLT - we can directly call it.
2040 if (Subtarget->isTargetELF() &&
2041 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2042 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2043 OpFlags = X86II::MO_PLT;
2044 } else if (Subtarget->isPICStyleStubAny() &&
2045 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2046 Subtarget->getDarwinVers() < 9) {
2047 // PC-relative references to external symbols should go through $stub,
2048 // unless we're building with the leopard linker or later, which
2049 // automatically synthesizes these stubs.
2050 OpFlags = X86II::MO_DARWIN_STUB;
2053 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2054 G->getOffset(), OpFlags);
2056 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2057 WasGlobalOrExternal = true;
2058 unsigned char OpFlags = 0;
2060 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2061 // symbols should go through the PLT.
2062 if (Subtarget->isTargetELF() &&
2063 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2064 OpFlags = X86II::MO_PLT;
2065 } else if (Subtarget->isPICStyleStubAny() &&
2066 Subtarget->getDarwinVers() < 9) {
2067 // PC-relative references to external symbols should go through $stub,
2068 // unless we're building with the leopard linker or later, which
2069 // automatically synthesizes these stubs.
2070 OpFlags = X86II::MO_DARWIN_STUB;
2073 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2077 if (isTailCall && !WasGlobalOrExternal) {
2078 // Force the address into a (call preserved) caller-saved register since
2079 // tailcall must happen after callee-saved registers are poped.
2080 // FIXME: Give it a special register class that contains caller-saved
2081 // register instead?
2082 unsigned TCReg = Is64Bit ? X86::R11 : X86::EAX;
2083 Chain = DAG.getCopyToReg(Chain, dl,
2084 DAG.getRegister(TCReg, getPointerTy()),
2086 Callee = DAG.getRegister(TCReg, getPointerTy());
2089 // Returns a chain & a flag for retval copy to use.
2090 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2091 SmallVector<SDValue, 8> Ops;
2093 if (!IsSibcall && isTailCall) {
2094 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2095 DAG.getIntPtrConstant(0, true), InFlag);
2096 InFlag = Chain.getValue(1);
2099 Ops.push_back(Chain);
2100 Ops.push_back(Callee);
2103 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2105 // Add argument registers to the end of the list so that they are known live
2107 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2108 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2109 RegsToPass[i].second.getValueType()));
2111 // Add an implicit use GOT pointer in EBX.
2112 if (!isTailCall && Subtarget->isPICStyleGOT())
2113 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2115 // Add an implicit use of AL for x86 vararg functions.
2116 if (Is64Bit && isVarArg)
2117 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2119 if (InFlag.getNode())
2120 Ops.push_back(InFlag);
2123 // If this is the first return lowered for this function, add the regs
2124 // to the liveout set for the function.
2125 if (MF.getRegInfo().liveout_empty()) {
2126 SmallVector<CCValAssign, 16> RVLocs;
2127 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2129 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2130 for (unsigned i = 0; i != RVLocs.size(); ++i)
2131 if (RVLocs[i].isRegLoc())
2132 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2135 assert(((Callee.getOpcode() == ISD::Register &&
2136 (cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
2137 cast<RegisterSDNode>(Callee)->getReg() == X86::R11)) ||
2138 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2139 Callee.getOpcode() == ISD::TargetGlobalAddress) &&
2140 "Expecting a global address, external symbol, or scratch register");
2142 return DAG.getNode(X86ISD::TC_RETURN, dl,
2143 NodeTys, &Ops[0], Ops.size());
2146 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2147 InFlag = Chain.getValue(1);
2149 // Create the CALLSEQ_END node.
2150 unsigned NumBytesForCalleeToPush;
2151 if (IsCalleePop(isVarArg, CallConv))
2152 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2153 else if (!Is64Bit && CallConv != CallingConv::Fast && IsStructRet)
2154 // If this is a call to a struct-return function, the callee
2155 // pops the hidden struct pointer, so we have to push it back.
2156 // This is common for Darwin/X86, Linux & Mingw32 targets.
2157 NumBytesForCalleeToPush = 4;
2159 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2161 // Returns a flag for retval copy to use.
2163 Chain = DAG.getCALLSEQ_END(Chain,
2164 DAG.getIntPtrConstant(NumBytes, true),
2165 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2168 InFlag = Chain.getValue(1);
2171 // Handle result values, copying them out of physregs into vregs that we
2173 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2174 Ins, dl, DAG, InVals);
2178 //===----------------------------------------------------------------------===//
2179 // Fast Calling Convention (tail call) implementation
2180 //===----------------------------------------------------------------------===//
2182 // Like std call, callee cleans arguments, convention except that ECX is
2183 // reserved for storing the tail called function address. Only 2 registers are
2184 // free for argument passing (inreg). Tail call optimization is performed
2186 // * tailcallopt is enabled
2187 // * caller/callee are fastcc
2188 // On X86_64 architecture with GOT-style position independent code only local
2189 // (within module) calls are supported at the moment.
2190 // To keep the stack aligned according to platform abi the function
2191 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2192 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2193 // If a tail called function callee has more arguments than the caller the
2194 // caller needs to make sure that there is room to move the RETADDR to. This is
2195 // achieved by reserving an area the size of the argument delta right after the
2196 // original REtADDR, but before the saved framepointer or the spilled registers
2197 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2209 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2210 /// for a 16 byte align requirement.
2211 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2212 SelectionDAG& DAG) {
2213 MachineFunction &MF = DAG.getMachineFunction();
2214 const TargetMachine &TM = MF.getTarget();
2215 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2216 unsigned StackAlignment = TFI.getStackAlignment();
2217 uint64_t AlignMask = StackAlignment - 1;
2218 int64_t Offset = StackSize;
2219 uint64_t SlotSize = TD->getPointerSize();
2220 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2221 // Number smaller than 12 so just add the difference.
2222 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2224 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2225 Offset = ((~AlignMask) & Offset) + StackAlignment +
2226 (StackAlignment-SlotSize);
2231 /// MatchingStackOffset - Return true if the given stack call argument is
2232 /// already available in the same position (relatively) of the caller's
2233 /// incoming argument stack.
2235 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2236 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2237 const X86InstrInfo *TII) {
2238 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
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 Bytes = Flags.getByValSize();
2259 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2260 if (Flags.isByVal())
2261 // ByVal argument is passed in as a pointer but it's now being
2262 // dereferenced. e.g.
2263 // define @foo(%struct.X* %A) {
2264 // tail call @bar(%struct.X* byval %A)
2267 SDValue Ptr = Ld->getBasePtr();
2268 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2271 FI = FINode->getIndex();
2275 assert(FI != INT_MAX);
2276 if (!MFI->isFixedObjectIndex(FI))
2278 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2281 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2282 /// for tail call optimization. Targets which want to do tail call
2283 /// optimization should implement this function.
2285 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2286 CallingConv::ID CalleeCC,
2288 const SmallVectorImpl<ISD::OutputArg> &Outs,
2289 const SmallVectorImpl<ISD::InputArg> &Ins,
2290 SelectionDAG& DAG) const {
2291 if (CalleeCC != CallingConv::Fast &&
2292 CalleeCC != CallingConv::C)
2295 // If -tailcallopt is specified, make fastcc functions tail-callable.
2296 const Function *CallerF = DAG.getMachineFunction().getFunction();
2297 if (GuaranteedTailCallOpt) {
2298 if (CalleeCC == CallingConv::Fast &&
2299 CallerF->getCallingConv() == CalleeCC)
2304 // Look for obvious safe cases to perform tail call optimization that does not
2305 // requite ABI changes. This is what gcc calls sibcall.
2307 // Do not tail call optimize vararg calls for now.
2311 // If the callee takes no arguments then go on to check the results of the
2313 if (!Outs.empty()) {
2314 // Check if stack adjustment is needed. For now, do not do this if any
2315 // argument is passed on the stack.
2316 SmallVector<CCValAssign, 16> ArgLocs;
2317 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2318 ArgLocs, *DAG.getContext());
2319 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2320 if (CCInfo.getNextStackOffset()) {
2321 MachineFunction &MF = DAG.getMachineFunction();
2322 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2324 if (Subtarget->isTargetWin64())
2325 // Win64 ABI has additional complications.
2328 // Check if the arguments are already laid out in the right way as
2329 // the caller's fixed stack objects.
2330 MachineFrameInfo *MFI = MF.getFrameInfo();
2331 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2332 const X86InstrInfo *TII =
2333 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2334 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2335 CCValAssign &VA = ArgLocs[i];
2336 EVT RegVT = VA.getLocVT();
2337 SDValue Arg = Outs[i].Val;
2338 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2339 if (VA.getLocInfo() == CCValAssign::Indirect)
2341 if (!VA.isRegLoc()) {
2342 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2354 X86TargetLowering::createFastISel(MachineFunction &mf, MachineModuleInfo *mmo,
2356 DenseMap<const Value *, unsigned> &vm,
2357 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2358 DenseMap<const AllocaInst *, int> &am
2360 , SmallSet<Instruction*, 8> &cil
2363 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2371 //===----------------------------------------------------------------------===//
2372 // Other Lowering Hooks
2373 //===----------------------------------------------------------------------===//
2376 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2377 MachineFunction &MF = DAG.getMachineFunction();
2378 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2379 int ReturnAddrIndex = FuncInfo->getRAIndex();
2381 if (ReturnAddrIndex == 0) {
2382 // Set up a frame object for the return address.
2383 uint64_t SlotSize = TD->getPointerSize();
2384 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2386 FuncInfo->setRAIndex(ReturnAddrIndex);
2389 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2393 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2394 bool hasSymbolicDisplacement) {
2395 // Offset should fit into 32 bit immediate field.
2396 if (!isInt32(Offset))
2399 // If we don't have a symbolic displacement - we don't have any extra
2401 if (!hasSymbolicDisplacement)
2404 // FIXME: Some tweaks might be needed for medium code model.
2405 if (M != CodeModel::Small && M != CodeModel::Kernel)
2408 // For small code model we assume that latest object is 16MB before end of 31
2409 // bits boundary. We may also accept pretty large negative constants knowing
2410 // that all objects are in the positive half of address space.
2411 if (M == CodeModel::Small && Offset < 16*1024*1024)
2414 // For kernel code model we know that all object resist in the negative half
2415 // of 32bits address space. We may not accept negative offsets, since they may
2416 // be just off and we may accept pretty large positive ones.
2417 if (M == CodeModel::Kernel && Offset > 0)
2423 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2424 /// specific condition code, returning the condition code and the LHS/RHS of the
2425 /// comparison to make.
2426 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2427 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2429 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2430 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2431 // X > -1 -> X == 0, jump !sign.
2432 RHS = DAG.getConstant(0, RHS.getValueType());
2433 return X86::COND_NS;
2434 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2435 // X < 0 -> X == 0, jump on sign.
2437 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2439 RHS = DAG.getConstant(0, RHS.getValueType());
2440 return X86::COND_LE;
2444 switch (SetCCOpcode) {
2445 default: llvm_unreachable("Invalid integer condition!");
2446 case ISD::SETEQ: return X86::COND_E;
2447 case ISD::SETGT: return X86::COND_G;
2448 case ISD::SETGE: return X86::COND_GE;
2449 case ISD::SETLT: return X86::COND_L;
2450 case ISD::SETLE: return X86::COND_LE;
2451 case ISD::SETNE: return X86::COND_NE;
2452 case ISD::SETULT: return X86::COND_B;
2453 case ISD::SETUGT: return X86::COND_A;
2454 case ISD::SETULE: return X86::COND_BE;
2455 case ISD::SETUGE: return X86::COND_AE;
2459 // First determine if it is required or is profitable to flip the operands.
2461 // If LHS is a foldable load, but RHS is not, flip the condition.
2462 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2463 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2464 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2465 std::swap(LHS, RHS);
2468 switch (SetCCOpcode) {
2474 std::swap(LHS, RHS);
2478 // On a floating point condition, the flags are set as follows:
2480 // 0 | 0 | 0 | X > Y
2481 // 0 | 0 | 1 | X < Y
2482 // 1 | 0 | 0 | X == Y
2483 // 1 | 1 | 1 | unordered
2484 switch (SetCCOpcode) {
2485 default: llvm_unreachable("Condcode should be pre-legalized away");
2487 case ISD::SETEQ: return X86::COND_E;
2488 case ISD::SETOLT: // flipped
2490 case ISD::SETGT: return X86::COND_A;
2491 case ISD::SETOLE: // flipped
2493 case ISD::SETGE: return X86::COND_AE;
2494 case ISD::SETUGT: // flipped
2496 case ISD::SETLT: return X86::COND_B;
2497 case ISD::SETUGE: // flipped
2499 case ISD::SETLE: return X86::COND_BE;
2501 case ISD::SETNE: return X86::COND_NE;
2502 case ISD::SETUO: return X86::COND_P;
2503 case ISD::SETO: return X86::COND_NP;
2505 case ISD::SETUNE: return X86::COND_INVALID;
2509 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2510 /// code. Current x86 isa includes the following FP cmov instructions:
2511 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2512 static bool hasFPCMov(unsigned X86CC) {
2528 /// isFPImmLegal - Returns true if the target can instruction select the
2529 /// specified FP immediate natively. If false, the legalizer will
2530 /// materialize the FP immediate as a load from a constant pool.
2531 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2532 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2533 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2539 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2540 /// the specified range (L, H].
2541 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2542 return (Val < 0) || (Val >= Low && Val < Hi);
2545 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2546 /// specified value.
2547 static bool isUndefOrEqual(int Val, int CmpVal) {
2548 if (Val < 0 || Val == CmpVal)
2553 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2554 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2555 /// the second operand.
2556 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2557 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2558 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2559 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2560 return (Mask[0] < 2 && Mask[1] < 2);
2564 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2565 SmallVector<int, 8> M;
2567 return ::isPSHUFDMask(M, N->getValueType(0));
2570 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2571 /// is suitable for input to PSHUFHW.
2572 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2573 if (VT != MVT::v8i16)
2576 // Lower quadword copied in order or undef.
2577 for (int i = 0; i != 4; ++i)
2578 if (Mask[i] >= 0 && Mask[i] != i)
2581 // Upper quadword shuffled.
2582 for (int i = 4; i != 8; ++i)
2583 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2589 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2590 SmallVector<int, 8> M;
2592 return ::isPSHUFHWMask(M, N->getValueType(0));
2595 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2596 /// is suitable for input to PSHUFLW.
2597 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2598 if (VT != MVT::v8i16)
2601 // Upper quadword copied in order.
2602 for (int i = 4; i != 8; ++i)
2603 if (Mask[i] >= 0 && Mask[i] != i)
2606 // Lower quadword shuffled.
2607 for (int i = 0; i != 4; ++i)
2614 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2615 SmallVector<int, 8> M;
2617 return ::isPSHUFLWMask(M, N->getValueType(0));
2620 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2621 /// is suitable for input to PALIGNR.
2622 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2624 int i, e = VT.getVectorNumElements();
2626 // Do not handle v2i64 / v2f64 shuffles with palignr.
2627 if (e < 4 || !hasSSSE3)
2630 for (i = 0; i != e; ++i)
2634 // All undef, not a palignr.
2638 // Determine if it's ok to perform a palignr with only the LHS, since we
2639 // don't have access to the actual shuffle elements to see if RHS is undef.
2640 bool Unary = Mask[i] < (int)e;
2641 bool NeedsUnary = false;
2643 int s = Mask[i] - i;
2645 // Check the rest of the elements to see if they are consecutive.
2646 for (++i; i != e; ++i) {
2651 Unary = Unary && (m < (int)e);
2652 NeedsUnary = NeedsUnary || (m < s);
2654 if (NeedsUnary && !Unary)
2656 if (Unary && m != ((s+i) & (e-1)))
2658 if (!Unary && m != (s+i))
2664 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2665 SmallVector<int, 8> M;
2667 return ::isPALIGNRMask(M, N->getValueType(0), true);
2670 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2671 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2672 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2673 int NumElems = VT.getVectorNumElements();
2674 if (NumElems != 2 && NumElems != 4)
2677 int Half = NumElems / 2;
2678 for (int i = 0; i < Half; ++i)
2679 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2681 for (int i = Half; i < NumElems; ++i)
2682 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2688 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2689 SmallVector<int, 8> M;
2691 return ::isSHUFPMask(M, N->getValueType(0));
2694 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2695 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2696 /// half elements to come from vector 1 (which would equal the dest.) and
2697 /// the upper half to come from vector 2.
2698 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2699 int NumElems = VT.getVectorNumElements();
2701 if (NumElems != 2 && NumElems != 4)
2704 int Half = NumElems / 2;
2705 for (int i = 0; i < Half; ++i)
2706 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2708 for (int i = Half; i < NumElems; ++i)
2709 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2714 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2715 SmallVector<int, 8> M;
2717 return isCommutedSHUFPMask(M, N->getValueType(0));
2720 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2721 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2722 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2723 if (N->getValueType(0).getVectorNumElements() != 4)
2726 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2727 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2728 isUndefOrEqual(N->getMaskElt(1), 7) &&
2729 isUndefOrEqual(N->getMaskElt(2), 2) &&
2730 isUndefOrEqual(N->getMaskElt(3), 3);
2733 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2734 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2736 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2737 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2742 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2743 isUndefOrEqual(N->getMaskElt(1), 3) &&
2744 isUndefOrEqual(N->getMaskElt(2), 2) &&
2745 isUndefOrEqual(N->getMaskElt(3), 3);
2748 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2749 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2750 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2751 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2753 if (NumElems != 2 && NumElems != 4)
2756 for (unsigned i = 0; i < NumElems/2; ++i)
2757 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2760 for (unsigned i = NumElems/2; i < NumElems; ++i)
2761 if (!isUndefOrEqual(N->getMaskElt(i), i))
2767 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2768 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2769 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2770 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2772 if (NumElems != 2 && NumElems != 4)
2775 for (unsigned i = 0; i < NumElems/2; ++i)
2776 if (!isUndefOrEqual(N->getMaskElt(i), i))
2779 for (unsigned i = 0; i < NumElems/2; ++i)
2780 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2786 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2787 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2788 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2789 bool V2IsSplat = false) {
2790 int NumElts = VT.getVectorNumElements();
2791 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2794 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2796 int BitI1 = Mask[i+1];
2797 if (!isUndefOrEqual(BitI, j))
2800 if (!isUndefOrEqual(BitI1, NumElts))
2803 if (!isUndefOrEqual(BitI1, j + NumElts))
2810 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2811 SmallVector<int, 8> M;
2813 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2816 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2817 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2818 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2819 bool V2IsSplat = false) {
2820 int NumElts = VT.getVectorNumElements();
2821 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2824 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2826 int BitI1 = Mask[i+1];
2827 if (!isUndefOrEqual(BitI, j + NumElts/2))
2830 if (isUndefOrEqual(BitI1, NumElts))
2833 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2840 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2841 SmallVector<int, 8> M;
2843 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2846 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2847 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2849 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2850 int NumElems = VT.getVectorNumElements();
2851 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2854 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2856 int BitI1 = Mask[i+1];
2857 if (!isUndefOrEqual(BitI, j))
2859 if (!isUndefOrEqual(BitI1, j))
2865 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2866 SmallVector<int, 8> M;
2868 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2871 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2872 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2874 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2875 int NumElems = VT.getVectorNumElements();
2876 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2879 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2881 int BitI1 = Mask[i+1];
2882 if (!isUndefOrEqual(BitI, j))
2884 if (!isUndefOrEqual(BitI1, j))
2890 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2891 SmallVector<int, 8> M;
2893 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2896 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2897 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2898 /// MOVSD, and MOVD, i.e. setting the lowest element.
2899 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2900 if (VT.getVectorElementType().getSizeInBits() < 32)
2903 int NumElts = VT.getVectorNumElements();
2905 if (!isUndefOrEqual(Mask[0], NumElts))
2908 for (int i = 1; i < NumElts; ++i)
2909 if (!isUndefOrEqual(Mask[i], i))
2915 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2916 SmallVector<int, 8> M;
2918 return ::isMOVLMask(M, N->getValueType(0));
2921 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2922 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2923 /// element of vector 2 and the other elements to come from vector 1 in order.
2924 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2925 bool V2IsSplat = false, bool V2IsUndef = false) {
2926 int NumOps = VT.getVectorNumElements();
2927 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2930 if (!isUndefOrEqual(Mask[0], 0))
2933 for (int i = 1; i < NumOps; ++i)
2934 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2935 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2936 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2942 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2943 bool V2IsUndef = false) {
2944 SmallVector<int, 8> M;
2946 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2949 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2950 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2951 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2952 if (N->getValueType(0).getVectorNumElements() != 4)
2955 // Expect 1, 1, 3, 3
2956 for (unsigned i = 0; i < 2; ++i) {
2957 int Elt = N->getMaskElt(i);
2958 if (Elt >= 0 && Elt != 1)
2963 for (unsigned i = 2; i < 4; ++i) {
2964 int Elt = N->getMaskElt(i);
2965 if (Elt >= 0 && Elt != 3)
2970 // Don't use movshdup if it can be done with a shufps.
2971 // FIXME: verify that matching u, u, 3, 3 is what we want.
2975 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2976 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2977 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2978 if (N->getValueType(0).getVectorNumElements() != 4)
2981 // Expect 0, 0, 2, 2
2982 for (unsigned i = 0; i < 2; ++i)
2983 if (N->getMaskElt(i) > 0)
2987 for (unsigned i = 2; i < 4; ++i) {
2988 int Elt = N->getMaskElt(i);
2989 if (Elt >= 0 && Elt != 2)
2994 // Don't use movsldup if it can be done with a shufps.
2998 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2999 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3000 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3001 int e = N->getValueType(0).getVectorNumElements() / 2;
3003 for (int i = 0; i < e; ++i)
3004 if (!isUndefOrEqual(N->getMaskElt(i), i))
3006 for (int i = 0; i < e; ++i)
3007 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3012 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3013 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3014 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3015 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3016 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3018 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3020 for (int i = 0; i < NumOperands; ++i) {
3021 int Val = SVOp->getMaskElt(NumOperands-i-1);
3022 if (Val < 0) Val = 0;
3023 if (Val >= NumOperands) Val -= NumOperands;
3025 if (i != NumOperands - 1)
3031 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3032 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3033 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3034 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3036 // 8 nodes, but we only care about the last 4.
3037 for (unsigned i = 7; i >= 4; --i) {
3038 int Val = SVOp->getMaskElt(i);
3047 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3048 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3049 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3050 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3052 // 8 nodes, but we only care about the first 4.
3053 for (int i = 3; i >= 0; --i) {
3054 int Val = SVOp->getMaskElt(i);
3063 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3064 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3065 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3066 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3067 EVT VVT = N->getValueType(0);
3068 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3072 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3073 Val = SVOp->getMaskElt(i);
3077 return (Val - i) * EltSize;
3080 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3082 bool X86::isZeroNode(SDValue Elt) {
3083 return ((isa<ConstantSDNode>(Elt) &&
3084 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
3085 (isa<ConstantFPSDNode>(Elt) &&
3086 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3089 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3090 /// their permute mask.
3091 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3092 SelectionDAG &DAG) {
3093 EVT VT = SVOp->getValueType(0);
3094 unsigned NumElems = VT.getVectorNumElements();
3095 SmallVector<int, 8> MaskVec;
3097 for (unsigned i = 0; i != NumElems; ++i) {
3098 int idx = SVOp->getMaskElt(i);
3100 MaskVec.push_back(idx);
3101 else if (idx < (int)NumElems)
3102 MaskVec.push_back(idx + NumElems);
3104 MaskVec.push_back(idx - NumElems);
3106 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3107 SVOp->getOperand(0), &MaskVec[0]);
3110 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3111 /// the two vector operands have swapped position.
3112 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3113 unsigned NumElems = VT.getVectorNumElements();
3114 for (unsigned i = 0; i != NumElems; ++i) {
3118 else if (idx < (int)NumElems)
3119 Mask[i] = idx + NumElems;
3121 Mask[i] = idx - NumElems;
3125 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3126 /// match movhlps. The lower half elements should come from upper half of
3127 /// V1 (and in order), and the upper half elements should come from the upper
3128 /// half of V2 (and in order).
3129 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3130 if (Op->getValueType(0).getVectorNumElements() != 4)
3132 for (unsigned i = 0, e = 2; i != e; ++i)
3133 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3135 for (unsigned i = 2; i != 4; ++i)
3136 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3141 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3142 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3144 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3145 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3147 N = N->getOperand(0).getNode();
3148 if (!ISD::isNON_EXTLoad(N))
3151 *LD = cast<LoadSDNode>(N);
3155 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3156 /// match movlp{s|d}. The lower half elements should come from lower half of
3157 /// V1 (and in order), and the upper half elements should come from the upper
3158 /// half of V2 (and in order). And since V1 will become the source of the
3159 /// MOVLP, it must be either a vector load or a scalar load to vector.
3160 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3161 ShuffleVectorSDNode *Op) {
3162 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3164 // Is V2 is a vector load, don't do this transformation. We will try to use
3165 // load folding shufps op.
3166 if (ISD::isNON_EXTLoad(V2))
3169 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3171 if (NumElems != 2 && NumElems != 4)
3173 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3174 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3176 for (unsigned i = NumElems/2; i != NumElems; ++i)
3177 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3182 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3184 static bool isSplatVector(SDNode *N) {
3185 if (N->getOpcode() != ISD::BUILD_VECTOR)
3188 SDValue SplatValue = N->getOperand(0);
3189 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3190 if (N->getOperand(i) != SplatValue)
3195 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3196 /// to an zero vector.
3197 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3198 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3199 SDValue V1 = N->getOperand(0);
3200 SDValue V2 = N->getOperand(1);
3201 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3202 for (unsigned i = 0; i != NumElems; ++i) {
3203 int Idx = N->getMaskElt(i);
3204 if (Idx >= (int)NumElems) {
3205 unsigned Opc = V2.getOpcode();
3206 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3208 if (Opc != ISD::BUILD_VECTOR ||
3209 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3211 } else if (Idx >= 0) {
3212 unsigned Opc = V1.getOpcode();
3213 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3215 if (Opc != ISD::BUILD_VECTOR ||
3216 !X86::isZeroNode(V1.getOperand(Idx)))
3223 /// getZeroVector - Returns a vector of specified type with all zero elements.
3225 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3227 assert(VT.isVector() && "Expected a vector type");
3229 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3230 // type. This ensures they get CSE'd.
3232 if (VT.getSizeInBits() == 64) { // MMX
3233 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3234 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3235 } else if (HasSSE2) { // SSE2
3236 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3237 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3239 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3240 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3242 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3245 /// getOnesVector - Returns a vector of specified type with all bits set.
3247 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3248 assert(VT.isVector() && "Expected a vector type");
3250 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3251 // type. This ensures they get CSE'd.
3252 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3254 if (VT.getSizeInBits() == 64) // MMX
3255 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3257 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3258 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3262 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3263 /// that point to V2 points to its first element.
3264 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3265 EVT VT = SVOp->getValueType(0);
3266 unsigned NumElems = VT.getVectorNumElements();
3268 bool Changed = false;
3269 SmallVector<int, 8> MaskVec;
3270 SVOp->getMask(MaskVec);
3272 for (unsigned i = 0; i != NumElems; ++i) {
3273 if (MaskVec[i] > (int)NumElems) {
3274 MaskVec[i] = NumElems;
3279 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3280 SVOp->getOperand(1), &MaskVec[0]);
3281 return SDValue(SVOp, 0);
3284 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3285 /// operation of specified width.
3286 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3288 unsigned NumElems = VT.getVectorNumElements();
3289 SmallVector<int, 8> Mask;
3290 Mask.push_back(NumElems);
3291 for (unsigned i = 1; i != NumElems; ++i)
3293 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3296 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3297 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3299 unsigned NumElems = VT.getVectorNumElements();
3300 SmallVector<int, 8> Mask;
3301 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3303 Mask.push_back(i + NumElems);
3305 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3308 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3309 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3311 unsigned NumElems = VT.getVectorNumElements();
3312 unsigned Half = NumElems/2;
3313 SmallVector<int, 8> Mask;
3314 for (unsigned i = 0; i != Half; ++i) {
3315 Mask.push_back(i + Half);
3316 Mask.push_back(i + NumElems + Half);
3318 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3321 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3322 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3324 if (SV->getValueType(0).getVectorNumElements() <= 4)
3325 return SDValue(SV, 0);
3327 EVT PVT = MVT::v4f32;
3328 EVT VT = SV->getValueType(0);
3329 DebugLoc dl = SV->getDebugLoc();
3330 SDValue V1 = SV->getOperand(0);
3331 int NumElems = VT.getVectorNumElements();
3332 int EltNo = SV->getSplatIndex();
3334 // unpack elements to the correct location
3335 while (NumElems > 4) {
3336 if (EltNo < NumElems/2) {
3337 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3339 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3340 EltNo -= NumElems/2;
3345 // Perform the splat.
3346 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3347 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3348 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3349 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3352 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3353 /// vector of zero or undef vector. This produces a shuffle where the low
3354 /// element of V2 is swizzled into the zero/undef vector, landing at element
3355 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3356 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3357 bool isZero, bool HasSSE2,
3358 SelectionDAG &DAG) {
3359 EVT VT = V2.getValueType();
3361 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3362 unsigned NumElems = VT.getVectorNumElements();
3363 SmallVector<int, 16> MaskVec;
3364 for (unsigned i = 0; i != NumElems; ++i)
3365 // If this is the insertion idx, put the low elt of V2 here.
3366 MaskVec.push_back(i == Idx ? NumElems : i);
3367 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3370 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3371 /// a shuffle that is zero.
3373 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3374 bool Low, SelectionDAG &DAG) {
3375 unsigned NumZeros = 0;
3376 for (int i = 0; i < NumElems; ++i) {
3377 unsigned Index = Low ? i : NumElems-i-1;
3378 int Idx = SVOp->getMaskElt(Index);
3383 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3384 if (Elt.getNode() && X86::isZeroNode(Elt))
3392 /// isVectorShift - Returns true if the shuffle can be implemented as a
3393 /// logical left or right shift of a vector.
3394 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3395 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3396 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3397 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3400 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3403 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3407 bool SeenV1 = false;
3408 bool SeenV2 = false;
3409 for (int i = NumZeros; i < NumElems; ++i) {
3410 int Val = isLeft ? (i - NumZeros) : i;
3411 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3423 if (SeenV1 && SeenV2)
3426 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3432 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3434 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3435 unsigned NumNonZero, unsigned NumZero,
3436 SelectionDAG &DAG, TargetLowering &TLI) {
3440 DebugLoc dl = Op.getDebugLoc();
3443 for (unsigned i = 0; i < 16; ++i) {
3444 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3445 if (ThisIsNonZero && First) {
3447 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3449 V = DAG.getUNDEF(MVT::v8i16);
3454 SDValue ThisElt(0, 0), LastElt(0, 0);
3455 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3456 if (LastIsNonZero) {
3457 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3458 MVT::i16, Op.getOperand(i-1));
3460 if (ThisIsNonZero) {
3461 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3462 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3463 ThisElt, DAG.getConstant(8, MVT::i8));
3465 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3469 if (ThisElt.getNode())
3470 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3471 DAG.getIntPtrConstant(i/2));
3475 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3478 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3480 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3481 unsigned NumNonZero, unsigned NumZero,
3482 SelectionDAG &DAG, TargetLowering &TLI) {
3486 DebugLoc dl = Op.getDebugLoc();
3489 for (unsigned i = 0; i < 8; ++i) {
3490 bool isNonZero = (NonZeros & (1 << i)) != 0;
3494 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3496 V = DAG.getUNDEF(MVT::v8i16);
3499 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3500 MVT::v8i16, V, Op.getOperand(i),
3501 DAG.getIntPtrConstant(i));
3508 /// getVShift - Return a vector logical shift node.
3510 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3511 unsigned NumBits, SelectionDAG &DAG,
3512 const TargetLowering &TLI, DebugLoc dl) {
3513 bool isMMX = VT.getSizeInBits() == 64;
3514 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3515 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3516 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3517 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3518 DAG.getNode(Opc, dl, ShVT, SrcOp,
3519 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3523 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3524 SelectionDAG &DAG) {
3526 // Check if the scalar load can be widened into a vector load. And if
3527 // the address is "base + cst" see if the cst can be "absorbed" into
3528 // the shuffle mask.
3529 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3530 SDValue Ptr = LD->getBasePtr();
3531 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3533 EVT PVT = LD->getValueType(0);
3534 if (PVT != MVT::i32 && PVT != MVT::f32)
3539 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3540 FI = FINode->getIndex();
3542 } else if (Ptr.getOpcode() == ISD::ADD &&
3543 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3544 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3545 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3546 Offset = Ptr.getConstantOperandVal(1);
3547 Ptr = Ptr.getOperand(0);
3552 SDValue Chain = LD->getChain();
3553 // Make sure the stack object alignment is at least 16.
3554 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3555 if (DAG.InferPtrAlignment(Ptr) < 16) {
3556 if (MFI->isFixedObjectIndex(FI)) {
3557 // Can't change the alignment. FIXME: It's possible to compute
3558 // the exact stack offset and reference FI + adjust offset instead.
3559 // If someone *really* cares about this. That's the way to implement it.
3562 MFI->setObjectAlignment(FI, 16);
3566 // (Offset % 16) must be multiple of 4. Then address is then
3567 // Ptr + (Offset & ~15).
3570 if ((Offset % 16) & 3)
3572 int64_t StartOffset = Offset & ~15;
3574 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3575 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3577 int EltNo = (Offset - StartOffset) >> 2;
3578 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3579 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3580 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3582 // Canonicalize it to a v4i32 shuffle.
3583 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3584 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3585 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3586 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3593 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3594 DebugLoc dl = Op.getDebugLoc();
3595 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3596 if (ISD::isBuildVectorAllZeros(Op.getNode())
3597 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3598 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3599 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3600 // eliminated on x86-32 hosts.
3601 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3604 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3605 return getOnesVector(Op.getValueType(), DAG, dl);
3606 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3609 EVT VT = Op.getValueType();
3610 EVT ExtVT = VT.getVectorElementType();
3611 unsigned EVTBits = ExtVT.getSizeInBits();
3613 unsigned NumElems = Op.getNumOperands();
3614 unsigned NumZero = 0;
3615 unsigned NumNonZero = 0;
3616 unsigned NonZeros = 0;
3617 bool IsAllConstants = true;
3618 SmallSet<SDValue, 8> Values;
3619 for (unsigned i = 0; i < NumElems; ++i) {
3620 SDValue Elt = Op.getOperand(i);
3621 if (Elt.getOpcode() == ISD::UNDEF)
3624 if (Elt.getOpcode() != ISD::Constant &&
3625 Elt.getOpcode() != ISD::ConstantFP)
3626 IsAllConstants = false;
3627 if (X86::isZeroNode(Elt))
3630 NonZeros |= (1 << i);
3635 if (NumNonZero == 0) {
3636 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3637 return DAG.getUNDEF(VT);
3640 // Special case for single non-zero, non-undef, element.
3641 if (NumNonZero == 1) {
3642 unsigned Idx = CountTrailingZeros_32(NonZeros);
3643 SDValue Item = Op.getOperand(Idx);
3645 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3646 // the value are obviously zero, truncate the value to i32 and do the
3647 // insertion that way. Only do this if the value is non-constant or if the
3648 // value is a constant being inserted into element 0. It is cheaper to do
3649 // a constant pool load than it is to do a movd + shuffle.
3650 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3651 (!IsAllConstants || Idx == 0)) {
3652 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3653 // Handle MMX and SSE both.
3654 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3655 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3657 // Truncate the value (which may itself be a constant) to i32, and
3658 // convert it to a vector with movd (S2V+shuffle to zero extend).
3659 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3660 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3661 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3662 Subtarget->hasSSE2(), DAG);
3664 // Now we have our 32-bit value zero extended in the low element of
3665 // a vector. If Idx != 0, swizzle it into place.
3667 SmallVector<int, 4> Mask;
3668 Mask.push_back(Idx);
3669 for (unsigned i = 1; i != VecElts; ++i)
3671 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3672 DAG.getUNDEF(Item.getValueType()),
3675 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3679 // If we have a constant or non-constant insertion into the low element of
3680 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3681 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3682 // depending on what the source datatype is.
3685 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3686 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3687 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3688 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3689 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3690 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3692 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3693 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3694 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3695 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3696 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3697 Subtarget->hasSSE2(), DAG);
3698 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3702 // Is it a vector logical left shift?
3703 if (NumElems == 2 && Idx == 1 &&
3704 X86::isZeroNode(Op.getOperand(0)) &&
3705 !X86::isZeroNode(Op.getOperand(1))) {
3706 unsigned NumBits = VT.getSizeInBits();
3707 return getVShift(true, VT,
3708 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3709 VT, Op.getOperand(1)),
3710 NumBits/2, DAG, *this, dl);
3713 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3716 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3717 // is a non-constant being inserted into an element other than the low one,
3718 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3719 // movd/movss) to move this into the low element, then shuffle it into
3721 if (EVTBits == 32) {
3722 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3724 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3725 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3726 Subtarget->hasSSE2(), DAG);
3727 SmallVector<int, 8> MaskVec;
3728 for (unsigned i = 0; i < NumElems; i++)
3729 MaskVec.push_back(i == Idx ? 0 : 1);
3730 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3734 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3735 if (Values.size() == 1) {
3736 if (EVTBits == 32) {
3737 // Instead of a shuffle like this:
3738 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3739 // Check if it's possible to issue this instead.
3740 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3741 unsigned Idx = CountTrailingZeros_32(NonZeros);
3742 SDValue Item = Op.getOperand(Idx);
3743 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3744 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3749 // A vector full of immediates; various special cases are already
3750 // handled, so this is best done with a single constant-pool load.
3754 // Let legalizer expand 2-wide build_vectors.
3755 if (EVTBits == 64) {
3756 if (NumNonZero == 1) {
3757 // One half is zero or undef.
3758 unsigned Idx = CountTrailingZeros_32(NonZeros);
3759 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3760 Op.getOperand(Idx));
3761 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3762 Subtarget->hasSSE2(), DAG);
3767 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3768 if (EVTBits == 8 && NumElems == 16) {
3769 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3771 if (V.getNode()) return V;
3774 if (EVTBits == 16 && NumElems == 8) {
3775 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3777 if (V.getNode()) return V;
3780 // If element VT is == 32 bits, turn it into a number of shuffles.
3781 SmallVector<SDValue, 8> V;
3783 if (NumElems == 4 && NumZero > 0) {
3784 for (unsigned i = 0; i < 4; ++i) {
3785 bool isZero = !(NonZeros & (1 << i));
3787 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3789 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3792 for (unsigned i = 0; i < 2; ++i) {
3793 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3796 V[i] = V[i*2]; // Must be a zero vector.
3799 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3802 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3805 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3810 SmallVector<int, 8> MaskVec;
3811 bool Reverse = (NonZeros & 0x3) == 2;
3812 for (unsigned i = 0; i < 2; ++i)
3813 MaskVec.push_back(Reverse ? 1-i : i);
3814 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3815 for (unsigned i = 0; i < 2; ++i)
3816 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3817 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3820 if (Values.size() > 2) {
3821 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3822 // values to be inserted is equal to the number of elements, in which case
3823 // use the unpack code below in the hopes of matching the consecutive elts
3824 // load merge pattern for shuffles.
3825 // FIXME: We could probably just check that here directly.
3826 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3827 getSubtarget()->hasSSE41()) {
3828 V[0] = DAG.getUNDEF(VT);
3829 for (unsigned i = 0; i < NumElems; ++i)
3830 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3831 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3832 Op.getOperand(i), DAG.getIntPtrConstant(i));
3835 // Expand into a number of unpckl*.
3837 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3838 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3839 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3840 for (unsigned i = 0; i < NumElems; ++i)
3841 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3843 while (NumElems != 0) {
3844 for (unsigned i = 0; i < NumElems; ++i)
3845 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3855 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3856 // We support concatenate two MMX registers and place them in a MMX
3857 // register. This is better than doing a stack convert.
3858 DebugLoc dl = Op.getDebugLoc();
3859 EVT ResVT = Op.getValueType();
3860 assert(Op.getNumOperands() == 2);
3861 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
3862 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
3864 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
3865 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3866 InVec = Op.getOperand(1);
3867 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3868 unsigned NumElts = ResVT.getVectorNumElements();
3869 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3870 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
3871 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
3873 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
3874 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3875 Mask[0] = 0; Mask[1] = 2;
3876 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
3878 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3881 // v8i16 shuffles - Prefer shuffles in the following order:
3882 // 1. [all] pshuflw, pshufhw, optional move
3883 // 2. [ssse3] 1 x pshufb
3884 // 3. [ssse3] 2 x pshufb + 1 x por
3885 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3887 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3888 SelectionDAG &DAG, X86TargetLowering &TLI) {
3889 SDValue V1 = SVOp->getOperand(0);
3890 SDValue V2 = SVOp->getOperand(1);
3891 DebugLoc dl = SVOp->getDebugLoc();
3892 SmallVector<int, 8> MaskVals;
3894 // Determine if more than 1 of the words in each of the low and high quadwords
3895 // of the result come from the same quadword of one of the two inputs. Undef
3896 // mask values count as coming from any quadword, for better codegen.
3897 SmallVector<unsigned, 4> LoQuad(4);
3898 SmallVector<unsigned, 4> HiQuad(4);
3899 BitVector InputQuads(4);
3900 for (unsigned i = 0; i < 8; ++i) {
3901 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3902 int EltIdx = SVOp->getMaskElt(i);
3903 MaskVals.push_back(EltIdx);
3912 InputQuads.set(EltIdx / 4);
3915 int BestLoQuad = -1;
3916 unsigned MaxQuad = 1;
3917 for (unsigned i = 0; i < 4; ++i) {
3918 if (LoQuad[i] > MaxQuad) {
3920 MaxQuad = LoQuad[i];
3924 int BestHiQuad = -1;
3926 for (unsigned i = 0; i < 4; ++i) {
3927 if (HiQuad[i] > MaxQuad) {
3929 MaxQuad = HiQuad[i];
3933 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3934 // of the two input vectors, shuffle them into one input vector so only a
3935 // single pshufb instruction is necessary. If There are more than 2 input
3936 // quads, disable the next transformation since it does not help SSSE3.
3937 bool V1Used = InputQuads[0] || InputQuads[1];
3938 bool V2Used = InputQuads[2] || InputQuads[3];
3939 if (TLI.getSubtarget()->hasSSSE3()) {
3940 if (InputQuads.count() == 2 && V1Used && V2Used) {
3941 BestLoQuad = InputQuads.find_first();
3942 BestHiQuad = InputQuads.find_next(BestLoQuad);
3944 if (InputQuads.count() > 2) {
3950 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3951 // the shuffle mask. If a quad is scored as -1, that means that it contains
3952 // words from all 4 input quadwords.
3954 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3955 SmallVector<int, 8> MaskV;
3956 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3957 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3958 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3959 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3960 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3961 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3963 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3964 // source words for the shuffle, to aid later transformations.
3965 bool AllWordsInNewV = true;
3966 bool InOrder[2] = { true, true };
3967 for (unsigned i = 0; i != 8; ++i) {
3968 int idx = MaskVals[i];
3970 InOrder[i/4] = false;
3971 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3973 AllWordsInNewV = false;
3977 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3978 if (AllWordsInNewV) {
3979 for (int i = 0; i != 8; ++i) {
3980 int idx = MaskVals[i];
3983 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3984 if ((idx != i) && idx < 4)
3986 if ((idx != i) && idx > 3)
3995 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3996 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3997 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3998 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3999 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4003 // If we have SSSE3, and all words of the result are from 1 input vector,
4004 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4005 // is present, fall back to case 4.
4006 if (TLI.getSubtarget()->hasSSSE3()) {
4007 SmallVector<SDValue,16> pshufbMask;
4009 // If we have elements from both input vectors, set the high bit of the
4010 // shuffle mask element to zero out elements that come from V2 in the V1
4011 // mask, and elements that come from V1 in the V2 mask, so that the two
4012 // results can be OR'd together.
4013 bool TwoInputs = V1Used && V2Used;
4014 for (unsigned i = 0; i != 8; ++i) {
4015 int EltIdx = MaskVals[i] * 2;
4016 if (TwoInputs && (EltIdx >= 16)) {
4017 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4018 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4021 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4022 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4024 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4025 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4026 DAG.getNode(ISD::BUILD_VECTOR, dl,
4027 MVT::v16i8, &pshufbMask[0], 16));
4029 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4031 // Calculate the shuffle mask for the second input, shuffle it, and
4032 // OR it with the first shuffled input.
4034 for (unsigned i = 0; i != 8; ++i) {
4035 int EltIdx = MaskVals[i] * 2;
4037 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4038 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4041 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4042 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4044 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4045 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4046 DAG.getNode(ISD::BUILD_VECTOR, dl,
4047 MVT::v16i8, &pshufbMask[0], 16));
4048 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4049 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4052 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4053 // and update MaskVals with new element order.
4054 BitVector InOrder(8);
4055 if (BestLoQuad >= 0) {
4056 SmallVector<int, 8> MaskV;
4057 for (int i = 0; i != 4; ++i) {
4058 int idx = MaskVals[i];
4060 MaskV.push_back(-1);
4062 } else if ((idx / 4) == BestLoQuad) {
4063 MaskV.push_back(idx & 3);
4066 MaskV.push_back(-1);
4069 for (unsigned i = 4; i != 8; ++i)
4071 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4075 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4076 // and update MaskVals with the new element order.
4077 if (BestHiQuad >= 0) {
4078 SmallVector<int, 8> MaskV;
4079 for (unsigned i = 0; i != 4; ++i)
4081 for (unsigned i = 4; i != 8; ++i) {
4082 int idx = MaskVals[i];
4084 MaskV.push_back(-1);
4086 } else if ((idx / 4) == BestHiQuad) {
4087 MaskV.push_back((idx & 3) + 4);
4090 MaskV.push_back(-1);
4093 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4097 // In case BestHi & BestLo were both -1, which means each quadword has a word
4098 // from each of the four input quadwords, calculate the InOrder bitvector now
4099 // before falling through to the insert/extract cleanup.
4100 if (BestLoQuad == -1 && BestHiQuad == -1) {
4102 for (int i = 0; i != 8; ++i)
4103 if (MaskVals[i] < 0 || MaskVals[i] == i)
4107 // The other elements are put in the right place using pextrw and pinsrw.
4108 for (unsigned i = 0; i != 8; ++i) {
4111 int EltIdx = MaskVals[i];
4114 SDValue ExtOp = (EltIdx < 8)
4115 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4116 DAG.getIntPtrConstant(EltIdx))
4117 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4118 DAG.getIntPtrConstant(EltIdx - 8));
4119 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4120 DAG.getIntPtrConstant(i));
4125 // v16i8 shuffles - Prefer shuffles in the following order:
4126 // 1. [ssse3] 1 x pshufb
4127 // 2. [ssse3] 2 x pshufb + 1 x por
4128 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4130 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4131 SelectionDAG &DAG, X86TargetLowering &TLI) {
4132 SDValue V1 = SVOp->getOperand(0);
4133 SDValue V2 = SVOp->getOperand(1);
4134 DebugLoc dl = SVOp->getDebugLoc();
4135 SmallVector<int, 16> MaskVals;
4136 SVOp->getMask(MaskVals);
4138 // If we have SSSE3, case 1 is generated when all result bytes come from
4139 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4140 // present, fall back to case 3.
4141 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4144 for (unsigned i = 0; i < 16; ++i) {
4145 int EltIdx = MaskVals[i];
4154 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4155 if (TLI.getSubtarget()->hasSSSE3()) {
4156 SmallVector<SDValue,16> pshufbMask;
4158 // If all result elements are from one input vector, then only translate
4159 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4161 // Otherwise, we have elements from both input vectors, and must zero out
4162 // elements that come from V2 in the first mask, and V1 in the second mask
4163 // so that we can OR them together.
4164 bool TwoInputs = !(V1Only || V2Only);
4165 for (unsigned i = 0; i != 16; ++i) {
4166 int EltIdx = MaskVals[i];
4167 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4168 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4171 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4173 // If all the elements are from V2, assign it to V1 and return after
4174 // building the first pshufb.
4177 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4178 DAG.getNode(ISD::BUILD_VECTOR, dl,
4179 MVT::v16i8, &pshufbMask[0], 16));
4183 // Calculate the shuffle mask for the second input, shuffle it, and
4184 // OR it with the first shuffled input.
4186 for (unsigned i = 0; i != 16; ++i) {
4187 int EltIdx = MaskVals[i];
4189 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4192 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4194 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4195 DAG.getNode(ISD::BUILD_VECTOR, dl,
4196 MVT::v16i8, &pshufbMask[0], 16));
4197 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4200 // No SSSE3 - Calculate in place words and then fix all out of place words
4201 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4202 // the 16 different words that comprise the two doublequadword input vectors.
4203 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4204 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4205 SDValue NewV = V2Only ? V2 : V1;
4206 for (int i = 0; i != 8; ++i) {
4207 int Elt0 = MaskVals[i*2];
4208 int Elt1 = MaskVals[i*2+1];
4210 // This word of the result is all undef, skip it.
4211 if (Elt0 < 0 && Elt1 < 0)
4214 // This word of the result is already in the correct place, skip it.
4215 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4217 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4220 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4221 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4224 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4225 // using a single extract together, load it and store it.
4226 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4227 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4228 DAG.getIntPtrConstant(Elt1 / 2));
4229 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4230 DAG.getIntPtrConstant(i));
4234 // If Elt1 is defined, extract it from the appropriate source. If the
4235 // source byte is not also odd, shift the extracted word left 8 bits
4236 // otherwise clear the bottom 8 bits if we need to do an or.
4238 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4239 DAG.getIntPtrConstant(Elt1 / 2));
4240 if ((Elt1 & 1) == 0)
4241 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4242 DAG.getConstant(8, TLI.getShiftAmountTy()));
4244 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4245 DAG.getConstant(0xFF00, MVT::i16));
4247 // If Elt0 is defined, extract it from the appropriate source. If the
4248 // source byte is not also even, shift the extracted word right 8 bits. If
4249 // Elt1 was also defined, OR the extracted values together before
4250 // inserting them in the result.
4252 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4253 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4254 if ((Elt0 & 1) != 0)
4255 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4256 DAG.getConstant(8, TLI.getShiftAmountTy()));
4258 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4259 DAG.getConstant(0x00FF, MVT::i16));
4260 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4263 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4264 DAG.getIntPtrConstant(i));
4266 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4269 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4270 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4271 /// done when every pair / quad of shuffle mask elements point to elements in
4272 /// the right sequence. e.g.
4273 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4275 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4277 TargetLowering &TLI, DebugLoc dl) {
4278 EVT VT = SVOp->getValueType(0);
4279 SDValue V1 = SVOp->getOperand(0);
4280 SDValue V2 = SVOp->getOperand(1);
4281 unsigned NumElems = VT.getVectorNumElements();
4282 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4283 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4284 EVT MaskEltVT = MaskVT.getVectorElementType();
4286 switch (VT.getSimpleVT().SimpleTy) {
4287 default: assert(false && "Unexpected!");
4288 case MVT::v4f32: NewVT = MVT::v2f64; break;
4289 case MVT::v4i32: NewVT = MVT::v2i64; break;
4290 case MVT::v8i16: NewVT = MVT::v4i32; break;
4291 case MVT::v16i8: NewVT = MVT::v4i32; break;
4294 if (NewWidth == 2) {
4300 int Scale = NumElems / NewWidth;
4301 SmallVector<int, 8> MaskVec;
4302 for (unsigned i = 0; i < NumElems; i += Scale) {
4304 for (int j = 0; j < Scale; ++j) {
4305 int EltIdx = SVOp->getMaskElt(i+j);
4309 StartIdx = EltIdx - (EltIdx % Scale);
4310 if (EltIdx != StartIdx + j)
4314 MaskVec.push_back(-1);
4316 MaskVec.push_back(StartIdx / Scale);
4319 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4320 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4321 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4324 /// getVZextMovL - Return a zero-extending vector move low node.
4326 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4327 SDValue SrcOp, SelectionDAG &DAG,
4328 const X86Subtarget *Subtarget, DebugLoc dl) {
4329 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4330 LoadSDNode *LD = NULL;
4331 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4332 LD = dyn_cast<LoadSDNode>(SrcOp);
4334 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4336 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4337 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4338 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4339 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4340 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4342 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4343 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4344 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4345 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4353 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4354 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4355 DAG.getNode(ISD::BIT_CONVERT, dl,
4359 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4362 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4363 SDValue V1 = SVOp->getOperand(0);
4364 SDValue V2 = SVOp->getOperand(1);
4365 DebugLoc dl = SVOp->getDebugLoc();
4366 EVT VT = SVOp->getValueType(0);
4368 SmallVector<std::pair<int, int>, 8> Locs;
4370 SmallVector<int, 8> Mask1(4U, -1);
4371 SmallVector<int, 8> PermMask;
4372 SVOp->getMask(PermMask);
4376 for (unsigned i = 0; i != 4; ++i) {
4377 int Idx = PermMask[i];
4379 Locs[i] = std::make_pair(-1, -1);
4381 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4383 Locs[i] = std::make_pair(0, NumLo);
4387 Locs[i] = std::make_pair(1, NumHi);
4389 Mask1[2+NumHi] = Idx;
4395 if (NumLo <= 2 && NumHi <= 2) {
4396 // If no more than two elements come from either vector. This can be
4397 // implemented with two shuffles. First shuffle gather the elements.
4398 // The second shuffle, which takes the first shuffle as both of its
4399 // vector operands, put the elements into the right order.
4400 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4402 SmallVector<int, 8> Mask2(4U, -1);
4404 for (unsigned i = 0; i != 4; ++i) {
4405 if (Locs[i].first == -1)
4408 unsigned Idx = (i < 2) ? 0 : 4;
4409 Idx += Locs[i].first * 2 + Locs[i].second;
4414 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4415 } else if (NumLo == 3 || NumHi == 3) {
4416 // Otherwise, we must have three elements from one vector, call it X, and
4417 // one element from the other, call it Y. First, use a shufps to build an
4418 // intermediate vector with the one element from Y and the element from X
4419 // that will be in the same half in the final destination (the indexes don't
4420 // matter). Then, use a shufps to build the final vector, taking the half
4421 // containing the element from Y from the intermediate, and the other half
4424 // Normalize it so the 3 elements come from V1.
4425 CommuteVectorShuffleMask(PermMask, VT);
4429 // Find the element from V2.
4431 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4432 int Val = PermMask[HiIndex];
4439 Mask1[0] = PermMask[HiIndex];
4441 Mask1[2] = PermMask[HiIndex^1];
4443 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4446 Mask1[0] = PermMask[0];
4447 Mask1[1] = PermMask[1];
4448 Mask1[2] = HiIndex & 1 ? 6 : 4;
4449 Mask1[3] = HiIndex & 1 ? 4 : 6;
4450 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4452 Mask1[0] = HiIndex & 1 ? 2 : 0;
4453 Mask1[1] = HiIndex & 1 ? 0 : 2;
4454 Mask1[2] = PermMask[2];
4455 Mask1[3] = PermMask[3];
4460 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4464 // Break it into (shuffle shuffle_hi, shuffle_lo).
4466 SmallVector<int,8> LoMask(4U, -1);
4467 SmallVector<int,8> HiMask(4U, -1);
4469 SmallVector<int,8> *MaskPtr = &LoMask;
4470 unsigned MaskIdx = 0;
4473 for (unsigned i = 0; i != 4; ++i) {
4480 int Idx = PermMask[i];
4482 Locs[i] = std::make_pair(-1, -1);
4483 } else if (Idx < 4) {
4484 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4485 (*MaskPtr)[LoIdx] = Idx;
4488 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4489 (*MaskPtr)[HiIdx] = Idx;
4494 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4495 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4496 SmallVector<int, 8> MaskOps;
4497 for (unsigned i = 0; i != 4; ++i) {
4498 if (Locs[i].first == -1) {
4499 MaskOps.push_back(-1);
4501 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4502 MaskOps.push_back(Idx);
4505 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4509 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4510 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4511 SDValue V1 = Op.getOperand(0);
4512 SDValue V2 = Op.getOperand(1);
4513 EVT VT = Op.getValueType();
4514 DebugLoc dl = Op.getDebugLoc();
4515 unsigned NumElems = VT.getVectorNumElements();
4516 bool isMMX = VT.getSizeInBits() == 64;
4517 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4518 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4519 bool V1IsSplat = false;
4520 bool V2IsSplat = false;
4522 if (isZeroShuffle(SVOp))
4523 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4525 // Promote splats to v4f32.
4526 if (SVOp->isSplat()) {
4527 if (isMMX || NumElems < 4)
4529 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4532 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4534 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4535 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4536 if (NewOp.getNode())
4537 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4538 LowerVECTOR_SHUFFLE(NewOp, DAG));
4539 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4540 // FIXME: Figure out a cleaner way to do this.
4541 // Try to make use of movq to zero out the top part.
4542 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4543 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4544 if (NewOp.getNode()) {
4545 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4546 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4547 DAG, Subtarget, dl);
4549 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4550 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4551 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4552 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4553 DAG, Subtarget, dl);
4557 if (X86::isPSHUFDMask(SVOp))
4560 // Check if this can be converted into a logical shift.
4561 bool isLeft = false;
4564 bool isShift = getSubtarget()->hasSSE2() &&
4565 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4566 if (isShift && ShVal.hasOneUse()) {
4567 // If the shifted value has multiple uses, it may be cheaper to use
4568 // v_set0 + movlhps or movhlps, etc.
4569 EVT EltVT = VT.getVectorElementType();
4570 ShAmt *= EltVT.getSizeInBits();
4571 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4574 if (X86::isMOVLMask(SVOp)) {
4577 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4578 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4583 // FIXME: fold these into legal mask.
4584 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4585 X86::isMOVSLDUPMask(SVOp) ||
4586 X86::isMOVHLPSMask(SVOp) ||
4587 X86::isMOVLHPSMask(SVOp) ||
4588 X86::isMOVLPMask(SVOp)))
4591 if (ShouldXformToMOVHLPS(SVOp) ||
4592 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4593 return CommuteVectorShuffle(SVOp, DAG);
4596 // No better options. Use a vshl / vsrl.
4597 EVT EltVT = VT.getVectorElementType();
4598 ShAmt *= EltVT.getSizeInBits();
4599 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4602 bool Commuted = false;
4603 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4604 // 1,1,1,1 -> v8i16 though.
4605 V1IsSplat = isSplatVector(V1.getNode());
4606 V2IsSplat = isSplatVector(V2.getNode());
4608 // Canonicalize the splat or undef, if present, to be on the RHS.
4609 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4610 Op = CommuteVectorShuffle(SVOp, DAG);
4611 SVOp = cast<ShuffleVectorSDNode>(Op);
4612 V1 = SVOp->getOperand(0);
4613 V2 = SVOp->getOperand(1);
4614 std::swap(V1IsSplat, V2IsSplat);
4615 std::swap(V1IsUndef, V2IsUndef);
4619 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4620 // Shuffling low element of v1 into undef, just return v1.
4623 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4624 // the instruction selector will not match, so get a canonical MOVL with
4625 // swapped operands to undo the commute.
4626 return getMOVL(DAG, dl, VT, V2, V1);
4629 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4630 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4631 X86::isUNPCKLMask(SVOp) ||
4632 X86::isUNPCKHMask(SVOp))
4636 // Normalize mask so all entries that point to V2 points to its first
4637 // element then try to match unpck{h|l} again. If match, return a
4638 // new vector_shuffle with the corrected mask.
4639 SDValue NewMask = NormalizeMask(SVOp, DAG);
4640 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4641 if (NSVOp != SVOp) {
4642 if (X86::isUNPCKLMask(NSVOp, true)) {
4644 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4651 // Commute is back and try unpck* again.
4652 // FIXME: this seems wrong.
4653 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4654 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4655 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4656 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4657 X86::isUNPCKLMask(NewSVOp) ||
4658 X86::isUNPCKHMask(NewSVOp))
4662 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4664 // Normalize the node to match x86 shuffle ops if needed
4665 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4666 return CommuteVectorShuffle(SVOp, DAG);
4668 // Check for legal shuffle and return?
4669 SmallVector<int, 16> PermMask;
4670 SVOp->getMask(PermMask);
4671 if (isShuffleMaskLegal(PermMask, VT))
4674 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4675 if (VT == MVT::v8i16) {
4676 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4677 if (NewOp.getNode())
4681 if (VT == MVT::v16i8) {
4682 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4683 if (NewOp.getNode())
4687 // Handle all 4 wide cases with a number of shuffles except for MMX.
4688 if (NumElems == 4 && !isMMX)
4689 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4695 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4696 SelectionDAG &DAG) {
4697 EVT VT = Op.getValueType();
4698 DebugLoc dl = Op.getDebugLoc();
4699 if (VT.getSizeInBits() == 8) {
4700 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4701 Op.getOperand(0), Op.getOperand(1));
4702 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4703 DAG.getValueType(VT));
4704 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4705 } else if (VT.getSizeInBits() == 16) {
4706 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4707 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4709 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4710 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4711 DAG.getNode(ISD::BIT_CONVERT, dl,
4715 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4716 Op.getOperand(0), Op.getOperand(1));
4717 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4718 DAG.getValueType(VT));
4719 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4720 } else if (VT == MVT::f32) {
4721 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4722 // the result back to FR32 register. It's only worth matching if the
4723 // result has a single use which is a store or a bitcast to i32. And in
4724 // the case of a store, it's not worth it if the index is a constant 0,
4725 // because a MOVSSmr can be used instead, which is smaller and faster.
4726 if (!Op.hasOneUse())
4728 SDNode *User = *Op.getNode()->use_begin();
4729 if ((User->getOpcode() != ISD::STORE ||
4730 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4731 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4732 (User->getOpcode() != ISD::BIT_CONVERT ||
4733 User->getValueType(0) != MVT::i32))
4735 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4736 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4739 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4740 } else if (VT == MVT::i32) {
4741 // ExtractPS works with constant index.
4742 if (isa<ConstantSDNode>(Op.getOperand(1)))
4750 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4751 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4754 if (Subtarget->hasSSE41()) {
4755 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4760 EVT VT = Op.getValueType();
4761 DebugLoc dl = Op.getDebugLoc();
4762 // TODO: handle v16i8.
4763 if (VT.getSizeInBits() == 16) {
4764 SDValue Vec = Op.getOperand(0);
4765 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4767 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4768 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4769 DAG.getNode(ISD::BIT_CONVERT, dl,
4772 // Transform it so it match pextrw which produces a 32-bit result.
4773 EVT EltVT = MVT::i32;
4774 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4775 Op.getOperand(0), Op.getOperand(1));
4776 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4777 DAG.getValueType(VT));
4778 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4779 } else if (VT.getSizeInBits() == 32) {
4780 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4784 // SHUFPS the element to the lowest double word, then movss.
4785 int Mask[4] = { Idx, -1, -1, -1 };
4786 EVT VVT = Op.getOperand(0).getValueType();
4787 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4788 DAG.getUNDEF(VVT), Mask);
4789 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4790 DAG.getIntPtrConstant(0));
4791 } else if (VT.getSizeInBits() == 64) {
4792 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4793 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4794 // to match extract_elt for f64.
4795 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4799 // UNPCKHPD the element to the lowest double word, then movsd.
4800 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4801 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4802 int Mask[2] = { 1, -1 };
4803 EVT VVT = Op.getOperand(0).getValueType();
4804 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4805 DAG.getUNDEF(VVT), Mask);
4806 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4807 DAG.getIntPtrConstant(0));
4814 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4815 EVT VT = Op.getValueType();
4816 EVT EltVT = VT.getVectorElementType();
4817 DebugLoc dl = Op.getDebugLoc();
4819 SDValue N0 = Op.getOperand(0);
4820 SDValue N1 = Op.getOperand(1);
4821 SDValue N2 = Op.getOperand(2);
4823 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4824 isa<ConstantSDNode>(N2)) {
4826 if (VT == MVT::v8i16)
4827 Opc = X86ISD::PINSRW;
4828 else if (VT == MVT::v4i16)
4829 Opc = X86ISD::MMX_PINSRW;
4830 else if (VT == MVT::v16i8)
4831 Opc = X86ISD::PINSRB;
4833 Opc = X86ISD::PINSRB;
4835 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4837 if (N1.getValueType() != MVT::i32)
4838 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4839 if (N2.getValueType() != MVT::i32)
4840 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4841 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4842 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4843 // Bits [7:6] of the constant are the source select. This will always be
4844 // zero here. The DAG Combiner may combine an extract_elt index into these
4845 // bits. For example (insert (extract, 3), 2) could be matched by putting
4846 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4847 // Bits [5:4] of the constant are the destination select. This is the
4848 // value of the incoming immediate.
4849 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4850 // combine either bitwise AND or insert of float 0.0 to set these bits.
4851 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4852 // Create this as a scalar to vector..
4853 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4854 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4855 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4856 // PINSR* works with constant index.
4863 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4864 EVT VT = Op.getValueType();
4865 EVT EltVT = VT.getVectorElementType();
4867 if (Subtarget->hasSSE41())
4868 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4870 if (EltVT == MVT::i8)
4873 DebugLoc dl = Op.getDebugLoc();
4874 SDValue N0 = Op.getOperand(0);
4875 SDValue N1 = Op.getOperand(1);
4876 SDValue N2 = Op.getOperand(2);
4878 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4879 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4880 // as its second argument.
4881 if (N1.getValueType() != MVT::i32)
4882 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4883 if (N2.getValueType() != MVT::i32)
4884 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4885 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
4886 dl, VT, N0, N1, N2);
4892 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4893 DebugLoc dl = Op.getDebugLoc();
4894 if (Op.getValueType() == MVT::v2f32)
4895 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4896 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4897 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4898 Op.getOperand(0))));
4900 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4901 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4903 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4904 EVT VT = MVT::v2i32;
4905 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4912 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4913 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4916 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4917 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4918 // one of the above mentioned nodes. It has to be wrapped because otherwise
4919 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4920 // be used to form addressing mode. These wrapped nodes will be selected
4923 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4924 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4926 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4928 unsigned char OpFlag = 0;
4929 unsigned WrapperKind = X86ISD::Wrapper;
4930 CodeModel::Model M = getTargetMachine().getCodeModel();
4932 if (Subtarget->isPICStyleRIPRel() &&
4933 (M == CodeModel::Small || M == CodeModel::Kernel))
4934 WrapperKind = X86ISD::WrapperRIP;
4935 else if (Subtarget->isPICStyleGOT())
4936 OpFlag = X86II::MO_GOTOFF;
4937 else if (Subtarget->isPICStyleStubPIC())
4938 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4940 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4942 CP->getOffset(), OpFlag);
4943 DebugLoc DL = CP->getDebugLoc();
4944 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4945 // With PIC, the address is actually $g + Offset.
4947 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4948 DAG.getNode(X86ISD::GlobalBaseReg,
4949 DebugLoc::getUnknownLoc(), getPointerTy()),
4956 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4957 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4959 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4961 unsigned char OpFlag = 0;
4962 unsigned WrapperKind = X86ISD::Wrapper;
4963 CodeModel::Model M = getTargetMachine().getCodeModel();
4965 if (Subtarget->isPICStyleRIPRel() &&
4966 (M == CodeModel::Small || M == CodeModel::Kernel))
4967 WrapperKind = X86ISD::WrapperRIP;
4968 else if (Subtarget->isPICStyleGOT())
4969 OpFlag = X86II::MO_GOTOFF;
4970 else if (Subtarget->isPICStyleStubPIC())
4971 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4973 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4975 DebugLoc DL = JT->getDebugLoc();
4976 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4978 // With PIC, the address is actually $g + Offset.
4980 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4981 DAG.getNode(X86ISD::GlobalBaseReg,
4982 DebugLoc::getUnknownLoc(), getPointerTy()),
4990 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4991 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4993 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4995 unsigned char OpFlag = 0;
4996 unsigned WrapperKind = X86ISD::Wrapper;
4997 CodeModel::Model M = getTargetMachine().getCodeModel();
4999 if (Subtarget->isPICStyleRIPRel() &&
5000 (M == CodeModel::Small || M == CodeModel::Kernel))
5001 WrapperKind = X86ISD::WrapperRIP;
5002 else if (Subtarget->isPICStyleGOT())
5003 OpFlag = X86II::MO_GOTOFF;
5004 else if (Subtarget->isPICStyleStubPIC())
5005 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5007 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5009 DebugLoc DL = Op.getDebugLoc();
5010 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5013 // With PIC, the address is actually $g + Offset.
5014 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5015 !Subtarget->is64Bit()) {
5016 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5017 DAG.getNode(X86ISD::GlobalBaseReg,
5018 DebugLoc::getUnknownLoc(),
5027 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
5028 // Create the TargetBlockAddressAddress node.
5029 unsigned char OpFlags =
5030 Subtarget->ClassifyBlockAddressReference();
5031 CodeModel::Model M = getTargetMachine().getCodeModel();
5032 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5033 DebugLoc dl = Op.getDebugLoc();
5034 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5035 /*isTarget=*/true, OpFlags);
5037 if (Subtarget->isPICStyleRIPRel() &&
5038 (M == CodeModel::Small || M == CodeModel::Kernel))
5039 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5041 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5043 // With PIC, the address is actually $g + Offset.
5044 if (isGlobalRelativeToPICBase(OpFlags)) {
5045 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5046 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5054 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5056 SelectionDAG &DAG) const {
5057 // Create the TargetGlobalAddress node, folding in the constant
5058 // offset if it is legal.
5059 unsigned char OpFlags =
5060 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5061 CodeModel::Model M = getTargetMachine().getCodeModel();
5063 if (OpFlags == X86II::MO_NO_FLAG &&
5064 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5065 // A direct static reference to a global.
5066 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5069 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5072 if (Subtarget->isPICStyleRIPRel() &&
5073 (M == CodeModel::Small || M == CodeModel::Kernel))
5074 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5076 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5078 // With PIC, the address is actually $g + Offset.
5079 if (isGlobalRelativeToPICBase(OpFlags)) {
5080 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5081 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5085 // For globals that require a load from a stub to get the address, emit the
5087 if (isGlobalStubReference(OpFlags))
5088 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5089 PseudoSourceValue::getGOT(), 0, false, false, 0);
5091 // If there was a non-zero offset that we didn't fold, create an explicit
5094 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5095 DAG.getConstant(Offset, getPointerTy()));
5101 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
5102 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5103 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5104 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5108 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5109 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5110 unsigned char OperandFlags) {
5111 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5112 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5113 DebugLoc dl = GA->getDebugLoc();
5114 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5115 GA->getValueType(0),
5119 SDValue Ops[] = { Chain, TGA, *InFlag };
5120 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5122 SDValue Ops[] = { Chain, TGA };
5123 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5126 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5127 MFI->setHasCalls(true);
5129 SDValue Flag = Chain.getValue(1);
5130 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5133 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5135 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5138 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5139 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5140 DAG.getNode(X86ISD::GlobalBaseReg,
5141 DebugLoc::getUnknownLoc(),
5143 InFlag = Chain.getValue(1);
5145 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5148 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5150 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5152 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5153 X86::RAX, X86II::MO_TLSGD);
5156 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5157 // "local exec" model.
5158 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5159 const EVT PtrVT, TLSModel::Model model,
5161 DebugLoc dl = GA->getDebugLoc();
5162 // Get the Thread Pointer
5163 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5164 DebugLoc::getUnknownLoc(), PtrVT,
5165 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5168 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5169 NULL, 0, false, false, 0);
5171 unsigned char OperandFlags = 0;
5172 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5174 unsigned WrapperKind = X86ISD::Wrapper;
5175 if (model == TLSModel::LocalExec) {
5176 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5177 } else if (is64Bit) {
5178 assert(model == TLSModel::InitialExec);
5179 OperandFlags = X86II::MO_GOTTPOFF;
5180 WrapperKind = X86ISD::WrapperRIP;
5182 assert(model == TLSModel::InitialExec);
5183 OperandFlags = X86II::MO_INDNTPOFF;
5186 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5188 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5189 GA->getOffset(), OperandFlags);
5190 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5192 if (model == TLSModel::InitialExec)
5193 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5194 PseudoSourceValue::getGOT(), 0, false, false, 0);
5196 // The address of the thread local variable is the add of the thread
5197 // pointer with the offset of the variable.
5198 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5202 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
5203 // TODO: implement the "local dynamic" model
5204 // TODO: implement the "initial exec"model for pic executables
5205 assert(Subtarget->isTargetELF() &&
5206 "TLS not implemented for non-ELF targets");
5207 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5208 const GlobalValue *GV = GA->getGlobal();
5210 // If GV is an alias then use the aliasee for determining
5211 // thread-localness.
5212 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5213 GV = GA->resolveAliasedGlobal(false);
5215 TLSModel::Model model = getTLSModel(GV,
5216 getTargetMachine().getRelocationModel());
5219 case TLSModel::GeneralDynamic:
5220 case TLSModel::LocalDynamic: // not implemented
5221 if (Subtarget->is64Bit())
5222 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5223 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5225 case TLSModel::InitialExec:
5226 case TLSModel::LocalExec:
5227 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5228 Subtarget->is64Bit());
5231 llvm_unreachable("Unreachable");
5236 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5237 /// take a 2 x i32 value to shift plus a shift amount.
5238 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
5239 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5240 EVT VT = Op.getValueType();
5241 unsigned VTBits = VT.getSizeInBits();
5242 DebugLoc dl = Op.getDebugLoc();
5243 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5244 SDValue ShOpLo = Op.getOperand(0);
5245 SDValue ShOpHi = Op.getOperand(1);
5246 SDValue ShAmt = Op.getOperand(2);
5247 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5248 DAG.getConstant(VTBits - 1, MVT::i8))
5249 : DAG.getConstant(0, VT);
5252 if (Op.getOpcode() == ISD::SHL_PARTS) {
5253 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5254 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5256 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5257 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5260 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5261 DAG.getConstant(VTBits, MVT::i8));
5262 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5263 AndNode, DAG.getConstant(0, MVT::i8));
5266 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5267 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5268 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5270 if (Op.getOpcode() == ISD::SHL_PARTS) {
5271 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5272 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5274 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5275 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5278 SDValue Ops[2] = { Lo, Hi };
5279 return DAG.getMergeValues(Ops, 2, dl);
5282 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5283 EVT SrcVT = Op.getOperand(0).getValueType();
5285 if (SrcVT.isVector()) {
5286 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5292 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5293 "Unknown SINT_TO_FP to lower!");
5295 // These are really Legal; return the operand so the caller accepts it as
5297 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5299 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5300 Subtarget->is64Bit()) {
5304 DebugLoc dl = Op.getDebugLoc();
5305 unsigned Size = SrcVT.getSizeInBits()/8;
5306 MachineFunction &MF = DAG.getMachineFunction();
5307 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5308 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5309 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5311 PseudoSourceValue::getFixedStack(SSFI), 0,
5313 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5316 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5318 SelectionDAG &DAG) {
5320 DebugLoc dl = Op.getDebugLoc();
5322 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5324 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5326 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5327 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5328 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5329 Tys, Ops, array_lengthof(Ops));
5332 Chain = Result.getValue(1);
5333 SDValue InFlag = Result.getValue(2);
5335 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5336 // shouldn't be necessary except that RFP cannot be live across
5337 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5338 MachineFunction &MF = DAG.getMachineFunction();
5339 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5340 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5341 Tys = DAG.getVTList(MVT::Other);
5343 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5345 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5346 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5347 PseudoSourceValue::getFixedStack(SSFI), 0,
5354 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5355 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5356 // This algorithm is not obvious. Here it is in C code, more or less:
5358 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5359 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5360 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5362 // Copy ints to xmm registers.
5363 __m128i xh = _mm_cvtsi32_si128( hi );
5364 __m128i xl = _mm_cvtsi32_si128( lo );
5366 // Combine into low half of a single xmm register.
5367 __m128i x = _mm_unpacklo_epi32( xh, xl );
5371 // Merge in appropriate exponents to give the integer bits the right
5373 x = _mm_unpacklo_epi32( x, exp );
5375 // Subtract away the biases to deal with the IEEE-754 double precision
5377 d = _mm_sub_pd( (__m128d) x, bias );
5379 // All conversions up to here are exact. The correctly rounded result is
5380 // calculated using the current rounding mode using the following
5382 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5383 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5384 // store doesn't really need to be here (except
5385 // maybe to zero the other double)
5390 DebugLoc dl = Op.getDebugLoc();
5391 LLVMContext *Context = DAG.getContext();
5393 // Build some magic constants.
5394 std::vector<Constant*> CV0;
5395 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5396 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5397 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5398 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5399 Constant *C0 = ConstantVector::get(CV0);
5400 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5402 std::vector<Constant*> CV1;
5404 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5406 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5407 Constant *C1 = ConstantVector::get(CV1);
5408 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5410 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5411 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5413 DAG.getIntPtrConstant(1)));
5414 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5415 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5417 DAG.getIntPtrConstant(0)));
5418 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5419 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5420 PseudoSourceValue::getConstantPool(), 0,
5422 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5423 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5424 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5425 PseudoSourceValue::getConstantPool(), 0,
5427 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5429 // Add the halves; easiest way is to swap them into another reg first.
5430 int ShufMask[2] = { 1, -1 };
5431 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5432 DAG.getUNDEF(MVT::v2f64), ShufMask);
5433 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5434 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5435 DAG.getIntPtrConstant(0));
5438 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5439 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5440 DebugLoc dl = Op.getDebugLoc();
5441 // FP constant to bias correct the final result.
5442 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5445 // Load the 32-bit value into an XMM register.
5446 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5447 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5449 DAG.getIntPtrConstant(0)));
5451 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5452 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5453 DAG.getIntPtrConstant(0));
5455 // Or the load with the bias.
5456 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5457 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5458 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5460 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5461 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5462 MVT::v2f64, Bias)));
5463 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5464 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5465 DAG.getIntPtrConstant(0));
5467 // Subtract the bias.
5468 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5470 // Handle final rounding.
5471 EVT DestVT = Op.getValueType();
5473 if (DestVT.bitsLT(MVT::f64)) {
5474 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5475 DAG.getIntPtrConstant(0));
5476 } else if (DestVT.bitsGT(MVT::f64)) {
5477 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5480 // Handle final rounding.
5484 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5485 SDValue N0 = Op.getOperand(0);
5486 DebugLoc dl = Op.getDebugLoc();
5488 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5489 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5490 // the optimization here.
5491 if (DAG.SignBitIsZero(N0))
5492 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5494 EVT SrcVT = N0.getValueType();
5495 if (SrcVT == MVT::i64) {
5496 // We only handle SSE2 f64 target here; caller can expand the rest.
5497 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5500 return LowerUINT_TO_FP_i64(Op, DAG);
5501 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5502 return LowerUINT_TO_FP_i32(Op, DAG);
5505 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5507 // Make a 64-bit buffer, and use it to build an FILD.
5508 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5509 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5510 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5511 getPointerTy(), StackSlot, WordOff);
5512 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5513 StackSlot, NULL, 0, false, false, 0);
5514 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5515 OffsetSlot, NULL, 0, false, false, 0);
5516 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5519 std::pair<SDValue,SDValue> X86TargetLowering::
5520 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5521 DebugLoc dl = Op.getDebugLoc();
5523 EVT DstTy = Op.getValueType();
5526 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5530 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5531 DstTy.getSimpleVT() >= MVT::i16 &&
5532 "Unknown FP_TO_SINT to lower!");
5534 // These are really Legal.
5535 if (DstTy == MVT::i32 &&
5536 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5537 return std::make_pair(SDValue(), SDValue());
5538 if (Subtarget->is64Bit() &&
5539 DstTy == MVT::i64 &&
5540 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5541 return std::make_pair(SDValue(), SDValue());
5543 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5545 MachineFunction &MF = DAG.getMachineFunction();
5546 unsigned MemSize = DstTy.getSizeInBits()/8;
5547 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5548 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5551 switch (DstTy.getSimpleVT().SimpleTy) {
5552 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5553 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5554 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5555 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5558 SDValue Chain = DAG.getEntryNode();
5559 SDValue Value = Op.getOperand(0);
5560 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5561 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5562 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5563 PseudoSourceValue::getFixedStack(SSFI), 0,
5565 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5567 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5569 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5570 Chain = Value.getValue(1);
5571 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5572 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5575 // Build the FP_TO_INT*_IN_MEM
5576 SDValue Ops[] = { Chain, Value, StackSlot };
5577 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5579 return std::make_pair(FIST, StackSlot);
5582 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5583 if (Op.getValueType().isVector()) {
5584 if (Op.getValueType() == MVT::v2i32 &&
5585 Op.getOperand(0).getValueType() == MVT::v2f64) {
5591 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5592 SDValue FIST = Vals.first, StackSlot = Vals.second;
5593 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5594 if (FIST.getNode() == 0) return Op;
5597 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5598 FIST, StackSlot, NULL, 0, false, false, 0);
5601 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5602 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5603 SDValue FIST = Vals.first, StackSlot = Vals.second;
5604 assert(FIST.getNode() && "Unexpected failure");
5607 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5608 FIST, StackSlot, NULL, 0, false, false, 0);
5611 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5612 LLVMContext *Context = DAG.getContext();
5613 DebugLoc dl = Op.getDebugLoc();
5614 EVT VT = Op.getValueType();
5617 EltVT = VT.getVectorElementType();
5618 std::vector<Constant*> CV;
5619 if (EltVT == MVT::f64) {
5620 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5624 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5630 Constant *C = ConstantVector::get(CV);
5631 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5632 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5633 PseudoSourceValue::getConstantPool(), 0,
5635 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5638 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5639 LLVMContext *Context = DAG.getContext();
5640 DebugLoc dl = Op.getDebugLoc();
5641 EVT VT = Op.getValueType();
5644 EltVT = VT.getVectorElementType();
5645 std::vector<Constant*> CV;
5646 if (EltVT == MVT::f64) {
5647 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5651 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5657 Constant *C = ConstantVector::get(CV);
5658 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5659 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5660 PseudoSourceValue::getConstantPool(), 0,
5662 if (VT.isVector()) {
5663 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5664 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5665 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5667 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5669 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5673 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5674 LLVMContext *Context = DAG.getContext();
5675 SDValue Op0 = Op.getOperand(0);
5676 SDValue Op1 = Op.getOperand(1);
5677 DebugLoc dl = Op.getDebugLoc();
5678 EVT VT = Op.getValueType();
5679 EVT SrcVT = Op1.getValueType();
5681 // If second operand is smaller, extend it first.
5682 if (SrcVT.bitsLT(VT)) {
5683 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5686 // And if it is bigger, shrink it first.
5687 if (SrcVT.bitsGT(VT)) {
5688 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5692 // At this point the operands and the result should have the same
5693 // type, and that won't be f80 since that is not custom lowered.
5695 // First get the sign bit of second operand.
5696 std::vector<Constant*> CV;
5697 if (SrcVT == MVT::f64) {
5698 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5699 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5701 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5702 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5703 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5704 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5706 Constant *C = ConstantVector::get(CV);
5707 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5708 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5709 PseudoSourceValue::getConstantPool(), 0,
5711 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5713 // Shift sign bit right or left if the two operands have different types.
5714 if (SrcVT.bitsGT(VT)) {
5715 // Op0 is MVT::f32, Op1 is MVT::f64.
5716 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5717 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5718 DAG.getConstant(32, MVT::i32));
5719 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5720 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5721 DAG.getIntPtrConstant(0));
5724 // Clear first operand sign bit.
5726 if (VT == MVT::f64) {
5727 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5728 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5730 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5731 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5732 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5733 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5735 C = ConstantVector::get(CV);
5736 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5737 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5738 PseudoSourceValue::getConstantPool(), 0,
5740 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5742 // Or the value with the sign bit.
5743 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5746 /// Emit nodes that will be selected as "test Op0,Op0", or something
5748 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5749 SelectionDAG &DAG) {
5750 DebugLoc dl = Op.getDebugLoc();
5752 // CF and OF aren't always set the way we want. Determine which
5753 // of these we need.
5754 bool NeedCF = false;
5755 bool NeedOF = false;
5757 case X86::COND_A: case X86::COND_AE:
5758 case X86::COND_B: case X86::COND_BE:
5761 case X86::COND_G: case X86::COND_GE:
5762 case X86::COND_L: case X86::COND_LE:
5763 case X86::COND_O: case X86::COND_NO:
5769 // See if we can use the EFLAGS value from the operand instead of
5770 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5771 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5772 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5773 unsigned Opcode = 0;
5774 unsigned NumOperands = 0;
5775 switch (Op.getNode()->getOpcode()) {
5777 // Due to an isel shortcoming, be conservative if this add is likely to
5778 // be selected as part of a load-modify-store instruction. When the root
5779 // node in a match is a store, isel doesn't know how to remap non-chain
5780 // non-flag uses of other nodes in the match, such as the ADD in this
5781 // case. This leads to the ADD being left around and reselected, with
5782 // the result being two adds in the output.
5783 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5784 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5785 if (UI->getOpcode() == ISD::STORE)
5787 if (ConstantSDNode *C =
5788 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5789 // An add of one will be selected as an INC.
5790 if (C->getAPIntValue() == 1) {
5791 Opcode = X86ISD::INC;
5795 // An add of negative one (subtract of one) will be selected as a DEC.
5796 if (C->getAPIntValue().isAllOnesValue()) {
5797 Opcode = X86ISD::DEC;
5802 // Otherwise use a regular EFLAGS-setting add.
5803 Opcode = X86ISD::ADD;
5807 // If the primary and result isn't used, don't bother using X86ISD::AND,
5808 // because a TEST instruction will be better.
5809 bool NonFlagUse = false;
5810 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5811 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
5813 unsigned UOpNo = UI.getOperandNo();
5814 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
5815 // Look pass truncate.
5816 UOpNo = User->use_begin().getOperandNo();
5817 User = *User->use_begin();
5819 if (User->getOpcode() != ISD::BRCOND &&
5820 User->getOpcode() != ISD::SETCC &&
5821 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
5833 // Due to the ISEL shortcoming noted above, be conservative if this op is
5834 // likely to be selected as part of a load-modify-store instruction.
5835 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5836 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5837 if (UI->getOpcode() == ISD::STORE)
5839 // Otherwise use a regular EFLAGS-setting instruction.
5840 switch (Op.getNode()->getOpcode()) {
5841 case ISD::SUB: Opcode = X86ISD::SUB; break;
5842 case ISD::OR: Opcode = X86ISD::OR; break;
5843 case ISD::XOR: Opcode = X86ISD::XOR; break;
5844 case ISD::AND: Opcode = X86ISD::AND; break;
5845 default: llvm_unreachable("unexpected operator!");
5856 return SDValue(Op.getNode(), 1);
5862 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5863 SmallVector<SDValue, 4> Ops;
5864 for (unsigned i = 0; i != NumOperands; ++i)
5865 Ops.push_back(Op.getOperand(i));
5866 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5867 DAG.ReplaceAllUsesWith(Op, New);
5868 return SDValue(New.getNode(), 1);
5872 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5873 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5874 DAG.getConstant(0, Op.getValueType()));
5877 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5879 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5880 SelectionDAG &DAG) {
5881 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5882 if (C->getAPIntValue() == 0)
5883 return EmitTest(Op0, X86CC, DAG);
5885 DebugLoc dl = Op0.getDebugLoc();
5886 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5889 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
5890 /// if it's possible.
5891 static SDValue LowerToBT(SDValue And, ISD::CondCode CC,
5892 DebugLoc dl, SelectionDAG &DAG) {
5893 SDValue Op0 = And.getOperand(0);
5894 SDValue Op1 = And.getOperand(1);
5895 if (Op0.getOpcode() == ISD::TRUNCATE)
5896 Op0 = Op0.getOperand(0);
5897 if (Op1.getOpcode() == ISD::TRUNCATE)
5898 Op1 = Op1.getOperand(0);
5901 if (Op1.getOpcode() == ISD::SHL) {
5902 if (ConstantSDNode *And10C = dyn_cast<ConstantSDNode>(Op1.getOperand(0)))
5903 if (And10C->getZExtValue() == 1) {
5905 RHS = Op1.getOperand(1);
5907 } else if (Op0.getOpcode() == ISD::SHL) {
5908 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
5909 if (And00C->getZExtValue() == 1) {
5911 RHS = Op0.getOperand(1);
5913 } else if (Op1.getOpcode() == ISD::Constant) {
5914 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
5915 SDValue AndLHS = Op0;
5916 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5917 LHS = AndLHS.getOperand(0);
5918 RHS = AndLHS.getOperand(1);
5922 if (LHS.getNode()) {
5923 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5924 // instruction. Since the shift amount is in-range-or-undefined, we know
5925 // that doing a bittest on the i16 value is ok. We extend to i32 because
5926 // the encoding for the i16 version is larger than the i32 version.
5927 if (LHS.getValueType() == MVT::i8)
5928 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5930 // If the operand types disagree, extend the shift amount to match. Since
5931 // BT ignores high bits (like shifts) we can use anyextend.
5932 if (LHS.getValueType() != RHS.getValueType())
5933 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5935 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5936 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5937 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5938 DAG.getConstant(Cond, MVT::i8), BT);
5944 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5945 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5946 SDValue Op0 = Op.getOperand(0);
5947 SDValue Op1 = Op.getOperand(1);
5948 DebugLoc dl = Op.getDebugLoc();
5949 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5951 // Optimize to BT if possible.
5952 // Lower (X & (1 << N)) == 0 to BT(X, N).
5953 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5954 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5955 if (Op0.getOpcode() == ISD::AND &&
5957 Op1.getOpcode() == ISD::Constant &&
5958 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5959 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5960 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
5961 if (NewSetCC.getNode())
5965 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
5966 if (Op0.getOpcode() == X86ISD::SETCC &&
5967 Op1.getOpcode() == ISD::Constant &&
5968 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
5969 cast<ConstantSDNode>(Op1)->isNullValue()) &&
5970 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5971 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
5972 bool Invert = (CC == ISD::SETNE) ^
5973 cast<ConstantSDNode>(Op1)->isNullValue();
5975 CCode = X86::GetOppositeBranchCondition(CCode);
5976 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5977 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
5980 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5981 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5982 if (X86CC == X86::COND_INVALID)
5985 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5987 // Use sbb x, x to materialize carry bit into a GPR.
5988 if (X86CC == X86::COND_B)
5989 return DAG.getNode(ISD::AND, dl, MVT::i8,
5990 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
5991 DAG.getConstant(X86CC, MVT::i8), Cond),
5992 DAG.getConstant(1, MVT::i8));
5994 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5995 DAG.getConstant(X86CC, MVT::i8), Cond);
5998 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
6000 SDValue Op0 = Op.getOperand(0);
6001 SDValue Op1 = Op.getOperand(1);
6002 SDValue CC = Op.getOperand(2);
6003 EVT VT = Op.getValueType();
6004 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6005 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6006 DebugLoc dl = Op.getDebugLoc();
6010 EVT VT0 = Op0.getValueType();
6011 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6012 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6015 switch (SetCCOpcode) {
6018 case ISD::SETEQ: SSECC = 0; break;
6020 case ISD::SETGT: Swap = true; // Fallthrough
6022 case ISD::SETOLT: SSECC = 1; break;
6024 case ISD::SETGE: Swap = true; // Fallthrough
6026 case ISD::SETOLE: SSECC = 2; break;
6027 case ISD::SETUO: SSECC = 3; break;
6029 case ISD::SETNE: SSECC = 4; break;
6030 case ISD::SETULE: Swap = true;
6031 case ISD::SETUGE: SSECC = 5; break;
6032 case ISD::SETULT: Swap = true;
6033 case ISD::SETUGT: SSECC = 6; break;
6034 case ISD::SETO: SSECC = 7; break;
6037 std::swap(Op0, Op1);
6039 // In the two special cases we can't handle, emit two comparisons.
6041 if (SetCCOpcode == ISD::SETUEQ) {
6043 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6044 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6045 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6047 else if (SetCCOpcode == ISD::SETONE) {
6049 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6050 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6051 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6053 llvm_unreachable("Illegal FP comparison");
6055 // Handle all other FP comparisons here.
6056 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6059 // We are handling one of the integer comparisons here. Since SSE only has
6060 // GT and EQ comparisons for integer, swapping operands and multiple
6061 // operations may be required for some comparisons.
6062 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6063 bool Swap = false, Invert = false, FlipSigns = false;
6065 switch (VT.getSimpleVT().SimpleTy) {
6068 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6070 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6072 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6073 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6076 switch (SetCCOpcode) {
6078 case ISD::SETNE: Invert = true;
6079 case ISD::SETEQ: Opc = EQOpc; break;
6080 case ISD::SETLT: Swap = true;
6081 case ISD::SETGT: Opc = GTOpc; break;
6082 case ISD::SETGE: Swap = true;
6083 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6084 case ISD::SETULT: Swap = true;
6085 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6086 case ISD::SETUGE: Swap = true;
6087 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6090 std::swap(Op0, Op1);
6092 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6093 // bits of the inputs before performing those operations.
6095 EVT EltVT = VT.getVectorElementType();
6096 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6098 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6099 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6101 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6102 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6105 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6107 // If the logical-not of the result is required, perform that now.
6109 Result = DAG.getNOT(dl, Result, VT);
6114 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6115 static bool isX86LogicalCmp(SDValue Op) {
6116 unsigned Opc = Op.getNode()->getOpcode();
6117 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6119 if (Op.getResNo() == 1 &&
6120 (Opc == X86ISD::ADD ||
6121 Opc == X86ISD::SUB ||
6122 Opc == X86ISD::SMUL ||
6123 Opc == X86ISD::UMUL ||
6124 Opc == X86ISD::INC ||
6125 Opc == X86ISD::DEC ||
6126 Opc == X86ISD::OR ||
6127 Opc == X86ISD::XOR ||
6128 Opc == X86ISD::AND))
6134 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
6135 bool addTest = true;
6136 SDValue Cond = Op.getOperand(0);
6137 DebugLoc dl = Op.getDebugLoc();
6140 if (Cond.getOpcode() == ISD::SETCC) {
6141 SDValue NewCond = LowerSETCC(Cond, DAG);
6142 if (NewCond.getNode())
6146 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6147 SDValue Op1 = Op.getOperand(1);
6148 SDValue Op2 = Op.getOperand(2);
6149 if (Cond.getOpcode() == X86ISD::SETCC &&
6150 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6151 SDValue Cmp = Cond.getOperand(1);
6152 if (Cmp.getOpcode() == X86ISD::CMP) {
6153 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6154 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6155 ConstantSDNode *RHSC =
6156 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6157 if (N1C && N1C->isAllOnesValue() &&
6158 N2C && N2C->isNullValue() &&
6159 RHSC && RHSC->isNullValue()) {
6160 SDValue CmpOp0 = Cmp.getOperand(0);
6161 Cmp = DAG.getNode(X86ISD::CMP, dl, CmpOp0.getValueType(),
6162 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6163 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6164 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6169 // Look pass (and (setcc_carry (cmp ...)), 1).
6170 if (Cond.getOpcode() == ISD::AND &&
6171 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6172 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6173 if (C && C->getAPIntValue() == 1)
6174 Cond = Cond.getOperand(0);
6177 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6178 // setting operand in place of the X86ISD::SETCC.
6179 if (Cond.getOpcode() == X86ISD::SETCC ||
6180 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6181 CC = Cond.getOperand(0);
6183 SDValue Cmp = Cond.getOperand(1);
6184 unsigned Opc = Cmp.getOpcode();
6185 EVT VT = Op.getValueType();
6187 bool IllegalFPCMov = false;
6188 if (VT.isFloatingPoint() && !VT.isVector() &&
6189 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6190 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6192 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6193 Opc == X86ISD::BT) { // FIXME
6200 // Look pass the truncate.
6201 if (Cond.getOpcode() == ISD::TRUNCATE)
6202 Cond = Cond.getOperand(0);
6204 // We know the result of AND is compared against zero. Try to match
6206 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6207 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6208 if (NewSetCC.getNode()) {
6209 CC = NewSetCC.getOperand(0);
6210 Cond = NewSetCC.getOperand(1);
6217 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6218 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6221 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6222 // condition is true.
6223 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6224 SDValue Ops[] = { Op2, Op1, CC, Cond };
6225 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6228 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6229 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6230 // from the AND / OR.
6231 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6232 Opc = Op.getOpcode();
6233 if (Opc != ISD::OR && Opc != ISD::AND)
6235 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6236 Op.getOperand(0).hasOneUse() &&
6237 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6238 Op.getOperand(1).hasOneUse());
6241 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6242 // 1 and that the SETCC node has a single use.
6243 static bool isXor1OfSetCC(SDValue Op) {
6244 if (Op.getOpcode() != ISD::XOR)
6246 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6247 if (N1C && N1C->getAPIntValue() == 1) {
6248 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6249 Op.getOperand(0).hasOneUse();
6254 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
6255 bool addTest = true;
6256 SDValue Chain = Op.getOperand(0);
6257 SDValue Cond = Op.getOperand(1);
6258 SDValue Dest = Op.getOperand(2);
6259 DebugLoc dl = Op.getDebugLoc();
6262 if (Cond.getOpcode() == ISD::SETCC) {
6263 SDValue NewCond = LowerSETCC(Cond, DAG);
6264 if (NewCond.getNode())
6268 // FIXME: LowerXALUO doesn't handle these!!
6269 else if (Cond.getOpcode() == X86ISD::ADD ||
6270 Cond.getOpcode() == X86ISD::SUB ||
6271 Cond.getOpcode() == X86ISD::SMUL ||
6272 Cond.getOpcode() == X86ISD::UMUL)
6273 Cond = LowerXALUO(Cond, DAG);
6276 // Look pass (and (setcc_carry (cmp ...)), 1).
6277 if (Cond.getOpcode() == ISD::AND &&
6278 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6279 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6280 if (C && C->getAPIntValue() == 1)
6281 Cond = Cond.getOperand(0);
6284 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6285 // setting operand in place of the X86ISD::SETCC.
6286 if (Cond.getOpcode() == X86ISD::SETCC ||
6287 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6288 CC = Cond.getOperand(0);
6290 SDValue Cmp = Cond.getOperand(1);
6291 unsigned Opc = Cmp.getOpcode();
6292 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6293 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6297 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6301 // These can only come from an arithmetic instruction with overflow,
6302 // e.g. SADDO, UADDO.
6303 Cond = Cond.getNode()->getOperand(1);
6310 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6311 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6312 if (CondOpc == ISD::OR) {
6313 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6314 // two branches instead of an explicit OR instruction with a
6316 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6317 isX86LogicalCmp(Cmp)) {
6318 CC = Cond.getOperand(0).getOperand(0);
6319 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6320 Chain, Dest, CC, Cmp);
6321 CC = Cond.getOperand(1).getOperand(0);
6325 } else { // ISD::AND
6326 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6327 // two branches instead of an explicit AND instruction with a
6328 // separate test. However, we only do this if this block doesn't
6329 // have a fall-through edge, because this requires an explicit
6330 // jmp when the condition is false.
6331 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6332 isX86LogicalCmp(Cmp) &&
6333 Op.getNode()->hasOneUse()) {
6334 X86::CondCode CCode =
6335 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6336 CCode = X86::GetOppositeBranchCondition(CCode);
6337 CC = DAG.getConstant(CCode, MVT::i8);
6338 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6339 // Look for an unconditional branch following this conditional branch.
6340 // We need this because we need to reverse the successors in order
6341 // to implement FCMP_OEQ.
6342 if (User.getOpcode() == ISD::BR) {
6343 SDValue FalseBB = User.getOperand(1);
6345 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6346 assert(NewBR == User);
6349 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6350 Chain, Dest, CC, Cmp);
6351 X86::CondCode CCode =
6352 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6353 CCode = X86::GetOppositeBranchCondition(CCode);
6354 CC = DAG.getConstant(CCode, MVT::i8);
6360 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6361 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6362 // It should be transformed during dag combiner except when the condition
6363 // is set by a arithmetics with overflow node.
6364 X86::CondCode CCode =
6365 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6366 CCode = X86::GetOppositeBranchCondition(CCode);
6367 CC = DAG.getConstant(CCode, MVT::i8);
6368 Cond = Cond.getOperand(0).getOperand(1);
6374 // Look pass the truncate.
6375 if (Cond.getOpcode() == ISD::TRUNCATE)
6376 Cond = Cond.getOperand(0);
6378 // We know the result of AND is compared against zero. Try to match
6380 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6381 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6382 if (NewSetCC.getNode()) {
6383 CC = NewSetCC.getOperand(0);
6384 Cond = NewSetCC.getOperand(1);
6391 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6392 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6394 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6395 Chain, Dest, CC, Cond);
6399 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6400 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6401 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6402 // that the guard pages used by the OS virtual memory manager are allocated in
6403 // correct sequence.
6405 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6406 SelectionDAG &DAG) {
6407 assert(Subtarget->isTargetCygMing() &&
6408 "This should be used only on Cygwin/Mingw targets");
6409 DebugLoc dl = Op.getDebugLoc();
6412 SDValue Chain = Op.getOperand(0);
6413 SDValue Size = Op.getOperand(1);
6414 // FIXME: Ensure alignment here
6418 EVT IntPtr = getPointerTy();
6419 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6421 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6422 Flag = Chain.getValue(1);
6424 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6426 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6427 Flag = Chain.getValue(1);
6429 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6431 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6432 return DAG.getMergeValues(Ops1, 2, dl);
6436 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6438 SDValue Dst, SDValue Src,
6439 SDValue Size, unsigned Align,
6441 uint64_t DstSVOff) {
6442 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6444 // If not DWORD aligned or size is more than the threshold, call the library.
6445 // The libc version is likely to be faster for these cases. It can use the
6446 // address value and run time information about the CPU.
6447 if ((Align & 3) != 0 ||
6449 ConstantSize->getZExtValue() >
6450 getSubtarget()->getMaxInlineSizeThreshold()) {
6451 SDValue InFlag(0, 0);
6453 // Check to see if there is a specialized entry-point for memory zeroing.
6454 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6456 if (const char *bzeroEntry = V &&
6457 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6458 EVT IntPtr = getPointerTy();
6459 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6460 TargetLowering::ArgListTy Args;
6461 TargetLowering::ArgListEntry Entry;
6463 Entry.Ty = IntPtrTy;
6464 Args.push_back(Entry);
6466 Args.push_back(Entry);
6467 std::pair<SDValue,SDValue> CallResult =
6468 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6469 false, false, false, false,
6470 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6471 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
6472 return CallResult.second;
6475 // Otherwise have the target-independent code call memset.
6479 uint64_t SizeVal = ConstantSize->getZExtValue();
6480 SDValue InFlag(0, 0);
6483 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6484 unsigned BytesLeft = 0;
6485 bool TwoRepStos = false;
6488 uint64_t Val = ValC->getZExtValue() & 255;
6490 // If the value is a constant, then we can potentially use larger sets.
6491 switch (Align & 3) {
6492 case 2: // WORD aligned
6495 Val = (Val << 8) | Val;
6497 case 0: // DWORD aligned
6500 Val = (Val << 8) | Val;
6501 Val = (Val << 16) | Val;
6502 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6505 Val = (Val << 32) | Val;
6508 default: // Byte aligned
6511 Count = DAG.getIntPtrConstant(SizeVal);
6515 if (AVT.bitsGT(MVT::i8)) {
6516 unsigned UBytes = AVT.getSizeInBits() / 8;
6517 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6518 BytesLeft = SizeVal % UBytes;
6521 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6523 InFlag = Chain.getValue(1);
6526 Count = DAG.getIntPtrConstant(SizeVal);
6527 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6528 InFlag = Chain.getValue(1);
6531 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6534 InFlag = Chain.getValue(1);
6535 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6538 InFlag = Chain.getValue(1);
6540 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6541 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6542 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6545 InFlag = Chain.getValue(1);
6547 EVT CVT = Count.getValueType();
6548 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6549 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6550 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6553 InFlag = Chain.getValue(1);
6554 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6555 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
6556 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6557 } else if (BytesLeft) {
6558 // Handle the last 1 - 7 bytes.
6559 unsigned Offset = SizeVal - BytesLeft;
6560 EVT AddrVT = Dst.getValueType();
6561 EVT SizeVT = Size.getValueType();
6563 Chain = DAG.getMemset(Chain, dl,
6564 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6565 DAG.getConstant(Offset, AddrVT)),
6567 DAG.getConstant(BytesLeft, SizeVT),
6568 Align, DstSV, DstSVOff + Offset);
6571 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6576 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6577 SDValue Chain, SDValue Dst, SDValue Src,
6578 SDValue Size, unsigned Align,
6580 const Value *DstSV, uint64_t DstSVOff,
6581 const Value *SrcSV, uint64_t SrcSVOff) {
6582 // This requires the copy size to be a constant, preferrably
6583 // within a subtarget-specific limit.
6584 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6587 uint64_t SizeVal = ConstantSize->getZExtValue();
6588 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6591 /// If not DWORD aligned, call the library.
6592 if ((Align & 3) != 0)
6597 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6600 unsigned UBytes = AVT.getSizeInBits() / 8;
6601 unsigned CountVal = SizeVal / UBytes;
6602 SDValue Count = DAG.getIntPtrConstant(CountVal);
6603 unsigned BytesLeft = SizeVal % UBytes;
6605 SDValue InFlag(0, 0);
6606 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6609 InFlag = Chain.getValue(1);
6610 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6613 InFlag = Chain.getValue(1);
6614 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6617 InFlag = Chain.getValue(1);
6619 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6620 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6621 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
6622 array_lengthof(Ops));
6624 SmallVector<SDValue, 4> Results;
6625 Results.push_back(RepMovs);
6627 // Handle the last 1 - 7 bytes.
6628 unsigned Offset = SizeVal - BytesLeft;
6629 EVT DstVT = Dst.getValueType();
6630 EVT SrcVT = Src.getValueType();
6631 EVT SizeVT = Size.getValueType();
6632 Results.push_back(DAG.getMemcpy(Chain, dl,
6633 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6634 DAG.getConstant(Offset, DstVT)),
6635 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6636 DAG.getConstant(Offset, SrcVT)),
6637 DAG.getConstant(BytesLeft, SizeVT),
6638 Align, AlwaysInline,
6639 DstSV, DstSVOff + Offset,
6640 SrcSV, SrcSVOff + Offset));
6643 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6644 &Results[0], Results.size());
6647 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6648 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6649 DebugLoc dl = Op.getDebugLoc();
6651 if (!Subtarget->is64Bit()) {
6652 // vastart just stores the address of the VarArgsFrameIndex slot into the
6653 // memory location argument.
6654 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6655 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6660 // gp_offset (0 - 6 * 8)
6661 // fp_offset (48 - 48 + 8 * 16)
6662 // overflow_arg_area (point to parameters coming in memory).
6664 SmallVector<SDValue, 8> MemOps;
6665 SDValue FIN = Op.getOperand(1);
6667 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6668 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6669 FIN, SV, 0, false, false, 0);
6670 MemOps.push_back(Store);
6673 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6674 FIN, DAG.getIntPtrConstant(4));
6675 Store = DAG.getStore(Op.getOperand(0), dl,
6676 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6677 FIN, SV, 0, false, false, 0);
6678 MemOps.push_back(Store);
6680 // Store ptr to overflow_arg_area
6681 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6682 FIN, DAG.getIntPtrConstant(4));
6683 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6684 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0,
6686 MemOps.push_back(Store);
6688 // Store ptr to reg_save_area.
6689 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6690 FIN, DAG.getIntPtrConstant(8));
6691 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6692 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0,
6694 MemOps.push_back(Store);
6695 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6696 &MemOps[0], MemOps.size());
6699 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6700 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6701 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6702 SDValue Chain = Op.getOperand(0);
6703 SDValue SrcPtr = Op.getOperand(1);
6704 SDValue SrcSV = Op.getOperand(2);
6706 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6710 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6711 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6712 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6713 SDValue Chain = Op.getOperand(0);
6714 SDValue DstPtr = Op.getOperand(1);
6715 SDValue SrcPtr = Op.getOperand(2);
6716 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6717 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6718 DebugLoc dl = Op.getDebugLoc();
6720 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6721 DAG.getIntPtrConstant(24), 8, false,
6722 DstSV, 0, SrcSV, 0);
6726 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6727 DebugLoc dl = Op.getDebugLoc();
6728 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6730 default: return SDValue(); // Don't custom lower most intrinsics.
6731 // Comparison intrinsics.
6732 case Intrinsic::x86_sse_comieq_ss:
6733 case Intrinsic::x86_sse_comilt_ss:
6734 case Intrinsic::x86_sse_comile_ss:
6735 case Intrinsic::x86_sse_comigt_ss:
6736 case Intrinsic::x86_sse_comige_ss:
6737 case Intrinsic::x86_sse_comineq_ss:
6738 case Intrinsic::x86_sse_ucomieq_ss:
6739 case Intrinsic::x86_sse_ucomilt_ss:
6740 case Intrinsic::x86_sse_ucomile_ss:
6741 case Intrinsic::x86_sse_ucomigt_ss:
6742 case Intrinsic::x86_sse_ucomige_ss:
6743 case Intrinsic::x86_sse_ucomineq_ss:
6744 case Intrinsic::x86_sse2_comieq_sd:
6745 case Intrinsic::x86_sse2_comilt_sd:
6746 case Intrinsic::x86_sse2_comile_sd:
6747 case Intrinsic::x86_sse2_comigt_sd:
6748 case Intrinsic::x86_sse2_comige_sd:
6749 case Intrinsic::x86_sse2_comineq_sd:
6750 case Intrinsic::x86_sse2_ucomieq_sd:
6751 case Intrinsic::x86_sse2_ucomilt_sd:
6752 case Intrinsic::x86_sse2_ucomile_sd:
6753 case Intrinsic::x86_sse2_ucomigt_sd:
6754 case Intrinsic::x86_sse2_ucomige_sd:
6755 case Intrinsic::x86_sse2_ucomineq_sd: {
6757 ISD::CondCode CC = ISD::SETCC_INVALID;
6760 case Intrinsic::x86_sse_comieq_ss:
6761 case Intrinsic::x86_sse2_comieq_sd:
6765 case Intrinsic::x86_sse_comilt_ss:
6766 case Intrinsic::x86_sse2_comilt_sd:
6770 case Intrinsic::x86_sse_comile_ss:
6771 case Intrinsic::x86_sse2_comile_sd:
6775 case Intrinsic::x86_sse_comigt_ss:
6776 case Intrinsic::x86_sse2_comigt_sd:
6780 case Intrinsic::x86_sse_comige_ss:
6781 case Intrinsic::x86_sse2_comige_sd:
6785 case Intrinsic::x86_sse_comineq_ss:
6786 case Intrinsic::x86_sse2_comineq_sd:
6790 case Intrinsic::x86_sse_ucomieq_ss:
6791 case Intrinsic::x86_sse2_ucomieq_sd:
6792 Opc = X86ISD::UCOMI;
6795 case Intrinsic::x86_sse_ucomilt_ss:
6796 case Intrinsic::x86_sse2_ucomilt_sd:
6797 Opc = X86ISD::UCOMI;
6800 case Intrinsic::x86_sse_ucomile_ss:
6801 case Intrinsic::x86_sse2_ucomile_sd:
6802 Opc = X86ISD::UCOMI;
6805 case Intrinsic::x86_sse_ucomigt_ss:
6806 case Intrinsic::x86_sse2_ucomigt_sd:
6807 Opc = X86ISD::UCOMI;
6810 case Intrinsic::x86_sse_ucomige_ss:
6811 case Intrinsic::x86_sse2_ucomige_sd:
6812 Opc = X86ISD::UCOMI;
6815 case Intrinsic::x86_sse_ucomineq_ss:
6816 case Intrinsic::x86_sse2_ucomineq_sd:
6817 Opc = X86ISD::UCOMI;
6822 SDValue LHS = Op.getOperand(1);
6823 SDValue RHS = Op.getOperand(2);
6824 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6825 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6826 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6827 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6828 DAG.getConstant(X86CC, MVT::i8), Cond);
6829 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6831 // ptest intrinsics. The intrinsic these come from are designed to return
6832 // an integer value, not just an instruction so lower it to the ptest
6833 // pattern and a setcc for the result.
6834 case Intrinsic::x86_sse41_ptestz:
6835 case Intrinsic::x86_sse41_ptestc:
6836 case Intrinsic::x86_sse41_ptestnzc:{
6839 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6840 case Intrinsic::x86_sse41_ptestz:
6842 X86CC = X86::COND_E;
6844 case Intrinsic::x86_sse41_ptestc:
6846 X86CC = X86::COND_B;
6848 case Intrinsic::x86_sse41_ptestnzc:
6850 X86CC = X86::COND_A;
6854 SDValue LHS = Op.getOperand(1);
6855 SDValue RHS = Op.getOperand(2);
6856 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6857 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6858 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6859 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6862 // Fix vector shift instructions where the last operand is a non-immediate
6864 case Intrinsic::x86_sse2_pslli_w:
6865 case Intrinsic::x86_sse2_pslli_d:
6866 case Intrinsic::x86_sse2_pslli_q:
6867 case Intrinsic::x86_sse2_psrli_w:
6868 case Intrinsic::x86_sse2_psrli_d:
6869 case Intrinsic::x86_sse2_psrli_q:
6870 case Intrinsic::x86_sse2_psrai_w:
6871 case Intrinsic::x86_sse2_psrai_d:
6872 case Intrinsic::x86_mmx_pslli_w:
6873 case Intrinsic::x86_mmx_pslli_d:
6874 case Intrinsic::x86_mmx_pslli_q:
6875 case Intrinsic::x86_mmx_psrli_w:
6876 case Intrinsic::x86_mmx_psrli_d:
6877 case Intrinsic::x86_mmx_psrli_q:
6878 case Intrinsic::x86_mmx_psrai_w:
6879 case Intrinsic::x86_mmx_psrai_d: {
6880 SDValue ShAmt = Op.getOperand(2);
6881 if (isa<ConstantSDNode>(ShAmt))
6884 unsigned NewIntNo = 0;
6885 EVT ShAmtVT = MVT::v4i32;
6887 case Intrinsic::x86_sse2_pslli_w:
6888 NewIntNo = Intrinsic::x86_sse2_psll_w;
6890 case Intrinsic::x86_sse2_pslli_d:
6891 NewIntNo = Intrinsic::x86_sse2_psll_d;
6893 case Intrinsic::x86_sse2_pslli_q:
6894 NewIntNo = Intrinsic::x86_sse2_psll_q;
6896 case Intrinsic::x86_sse2_psrli_w:
6897 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6899 case Intrinsic::x86_sse2_psrli_d:
6900 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6902 case Intrinsic::x86_sse2_psrli_q:
6903 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6905 case Intrinsic::x86_sse2_psrai_w:
6906 NewIntNo = Intrinsic::x86_sse2_psra_w;
6908 case Intrinsic::x86_sse2_psrai_d:
6909 NewIntNo = Intrinsic::x86_sse2_psra_d;
6912 ShAmtVT = MVT::v2i32;
6914 case Intrinsic::x86_mmx_pslli_w:
6915 NewIntNo = Intrinsic::x86_mmx_psll_w;
6917 case Intrinsic::x86_mmx_pslli_d:
6918 NewIntNo = Intrinsic::x86_mmx_psll_d;
6920 case Intrinsic::x86_mmx_pslli_q:
6921 NewIntNo = Intrinsic::x86_mmx_psll_q;
6923 case Intrinsic::x86_mmx_psrli_w:
6924 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6926 case Intrinsic::x86_mmx_psrli_d:
6927 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6929 case Intrinsic::x86_mmx_psrli_q:
6930 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6932 case Intrinsic::x86_mmx_psrai_w:
6933 NewIntNo = Intrinsic::x86_mmx_psra_w;
6935 case Intrinsic::x86_mmx_psrai_d:
6936 NewIntNo = Intrinsic::x86_mmx_psra_d;
6938 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6944 // The vector shift intrinsics with scalars uses 32b shift amounts but
6945 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
6949 ShOps[1] = DAG.getConstant(0, MVT::i32);
6950 if (ShAmtVT == MVT::v4i32) {
6951 ShOps[2] = DAG.getUNDEF(MVT::i32);
6952 ShOps[3] = DAG.getUNDEF(MVT::i32);
6953 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
6955 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
6958 EVT VT = Op.getValueType();
6959 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
6960 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6961 DAG.getConstant(NewIntNo, MVT::i32),
6962 Op.getOperand(1), ShAmt);
6967 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6968 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6969 DebugLoc dl = Op.getDebugLoc();
6972 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6974 DAG.getConstant(TD->getPointerSize(),
6975 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6976 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6977 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6979 NULL, 0, false, false, 0);
6982 // Just load the return address.
6983 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6984 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6985 RetAddrFI, NULL, 0, false, false, 0);
6988 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6989 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6990 MFI->setFrameAddressIsTaken(true);
6991 EVT VT = Op.getValueType();
6992 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6993 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6994 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6995 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6997 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7002 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7003 SelectionDAG &DAG) {
7004 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7007 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
7009 MachineFunction &MF = DAG.getMachineFunction();
7010 SDValue Chain = Op.getOperand(0);
7011 SDValue Offset = Op.getOperand(1);
7012 SDValue Handler = Op.getOperand(2);
7013 DebugLoc dl = Op.getDebugLoc();
7015 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7017 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7019 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7020 DAG.getIntPtrConstant(-TD->getPointerSize()));
7021 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7022 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7023 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7024 MF.getRegInfo().addLiveOut(StoreAddrReg);
7026 return DAG.getNode(X86ISD::EH_RETURN, dl,
7028 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7031 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7032 SelectionDAG &DAG) {
7033 SDValue Root = Op.getOperand(0);
7034 SDValue Trmp = Op.getOperand(1); // trampoline
7035 SDValue FPtr = Op.getOperand(2); // nested function
7036 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7037 DebugLoc dl = Op.getDebugLoc();
7039 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7041 if (Subtarget->is64Bit()) {
7042 SDValue OutChains[6];
7044 // Large code-model.
7045 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7046 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7048 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7049 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7051 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7053 // Load the pointer to the nested function into R11.
7054 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7055 SDValue Addr = Trmp;
7056 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7057 Addr, TrmpAddr, 0, false, false, 0);
7059 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7060 DAG.getConstant(2, MVT::i64));
7061 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7064 // Load the 'nest' parameter value into R10.
7065 // R10 is specified in X86CallingConv.td
7066 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7067 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7068 DAG.getConstant(10, MVT::i64));
7069 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7070 Addr, TrmpAddr, 10, false, false, 0);
7072 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7073 DAG.getConstant(12, MVT::i64));
7074 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7077 // Jump to the nested function.
7078 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7079 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7080 DAG.getConstant(20, MVT::i64));
7081 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7082 Addr, TrmpAddr, 20, false, false, 0);
7084 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7085 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7086 DAG.getConstant(22, MVT::i64));
7087 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7088 TrmpAddr, 22, false, false, 0);
7091 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7092 return DAG.getMergeValues(Ops, 2, dl);
7094 const Function *Func =
7095 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7096 CallingConv::ID CC = Func->getCallingConv();
7101 llvm_unreachable("Unsupported calling convention");
7102 case CallingConv::C:
7103 case CallingConv::X86_StdCall: {
7104 // Pass 'nest' parameter in ECX.
7105 // Must be kept in sync with X86CallingConv.td
7108 // Check that ECX wasn't needed by an 'inreg' parameter.
7109 const FunctionType *FTy = Func->getFunctionType();
7110 const AttrListPtr &Attrs = Func->getAttributes();
7112 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7113 unsigned InRegCount = 0;
7116 for (FunctionType::param_iterator I = FTy->param_begin(),
7117 E = FTy->param_end(); I != E; ++I, ++Idx)
7118 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7119 // FIXME: should only count parameters that are lowered to integers.
7120 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7122 if (InRegCount > 2) {
7123 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
7128 case CallingConv::X86_FastCall:
7129 case CallingConv::Fast:
7130 // Pass 'nest' parameter in EAX.
7131 // Must be kept in sync with X86CallingConv.td
7136 SDValue OutChains[4];
7139 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7140 DAG.getConstant(10, MVT::i32));
7141 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7143 // This is storing the opcode for MOV32ri.
7144 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7145 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7146 OutChains[0] = DAG.getStore(Root, dl,
7147 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7148 Trmp, TrmpAddr, 0, false, false, 0);
7150 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7151 DAG.getConstant(1, MVT::i32));
7152 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7155 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7156 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7157 DAG.getConstant(5, MVT::i32));
7158 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7159 TrmpAddr, 5, false, false, 1);
7161 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7162 DAG.getConstant(6, MVT::i32));
7163 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7167 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7168 return DAG.getMergeValues(Ops, 2, dl);
7172 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
7174 The rounding mode is in bits 11:10 of FPSR, and has the following
7181 FLT_ROUNDS, on the other hand, expects the following:
7188 To perform the conversion, we do:
7189 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7192 MachineFunction &MF = DAG.getMachineFunction();
7193 const TargetMachine &TM = MF.getTarget();
7194 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7195 unsigned StackAlignment = TFI.getStackAlignment();
7196 EVT VT = Op.getValueType();
7197 DebugLoc dl = Op.getDebugLoc();
7199 // Save FP Control Word to stack slot
7200 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7201 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7203 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7204 DAG.getEntryNode(), StackSlot);
7206 // Load FP Control Word from stack slot
7207 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7210 // Transform as necessary
7212 DAG.getNode(ISD::SRL, dl, MVT::i16,
7213 DAG.getNode(ISD::AND, dl, MVT::i16,
7214 CWD, DAG.getConstant(0x800, MVT::i16)),
7215 DAG.getConstant(11, MVT::i8));
7217 DAG.getNode(ISD::SRL, dl, MVT::i16,
7218 DAG.getNode(ISD::AND, dl, MVT::i16,
7219 CWD, DAG.getConstant(0x400, MVT::i16)),
7220 DAG.getConstant(9, MVT::i8));
7223 DAG.getNode(ISD::AND, dl, MVT::i16,
7224 DAG.getNode(ISD::ADD, dl, MVT::i16,
7225 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7226 DAG.getConstant(1, MVT::i16)),
7227 DAG.getConstant(3, MVT::i16));
7230 return DAG.getNode((VT.getSizeInBits() < 16 ?
7231 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7234 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
7235 EVT VT = Op.getValueType();
7237 unsigned NumBits = VT.getSizeInBits();
7238 DebugLoc dl = Op.getDebugLoc();
7240 Op = Op.getOperand(0);
7241 if (VT == MVT::i8) {
7242 // Zero extend to i32 since there is not an i8 bsr.
7244 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7247 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7248 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7249 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7251 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7254 DAG.getConstant(NumBits+NumBits-1, OpVT),
7255 DAG.getConstant(X86::COND_E, MVT::i8),
7258 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7260 // Finally xor with NumBits-1.
7261 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7264 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7268 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
7269 EVT VT = Op.getValueType();
7271 unsigned NumBits = VT.getSizeInBits();
7272 DebugLoc dl = Op.getDebugLoc();
7274 Op = Op.getOperand(0);
7275 if (VT == MVT::i8) {
7277 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7280 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7281 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7282 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7284 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7287 DAG.getConstant(NumBits, OpVT),
7288 DAG.getConstant(X86::COND_E, MVT::i8),
7291 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7294 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7298 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
7299 EVT VT = Op.getValueType();
7300 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7301 DebugLoc dl = Op.getDebugLoc();
7303 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7304 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7305 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7306 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7307 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7309 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7310 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7311 // return AloBlo + AloBhi + AhiBlo;
7313 SDValue A = Op.getOperand(0);
7314 SDValue B = Op.getOperand(1);
7316 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7317 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7318 A, DAG.getConstant(32, MVT::i32));
7319 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7320 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7321 B, DAG.getConstant(32, MVT::i32));
7322 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7323 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7325 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7326 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7328 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7329 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7331 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7332 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7333 AloBhi, DAG.getConstant(32, MVT::i32));
7334 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7335 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7336 AhiBlo, DAG.getConstant(32, MVT::i32));
7337 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7338 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7343 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
7344 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7345 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7346 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7347 // has only one use.
7348 SDNode *N = Op.getNode();
7349 SDValue LHS = N->getOperand(0);
7350 SDValue RHS = N->getOperand(1);
7351 unsigned BaseOp = 0;
7353 DebugLoc dl = Op.getDebugLoc();
7355 switch (Op.getOpcode()) {
7356 default: llvm_unreachable("Unknown ovf instruction!");
7358 // A subtract of one will be selected as a INC. Note that INC doesn't
7359 // set CF, so we can't do this for UADDO.
7360 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7361 if (C->getAPIntValue() == 1) {
7362 BaseOp = X86ISD::INC;
7366 BaseOp = X86ISD::ADD;
7370 BaseOp = X86ISD::ADD;
7374 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7375 // set CF, so we can't do this for USUBO.
7376 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7377 if (C->getAPIntValue() == 1) {
7378 BaseOp = X86ISD::DEC;
7382 BaseOp = X86ISD::SUB;
7386 BaseOp = X86ISD::SUB;
7390 BaseOp = X86ISD::SMUL;
7394 BaseOp = X86ISD::UMUL;
7399 // Also sets EFLAGS.
7400 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7401 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7404 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7405 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7407 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7411 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
7412 EVT T = Op.getValueType();
7413 DebugLoc dl = Op.getDebugLoc();
7416 switch(T.getSimpleVT().SimpleTy) {
7418 assert(false && "Invalid value type!");
7419 case MVT::i8: Reg = X86::AL; size = 1; break;
7420 case MVT::i16: Reg = X86::AX; size = 2; break;
7421 case MVT::i32: Reg = X86::EAX; size = 4; break;
7423 assert(Subtarget->is64Bit() && "Node not type legal!");
7424 Reg = X86::RAX; size = 8;
7427 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7428 Op.getOperand(2), SDValue());
7429 SDValue Ops[] = { cpIn.getValue(0),
7432 DAG.getTargetConstant(size, MVT::i8),
7434 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7435 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7437 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7441 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7442 SelectionDAG &DAG) {
7443 assert(Subtarget->is64Bit() && "Result not type legalized?");
7444 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7445 SDValue TheChain = Op.getOperand(0);
7446 DebugLoc dl = Op.getDebugLoc();
7447 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7448 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7449 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7451 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7452 DAG.getConstant(32, MVT::i8));
7454 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7457 return DAG.getMergeValues(Ops, 2, dl);
7460 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7461 SDNode *Node = Op.getNode();
7462 DebugLoc dl = Node->getDebugLoc();
7463 EVT T = Node->getValueType(0);
7464 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7465 DAG.getConstant(0, T), Node->getOperand(2));
7466 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7467 cast<AtomicSDNode>(Node)->getMemoryVT(),
7468 Node->getOperand(0),
7469 Node->getOperand(1), negOp,
7470 cast<AtomicSDNode>(Node)->getSrcValue(),
7471 cast<AtomicSDNode>(Node)->getAlignment());
7474 /// LowerOperation - Provide custom lowering hooks for some operations.
7476 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7477 switch (Op.getOpcode()) {
7478 default: llvm_unreachable("Should not custom lower this!");
7479 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7480 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7481 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7482 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7483 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7484 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7485 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7486 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7487 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7488 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7489 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7490 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7491 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7492 case ISD::SHL_PARTS:
7493 case ISD::SRA_PARTS:
7494 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7495 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7496 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7497 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7498 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7499 case ISD::FABS: return LowerFABS(Op, DAG);
7500 case ISD::FNEG: return LowerFNEG(Op, DAG);
7501 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7502 case ISD::SETCC: return LowerSETCC(Op, DAG);
7503 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7504 case ISD::SELECT: return LowerSELECT(Op, DAG);
7505 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7506 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7507 case ISD::VASTART: return LowerVASTART(Op, DAG);
7508 case ISD::VAARG: return LowerVAARG(Op, DAG);
7509 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7510 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7511 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7512 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7513 case ISD::FRAME_TO_ARGS_OFFSET:
7514 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7515 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7516 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7517 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7518 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7519 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7520 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7521 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7527 case ISD::UMULO: return LowerXALUO(Op, DAG);
7528 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7532 void X86TargetLowering::
7533 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7534 SelectionDAG &DAG, unsigned NewOp) {
7535 EVT T = Node->getValueType(0);
7536 DebugLoc dl = Node->getDebugLoc();
7537 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7539 SDValue Chain = Node->getOperand(0);
7540 SDValue In1 = Node->getOperand(1);
7541 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7542 Node->getOperand(2), DAG.getIntPtrConstant(0));
7543 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7544 Node->getOperand(2), DAG.getIntPtrConstant(1));
7545 SDValue Ops[] = { Chain, In1, In2L, In2H };
7546 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7548 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7549 cast<MemSDNode>(Node)->getMemOperand());
7550 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7551 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7552 Results.push_back(Result.getValue(2));
7555 /// ReplaceNodeResults - Replace a node with an illegal result type
7556 /// with a new node built out of custom code.
7557 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7558 SmallVectorImpl<SDValue>&Results,
7559 SelectionDAG &DAG) {
7560 DebugLoc dl = N->getDebugLoc();
7561 switch (N->getOpcode()) {
7563 assert(false && "Do not know how to custom type legalize this operation!");
7565 case ISD::FP_TO_SINT: {
7566 std::pair<SDValue,SDValue> Vals =
7567 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7568 SDValue FIST = Vals.first, StackSlot = Vals.second;
7569 if (FIST.getNode() != 0) {
7570 EVT VT = N->getValueType(0);
7571 // Return a load from the stack slot.
7572 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7577 case ISD::READCYCLECOUNTER: {
7578 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7579 SDValue TheChain = N->getOperand(0);
7580 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7581 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7583 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7585 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7586 SDValue Ops[] = { eax, edx };
7587 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7588 Results.push_back(edx.getValue(1));
7591 case ISD::ATOMIC_CMP_SWAP: {
7592 EVT T = N->getValueType(0);
7593 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7594 SDValue cpInL, cpInH;
7595 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7596 DAG.getConstant(0, MVT::i32));
7597 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7598 DAG.getConstant(1, MVT::i32));
7599 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7600 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7602 SDValue swapInL, swapInH;
7603 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7604 DAG.getConstant(0, MVT::i32));
7605 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7606 DAG.getConstant(1, MVT::i32));
7607 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7609 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7610 swapInL.getValue(1));
7611 SDValue Ops[] = { swapInH.getValue(0),
7613 swapInH.getValue(1) };
7614 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7615 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7616 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7617 MVT::i32, Result.getValue(1));
7618 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7619 MVT::i32, cpOutL.getValue(2));
7620 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7621 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7622 Results.push_back(cpOutH.getValue(1));
7625 case ISD::ATOMIC_LOAD_ADD:
7626 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7628 case ISD::ATOMIC_LOAD_AND:
7629 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7631 case ISD::ATOMIC_LOAD_NAND:
7632 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7634 case ISD::ATOMIC_LOAD_OR:
7635 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7637 case ISD::ATOMIC_LOAD_SUB:
7638 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7640 case ISD::ATOMIC_LOAD_XOR:
7641 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7643 case ISD::ATOMIC_SWAP:
7644 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7649 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7651 default: return NULL;
7652 case X86ISD::BSF: return "X86ISD::BSF";
7653 case X86ISD::BSR: return "X86ISD::BSR";
7654 case X86ISD::SHLD: return "X86ISD::SHLD";
7655 case X86ISD::SHRD: return "X86ISD::SHRD";
7656 case X86ISD::FAND: return "X86ISD::FAND";
7657 case X86ISD::FOR: return "X86ISD::FOR";
7658 case X86ISD::FXOR: return "X86ISD::FXOR";
7659 case X86ISD::FSRL: return "X86ISD::FSRL";
7660 case X86ISD::FILD: return "X86ISD::FILD";
7661 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7662 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7663 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7664 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7665 case X86ISD::FLD: return "X86ISD::FLD";
7666 case X86ISD::FST: return "X86ISD::FST";
7667 case X86ISD::CALL: return "X86ISD::CALL";
7668 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7669 case X86ISD::BT: return "X86ISD::BT";
7670 case X86ISD::CMP: return "X86ISD::CMP";
7671 case X86ISD::COMI: return "X86ISD::COMI";
7672 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7673 case X86ISD::SETCC: return "X86ISD::SETCC";
7674 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7675 case X86ISD::CMOV: return "X86ISD::CMOV";
7676 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7677 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7678 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7679 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7680 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7681 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7682 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7683 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7684 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7685 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7686 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7687 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7688 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
7689 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7690 case X86ISD::FMAX: return "X86ISD::FMAX";
7691 case X86ISD::FMIN: return "X86ISD::FMIN";
7692 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7693 case X86ISD::FRCP: return "X86ISD::FRCP";
7694 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7695 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7696 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7697 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7698 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7699 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7700 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7701 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7702 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7703 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7704 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7705 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7706 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7707 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7708 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7709 case X86ISD::VSHL: return "X86ISD::VSHL";
7710 case X86ISD::VSRL: return "X86ISD::VSRL";
7711 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7712 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7713 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7714 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7715 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7716 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7717 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7718 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7719 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7720 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7721 case X86ISD::ADD: return "X86ISD::ADD";
7722 case X86ISD::SUB: return "X86ISD::SUB";
7723 case X86ISD::SMUL: return "X86ISD::SMUL";
7724 case X86ISD::UMUL: return "X86ISD::UMUL";
7725 case X86ISD::INC: return "X86ISD::INC";
7726 case X86ISD::DEC: return "X86ISD::DEC";
7727 case X86ISD::OR: return "X86ISD::OR";
7728 case X86ISD::XOR: return "X86ISD::XOR";
7729 case X86ISD::AND: return "X86ISD::AND";
7730 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7731 case X86ISD::PTEST: return "X86ISD::PTEST";
7732 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7733 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
7737 // isLegalAddressingMode - Return true if the addressing mode represented
7738 // by AM is legal for this target, for a load/store of the specified type.
7739 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7740 const Type *Ty) const {
7741 // X86 supports extremely general addressing modes.
7742 CodeModel::Model M = getTargetMachine().getCodeModel();
7744 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7745 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7750 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7752 // If a reference to this global requires an extra load, we can't fold it.
7753 if (isGlobalStubReference(GVFlags))
7756 // If BaseGV requires a register for the PIC base, we cannot also have a
7757 // BaseReg specified.
7758 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7761 // If lower 4G is not available, then we must use rip-relative addressing.
7762 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7772 // These scales always work.
7777 // These scales are formed with basereg+scalereg. Only accept if there is
7782 default: // Other stuff never works.
7790 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7791 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7793 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7794 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7795 if (NumBits1 <= NumBits2)
7800 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7801 if (!VT1.isInteger() || !VT2.isInteger())
7803 unsigned NumBits1 = VT1.getSizeInBits();
7804 unsigned NumBits2 = VT2.getSizeInBits();
7805 if (NumBits1 <= NumBits2)
7810 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7811 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7812 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7815 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7816 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7817 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7820 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7821 // i16 instructions are longer (0x66 prefix) and potentially slower.
7822 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7825 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7826 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7827 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7828 /// are assumed to be legal.
7830 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7832 // Only do shuffles on 128-bit vector types for now.
7833 if (VT.getSizeInBits() == 64)
7836 // FIXME: pshufb, blends, shifts.
7837 return (VT.getVectorNumElements() == 2 ||
7838 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7839 isMOVLMask(M, VT) ||
7840 isSHUFPMask(M, VT) ||
7841 isPSHUFDMask(M, VT) ||
7842 isPSHUFHWMask(M, VT) ||
7843 isPSHUFLWMask(M, VT) ||
7844 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7845 isUNPCKLMask(M, VT) ||
7846 isUNPCKHMask(M, VT) ||
7847 isUNPCKL_v_undef_Mask(M, VT) ||
7848 isUNPCKH_v_undef_Mask(M, VT));
7852 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7854 unsigned NumElts = VT.getVectorNumElements();
7855 // FIXME: This collection of masks seems suspect.
7858 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7859 return (isMOVLMask(Mask, VT) ||
7860 isCommutedMOVLMask(Mask, VT, true) ||
7861 isSHUFPMask(Mask, VT) ||
7862 isCommutedSHUFPMask(Mask, VT));
7867 //===----------------------------------------------------------------------===//
7868 // X86 Scheduler Hooks
7869 //===----------------------------------------------------------------------===//
7871 // private utility function
7873 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7874 MachineBasicBlock *MBB,
7882 TargetRegisterClass *RC,
7883 bool invSrc) const {
7884 // For the atomic bitwise operator, we generate
7887 // ld t1 = [bitinstr.addr]
7888 // op t2 = t1, [bitinstr.val]
7890 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7892 // fallthrough -->nextMBB
7893 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7894 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7895 MachineFunction::iterator MBBIter = MBB;
7898 /// First build the CFG
7899 MachineFunction *F = MBB->getParent();
7900 MachineBasicBlock *thisMBB = MBB;
7901 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7902 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7903 F->insert(MBBIter, newMBB);
7904 F->insert(MBBIter, nextMBB);
7906 // Move all successors to thisMBB to nextMBB
7907 nextMBB->transferSuccessors(thisMBB);
7909 // Update thisMBB to fall through to newMBB
7910 thisMBB->addSuccessor(newMBB);
7912 // newMBB jumps to itself and fall through to nextMBB
7913 newMBB->addSuccessor(nextMBB);
7914 newMBB->addSuccessor(newMBB);
7916 // Insert instructions into newMBB based on incoming instruction
7917 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7918 "unexpected number of operands");
7919 DebugLoc dl = bInstr->getDebugLoc();
7920 MachineOperand& destOper = bInstr->getOperand(0);
7921 MachineOperand* argOpers[2 + X86AddrNumOperands];
7922 int numArgs = bInstr->getNumOperands() - 1;
7923 for (int i=0; i < numArgs; ++i)
7924 argOpers[i] = &bInstr->getOperand(i+1);
7926 // x86 address has 4 operands: base, index, scale, and displacement
7927 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7928 int valArgIndx = lastAddrIndx + 1;
7930 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7931 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7932 for (int i=0; i <= lastAddrIndx; ++i)
7933 (*MIB).addOperand(*argOpers[i]);
7935 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7937 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7942 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7943 assert((argOpers[valArgIndx]->isReg() ||
7944 argOpers[valArgIndx]->isImm()) &&
7946 if (argOpers[valArgIndx]->isReg())
7947 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7949 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7951 (*MIB).addOperand(*argOpers[valArgIndx]);
7953 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7956 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7957 for (int i=0; i <= lastAddrIndx; ++i)
7958 (*MIB).addOperand(*argOpers[i]);
7960 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7961 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7962 bInstr->memoperands_end());
7964 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7968 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
7970 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7974 // private utility function: 64 bit atomics on 32 bit host.
7976 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7977 MachineBasicBlock *MBB,
7982 bool invSrc) const {
7983 // For the atomic bitwise operator, we generate
7984 // thisMBB (instructions are in pairs, except cmpxchg8b)
7985 // ld t1,t2 = [bitinstr.addr]
7987 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7988 // op t5, t6 <- out1, out2, [bitinstr.val]
7989 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7990 // mov ECX, EBX <- t5, t6
7991 // mov EAX, EDX <- t1, t2
7992 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7993 // mov t3, t4 <- EAX, EDX
7995 // result in out1, out2
7996 // fallthrough -->nextMBB
7998 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7999 const unsigned LoadOpc = X86::MOV32rm;
8000 const unsigned copyOpc = X86::MOV32rr;
8001 const unsigned NotOpc = X86::NOT32r;
8002 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8003 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8004 MachineFunction::iterator MBBIter = MBB;
8007 /// First build the CFG
8008 MachineFunction *F = MBB->getParent();
8009 MachineBasicBlock *thisMBB = MBB;
8010 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8011 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8012 F->insert(MBBIter, newMBB);
8013 F->insert(MBBIter, nextMBB);
8015 // Move all successors to thisMBB to nextMBB
8016 nextMBB->transferSuccessors(thisMBB);
8018 // Update thisMBB to fall through to newMBB
8019 thisMBB->addSuccessor(newMBB);
8021 // newMBB jumps to itself and fall through to nextMBB
8022 newMBB->addSuccessor(nextMBB);
8023 newMBB->addSuccessor(newMBB);
8025 DebugLoc dl = bInstr->getDebugLoc();
8026 // Insert instructions into newMBB based on incoming instruction
8027 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8028 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
8029 "unexpected number of operands");
8030 MachineOperand& dest1Oper = bInstr->getOperand(0);
8031 MachineOperand& dest2Oper = bInstr->getOperand(1);
8032 MachineOperand* argOpers[2 + X86AddrNumOperands];
8033 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
8034 argOpers[i] = &bInstr->getOperand(i+2);
8036 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8037 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8039 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8040 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8041 for (int i=0; i <= lastAddrIndx; ++i)
8042 (*MIB).addOperand(*argOpers[i]);
8043 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8044 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8045 // add 4 to displacement.
8046 for (int i=0; i <= lastAddrIndx-2; ++i)
8047 (*MIB).addOperand(*argOpers[i]);
8048 MachineOperand newOp3 = *(argOpers[3]);
8050 newOp3.setImm(newOp3.getImm()+4);
8052 newOp3.setOffset(newOp3.getOffset()+4);
8053 (*MIB).addOperand(newOp3);
8054 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8056 // t3/4 are defined later, at the bottom of the loop
8057 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8058 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8059 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8060 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8061 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8062 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8064 // The subsequent operations should be using the destination registers of
8065 //the PHI instructions.
8067 t1 = F->getRegInfo().createVirtualRegister(RC);
8068 t2 = F->getRegInfo().createVirtualRegister(RC);
8069 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8070 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8072 t1 = dest1Oper.getReg();
8073 t2 = dest2Oper.getReg();
8076 int valArgIndx = lastAddrIndx + 1;
8077 assert((argOpers[valArgIndx]->isReg() ||
8078 argOpers[valArgIndx]->isImm()) &&
8080 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8081 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8082 if (argOpers[valArgIndx]->isReg())
8083 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8085 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8086 if (regOpcL != X86::MOV32rr)
8088 (*MIB).addOperand(*argOpers[valArgIndx]);
8089 assert(argOpers[valArgIndx + 1]->isReg() ==
8090 argOpers[valArgIndx]->isReg());
8091 assert(argOpers[valArgIndx + 1]->isImm() ==
8092 argOpers[valArgIndx]->isImm());
8093 if (argOpers[valArgIndx + 1]->isReg())
8094 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8096 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8097 if (regOpcH != X86::MOV32rr)
8099 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8101 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8103 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8106 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8108 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8111 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8112 for (int i=0; i <= lastAddrIndx; ++i)
8113 (*MIB).addOperand(*argOpers[i]);
8115 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8116 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8117 bInstr->memoperands_end());
8119 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8120 MIB.addReg(X86::EAX);
8121 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8122 MIB.addReg(X86::EDX);
8125 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8127 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8131 // private utility function
8133 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8134 MachineBasicBlock *MBB,
8135 unsigned cmovOpc) const {
8136 // For the atomic min/max operator, we generate
8139 // ld t1 = [min/max.addr]
8140 // mov t2 = [min/max.val]
8142 // cmov[cond] t2 = t1
8144 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8146 // fallthrough -->nextMBB
8148 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8149 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8150 MachineFunction::iterator MBBIter = MBB;
8153 /// First build the CFG
8154 MachineFunction *F = MBB->getParent();
8155 MachineBasicBlock *thisMBB = MBB;
8156 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8157 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8158 F->insert(MBBIter, newMBB);
8159 F->insert(MBBIter, nextMBB);
8161 // Move all successors of thisMBB to nextMBB
8162 nextMBB->transferSuccessors(thisMBB);
8164 // Update thisMBB to fall through to newMBB
8165 thisMBB->addSuccessor(newMBB);
8167 // newMBB jumps to newMBB and fall through to nextMBB
8168 newMBB->addSuccessor(nextMBB);
8169 newMBB->addSuccessor(newMBB);
8171 DebugLoc dl = mInstr->getDebugLoc();
8172 // Insert instructions into newMBB based on incoming instruction
8173 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8174 "unexpected number of operands");
8175 MachineOperand& destOper = mInstr->getOperand(0);
8176 MachineOperand* argOpers[2 + X86AddrNumOperands];
8177 int numArgs = mInstr->getNumOperands() - 1;
8178 for (int i=0; i < numArgs; ++i)
8179 argOpers[i] = &mInstr->getOperand(i+1);
8181 // x86 address has 4 operands: base, index, scale, and displacement
8182 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8183 int valArgIndx = lastAddrIndx + 1;
8185 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8186 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8187 for (int i=0; i <= lastAddrIndx; ++i)
8188 (*MIB).addOperand(*argOpers[i]);
8190 // We only support register and immediate values
8191 assert((argOpers[valArgIndx]->isReg() ||
8192 argOpers[valArgIndx]->isImm()) &&
8195 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8196 if (argOpers[valArgIndx]->isReg())
8197 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8199 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8200 (*MIB).addOperand(*argOpers[valArgIndx]);
8202 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8205 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8210 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8211 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8215 // Cmp and exchange if none has modified the memory location
8216 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8217 for (int i=0; i <= lastAddrIndx; ++i)
8218 (*MIB).addOperand(*argOpers[i]);
8220 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8221 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8222 mInstr->memoperands_end());
8224 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8225 MIB.addReg(X86::EAX);
8228 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8230 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8234 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8235 // all of this code can be replaced with that in the .td file.
8237 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8238 unsigned numArgs, bool memArg) const {
8240 MachineFunction *F = BB->getParent();
8241 DebugLoc dl = MI->getDebugLoc();
8242 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8246 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8248 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8250 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8252 for (unsigned i = 0; i < numArgs; ++i) {
8253 MachineOperand &Op = MI->getOperand(i+1);
8255 if (!(Op.isReg() && Op.isImplicit()))
8259 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8262 F->DeleteMachineInstr(MI);
8268 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8270 MachineBasicBlock *MBB) const {
8271 // Emit code to save XMM registers to the stack. The ABI says that the
8272 // number of registers to save is given in %al, so it's theoretically
8273 // possible to do an indirect jump trick to avoid saving all of them,
8274 // however this code takes a simpler approach and just executes all
8275 // of the stores if %al is non-zero. It's less code, and it's probably
8276 // easier on the hardware branch predictor, and stores aren't all that
8277 // expensive anyway.
8279 // Create the new basic blocks. One block contains all the XMM stores,
8280 // and one block is the final destination regardless of whether any
8281 // stores were performed.
8282 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8283 MachineFunction *F = MBB->getParent();
8284 MachineFunction::iterator MBBIter = MBB;
8286 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8287 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8288 F->insert(MBBIter, XMMSaveMBB);
8289 F->insert(MBBIter, EndMBB);
8292 // Move any original successors of MBB to the end block.
8293 EndMBB->transferSuccessors(MBB);
8294 // The original block will now fall through to the XMM save block.
8295 MBB->addSuccessor(XMMSaveMBB);
8296 // The XMMSaveMBB will fall through to the end block.
8297 XMMSaveMBB->addSuccessor(EndMBB);
8299 // Now add the instructions.
8300 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8301 DebugLoc DL = MI->getDebugLoc();
8303 unsigned CountReg = MI->getOperand(0).getReg();
8304 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8305 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8307 if (!Subtarget->isTargetWin64()) {
8308 // If %al is 0, branch around the XMM save block.
8309 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8310 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8311 MBB->addSuccessor(EndMBB);
8314 // In the XMM save block, save all the XMM argument registers.
8315 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8316 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8317 MachineMemOperand *MMO =
8318 F->getMachineMemOperand(
8319 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8320 MachineMemOperand::MOStore, Offset,
8321 /*Size=*/16, /*Align=*/16);
8322 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8323 .addFrameIndex(RegSaveFrameIndex)
8324 .addImm(/*Scale=*/1)
8325 .addReg(/*IndexReg=*/0)
8326 .addImm(/*Disp=*/Offset)
8327 .addReg(/*Segment=*/0)
8328 .addReg(MI->getOperand(i).getReg())
8329 .addMemOperand(MMO);
8332 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8338 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8339 MachineBasicBlock *BB,
8340 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8341 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8342 DebugLoc DL = MI->getDebugLoc();
8344 // To "insert" a SELECT_CC instruction, we actually have to insert the
8345 // diamond control-flow pattern. The incoming instruction knows the
8346 // destination vreg to set, the condition code register to branch on, the
8347 // true/false values to select between, and a branch opcode to use.
8348 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8349 MachineFunction::iterator It = BB;
8355 // cmpTY ccX, r1, r2
8357 // fallthrough --> copy0MBB
8358 MachineBasicBlock *thisMBB = BB;
8359 MachineFunction *F = BB->getParent();
8360 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8361 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8363 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8364 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8365 F->insert(It, copy0MBB);
8366 F->insert(It, sinkMBB);
8367 // Update machine-CFG edges by first adding all successors of the current
8368 // block to the new block which will contain the Phi node for the select.
8369 // Also inform sdisel of the edge changes.
8370 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8371 E = BB->succ_end(); I != E; ++I) {
8372 EM->insert(std::make_pair(*I, sinkMBB));
8373 sinkMBB->addSuccessor(*I);
8375 // Next, remove all successors of the current block, and add the true
8376 // and fallthrough blocks as its successors.
8377 while (!BB->succ_empty())
8378 BB->removeSuccessor(BB->succ_begin());
8379 // Add the true and fallthrough blocks as its successors.
8380 BB->addSuccessor(copy0MBB);
8381 BB->addSuccessor(sinkMBB);
8384 // %FalseValue = ...
8385 // # fallthrough to sinkMBB
8388 // Update machine-CFG edges
8389 BB->addSuccessor(sinkMBB);
8392 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8395 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8396 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8397 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8399 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8404 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8405 MachineBasicBlock *BB,
8406 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8407 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8408 DebugLoc DL = MI->getDebugLoc();
8409 MachineFunction *F = BB->getParent();
8411 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8412 // non-trivial part is impdef of ESP.
8413 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8416 BuildMI(BB, DL, TII->get(X86::CALLpcrel32))
8417 .addExternalSymbol("_alloca")
8418 .addReg(X86::EAX, RegState::Implicit)
8419 .addReg(X86::ESP, RegState::Implicit)
8420 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8421 .addReg(X86::ESP, RegState::Define | RegState::Implicit);
8423 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8428 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8429 MachineBasicBlock *BB,
8430 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8431 switch (MI->getOpcode()) {
8432 default: assert(false && "Unexpected instr type to insert");
8433 case X86::MINGW_ALLOCA:
8434 return EmitLoweredMingwAlloca(MI, BB, EM);
8436 case X86::CMOV_V1I64:
8437 case X86::CMOV_FR32:
8438 case X86::CMOV_FR64:
8439 case X86::CMOV_V4F32:
8440 case X86::CMOV_V2F64:
8441 case X86::CMOV_V2I64:
8442 return EmitLoweredSelect(MI, BB, EM);
8444 case X86::FP32_TO_INT16_IN_MEM:
8445 case X86::FP32_TO_INT32_IN_MEM:
8446 case X86::FP32_TO_INT64_IN_MEM:
8447 case X86::FP64_TO_INT16_IN_MEM:
8448 case X86::FP64_TO_INT32_IN_MEM:
8449 case X86::FP64_TO_INT64_IN_MEM:
8450 case X86::FP80_TO_INT16_IN_MEM:
8451 case X86::FP80_TO_INT32_IN_MEM:
8452 case X86::FP80_TO_INT64_IN_MEM: {
8453 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8454 DebugLoc DL = MI->getDebugLoc();
8456 // Change the floating point control register to use "round towards zero"
8457 // mode when truncating to an integer value.
8458 MachineFunction *F = BB->getParent();
8459 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8460 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8462 // Load the old value of the high byte of the control word...
8464 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8465 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8468 // Set the high part to be round to zero...
8469 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8472 // Reload the modified control word now...
8473 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8475 // Restore the memory image of control word to original value
8476 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8479 // Get the X86 opcode to use.
8481 switch (MI->getOpcode()) {
8482 default: llvm_unreachable("illegal opcode!");
8483 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8484 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8485 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8486 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8487 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8488 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8489 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8490 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8491 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8495 MachineOperand &Op = MI->getOperand(0);
8497 AM.BaseType = X86AddressMode::RegBase;
8498 AM.Base.Reg = Op.getReg();
8500 AM.BaseType = X86AddressMode::FrameIndexBase;
8501 AM.Base.FrameIndex = Op.getIndex();
8503 Op = MI->getOperand(1);
8505 AM.Scale = Op.getImm();
8506 Op = MI->getOperand(2);
8508 AM.IndexReg = Op.getImm();
8509 Op = MI->getOperand(3);
8510 if (Op.isGlobal()) {
8511 AM.GV = Op.getGlobal();
8513 AM.Disp = Op.getImm();
8515 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8516 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8518 // Reload the original control word now.
8519 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8521 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8524 // String/text processing lowering.
8525 case X86::PCMPISTRM128REG:
8526 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8527 case X86::PCMPISTRM128MEM:
8528 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8529 case X86::PCMPESTRM128REG:
8530 return EmitPCMP(MI, BB, 5, false /* in mem */);
8531 case X86::PCMPESTRM128MEM:
8532 return EmitPCMP(MI, BB, 5, true /* in mem */);
8535 case X86::ATOMAND32:
8536 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8537 X86::AND32ri, X86::MOV32rm,
8538 X86::LCMPXCHG32, X86::MOV32rr,
8539 X86::NOT32r, X86::EAX,
8540 X86::GR32RegisterClass);
8542 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8543 X86::OR32ri, X86::MOV32rm,
8544 X86::LCMPXCHG32, X86::MOV32rr,
8545 X86::NOT32r, X86::EAX,
8546 X86::GR32RegisterClass);
8547 case X86::ATOMXOR32:
8548 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8549 X86::XOR32ri, X86::MOV32rm,
8550 X86::LCMPXCHG32, X86::MOV32rr,
8551 X86::NOT32r, X86::EAX,
8552 X86::GR32RegisterClass);
8553 case X86::ATOMNAND32:
8554 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8555 X86::AND32ri, X86::MOV32rm,
8556 X86::LCMPXCHG32, X86::MOV32rr,
8557 X86::NOT32r, X86::EAX,
8558 X86::GR32RegisterClass, true);
8559 case X86::ATOMMIN32:
8560 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8561 case X86::ATOMMAX32:
8562 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8563 case X86::ATOMUMIN32:
8564 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8565 case X86::ATOMUMAX32:
8566 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8568 case X86::ATOMAND16:
8569 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8570 X86::AND16ri, X86::MOV16rm,
8571 X86::LCMPXCHG16, X86::MOV16rr,
8572 X86::NOT16r, X86::AX,
8573 X86::GR16RegisterClass);
8575 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8576 X86::OR16ri, X86::MOV16rm,
8577 X86::LCMPXCHG16, X86::MOV16rr,
8578 X86::NOT16r, X86::AX,
8579 X86::GR16RegisterClass);
8580 case X86::ATOMXOR16:
8581 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8582 X86::XOR16ri, X86::MOV16rm,
8583 X86::LCMPXCHG16, X86::MOV16rr,
8584 X86::NOT16r, X86::AX,
8585 X86::GR16RegisterClass);
8586 case X86::ATOMNAND16:
8587 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8588 X86::AND16ri, X86::MOV16rm,
8589 X86::LCMPXCHG16, X86::MOV16rr,
8590 X86::NOT16r, X86::AX,
8591 X86::GR16RegisterClass, true);
8592 case X86::ATOMMIN16:
8593 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8594 case X86::ATOMMAX16:
8595 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8596 case X86::ATOMUMIN16:
8597 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8598 case X86::ATOMUMAX16:
8599 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8602 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8603 X86::AND8ri, X86::MOV8rm,
8604 X86::LCMPXCHG8, X86::MOV8rr,
8605 X86::NOT8r, X86::AL,
8606 X86::GR8RegisterClass);
8608 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8609 X86::OR8ri, X86::MOV8rm,
8610 X86::LCMPXCHG8, X86::MOV8rr,
8611 X86::NOT8r, X86::AL,
8612 X86::GR8RegisterClass);
8614 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8615 X86::XOR8ri, X86::MOV8rm,
8616 X86::LCMPXCHG8, X86::MOV8rr,
8617 X86::NOT8r, X86::AL,
8618 X86::GR8RegisterClass);
8619 case X86::ATOMNAND8:
8620 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8621 X86::AND8ri, X86::MOV8rm,
8622 X86::LCMPXCHG8, X86::MOV8rr,
8623 X86::NOT8r, X86::AL,
8624 X86::GR8RegisterClass, true);
8625 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8626 // This group is for 64-bit host.
8627 case X86::ATOMAND64:
8628 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8629 X86::AND64ri32, X86::MOV64rm,
8630 X86::LCMPXCHG64, X86::MOV64rr,
8631 X86::NOT64r, X86::RAX,
8632 X86::GR64RegisterClass);
8634 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8635 X86::OR64ri32, X86::MOV64rm,
8636 X86::LCMPXCHG64, X86::MOV64rr,
8637 X86::NOT64r, X86::RAX,
8638 X86::GR64RegisterClass);
8639 case X86::ATOMXOR64:
8640 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8641 X86::XOR64ri32, X86::MOV64rm,
8642 X86::LCMPXCHG64, X86::MOV64rr,
8643 X86::NOT64r, X86::RAX,
8644 X86::GR64RegisterClass);
8645 case X86::ATOMNAND64:
8646 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8647 X86::AND64ri32, X86::MOV64rm,
8648 X86::LCMPXCHG64, X86::MOV64rr,
8649 X86::NOT64r, X86::RAX,
8650 X86::GR64RegisterClass, true);
8651 case X86::ATOMMIN64:
8652 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8653 case X86::ATOMMAX64:
8654 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8655 case X86::ATOMUMIN64:
8656 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8657 case X86::ATOMUMAX64:
8658 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8660 // This group does 64-bit operations on a 32-bit host.
8661 case X86::ATOMAND6432:
8662 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8663 X86::AND32rr, X86::AND32rr,
8664 X86::AND32ri, X86::AND32ri,
8666 case X86::ATOMOR6432:
8667 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8668 X86::OR32rr, X86::OR32rr,
8669 X86::OR32ri, X86::OR32ri,
8671 case X86::ATOMXOR6432:
8672 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8673 X86::XOR32rr, X86::XOR32rr,
8674 X86::XOR32ri, X86::XOR32ri,
8676 case X86::ATOMNAND6432:
8677 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8678 X86::AND32rr, X86::AND32rr,
8679 X86::AND32ri, X86::AND32ri,
8681 case X86::ATOMADD6432:
8682 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8683 X86::ADD32rr, X86::ADC32rr,
8684 X86::ADD32ri, X86::ADC32ri,
8686 case X86::ATOMSUB6432:
8687 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8688 X86::SUB32rr, X86::SBB32rr,
8689 X86::SUB32ri, X86::SBB32ri,
8691 case X86::ATOMSWAP6432:
8692 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8693 X86::MOV32rr, X86::MOV32rr,
8694 X86::MOV32ri, X86::MOV32ri,
8696 case X86::VASTART_SAVE_XMM_REGS:
8697 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8701 //===----------------------------------------------------------------------===//
8702 // X86 Optimization Hooks
8703 //===----------------------------------------------------------------------===//
8705 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8709 const SelectionDAG &DAG,
8710 unsigned Depth) const {
8711 unsigned Opc = Op.getOpcode();
8712 assert((Opc >= ISD::BUILTIN_OP_END ||
8713 Opc == ISD::INTRINSIC_WO_CHAIN ||
8714 Opc == ISD::INTRINSIC_W_CHAIN ||
8715 Opc == ISD::INTRINSIC_VOID) &&
8716 "Should use MaskedValueIsZero if you don't know whether Op"
8717 " is a target node!");
8719 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8731 // These nodes' second result is a boolean.
8732 if (Op.getResNo() == 0)
8736 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8737 Mask.getBitWidth() - 1);
8742 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8743 /// node is a GlobalAddress + offset.
8744 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8745 GlobalValue* &GA, int64_t &Offset) const{
8746 if (N->getOpcode() == X86ISD::Wrapper) {
8747 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8748 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8749 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8753 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8756 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8757 EVT EltVT, LoadSDNode *&LDBase,
8758 unsigned &LastLoadedElt,
8759 SelectionDAG &DAG, MachineFrameInfo *MFI,
8760 const TargetLowering &TLI) {
8762 LastLoadedElt = -1U;
8763 for (unsigned i = 0; i < NumElems; ++i) {
8764 if (N->getMaskElt(i) < 0) {
8770 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8771 if (!Elt.getNode() ||
8772 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8775 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8777 LDBase = cast<LoadSDNode>(Elt.getNode());
8781 if (Elt.getOpcode() == ISD::UNDEF)
8784 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8785 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
8792 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8793 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8794 /// if the load addresses are consecutive, non-overlapping, and in the right
8795 /// order. In the case of v2i64, it will see if it can rewrite the
8796 /// shuffle to be an appropriate build vector so it can take advantage of
8797 // performBuildVectorCombine.
8798 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8799 const TargetLowering &TLI) {
8800 DebugLoc dl = N->getDebugLoc();
8801 EVT VT = N->getValueType(0);
8802 EVT EltVT = VT.getVectorElementType();
8803 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8804 unsigned NumElems = VT.getVectorNumElements();
8806 if (VT.getSizeInBits() != 128)
8809 // Try to combine a vector_shuffle into a 128-bit load.
8810 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8811 LoadSDNode *LD = NULL;
8812 unsigned LastLoadedElt;
8813 if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
8817 if (LastLoadedElt == NumElems - 1) {
8818 if (DAG.InferPtrAlignment(LD->getBasePtr()) >= 16)
8819 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8820 LD->getSrcValue(), LD->getSrcValueOffset(),
8821 LD->isVolatile(), LD->isNonTemporal(), 0);
8822 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8823 LD->getSrcValue(), LD->getSrcValueOffset(),
8824 LD->isVolatile(), LD->isNonTemporal(),
8825 LD->getAlignment());
8826 } else if (NumElems == 4 && LastLoadedElt == 1) {
8827 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8828 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8829 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8830 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8835 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8836 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8837 const X86Subtarget *Subtarget) {
8838 DebugLoc DL = N->getDebugLoc();
8839 SDValue Cond = N->getOperand(0);
8840 // Get the LHS/RHS of the select.
8841 SDValue LHS = N->getOperand(1);
8842 SDValue RHS = N->getOperand(2);
8844 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8845 // instructions match the semantics of the common C idiom x<y?x:y but not
8846 // x<=y?x:y, because of how they handle negative zero (which can be
8847 // ignored in unsafe-math mode).
8848 if (Subtarget->hasSSE2() &&
8849 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8850 Cond.getOpcode() == ISD::SETCC) {
8851 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8853 unsigned Opcode = 0;
8854 // Check for x CC y ? x : y.
8855 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
8856 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
8860 // Converting this to a min 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::FMIN;
8873 // Converting this to a min would handle comparisons between positive
8874 // and negative zero incorrectly.
8875 if (!UnsafeFPMath &&
8876 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
8878 Opcode = X86ISD::FMIN;
8881 // Converting this to a min would handle both negative zeros and NaNs
8882 // incorrectly, but we can swap the operands to fix both.
8883 std::swap(LHS, RHS);
8887 Opcode = X86ISD::FMIN;
8891 // Converting this to a max would handle comparisons between positive
8892 // and negative zero incorrectly.
8893 if (!UnsafeFPMath &&
8894 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
8896 Opcode = X86ISD::FMAX;
8899 // Converting this to a max would handle NaNs incorrectly, and swapping
8900 // the operands would cause it to handle comparisons between positive
8901 // and negative zero incorrectly.
8902 if (!FiniteOnlyFPMath() &&
8903 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))) {
8904 if (!UnsafeFPMath &&
8905 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8907 std::swap(LHS, RHS);
8909 Opcode = X86ISD::FMAX;
8912 // Converting this to a max would handle both negative zeros and NaNs
8913 // incorrectly, but we can swap the operands to fix both.
8914 std::swap(LHS, RHS);
8918 Opcode = X86ISD::FMAX;
8921 // Check for x CC y ? y : x -- a min/max with reversed arms.
8922 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
8923 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
8927 // Converting this to a min would handle comparisons between positive
8928 // and negative zero incorrectly, and swapping the operands would
8929 // cause it to handle NaNs incorrectly.
8930 if (!UnsafeFPMath &&
8931 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
8932 if (!FiniteOnlyFPMath() &&
8933 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
8935 std::swap(LHS, RHS);
8937 Opcode = X86ISD::FMIN;
8940 // Converting this to a min would handle NaNs incorrectly.
8941 if (!UnsafeFPMath &&
8942 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
8944 Opcode = X86ISD::FMIN;
8947 // Converting this to a min would handle both negative zeros and NaNs
8948 // incorrectly, but we can swap the operands to fix both.
8949 std::swap(LHS, RHS);
8953 Opcode = X86ISD::FMIN;
8957 // Converting this to a max would handle NaNs incorrectly.
8958 if (!FiniteOnlyFPMath() &&
8959 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
8961 Opcode = X86ISD::FMAX;
8964 // Converting this to a max would handle comparisons between positive
8965 // and negative zero incorrectly, and swapping the operands would
8966 // cause it to handle NaNs incorrectly.
8967 if (!UnsafeFPMath &&
8968 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
8969 if (!FiniteOnlyFPMath() &&
8970 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
8972 std::swap(LHS, RHS);
8974 Opcode = X86ISD::FMAX;
8977 // Converting this to a max would handle both negative zeros and NaNs
8978 // incorrectly, but we can swap the operands to fix both.
8979 std::swap(LHS, RHS);
8983 Opcode = X86ISD::FMAX;
8989 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8992 // If this is a select between two integer constants, try to do some
8994 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8995 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8996 // Don't do this for crazy integer types.
8997 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8998 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8999 // so that TrueC (the true value) is larger than FalseC.
9000 bool NeedsCondInvert = false;
9002 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9003 // Efficiently invertible.
9004 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9005 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9006 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9007 NeedsCondInvert = true;
9008 std::swap(TrueC, FalseC);
9011 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9012 if (FalseC->getAPIntValue() == 0 &&
9013 TrueC->getAPIntValue().isPowerOf2()) {
9014 if (NeedsCondInvert) // Invert the condition if needed.
9015 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9016 DAG.getConstant(1, Cond.getValueType()));
9018 // Zero extend the condition if needed.
9019 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9021 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9022 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9023 DAG.getConstant(ShAmt, MVT::i8));
9026 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9027 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9028 if (NeedsCondInvert) // Invert the condition if needed.
9029 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9030 DAG.getConstant(1, Cond.getValueType()));
9032 // Zero extend the condition if needed.
9033 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9034 FalseC->getValueType(0), Cond);
9035 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9036 SDValue(FalseC, 0));
9039 // Optimize cases that will turn into an LEA instruction. This requires
9040 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9041 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9042 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9043 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9045 bool isFastMultiplier = false;
9047 switch ((unsigned char)Diff) {
9049 case 1: // result = add base, cond
9050 case 2: // result = lea base( , cond*2)
9051 case 3: // result = lea base(cond, cond*2)
9052 case 4: // result = lea base( , cond*4)
9053 case 5: // result = lea base(cond, cond*4)
9054 case 8: // result = lea base( , cond*8)
9055 case 9: // result = lea base(cond, cond*8)
9056 isFastMultiplier = true;
9061 if (isFastMultiplier) {
9062 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9063 if (NeedsCondInvert) // Invert the condition if needed.
9064 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9065 DAG.getConstant(1, Cond.getValueType()));
9067 // Zero extend the condition if needed.
9068 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9070 // Scale the condition by the difference.
9072 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9073 DAG.getConstant(Diff, Cond.getValueType()));
9075 // Add the base if non-zero.
9076 if (FalseC->getAPIntValue() != 0)
9077 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9078 SDValue(FalseC, 0));
9088 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9089 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9090 TargetLowering::DAGCombinerInfo &DCI) {
9091 DebugLoc DL = N->getDebugLoc();
9093 // If the flag operand isn't dead, don't touch this CMOV.
9094 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9097 // If this is a select between two integer constants, try to do some
9098 // optimizations. Note that the operands are ordered the opposite of SELECT
9100 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9101 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9102 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9103 // larger than FalseC (the false value).
9104 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9106 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9107 CC = X86::GetOppositeBranchCondition(CC);
9108 std::swap(TrueC, FalseC);
9111 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9112 // This is efficient for any integer data type (including i8/i16) and
9114 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9115 SDValue Cond = N->getOperand(3);
9116 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9117 DAG.getConstant(CC, MVT::i8), Cond);
9119 // Zero extend the condition if needed.
9120 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9122 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9123 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9124 DAG.getConstant(ShAmt, MVT::i8));
9125 if (N->getNumValues() == 2) // Dead flag value?
9126 return DCI.CombineTo(N, Cond, SDValue());
9130 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9131 // for any integer data type, including i8/i16.
9132 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9133 SDValue Cond = N->getOperand(3);
9134 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9135 DAG.getConstant(CC, MVT::i8), Cond);
9137 // Zero extend the condition if needed.
9138 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9139 FalseC->getValueType(0), Cond);
9140 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9141 SDValue(FalseC, 0));
9143 if (N->getNumValues() == 2) // Dead flag value?
9144 return DCI.CombineTo(N, Cond, SDValue());
9148 // Optimize cases that will turn into an LEA instruction. This requires
9149 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9150 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9151 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9152 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9154 bool isFastMultiplier = false;
9156 switch ((unsigned char)Diff) {
9158 case 1: // result = add base, cond
9159 case 2: // result = lea base( , cond*2)
9160 case 3: // result = lea base(cond, cond*2)
9161 case 4: // result = lea base( , cond*4)
9162 case 5: // result = lea base(cond, cond*4)
9163 case 8: // result = lea base( , cond*8)
9164 case 9: // result = lea base(cond, cond*8)
9165 isFastMultiplier = true;
9170 if (isFastMultiplier) {
9171 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9172 SDValue Cond = N->getOperand(3);
9173 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9174 DAG.getConstant(CC, MVT::i8), Cond);
9175 // Zero extend the condition if needed.
9176 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9178 // Scale the condition by the difference.
9180 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9181 DAG.getConstant(Diff, Cond.getValueType()));
9183 // Add the base if non-zero.
9184 if (FalseC->getAPIntValue() != 0)
9185 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9186 SDValue(FalseC, 0));
9187 if (N->getNumValues() == 2) // Dead flag value?
9188 return DCI.CombineTo(N, Cond, SDValue());
9198 /// PerformMulCombine - Optimize a single multiply with constant into two
9199 /// in order to implement it with two cheaper instructions, e.g.
9200 /// LEA + SHL, LEA + LEA.
9201 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9202 TargetLowering::DAGCombinerInfo &DCI) {
9203 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9206 EVT VT = N->getValueType(0);
9210 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9213 uint64_t MulAmt = C->getZExtValue();
9214 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9217 uint64_t MulAmt1 = 0;
9218 uint64_t MulAmt2 = 0;
9219 if ((MulAmt % 9) == 0) {
9221 MulAmt2 = MulAmt / 9;
9222 } else if ((MulAmt % 5) == 0) {
9224 MulAmt2 = MulAmt / 5;
9225 } else if ((MulAmt % 3) == 0) {
9227 MulAmt2 = MulAmt / 3;
9230 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9231 DebugLoc DL = N->getDebugLoc();
9233 if (isPowerOf2_64(MulAmt2) &&
9234 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9235 // If second multiplifer is pow2, issue it first. We want the multiply by
9236 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9238 std::swap(MulAmt1, MulAmt2);
9241 if (isPowerOf2_64(MulAmt1))
9242 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9243 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9245 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9246 DAG.getConstant(MulAmt1, VT));
9248 if (isPowerOf2_64(MulAmt2))
9249 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9250 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9252 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9253 DAG.getConstant(MulAmt2, VT));
9255 // Do not add new nodes to DAG combiner worklist.
9256 DCI.CombineTo(N, NewMul, false);
9261 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9262 SDValue N0 = N->getOperand(0);
9263 SDValue N1 = N->getOperand(1);
9264 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9265 EVT VT = N0.getValueType();
9267 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9268 // since the result of setcc_c is all zero's or all ones.
9269 if (N1C && N0.getOpcode() == ISD::AND &&
9270 N0.getOperand(1).getOpcode() == ISD::Constant) {
9271 SDValue N00 = N0.getOperand(0);
9272 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9273 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9274 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9275 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9276 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9277 APInt ShAmt = N1C->getAPIntValue();
9278 Mask = Mask.shl(ShAmt);
9280 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9281 N00, DAG.getConstant(Mask, VT));
9288 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9290 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9291 const X86Subtarget *Subtarget) {
9292 EVT VT = N->getValueType(0);
9293 if (!VT.isVector() && VT.isInteger() &&
9294 N->getOpcode() == ISD::SHL)
9295 return PerformSHLCombine(N, DAG);
9297 // On X86 with SSE2 support, we can transform this to a vector shift if
9298 // all elements are shifted by the same amount. We can't do this in legalize
9299 // because the a constant vector is typically transformed to a constant pool
9300 // so we have no knowledge of the shift amount.
9301 if (!Subtarget->hasSSE2())
9304 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9307 SDValue ShAmtOp = N->getOperand(1);
9308 EVT EltVT = VT.getVectorElementType();
9309 DebugLoc DL = N->getDebugLoc();
9310 SDValue BaseShAmt = SDValue();
9311 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9312 unsigned NumElts = VT.getVectorNumElements();
9314 for (; i != NumElts; ++i) {
9315 SDValue Arg = ShAmtOp.getOperand(i);
9316 if (Arg.getOpcode() == ISD::UNDEF) continue;
9320 for (; i != NumElts; ++i) {
9321 SDValue Arg = ShAmtOp.getOperand(i);
9322 if (Arg.getOpcode() == ISD::UNDEF) continue;
9323 if (Arg != BaseShAmt) {
9327 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9328 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9329 SDValue InVec = ShAmtOp.getOperand(0);
9330 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9331 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9333 for (; i != NumElts; ++i) {
9334 SDValue Arg = InVec.getOperand(i);
9335 if (Arg.getOpcode() == ISD::UNDEF) continue;
9339 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9340 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9341 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9342 if (C->getZExtValue() == SplatIdx)
9343 BaseShAmt = InVec.getOperand(1);
9346 if (BaseShAmt.getNode() == 0)
9347 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9348 DAG.getIntPtrConstant(0));
9352 // The shift amount is an i32.
9353 if (EltVT.bitsGT(MVT::i32))
9354 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9355 else if (EltVT.bitsLT(MVT::i32))
9356 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9358 // The shift amount is identical so we can do a vector shift.
9359 SDValue ValOp = N->getOperand(0);
9360 switch (N->getOpcode()) {
9362 llvm_unreachable("Unknown shift opcode!");
9365 if (VT == MVT::v2i64)
9366 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9367 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9369 if (VT == MVT::v4i32)
9370 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9371 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9373 if (VT == MVT::v8i16)
9374 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9375 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9379 if (VT == MVT::v4i32)
9380 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9381 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9383 if (VT == MVT::v8i16)
9384 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9385 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9389 if (VT == MVT::v2i64)
9390 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9391 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9393 if (VT == MVT::v4i32)
9394 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9395 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9397 if (VT == MVT::v8i16)
9398 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9399 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9406 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9407 const X86Subtarget *Subtarget) {
9408 EVT VT = N->getValueType(0);
9409 if (VT != MVT::i64 || !Subtarget->is64Bit())
9412 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9413 SDValue N0 = N->getOperand(0);
9414 SDValue N1 = N->getOperand(1);
9415 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9417 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9420 SDValue ShAmt0 = N0.getOperand(1);
9421 if (ShAmt0.getValueType() != MVT::i8)
9423 SDValue ShAmt1 = N1.getOperand(1);
9424 if (ShAmt1.getValueType() != MVT::i8)
9426 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9427 ShAmt0 = ShAmt0.getOperand(0);
9428 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9429 ShAmt1 = ShAmt1.getOperand(0);
9431 DebugLoc DL = N->getDebugLoc();
9432 unsigned Opc = X86ISD::SHLD;
9433 SDValue Op0 = N0.getOperand(0);
9434 SDValue Op1 = N1.getOperand(0);
9435 if (ShAmt0.getOpcode() == ISD::SUB) {
9437 std::swap(Op0, Op1);
9438 std::swap(ShAmt0, ShAmt1);
9441 if (ShAmt1.getOpcode() == ISD::SUB) {
9442 SDValue Sum = ShAmt1.getOperand(0);
9443 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9444 if (SumC->getSExtValue() == 64 &&
9445 ShAmt1.getOperand(1) == ShAmt0)
9446 return DAG.getNode(Opc, DL, VT,
9448 DAG.getNode(ISD::TRUNCATE, DL,
9451 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9452 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9454 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
9455 return DAG.getNode(Opc, DL, VT,
9456 N0.getOperand(0), N1.getOperand(0),
9457 DAG.getNode(ISD::TRUNCATE, DL,
9464 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9465 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9466 const X86Subtarget *Subtarget) {
9467 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9468 // the FP state in cases where an emms may be missing.
9469 // A preferable solution to the general problem is to figure out the right
9470 // places to insert EMMS. This qualifies as a quick hack.
9472 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9473 StoreSDNode *St = cast<StoreSDNode>(N);
9474 EVT VT = St->getValue().getValueType();
9475 if (VT.getSizeInBits() != 64)
9478 const Function *F = DAG.getMachineFunction().getFunction();
9479 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9480 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9481 && Subtarget->hasSSE2();
9482 if ((VT.isVector() ||
9483 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9484 isa<LoadSDNode>(St->getValue()) &&
9485 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9486 St->getChain().hasOneUse() && !St->isVolatile()) {
9487 SDNode* LdVal = St->getValue().getNode();
9489 int TokenFactorIndex = -1;
9490 SmallVector<SDValue, 8> Ops;
9491 SDNode* ChainVal = St->getChain().getNode();
9492 // Must be a store of a load. We currently handle two cases: the load
9493 // is a direct child, and it's under an intervening TokenFactor. It is
9494 // possible to dig deeper under nested TokenFactors.
9495 if (ChainVal == LdVal)
9496 Ld = cast<LoadSDNode>(St->getChain());
9497 else if (St->getValue().hasOneUse() &&
9498 ChainVal->getOpcode() == ISD::TokenFactor) {
9499 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9500 if (ChainVal->getOperand(i).getNode() == LdVal) {
9501 TokenFactorIndex = i;
9502 Ld = cast<LoadSDNode>(St->getValue());
9504 Ops.push_back(ChainVal->getOperand(i));
9508 if (!Ld || !ISD::isNormalLoad(Ld))
9511 // If this is not the MMX case, i.e. we are just turning i64 load/store
9512 // into f64 load/store, avoid the transformation if there are multiple
9513 // uses of the loaded value.
9514 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9517 DebugLoc LdDL = Ld->getDebugLoc();
9518 DebugLoc StDL = N->getDebugLoc();
9519 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9520 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9522 if (Subtarget->is64Bit() || F64IsLegal) {
9523 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9524 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9525 Ld->getBasePtr(), Ld->getSrcValue(),
9526 Ld->getSrcValueOffset(), Ld->isVolatile(),
9527 Ld->isNonTemporal(), Ld->getAlignment());
9528 SDValue NewChain = NewLd.getValue(1);
9529 if (TokenFactorIndex != -1) {
9530 Ops.push_back(NewChain);
9531 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9534 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9535 St->getSrcValue(), St->getSrcValueOffset(),
9536 St->isVolatile(), St->isNonTemporal(),
9537 St->getAlignment());
9540 // Otherwise, lower to two pairs of 32-bit loads / stores.
9541 SDValue LoAddr = Ld->getBasePtr();
9542 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9543 DAG.getConstant(4, MVT::i32));
9545 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9546 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9547 Ld->isVolatile(), Ld->isNonTemporal(),
9548 Ld->getAlignment());
9549 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9550 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9551 Ld->isVolatile(), Ld->isNonTemporal(),
9552 MinAlign(Ld->getAlignment(), 4));
9554 SDValue NewChain = LoLd.getValue(1);
9555 if (TokenFactorIndex != -1) {
9556 Ops.push_back(LoLd);
9557 Ops.push_back(HiLd);
9558 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9562 LoAddr = St->getBasePtr();
9563 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9564 DAG.getConstant(4, MVT::i32));
9566 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9567 St->getSrcValue(), St->getSrcValueOffset(),
9568 St->isVolatile(), St->isNonTemporal(),
9569 St->getAlignment());
9570 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9572 St->getSrcValueOffset() + 4,
9574 St->isNonTemporal(),
9575 MinAlign(St->getAlignment(), 4));
9576 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9581 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9582 /// X86ISD::FXOR nodes.
9583 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9584 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9585 // F[X]OR(0.0, x) -> x
9586 // F[X]OR(x, 0.0) -> x
9587 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9588 if (C->getValueAPF().isPosZero())
9589 return N->getOperand(1);
9590 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9591 if (C->getValueAPF().isPosZero())
9592 return N->getOperand(0);
9596 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9597 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9598 // FAND(0.0, x) -> 0.0
9599 // FAND(x, 0.0) -> 0.0
9600 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9601 if (C->getValueAPF().isPosZero())
9602 return N->getOperand(0);
9603 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9604 if (C->getValueAPF().isPosZero())
9605 return N->getOperand(1);
9609 static SDValue PerformBTCombine(SDNode *N,
9611 TargetLowering::DAGCombinerInfo &DCI) {
9612 // BT ignores high bits in the bit index operand.
9613 SDValue Op1 = N->getOperand(1);
9614 if (Op1.hasOneUse()) {
9615 unsigned BitWidth = Op1.getValueSizeInBits();
9616 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9617 APInt KnownZero, KnownOne;
9618 TargetLowering::TargetLoweringOpt TLO(DAG);
9619 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9620 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9621 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9622 DCI.CommitTargetLoweringOpt(TLO);
9627 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9628 SDValue Op = N->getOperand(0);
9629 if (Op.getOpcode() == ISD::BIT_CONVERT)
9630 Op = Op.getOperand(0);
9631 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9632 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9633 VT.getVectorElementType().getSizeInBits() ==
9634 OpVT.getVectorElementType().getSizeInBits()) {
9635 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9640 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9641 // Locked instructions, in turn, have implicit fence semantics (all memory
9642 // operations are flushed before issuing the locked instruction, and the
9643 // are not buffered), so we can fold away the common pattern of
9644 // fence-atomic-fence.
9645 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9646 SDValue atomic = N->getOperand(0);
9647 switch (atomic.getOpcode()) {
9648 case ISD::ATOMIC_CMP_SWAP:
9649 case ISD::ATOMIC_SWAP:
9650 case ISD::ATOMIC_LOAD_ADD:
9651 case ISD::ATOMIC_LOAD_SUB:
9652 case ISD::ATOMIC_LOAD_AND:
9653 case ISD::ATOMIC_LOAD_OR:
9654 case ISD::ATOMIC_LOAD_XOR:
9655 case ISD::ATOMIC_LOAD_NAND:
9656 case ISD::ATOMIC_LOAD_MIN:
9657 case ISD::ATOMIC_LOAD_MAX:
9658 case ISD::ATOMIC_LOAD_UMIN:
9659 case ISD::ATOMIC_LOAD_UMAX:
9665 SDValue fence = atomic.getOperand(0);
9666 if (fence.getOpcode() != ISD::MEMBARRIER)
9669 switch (atomic.getOpcode()) {
9670 case ISD::ATOMIC_CMP_SWAP:
9671 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9672 atomic.getOperand(1), atomic.getOperand(2),
9673 atomic.getOperand(3));
9674 case ISD::ATOMIC_SWAP:
9675 case ISD::ATOMIC_LOAD_ADD:
9676 case ISD::ATOMIC_LOAD_SUB:
9677 case ISD::ATOMIC_LOAD_AND:
9678 case ISD::ATOMIC_LOAD_OR:
9679 case ISD::ATOMIC_LOAD_XOR:
9680 case ISD::ATOMIC_LOAD_NAND:
9681 case ISD::ATOMIC_LOAD_MIN:
9682 case ISD::ATOMIC_LOAD_MAX:
9683 case ISD::ATOMIC_LOAD_UMIN:
9684 case ISD::ATOMIC_LOAD_UMAX:
9685 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9686 atomic.getOperand(1), atomic.getOperand(2));
9692 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9693 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9694 // (and (i32 x86isd::setcc_carry), 1)
9695 // This eliminates the zext. This transformation is necessary because
9696 // ISD::SETCC is always legalized to i8.
9697 DebugLoc dl = N->getDebugLoc();
9698 SDValue N0 = N->getOperand(0);
9699 EVT VT = N->getValueType(0);
9700 if (N0.getOpcode() == ISD::AND &&
9702 N0.getOperand(0).hasOneUse()) {
9703 SDValue N00 = N0.getOperand(0);
9704 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9706 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9707 if (!C || C->getZExtValue() != 1)
9709 return DAG.getNode(ISD::AND, dl, VT,
9710 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9711 N00.getOperand(0), N00.getOperand(1)),
9712 DAG.getConstant(1, VT));
9718 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9719 DAGCombinerInfo &DCI) const {
9720 SelectionDAG &DAG = DCI.DAG;
9721 switch (N->getOpcode()) {
9723 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9724 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9725 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9726 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9729 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9730 case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
9731 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9733 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9734 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9735 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9736 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9737 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9738 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9744 //===----------------------------------------------------------------------===//
9745 // X86 Inline Assembly Support
9746 //===----------------------------------------------------------------------===//
9748 static bool LowerToBSwap(CallInst *CI) {
9749 // FIXME: this should verify that we are targetting a 486 or better. If not,
9750 // we will turn this bswap into something that will be lowered to logical ops
9751 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9752 // so don't worry about this.
9754 // Verify this is a simple bswap.
9755 if (CI->getNumOperands() != 2 ||
9756 CI->getType() != CI->getOperand(1)->getType() ||
9757 !CI->getType()->isIntegerTy())
9760 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9761 if (!Ty || Ty->getBitWidth() % 16 != 0)
9764 // Okay, we can do this xform, do so now.
9765 const Type *Tys[] = { Ty };
9766 Module *M = CI->getParent()->getParent()->getParent();
9767 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9769 Value *Op = CI->getOperand(1);
9770 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9772 CI->replaceAllUsesWith(Op);
9773 CI->eraseFromParent();
9777 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9778 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9779 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9781 std::string AsmStr = IA->getAsmString();
9783 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9784 SmallVector<StringRef, 4> AsmPieces;
9785 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9787 switch (AsmPieces.size()) {
9788 default: return false;
9790 AsmStr = AsmPieces[0];
9792 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9795 if (AsmPieces.size() == 2 &&
9796 (AsmPieces[0] == "bswap" ||
9797 AsmPieces[0] == "bswapq" ||
9798 AsmPieces[0] == "bswapl") &&
9799 (AsmPieces[1] == "$0" ||
9800 AsmPieces[1] == "${0:q}")) {
9801 // No need to check constraints, nothing other than the equivalent of
9802 // "=r,0" would be valid here.
9803 return LowerToBSwap(CI);
9805 // rorw $$8, ${0:w} --> llvm.bswap.i16
9806 if (CI->getType()->isIntegerTy(16) &&
9807 AsmPieces.size() == 3 &&
9808 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
9809 AsmPieces[1] == "$$8," &&
9810 AsmPieces[2] == "${0:w}" &&
9811 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
9813 SplitString(IA->getConstraintString().substr(5), AsmPieces, ",");
9814 std::sort(AsmPieces.begin(), AsmPieces.end());
9815 if (AsmPieces.size() == 4 &&
9816 AsmPieces[0] == "~{cc}" &&
9817 AsmPieces[1] == "~{dirflag}" &&
9818 AsmPieces[2] == "~{flags}" &&
9819 AsmPieces[3] == "~{fpsr}") {
9820 return LowerToBSwap(CI);
9825 if (CI->getType()->isIntegerTy(64) &&
9826 Constraints.size() >= 2 &&
9827 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9828 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9829 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9830 SmallVector<StringRef, 4> Words;
9831 SplitString(AsmPieces[0], Words, " \t");
9832 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9834 SplitString(AsmPieces[1], Words, " \t");
9835 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9837 SplitString(AsmPieces[2], Words, " \t,");
9838 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9839 Words[2] == "%edx") {
9840 return LowerToBSwap(CI);
9852 /// getConstraintType - Given a constraint letter, return the type of
9853 /// constraint it is for this target.
9854 X86TargetLowering::ConstraintType
9855 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
9856 if (Constraint.size() == 1) {
9857 switch (Constraint[0]) {
9869 return C_RegisterClass;
9877 return TargetLowering::getConstraintType(Constraint);
9880 /// LowerXConstraint - try to replace an X constraint, which matches anything,
9881 /// with another that has more specific requirements based on the type of the
9882 /// corresponding operand.
9883 const char *X86TargetLowering::
9884 LowerXConstraint(EVT ConstraintVT) const {
9885 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
9886 // 'f' like normal targets.
9887 if (ConstraintVT.isFloatingPoint()) {
9888 if (Subtarget->hasSSE2())
9890 if (Subtarget->hasSSE1())
9894 return TargetLowering::LowerXConstraint(ConstraintVT);
9897 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9898 /// vector. If it is invalid, don't add anything to Ops.
9899 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9902 std::vector<SDValue>&Ops,
9903 SelectionDAG &DAG) const {
9904 SDValue Result(0, 0);
9906 switch (Constraint) {
9909 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9910 if (C->getZExtValue() <= 31) {
9911 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9917 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9918 if (C->getZExtValue() <= 63) {
9919 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9925 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9926 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9927 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9933 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9934 if (C->getZExtValue() <= 255) {
9935 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9941 // 32-bit signed value
9942 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9943 const ConstantInt *CI = C->getConstantIntValue();
9944 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9945 C->getSExtValue())) {
9946 // Widen to 64 bits here to get it sign extended.
9947 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9950 // FIXME gcc accepts some relocatable values here too, but only in certain
9951 // memory models; it's complicated.
9956 // 32-bit unsigned value
9957 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9958 const ConstantInt *CI = C->getConstantIntValue();
9959 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9960 C->getZExtValue())) {
9961 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9965 // FIXME gcc accepts some relocatable values here too, but only in certain
9966 // memory models; it's complicated.
9970 // Literal immediates are always ok.
9971 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
9972 // Widen to 64 bits here to get it sign extended.
9973 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
9977 // If we are in non-pic codegen mode, we allow the address of a global (with
9978 // an optional displacement) to be used with 'i'.
9979 GlobalAddressSDNode *GA = 0;
9982 // Match either (GA), (GA+C), (GA+C1+C2), etc.
9984 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
9985 Offset += GA->getOffset();
9987 } else if (Op.getOpcode() == ISD::ADD) {
9988 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9989 Offset += C->getZExtValue();
9990 Op = Op.getOperand(0);
9993 } else if (Op.getOpcode() == ISD::SUB) {
9994 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9995 Offset += -C->getZExtValue();
9996 Op = Op.getOperand(0);
10001 // Otherwise, this isn't something we can handle, reject it.
10005 GlobalValue *GV = GA->getGlobal();
10006 // If we require an extra load to get this address, as in PIC mode, we
10007 // can't accept it.
10008 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10009 getTargetMachine())))
10013 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
10015 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
10021 if (Result.getNode()) {
10022 Ops.push_back(Result);
10025 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
10029 std::vector<unsigned> X86TargetLowering::
10030 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10032 if (Constraint.size() == 1) {
10033 // FIXME: not handling fp-stack yet!
10034 switch (Constraint[0]) { // GCC X86 Constraint Letters
10035 default: break; // Unknown constraint letter
10036 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10037 if (Subtarget->is64Bit()) {
10038 if (VT == MVT::i32)
10039 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10040 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10041 X86::R10D,X86::R11D,X86::R12D,
10042 X86::R13D,X86::R14D,X86::R15D,
10043 X86::EBP, X86::ESP, 0);
10044 else if (VT == MVT::i16)
10045 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10046 X86::SI, X86::DI, X86::R8W,X86::R9W,
10047 X86::R10W,X86::R11W,X86::R12W,
10048 X86::R13W,X86::R14W,X86::R15W,
10049 X86::BP, X86::SP, 0);
10050 else if (VT == MVT::i8)
10051 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10052 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10053 X86::R10B,X86::R11B,X86::R12B,
10054 X86::R13B,X86::R14B,X86::R15B,
10055 X86::BPL, X86::SPL, 0);
10057 else if (VT == MVT::i64)
10058 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10059 X86::RSI, X86::RDI, X86::R8, X86::R9,
10060 X86::R10, X86::R11, X86::R12,
10061 X86::R13, X86::R14, X86::R15,
10062 X86::RBP, X86::RSP, 0);
10066 // 32-bit fallthrough
10067 case 'Q': // Q_REGS
10068 if (VT == MVT::i32)
10069 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10070 else if (VT == MVT::i16)
10071 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10072 else if (VT == MVT::i8)
10073 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10074 else if (VT == MVT::i64)
10075 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10080 return std::vector<unsigned>();
10083 std::pair<unsigned, const TargetRegisterClass*>
10084 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10086 // First, see if this is a constraint that directly corresponds to an LLVM
10088 if (Constraint.size() == 1) {
10089 // GCC Constraint Letters
10090 switch (Constraint[0]) {
10092 case 'r': // GENERAL_REGS
10093 case 'l': // INDEX_REGS
10095 return std::make_pair(0U, X86::GR8RegisterClass);
10096 if (VT == MVT::i16)
10097 return std::make_pair(0U, X86::GR16RegisterClass);
10098 if (VT == MVT::i32 || !Subtarget->is64Bit())
10099 return std::make_pair(0U, X86::GR32RegisterClass);
10100 return std::make_pair(0U, X86::GR64RegisterClass);
10101 case 'R': // LEGACY_REGS
10103 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10104 if (VT == MVT::i16)
10105 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10106 if (VT == MVT::i32 || !Subtarget->is64Bit())
10107 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10108 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10109 case 'f': // FP Stack registers.
10110 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10111 // value to the correct fpstack register class.
10112 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10113 return std::make_pair(0U, X86::RFP32RegisterClass);
10114 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10115 return std::make_pair(0U, X86::RFP64RegisterClass);
10116 return std::make_pair(0U, X86::RFP80RegisterClass);
10117 case 'y': // MMX_REGS if MMX allowed.
10118 if (!Subtarget->hasMMX()) break;
10119 return std::make_pair(0U, X86::VR64RegisterClass);
10120 case 'Y': // SSE_REGS if SSE2 allowed
10121 if (!Subtarget->hasSSE2()) break;
10123 case 'x': // SSE_REGS if SSE1 allowed
10124 if (!Subtarget->hasSSE1()) break;
10126 switch (VT.getSimpleVT().SimpleTy) {
10128 // Scalar SSE types.
10131 return std::make_pair(0U, X86::FR32RegisterClass);
10134 return std::make_pair(0U, X86::FR64RegisterClass);
10142 return std::make_pair(0U, X86::VR128RegisterClass);
10148 // Use the default implementation in TargetLowering to convert the register
10149 // constraint into a member of a register class.
10150 std::pair<unsigned, const TargetRegisterClass*> Res;
10151 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10153 // Not found as a standard register?
10154 if (Res.second == 0) {
10155 // Map st(0) -> st(7) -> ST0
10156 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10157 tolower(Constraint[1]) == 's' &&
10158 tolower(Constraint[2]) == 't' &&
10159 Constraint[3] == '(' &&
10160 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10161 Constraint[5] == ')' &&
10162 Constraint[6] == '}') {
10164 Res.first = X86::ST0+Constraint[4]-'0';
10165 Res.second = X86::RFP80RegisterClass;
10169 // GCC allows "st(0)" to be called just plain "st".
10170 if (StringRef("{st}").equals_lower(Constraint)) {
10171 Res.first = X86::ST0;
10172 Res.second = X86::RFP80RegisterClass;
10177 if (StringRef("{flags}").equals_lower(Constraint)) {
10178 Res.first = X86::EFLAGS;
10179 Res.second = X86::CCRRegisterClass;
10183 // 'A' means EAX + EDX.
10184 if (Constraint == "A") {
10185 Res.first = X86::EAX;
10186 Res.second = X86::GR32_ADRegisterClass;
10192 // Otherwise, check to see if this is a register class of the wrong value
10193 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10194 // turn into {ax},{dx}.
10195 if (Res.second->hasType(VT))
10196 return Res; // Correct type already, nothing to do.
10198 // All of the single-register GCC register classes map their values onto
10199 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10200 // really want an 8-bit or 32-bit register, map to the appropriate register
10201 // class and return the appropriate register.
10202 if (Res.second == X86::GR16RegisterClass) {
10203 if (VT == MVT::i8) {
10204 unsigned DestReg = 0;
10205 switch (Res.first) {
10207 case X86::AX: DestReg = X86::AL; break;
10208 case X86::DX: DestReg = X86::DL; break;
10209 case X86::CX: DestReg = X86::CL; break;
10210 case X86::BX: DestReg = X86::BL; break;
10213 Res.first = DestReg;
10214 Res.second = X86::GR8RegisterClass;
10216 } else if (VT == MVT::i32) {
10217 unsigned DestReg = 0;
10218 switch (Res.first) {
10220 case X86::AX: DestReg = X86::EAX; break;
10221 case X86::DX: DestReg = X86::EDX; break;
10222 case X86::CX: DestReg = X86::ECX; break;
10223 case X86::BX: DestReg = X86::EBX; break;
10224 case X86::SI: DestReg = X86::ESI; break;
10225 case X86::DI: DestReg = X86::EDI; break;
10226 case X86::BP: DestReg = X86::EBP; break;
10227 case X86::SP: DestReg = X86::ESP; break;
10230 Res.first = DestReg;
10231 Res.second = X86::GR32RegisterClass;
10233 } else if (VT == MVT::i64) {
10234 unsigned DestReg = 0;
10235 switch (Res.first) {
10237 case X86::AX: DestReg = X86::RAX; break;
10238 case X86::DX: DestReg = X86::RDX; break;
10239 case X86::CX: DestReg = X86::RCX; break;
10240 case X86::BX: DestReg = X86::RBX; break;
10241 case X86::SI: DestReg = X86::RSI; break;
10242 case X86::DI: DestReg = X86::RDI; break;
10243 case X86::BP: DestReg = X86::RBP; break;
10244 case X86::SP: DestReg = X86::RSP; break;
10247 Res.first = DestReg;
10248 Res.second = X86::GR64RegisterClass;
10251 } else if (Res.second == X86::FR32RegisterClass ||
10252 Res.second == X86::FR64RegisterClass ||
10253 Res.second == X86::VR128RegisterClass) {
10254 // Handle references to XMM physical registers that got mapped into the
10255 // wrong class. This can happen with constraints like {xmm0} where the
10256 // target independent register mapper will just pick the first match it can
10257 // find, ignoring the required type.
10258 if (VT == MVT::f32)
10259 Res.second = X86::FR32RegisterClass;
10260 else if (VT == MVT::f64)
10261 Res.second = X86::FR64RegisterClass;
10262 else if (X86::VR128RegisterClass->hasType(VT))
10263 Res.second = X86::VR128RegisterClass;
10269 //===----------------------------------------------------------------------===//
10270 // X86 Widen vector type
10271 //===----------------------------------------------------------------------===//
10273 /// getWidenVectorType: given a vector type, returns the type to widen
10274 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
10275 /// If there is no vector type that we want to widen to, returns MVT::Other
10276 /// When and where to widen is target dependent based on the cost of
10277 /// scalarizing vs using the wider vector type.
10279 EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
10280 assert(VT.isVector());
10281 if (isTypeLegal(VT))
10284 // TODO: In computeRegisterProperty, we can compute the list of legal vector
10285 // type based on element type. This would speed up our search (though
10286 // it may not be worth it since the size of the list is relatively
10288 EVT EltVT = VT.getVectorElementType();
10289 unsigned NElts = VT.getVectorNumElements();
10291 // On X86, it make sense to widen any vector wider than 1
10295 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
10296 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
10297 EVT SVT = (MVT::SimpleValueType)nVT;
10299 if (isTypeLegal(SVT) &&
10300 SVT.getVectorElementType() == EltVT &&
10301 SVT.getVectorNumElements() > NElts)