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 //===----------------------------------------------------------------------===//
16 #include "X86InstrBuilder.h"
17 #include "X86ISelLowering.h"
18 #include "X86TargetMachine.h"
19 #include "llvm/CallingConv.h"
20 #include "llvm/Constants.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/GlobalAlias.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/Function.h"
25 #include "llvm/Instructions.h"
26 #include "llvm/Intrinsics.h"
27 #include "llvm/LLVMContext.h"
28 #include "llvm/ADT/BitVector.h"
29 #include "llvm/ADT/VectorExtras.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/PseudoSourceValue.h"
36 #include "llvm/Support/MathExtras.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/ErrorHandling.h"
39 #include "llvm/Target/TargetLoweringObjectFile.h"
40 #include "llvm/Target/TargetOptions.h"
41 #include "llvm/ADT/SmallSet.h"
42 #include "llvm/ADT/StringExtras.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/Support/raw_ostream.h"
48 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
50 // Forward declarations.
51 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
54 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
55 switch (TM.getSubtarget<X86Subtarget>().TargetType) {
56 default: llvm_unreachable("unknown subtarget type");
57 case X86Subtarget::isDarwin:
58 return new TargetLoweringObjectFileMachO();
59 case X86Subtarget::isELF:
60 return new TargetLoweringObjectFileELF();
61 case X86Subtarget::isMingw:
62 case X86Subtarget::isCygwin:
63 case X86Subtarget::isWindows:
64 return new TargetLoweringObjectFileCOFF();
69 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
70 : TargetLowering(TM, createTLOF(TM)) {
71 Subtarget = &TM.getSubtarget<X86Subtarget>();
72 X86ScalarSSEf64 = Subtarget->hasSSE2();
73 X86ScalarSSEf32 = Subtarget->hasSSE1();
74 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
76 RegInfo = TM.getRegisterInfo();
79 // Set up the TargetLowering object.
81 // X86 is weird, it always uses i8 for shift amounts and setcc results.
82 setShiftAmountType(MVT::i8);
83 setBooleanContents(ZeroOrOneBooleanContent);
84 setSchedulingPreference(SchedulingForRegPressure);
85 setStackPointerRegisterToSaveRestore(X86StackPtr);
87 if (Subtarget->isTargetDarwin()) {
88 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
89 setUseUnderscoreSetJmp(false);
90 setUseUnderscoreLongJmp(false);
91 } else if (Subtarget->isTargetMingw()) {
92 // MS runtime is weird: it exports _setjmp, but longjmp!
93 setUseUnderscoreSetJmp(true);
94 setUseUnderscoreLongJmp(false);
96 setUseUnderscoreSetJmp(true);
97 setUseUnderscoreLongJmp(true);
100 // Set up the register classes.
101 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
102 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
103 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
104 if (Subtarget->is64Bit())
105 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
107 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
109 // We don't accept any truncstore of integer registers.
110 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
111 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
112 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
113 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
114 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
115 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
117 // SETOEQ and SETUNE require checking two conditions.
118 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
119 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
120 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
121 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
122 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
123 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
125 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
127 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
128 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
129 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
131 if (Subtarget->is64Bit()) {
132 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
133 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
134 } else if (!UseSoftFloat) {
135 if (X86ScalarSSEf64) {
136 // We have an impenetrably clever algorithm for ui64->double only.
137 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
139 // We have an algorithm for SSE2, and we turn this into a 64-bit
140 // FILD for other targets.
141 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
144 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
146 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
147 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
150 // SSE has no i16 to fp conversion, only i32
151 if (X86ScalarSSEf32) {
152 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
153 // f32 and f64 cases are Legal, f80 case is not
154 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
156 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
157 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
160 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
161 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
164 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
165 // are Legal, f80 is custom lowered.
166 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
167 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
169 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
171 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
172 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
174 if (X86ScalarSSEf32) {
175 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
176 // f32 and f64 cases are Legal, f80 case is not
177 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
179 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
180 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
183 // Handle FP_TO_UINT by promoting the destination to a larger signed
185 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
186 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
187 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
189 if (Subtarget->is64Bit()) {
190 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
191 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
192 } else if (!UseSoftFloat) {
193 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
194 // Expand FP_TO_UINT into a select.
195 // FIXME: We would like to use a Custom expander here eventually to do
196 // the optimal thing for SSE vs. the default expansion in the legalizer.
197 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
199 // With SSE3 we can use fisttpll to convert to a signed i64; without
200 // SSE, we're stuck with a fistpll.
201 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
204 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
205 if (!X86ScalarSSEf64) {
206 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
207 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
210 // Scalar integer divide and remainder are lowered to use operations that
211 // produce two results, to match the available instructions. This exposes
212 // the two-result form to trivial CSE, which is able to combine x/y and x%y
213 // into a single instruction.
215 // Scalar integer multiply-high is also lowered to use two-result
216 // operations, to match the available instructions. However, plain multiply
217 // (low) operations are left as Legal, as there are single-result
218 // instructions for this in x86. Using the two-result multiply instructions
219 // when both high and low results are needed must be arranged by dagcombine.
220 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
221 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
222 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
223 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
224 setOperationAction(ISD::SREM , MVT::i8 , Expand);
225 setOperationAction(ISD::UREM , MVT::i8 , Expand);
226 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
227 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
228 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
229 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
230 setOperationAction(ISD::SREM , MVT::i16 , Expand);
231 setOperationAction(ISD::UREM , MVT::i16 , Expand);
232 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
233 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
234 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
235 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
236 setOperationAction(ISD::SREM , MVT::i32 , Expand);
237 setOperationAction(ISD::UREM , MVT::i32 , Expand);
238 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
239 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
240 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
241 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
242 setOperationAction(ISD::SREM , MVT::i64 , Expand);
243 setOperationAction(ISD::UREM , MVT::i64 , Expand);
245 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
246 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
247 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
248 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
249 if (Subtarget->is64Bit())
250 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
251 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
252 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
253 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
254 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
255 setOperationAction(ISD::FREM , MVT::f32 , Expand);
256 setOperationAction(ISD::FREM , MVT::f64 , Expand);
257 setOperationAction(ISD::FREM , MVT::f80 , Expand);
258 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
260 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
261 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
262 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
263 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
264 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
265 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
266 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
267 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
268 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
269 if (Subtarget->is64Bit()) {
270 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
271 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
272 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
275 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
276 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
278 // These should be promoted to a larger select which is supported.
279 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
280 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
281 // X86 wants to expand cmov itself.
282 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
283 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
284 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
285 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
286 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
287 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
288 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
289 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
290 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
291 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
292 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
293 if (Subtarget->is64Bit()) {
294 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
295 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
297 // X86 ret instruction may pop stack.
298 setOperationAction(ISD::RET , MVT::Other, Custom);
299 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
302 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
303 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
304 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
305 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
306 if (Subtarget->is64Bit())
307 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
308 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
309 if (Subtarget->is64Bit()) {
310 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
311 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
312 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
313 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
315 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
316 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
317 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
318 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
319 if (Subtarget->is64Bit()) {
320 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
321 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
322 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
325 if (Subtarget->hasSSE1())
326 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
328 if (!Subtarget->hasSSE2())
329 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
331 // Expand certain atomics
332 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
333 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
334 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
335 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
337 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
338 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
339 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
340 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
342 if (!Subtarget->is64Bit()) {
343 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
344 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
345 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
346 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
347 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
348 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
349 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
352 // Use the default ISD::DBG_STOPPOINT, ISD::DECLARE expansion.
353 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
354 // FIXME - use subtarget debug flags
355 if (!Subtarget->isTargetDarwin() &&
356 !Subtarget->isTargetELF() &&
357 !Subtarget->isTargetCygMing()) {
358 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
359 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
362 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
363 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
364 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
365 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
366 if (Subtarget->is64Bit()) {
367 setExceptionPointerRegister(X86::RAX);
368 setExceptionSelectorRegister(X86::RDX);
370 setExceptionPointerRegister(X86::EAX);
371 setExceptionSelectorRegister(X86::EDX);
373 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
374 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
376 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
378 setOperationAction(ISD::TRAP, MVT::Other, Legal);
380 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
381 setOperationAction(ISD::VASTART , MVT::Other, Custom);
382 setOperationAction(ISD::VAEND , MVT::Other, Expand);
383 if (Subtarget->is64Bit()) {
384 setOperationAction(ISD::VAARG , MVT::Other, Custom);
385 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
387 setOperationAction(ISD::VAARG , MVT::Other, Expand);
388 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
391 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
392 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
393 if (Subtarget->is64Bit())
394 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
395 if (Subtarget->isTargetCygMing())
396 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
398 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
400 if (!UseSoftFloat && X86ScalarSSEf64) {
401 // f32 and f64 use SSE.
402 // Set up the FP register classes.
403 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
404 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
406 // Use ANDPD to simulate FABS.
407 setOperationAction(ISD::FABS , MVT::f64, Custom);
408 setOperationAction(ISD::FABS , MVT::f32, Custom);
410 // Use XORP to simulate FNEG.
411 setOperationAction(ISD::FNEG , MVT::f64, Custom);
412 setOperationAction(ISD::FNEG , MVT::f32, Custom);
414 // Use ANDPD and ORPD to simulate FCOPYSIGN.
415 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
416 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
418 // We don't support sin/cos/fmod
419 setOperationAction(ISD::FSIN , MVT::f64, Expand);
420 setOperationAction(ISD::FCOS , MVT::f64, Expand);
421 setOperationAction(ISD::FSIN , MVT::f32, Expand);
422 setOperationAction(ISD::FCOS , MVT::f32, Expand);
424 // Expand FP immediates into loads from the stack, except for the special
426 addLegalFPImmediate(APFloat(+0.0)); // xorpd
427 addLegalFPImmediate(APFloat(+0.0f)); // xorps
428 } else if (!UseSoftFloat && X86ScalarSSEf32) {
429 // Use SSE for f32, x87 for f64.
430 // Set up the FP register classes.
431 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
432 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
434 // Use ANDPS to simulate FABS.
435 setOperationAction(ISD::FABS , MVT::f32, Custom);
437 // Use XORP to simulate FNEG.
438 setOperationAction(ISD::FNEG , MVT::f32, Custom);
440 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
442 // Use ANDPS and ORPS to simulate FCOPYSIGN.
443 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
444 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
446 // We don't support sin/cos/fmod
447 setOperationAction(ISD::FSIN , MVT::f32, Expand);
448 setOperationAction(ISD::FCOS , MVT::f32, Expand);
450 // Special cases we handle for FP constants.
451 addLegalFPImmediate(APFloat(+0.0f)); // xorps
452 addLegalFPImmediate(APFloat(+0.0)); // FLD0
453 addLegalFPImmediate(APFloat(+1.0)); // FLD1
454 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
455 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
458 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
459 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
461 } else if (!UseSoftFloat) {
462 // f32 and f64 in x87.
463 // Set up the FP register classes.
464 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
465 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
467 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
468 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
469 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
470 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
473 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
474 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
476 addLegalFPImmediate(APFloat(+0.0)); // FLD0
477 addLegalFPImmediate(APFloat(+1.0)); // FLD1
478 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
479 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
480 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
481 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
482 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
483 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
486 // Long double always uses X87.
488 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
489 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
490 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
493 APFloat TmpFlt(+0.0);
494 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
496 addLegalFPImmediate(TmpFlt); // FLD0
498 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
499 APFloat TmpFlt2(+1.0);
500 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
502 addLegalFPImmediate(TmpFlt2); // FLD1
503 TmpFlt2.changeSign();
504 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
508 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
509 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
513 // Always use a library call for pow.
514 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
515 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
516 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
518 setOperationAction(ISD::FLOG, MVT::f80, Expand);
519 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
520 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
521 setOperationAction(ISD::FEXP, MVT::f80, Expand);
522 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
524 // First set operation action for all vector types to either promote
525 // (for widening) or expand (for scalarization). Then we will selectively
526 // turn on ones that can be effectively codegen'd.
527 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
528 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
529 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
530 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
531 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
532 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
533 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
534 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
535 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
536 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
537 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
538 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
539 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
540 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
541 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
542 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
543 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
544 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
545 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
546 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
547 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
548 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
549 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
550 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
579 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
580 // with -msoft-float, disable use of MMX as well.
581 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
582 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
583 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
584 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
585 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
586 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
588 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
589 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
590 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
591 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
593 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
594 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
595 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
596 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
598 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
599 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
601 setOperationAction(ISD::AND, MVT::v8i8, Promote);
602 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
603 setOperationAction(ISD::AND, MVT::v4i16, Promote);
604 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
605 setOperationAction(ISD::AND, MVT::v2i32, Promote);
606 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
607 setOperationAction(ISD::AND, MVT::v1i64, Legal);
609 setOperationAction(ISD::OR, MVT::v8i8, Promote);
610 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
611 setOperationAction(ISD::OR, MVT::v4i16, Promote);
612 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
613 setOperationAction(ISD::OR, MVT::v2i32, Promote);
614 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
615 setOperationAction(ISD::OR, MVT::v1i64, Legal);
617 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
618 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
619 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
620 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
621 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
622 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
623 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
625 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
626 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
627 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
628 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
629 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
630 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
631 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
632 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
633 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
635 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
636 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
637 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
638 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
639 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
641 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
642 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
643 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
644 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
646 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
647 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
648 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
649 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
651 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
653 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Expand);
654 setOperationAction(ISD::TRUNCATE, MVT::v8i8, Expand);
655 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
656 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
657 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
658 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
659 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
660 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
661 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
664 if (!UseSoftFloat && Subtarget->hasSSE1()) {
665 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
667 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
668 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
669 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
670 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
671 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
672 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
673 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
674 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
675 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
676 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
677 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
678 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
681 if (!UseSoftFloat && Subtarget->hasSSE2()) {
682 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
684 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
685 // registers cannot be used even for integer operations.
686 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
687 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
688 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
689 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
691 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
692 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
693 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
694 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
695 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
696 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
697 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
698 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
699 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
700 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
701 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
702 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
703 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
704 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
705 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
706 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
708 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
709 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
710 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
711 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
713 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
714 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
715 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
716 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
717 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
719 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
720 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
721 MVT VT = (MVT::SimpleValueType)i;
722 // Do not attempt to custom lower non-power-of-2 vectors
723 if (!isPowerOf2_32(VT.getVectorNumElements()))
725 // Do not attempt to custom lower non-128-bit vectors
726 if (!VT.is128BitVector())
728 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
729 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
730 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
733 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
734 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
735 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
736 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
737 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
738 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
740 if (Subtarget->is64Bit()) {
741 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
742 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
745 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
746 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
747 MVT VT = (MVT::SimpleValueType)i;
749 // Do not attempt to promote non-128-bit vectors
750 if (!VT.is128BitVector()) {
753 setOperationAction(ISD::AND, VT, Promote);
754 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
755 setOperationAction(ISD::OR, VT, Promote);
756 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
757 setOperationAction(ISD::XOR, VT, Promote);
758 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
759 setOperationAction(ISD::LOAD, VT, Promote);
760 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
761 setOperationAction(ISD::SELECT, VT, Promote);
762 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
765 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
767 // Custom lower v2i64 and v2f64 selects.
768 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
769 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
770 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
771 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
773 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
774 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
775 if (!DisableMMX && Subtarget->hasMMX()) {
776 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
777 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
781 if (Subtarget->hasSSE41()) {
782 // FIXME: Do we need to handle scalar-to-vector here?
783 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
785 // i8 and i16 vectors are custom , because the source register and source
786 // source memory operand types are not the same width. f32 vectors are
787 // custom since the immediate controlling the insert encodes additional
789 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
790 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
791 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
792 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
794 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
795 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
796 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
797 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
799 if (Subtarget->is64Bit()) {
800 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
801 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
805 if (Subtarget->hasSSE42()) {
806 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
809 if (!UseSoftFloat && Subtarget->hasAVX()) {
810 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
811 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
812 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
813 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
815 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
816 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
817 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
818 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
819 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
820 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
821 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
822 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
823 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
824 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
825 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
826 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
827 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
828 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
829 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
831 // Operations to consider commented out -v16i16 v32i8
832 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
833 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
834 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
835 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
836 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
837 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
838 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
839 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
840 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
841 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
842 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
843 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
844 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
845 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
847 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
848 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
849 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
850 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
852 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
853 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
854 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
855 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
856 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
858 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
859 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
860 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
861 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
862 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
863 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
866 // Not sure we want to do this since there are no 256-bit integer
869 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
870 // This includes 256-bit vectors
871 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
872 MVT VT = (MVT::SimpleValueType)i;
874 // Do not attempt to custom lower non-power-of-2 vectors
875 if (!isPowerOf2_32(VT.getVectorNumElements()))
878 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
879 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
880 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
883 if (Subtarget->is64Bit()) {
884 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
885 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
890 // Not sure we want to do this since there are no 256-bit integer
893 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
894 // Including 256-bit vectors
895 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
896 MVT VT = (MVT::SimpleValueType)i;
898 if (!VT.is256BitVector()) {
901 setOperationAction(ISD::AND, VT, Promote);
902 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
903 setOperationAction(ISD::OR, VT, Promote);
904 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
905 setOperationAction(ISD::XOR, VT, Promote);
906 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
907 setOperationAction(ISD::LOAD, VT, Promote);
908 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
909 setOperationAction(ISD::SELECT, VT, Promote);
910 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
913 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
917 // We want to custom lower some of our intrinsics.
918 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
920 // Add/Sub/Mul with overflow operations are custom lowered.
921 setOperationAction(ISD::SADDO, MVT::i32, Custom);
922 setOperationAction(ISD::SADDO, MVT::i64, Custom);
923 setOperationAction(ISD::UADDO, MVT::i32, Custom);
924 setOperationAction(ISD::UADDO, MVT::i64, Custom);
925 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
926 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
927 setOperationAction(ISD::USUBO, MVT::i32, Custom);
928 setOperationAction(ISD::USUBO, MVT::i64, Custom);
929 setOperationAction(ISD::SMULO, MVT::i32, Custom);
930 setOperationAction(ISD::SMULO, MVT::i64, Custom);
932 if (!Subtarget->is64Bit()) {
933 // These libcalls are not available in 32-bit.
934 setLibcallName(RTLIB::SHL_I128, 0);
935 setLibcallName(RTLIB::SRL_I128, 0);
936 setLibcallName(RTLIB::SRA_I128, 0);
939 // We have target-specific dag combine patterns for the following nodes:
940 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
941 setTargetDAGCombine(ISD::BUILD_VECTOR);
942 setTargetDAGCombine(ISD::SELECT);
943 setTargetDAGCombine(ISD::SHL);
944 setTargetDAGCombine(ISD::SRA);
945 setTargetDAGCombine(ISD::SRL);
946 setTargetDAGCombine(ISD::STORE);
947 setTargetDAGCombine(ISD::MEMBARRIER);
948 if (Subtarget->is64Bit())
949 setTargetDAGCombine(ISD::MUL);
951 computeRegisterProperties();
953 // FIXME: These should be based on subtarget info. Plus, the values should
954 // be smaller when we are in optimizing for size mode.
955 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
956 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
957 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
958 allowUnalignedMemoryAccesses = true; // x86 supports it!
959 setPrefLoopAlignment(16);
960 benefitFromCodePlacementOpt = true;
964 MVT X86TargetLowering::getSetCCResultType(MVT VT) const {
969 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
970 /// the desired ByVal argument alignment.
971 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
974 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
975 if (VTy->getBitWidth() == 128)
977 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
978 unsigned EltAlign = 0;
979 getMaxByValAlign(ATy->getElementType(), EltAlign);
980 if (EltAlign > MaxAlign)
982 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
983 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
984 unsigned EltAlign = 0;
985 getMaxByValAlign(STy->getElementType(i), EltAlign);
986 if (EltAlign > MaxAlign)
995 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
996 /// function arguments in the caller parameter area. For X86, aggregates
997 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
998 /// are at 4-byte boundaries.
999 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1000 if (Subtarget->is64Bit()) {
1001 // Max of 8 and alignment of type.
1002 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1009 if (Subtarget->hasSSE1())
1010 getMaxByValAlign(Ty, Align);
1014 /// getOptimalMemOpType - Returns the target specific optimal type for load
1015 /// and store operations as a result of memset, memcpy, and memmove
1016 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
1019 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1020 bool isSrcConst, bool isSrcStr,
1021 SelectionDAG &DAG) const {
1022 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1023 // linux. This is because the stack realignment code can't handle certain
1024 // cases like PR2962. This should be removed when PR2962 is fixed.
1025 const Function *F = DAG.getMachineFunction().getFunction();
1026 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1027 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1028 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1030 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1033 if (Subtarget->is64Bit() && Size >= 8)
1038 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1040 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1041 SelectionDAG &DAG) const {
1042 if (usesGlobalOffsetTable())
1043 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
1044 if (!Subtarget->is64Bit())
1045 // This doesn't have DebugLoc associated with it, but is not really the
1046 // same as a Register.
1047 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1052 /// getFunctionAlignment - Return the Log2 alignment of this function.
1053 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1054 return F->hasFnAttr(Attribute::OptimizeForSize) ? 1 : 4;
1057 //===----------------------------------------------------------------------===//
1058 // Return Value Calling Convention Implementation
1059 //===----------------------------------------------------------------------===//
1061 #include "X86GenCallingConv.inc"
1063 /// LowerRET - Lower an ISD::RET node.
1064 SDValue X86TargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
1065 DebugLoc dl = Op.getDebugLoc();
1066 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
1068 SmallVector<CCValAssign, 16> RVLocs;
1069 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
1070 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1071 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs, *DAG.getContext());
1072 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_X86);
1074 // If this is the first return lowered for this function, add the regs to the
1075 // liveout set for the function.
1076 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1077 for (unsigned i = 0; i != RVLocs.size(); ++i)
1078 if (RVLocs[i].isRegLoc())
1079 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1081 SDValue Chain = Op.getOperand(0);
1083 // Handle tail call return.
1084 Chain = GetPossiblePreceedingTailCall(Chain, X86ISD::TAILCALL);
1085 if (Chain.getOpcode() == X86ISD::TAILCALL) {
1086 SDValue TailCall = Chain;
1087 SDValue TargetAddress = TailCall.getOperand(1);
1088 SDValue StackAdjustment = TailCall.getOperand(2);
1089 assert(((TargetAddress.getOpcode() == ISD::Register &&
1090 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::EAX ||
1091 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R11)) ||
1092 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
1093 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
1094 "Expecting an global address, external symbol, or register");
1095 assert(StackAdjustment.getOpcode() == ISD::Constant &&
1096 "Expecting a const value");
1098 SmallVector<SDValue,8> Operands;
1099 Operands.push_back(Chain.getOperand(0));
1100 Operands.push_back(TargetAddress);
1101 Operands.push_back(StackAdjustment);
1102 // Copy registers used by the call. Last operand is a flag so it is not
1104 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
1105 Operands.push_back(Chain.getOperand(i));
1107 return DAG.getNode(X86ISD::TC_RETURN, dl, MVT::Other, &Operands[0],
1114 SmallVector<SDValue, 6> RetOps;
1115 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1116 // Operand #1 = Bytes To Pop
1117 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
1119 // Copy the result values into the output registers.
1120 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1121 CCValAssign &VA = RVLocs[i];
1122 assert(VA.isRegLoc() && "Can only return in registers!");
1123 SDValue ValToCopy = Op.getOperand(i*2+1);
1125 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1126 // the RET instruction and handled by the FP Stackifier.
1127 if (VA.getLocReg() == X86::ST0 ||
1128 VA.getLocReg() == X86::ST1) {
1129 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1130 // change the value to the FP stack register class.
1131 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1132 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1133 RetOps.push_back(ValToCopy);
1134 // Don't emit a copytoreg.
1138 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1139 // which is returned in RAX / RDX.
1140 if (Subtarget->is64Bit()) {
1141 MVT ValVT = ValToCopy.getValueType();
1142 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1143 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1144 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1145 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1149 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1150 Flag = Chain.getValue(1);
1153 // The x86-64 ABI for returning structs by value requires that we copy
1154 // the sret argument into %rax for the return. We saved the argument into
1155 // a virtual register in the entry block, so now we copy the value out
1157 if (Subtarget->is64Bit() &&
1158 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1159 MachineFunction &MF = DAG.getMachineFunction();
1160 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1161 unsigned Reg = FuncInfo->getSRetReturnReg();
1163 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1164 FuncInfo->setSRetReturnReg(Reg);
1166 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1168 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1169 Flag = Chain.getValue(1);
1172 RetOps[0] = Chain; // Update chain.
1174 // Add the flag if we have it.
1176 RetOps.push_back(Flag);
1178 return DAG.getNode(X86ISD::RET_FLAG, dl,
1179 MVT::Other, &RetOps[0], RetOps.size());
1183 /// LowerCallResult - Lower the result values of an ISD::CALL into the
1184 /// appropriate copies out of appropriate physical registers. This assumes that
1185 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
1186 /// being lowered. The returns a SDNode with the same number of values as the
1188 SDNode *X86TargetLowering::
1189 LowerCallResult(SDValue Chain, SDValue InFlag, CallSDNode *TheCall,
1190 unsigned CallingConv, SelectionDAG &DAG) {
1192 DebugLoc dl = TheCall->getDebugLoc();
1193 // Assign locations to each value returned by this call.
1194 SmallVector<CCValAssign, 16> RVLocs;
1195 bool isVarArg = TheCall->isVarArg();
1196 bool Is64Bit = Subtarget->is64Bit();
1197 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(),
1198 RVLocs, *DAG.getContext());
1199 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
1201 SmallVector<SDValue, 8> ResultVals;
1203 // Copy all of the result registers out of their specified physreg.
1204 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1205 CCValAssign &VA = RVLocs[i];
1206 MVT CopyVT = VA.getValVT();
1208 // If this is x86-64, and we disabled SSE, we can't return FP values
1209 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1210 ((Is64Bit || TheCall->isInreg()) && !Subtarget->hasSSE1())) {
1211 llvm_report_error("SSE register return with SSE disabled");
1214 // If this is a call to a function that returns an fp value on the floating
1215 // point stack, but where we prefer to use the value in xmm registers, copy
1216 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1217 if ((VA.getLocReg() == X86::ST0 ||
1218 VA.getLocReg() == X86::ST1) &&
1219 isScalarFPTypeInSSEReg(VA.getValVT())) {
1224 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1225 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1226 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1227 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1228 MVT::v2i64, InFlag).getValue(1);
1229 Val = Chain.getValue(0);
1230 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1231 Val, DAG.getConstant(0, MVT::i64));
1233 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1234 MVT::i64, InFlag).getValue(1);
1235 Val = Chain.getValue(0);
1237 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1239 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1240 CopyVT, InFlag).getValue(1);
1241 Val = Chain.getValue(0);
1243 InFlag = Chain.getValue(2);
1245 if (CopyVT != VA.getValVT()) {
1246 // Round the F80 the right size, which also moves to the appropriate xmm
1248 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1249 // This truncation won't change the value.
1250 DAG.getIntPtrConstant(1));
1253 ResultVals.push_back(Val);
1256 // Merge everything together with a MERGE_VALUES node.
1257 ResultVals.push_back(Chain);
1258 return DAG.getNode(ISD::MERGE_VALUES, dl, TheCall->getVTList(),
1259 &ResultVals[0], ResultVals.size()).getNode();
1263 //===----------------------------------------------------------------------===//
1264 // C & StdCall & Fast Calling Convention implementation
1265 //===----------------------------------------------------------------------===//
1266 // StdCall calling convention seems to be standard for many Windows' API
1267 // routines and around. It differs from C calling convention just a little:
1268 // callee should clean up the stack, not caller. Symbols should be also
1269 // decorated in some fancy way :) It doesn't support any vector arguments.
1270 // For info on fast calling convention see Fast Calling Convention (tail call)
1271 // implementation LowerX86_32FastCCCallTo.
1273 /// CallIsStructReturn - Determines whether a CALL node uses struct return
1275 static bool CallIsStructReturn(CallSDNode *TheCall) {
1276 unsigned NumOps = TheCall->getNumArgs();
1280 return TheCall->getArgFlags(0).isSRet();
1283 /// ArgsAreStructReturn - Determines whether a function uses struct
1284 /// return semantics.
1285 static bool ArgsAreStructReturn(SDValue Op) {
1286 unsigned NumArgs = Op.getNode()->getNumValues() - 1;
1290 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
1293 /// IsCalleePop - Determines whether the callee is required to pop its
1294 /// own arguments. Callee pop is necessary to support tail calls.
1295 bool X86TargetLowering::IsCalleePop(bool IsVarArg, unsigned CallingConv) {
1299 switch (CallingConv) {
1302 case CallingConv::X86_StdCall:
1303 return !Subtarget->is64Bit();
1304 case CallingConv::X86_FastCall:
1305 return !Subtarget->is64Bit();
1306 case CallingConv::Fast:
1307 return PerformTailCallOpt;
1311 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1312 /// given CallingConvention value.
1313 CCAssignFn *X86TargetLowering::CCAssignFnForNode(unsigned CC) const {
1314 if (Subtarget->is64Bit()) {
1315 if (Subtarget->isTargetWin64())
1316 return CC_X86_Win64_C;
1321 if (CC == CallingConv::X86_FastCall)
1322 return CC_X86_32_FastCall;
1323 else if (CC == CallingConv::Fast)
1324 return CC_X86_32_FastCC;
1329 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1330 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1332 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDValue Op) {
1333 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1334 if (CC == CallingConv::X86_FastCall)
1336 else if (CC == CallingConv::X86_StdCall)
1342 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1343 /// by "Src" to address "Dst" with size and alignment information specified by
1344 /// the specific parameter attribute. The copy will be passed as a byval
1345 /// function parameter.
1347 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1348 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1350 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1351 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1352 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1355 SDValue X86TargetLowering::LowerMemArgument(SDValue Op, SelectionDAG &DAG,
1356 const CCValAssign &VA,
1357 MachineFrameInfo *MFI,
1359 SDValue Root, unsigned i) {
1360 // Create the nodes corresponding to a load from this parameter slot.
1361 ISD::ArgFlagsTy Flags =
1362 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1363 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1364 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1366 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1367 // changed with more analysis.
1368 // In case of tail call optimization mark all arguments mutable. Since they
1369 // could be overwritten by lowering of arguments in case of a tail call.
1370 int FI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
1371 VA.getLocMemOffset(), isImmutable);
1372 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1373 if (Flags.isByVal())
1375 return DAG.getLoad(VA.getValVT(), Op.getDebugLoc(), Root, FIN,
1376 PseudoSourceValue::getFixedStack(FI), 0);
1380 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1381 MachineFunction &MF = DAG.getMachineFunction();
1382 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1383 DebugLoc dl = Op.getDebugLoc();
1385 const Function* Fn = MF.getFunction();
1386 if (Fn->hasExternalLinkage() &&
1387 Subtarget->isTargetCygMing() &&
1388 Fn->getName() == "main")
1389 FuncInfo->setForceFramePointer(true);
1391 // Decorate the function name.
1392 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1394 MachineFrameInfo *MFI = MF.getFrameInfo();
1395 SDValue Root = Op.getOperand(0);
1396 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1397 unsigned CC = MF.getFunction()->getCallingConv();
1398 bool Is64Bit = Subtarget->is64Bit();
1399 bool IsWin64 = Subtarget->isTargetWin64();
1401 assert(!(isVarArg && CC == CallingConv::Fast) &&
1402 "Var args not supported with calling convention fastcc");
1404 // Assign locations to all of the incoming arguments.
1405 SmallVector<CCValAssign, 16> ArgLocs;
1406 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs, *DAG.getContext());
1407 CCInfo.AnalyzeFormalArguments(Op.getNode(), CCAssignFnForNode(CC));
1409 SmallVector<SDValue, 8> ArgValues;
1410 unsigned LastVal = ~0U;
1411 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1412 CCValAssign &VA = ArgLocs[i];
1413 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1415 assert(VA.getValNo() != LastVal &&
1416 "Don't support value assigned to multiple locs yet");
1417 LastVal = VA.getValNo();
1419 if (VA.isRegLoc()) {
1420 MVT RegVT = VA.getLocVT();
1421 TargetRegisterClass *RC = NULL;
1422 if (RegVT == MVT::i32)
1423 RC = X86::GR32RegisterClass;
1424 else if (Is64Bit && RegVT == MVT::i64)
1425 RC = X86::GR64RegisterClass;
1426 else if (RegVT == MVT::f32)
1427 RC = X86::FR32RegisterClass;
1428 else if (RegVT == MVT::f64)
1429 RC = X86::FR64RegisterClass;
1430 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1431 RC = X86::VR128RegisterClass;
1432 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1433 RC = X86::VR64RegisterClass;
1435 llvm_unreachable("Unknown argument type!");
1437 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1438 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, RegVT);
1440 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1441 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1443 if (VA.getLocInfo() == CCValAssign::SExt)
1444 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1445 DAG.getValueType(VA.getValVT()));
1446 else if (VA.getLocInfo() == CCValAssign::ZExt)
1447 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1448 DAG.getValueType(VA.getValVT()));
1449 else if (VA.getLocInfo() == CCValAssign::BCvt)
1450 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, ArgValue,
1451 DAG.getValueType(VA.getValVT()));
1453 if (VA.getLocInfo() != CCValAssign::Full &&
1454 VA.getLocInfo() != CCValAssign::BCvt) {
1455 // Handle MMX values passed in XMM regs.
1456 if (RegVT.isVector()) {
1457 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1458 ArgValue, DAG.getConstant(0, MVT::i64));
1459 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1461 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1464 ArgValues.push_back(ArgValue);
1466 assert(VA.isMemLoc());
1467 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1471 // The x86-64 ABI for returning structs by value requires that we copy
1472 // the sret argument into %rax for the return. Save the argument into
1473 // a virtual register so that we can access it from the return points.
1474 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1475 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1476 unsigned Reg = FuncInfo->getSRetReturnReg();
1478 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1479 FuncInfo->setSRetReturnReg(Reg);
1481 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, ArgValues[0]);
1482 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Root);
1485 unsigned StackSize = CCInfo.getNextStackOffset();
1486 // align stack specially for tail calls
1487 if (PerformTailCallOpt && CC == CallingConv::Fast)
1488 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1490 // If the function takes variable number of arguments, make a frame index for
1491 // the start of the first vararg value... for expansion of llvm.va_start.
1493 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1494 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1497 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1499 // FIXME: We should really autogenerate these arrays
1500 static const unsigned GPR64ArgRegsWin64[] = {
1501 X86::RCX, X86::RDX, X86::R8, X86::R9
1503 static const unsigned XMMArgRegsWin64[] = {
1504 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1506 static const unsigned GPR64ArgRegs64Bit[] = {
1507 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1509 static const unsigned XMMArgRegs64Bit[] = {
1510 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1511 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1513 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1516 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1517 GPR64ArgRegs = GPR64ArgRegsWin64;
1518 XMMArgRegs = XMMArgRegsWin64;
1520 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1521 GPR64ArgRegs = GPR64ArgRegs64Bit;
1522 XMMArgRegs = XMMArgRegs64Bit;
1524 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1526 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1529 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1530 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1531 "SSE register cannot be used when SSE is disabled!");
1532 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1533 "SSE register cannot be used when SSE is disabled!");
1534 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1535 // Kernel mode asks for SSE to be disabled, so don't push them
1537 TotalNumXMMRegs = 0;
1539 // For X86-64, if there are vararg parameters that are passed via
1540 // registers, then we must store them to their spots on the stack so they
1541 // may be loaded by deferencing the result of va_next.
1542 VarArgsGPOffset = NumIntRegs * 8;
1543 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1544 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1545 TotalNumXMMRegs * 16, 16);
1547 // Store the integer parameter registers.
1548 SmallVector<SDValue, 8> MemOps;
1549 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1550 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1551 DAG.getIntPtrConstant(VarArgsGPOffset));
1552 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1553 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1554 X86::GR64RegisterClass);
1555 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::i64);
1557 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1558 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1559 MemOps.push_back(Store);
1560 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1561 DAG.getIntPtrConstant(8));
1564 // Now store the XMM (fp + vector) parameter registers.
1565 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1566 DAG.getIntPtrConstant(VarArgsFPOffset));
1567 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1568 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1569 X86::VR128RegisterClass);
1570 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::v4f32);
1572 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1573 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1574 MemOps.push_back(Store);
1575 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1576 DAG.getIntPtrConstant(16));
1578 if (!MemOps.empty())
1579 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1580 &MemOps[0], MemOps.size());
1584 ArgValues.push_back(Root);
1586 // Some CCs need callee pop.
1587 if (IsCalleePop(isVarArg, CC)) {
1588 BytesToPopOnReturn = StackSize; // Callee pops everything.
1589 BytesCallerReserves = 0;
1591 BytesToPopOnReturn = 0; // Callee pops nothing.
1592 // If this is an sret function, the return should pop the hidden pointer.
1593 if (!Is64Bit && CC != CallingConv::Fast && ArgsAreStructReturn(Op))
1594 BytesToPopOnReturn = 4;
1595 BytesCallerReserves = StackSize;
1599 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1600 if (CC == CallingConv::X86_FastCall)
1601 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1604 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1606 // Return the new list of results.
1607 return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getNode()->getVTList(),
1608 &ArgValues[0], ArgValues.size()).getValue(Op.getResNo());
1612 X86TargetLowering::LowerMemOpCallTo(CallSDNode *TheCall, SelectionDAG &DAG,
1613 const SDValue &StackPtr,
1614 const CCValAssign &VA,
1616 SDValue Arg, ISD::ArgFlagsTy Flags) {
1617 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1618 DebugLoc dl = TheCall->getDebugLoc();
1619 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1620 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1621 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1622 if (Flags.isByVal()) {
1623 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1625 return DAG.getStore(Chain, dl, Arg, PtrOff,
1626 PseudoSourceValue::getStack(), LocMemOffset);
1629 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1630 /// optimization is performed and it is required.
1632 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1633 SDValue &OutRetAddr,
1639 if (!IsTailCall || FPDiff==0) return Chain;
1641 // Adjust the Return address stack slot.
1642 MVT VT = getPointerTy();
1643 OutRetAddr = getReturnAddressFrameIndex(DAG);
1645 // Load the "old" Return address.
1646 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1647 return SDValue(OutRetAddr.getNode(), 1);
1650 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1651 /// optimization is performed and it is required (FPDiff!=0).
1653 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1654 SDValue Chain, SDValue RetAddrFrIdx,
1655 bool Is64Bit, int FPDiff, DebugLoc dl) {
1656 // Store the return address to the appropriate stack slot.
1657 if (!FPDiff) return Chain;
1658 // Calculate the new stack slot for the return address.
1659 int SlotSize = Is64Bit ? 8 : 4;
1660 int NewReturnAddrFI =
1661 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1662 MVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1663 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1664 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1665 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1669 SDValue X86TargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
1670 MachineFunction &MF = DAG.getMachineFunction();
1671 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
1672 SDValue Chain = TheCall->getChain();
1673 unsigned CC = TheCall->getCallingConv();
1674 bool isVarArg = TheCall->isVarArg();
1675 bool IsTailCall = TheCall->isTailCall() &&
1676 CC == CallingConv::Fast && PerformTailCallOpt;
1677 SDValue Callee = TheCall->getCallee();
1678 bool Is64Bit = Subtarget->is64Bit();
1679 bool IsStructRet = CallIsStructReturn(TheCall);
1680 DebugLoc dl = TheCall->getDebugLoc();
1682 assert(!(isVarArg && CC == CallingConv::Fast) &&
1683 "Var args not supported with calling convention fastcc");
1685 // Analyze operands of the call, assigning locations to each operand.
1686 SmallVector<CCValAssign, 16> ArgLocs;
1687 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs, *DAG.getContext());
1688 CCInfo.AnalyzeCallOperands(TheCall, CCAssignFnForNode(CC));
1690 // Get a count of how many bytes are to be pushed on the stack.
1691 unsigned NumBytes = CCInfo.getNextStackOffset();
1692 if (PerformTailCallOpt && CC == CallingConv::Fast)
1693 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1697 // Lower arguments at fp - stackoffset + fpdiff.
1698 unsigned NumBytesCallerPushed =
1699 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1700 FPDiff = NumBytesCallerPushed - NumBytes;
1702 // Set the delta of movement of the returnaddr stackslot.
1703 // But only set if delta is greater than previous delta.
1704 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1705 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1708 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1710 SDValue RetAddrFrIdx;
1711 // Load return adress for tail calls.
1712 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, IsTailCall, Is64Bit,
1715 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1716 SmallVector<SDValue, 8> MemOpChains;
1719 // Walk the register/memloc assignments, inserting copies/loads. In the case
1720 // of tail call optimization arguments are handle later.
1721 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1722 CCValAssign &VA = ArgLocs[i];
1723 MVT RegVT = VA.getLocVT();
1724 SDValue Arg = TheCall->getArg(i);
1725 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1726 bool isByVal = Flags.isByVal();
1728 // Promote the value if needed.
1729 switch (VA.getLocInfo()) {
1730 default: llvm_unreachable("Unknown loc info!");
1731 case CCValAssign::Full: break;
1732 case CCValAssign::SExt:
1733 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1735 case CCValAssign::ZExt:
1736 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1738 case CCValAssign::AExt:
1739 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1740 // Special case: passing MMX values in XMM registers.
1741 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1742 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1743 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1745 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1747 case CCValAssign::BCvt:
1748 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1752 if (VA.isRegLoc()) {
1753 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1755 if (!IsTailCall || (IsTailCall && isByVal)) {
1756 assert(VA.isMemLoc());
1757 if (StackPtr.getNode() == 0)
1758 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1760 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA,
1761 Chain, Arg, Flags));
1766 if (!MemOpChains.empty())
1767 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1768 &MemOpChains[0], MemOpChains.size());
1770 // Build a sequence of copy-to-reg nodes chained together with token chain
1771 // and flag operands which copy the outgoing args into registers.
1773 // Tail call byval lowering might overwrite argument registers so in case of
1774 // tail call optimization the copies to registers are lowered later.
1776 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1777 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1778 RegsToPass[i].second, InFlag);
1779 InFlag = Chain.getValue(1);
1783 if (Subtarget->isPICStyleGOT()) {
1784 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1787 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1788 DAG.getNode(X86ISD::GlobalBaseReg,
1789 DebugLoc::getUnknownLoc(),
1792 InFlag = Chain.getValue(1);
1794 // If we are tail calling and generating PIC/GOT style code load the
1795 // address of the callee into ECX. The value in ecx is used as target of
1796 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1797 // for tail calls on PIC/GOT architectures. Normally we would just put the
1798 // address of GOT into ebx and then call target@PLT. But for tail calls
1799 // ebx would be restored (since ebx is callee saved) before jumping to the
1802 // Note: The actual moving to ECX is done further down.
1803 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1804 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1805 !G->getGlobal()->hasProtectedVisibility())
1806 Callee = LowerGlobalAddress(Callee, DAG);
1807 else if (isa<ExternalSymbolSDNode>(Callee))
1808 Callee = LowerExternalSymbol(Callee, DAG);
1812 if (Is64Bit && isVarArg) {
1813 // From AMD64 ABI document:
1814 // For calls that may call functions that use varargs or stdargs
1815 // (prototype-less calls or calls to functions containing ellipsis (...) in
1816 // the declaration) %al is used as hidden argument to specify the number
1817 // of SSE registers used. The contents of %al do not need to match exactly
1818 // the number of registers, but must be an ubound on the number of SSE
1819 // registers used and is in the range 0 - 8 inclusive.
1821 // FIXME: Verify this on Win64
1822 // Count the number of XMM registers allocated.
1823 static const unsigned XMMArgRegs[] = {
1824 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1825 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1827 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1828 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1829 && "SSE registers cannot be used when SSE is disabled");
1831 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1832 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1833 InFlag = Chain.getValue(1);
1837 // For tail calls lower the arguments to the 'real' stack slot.
1839 SmallVector<SDValue, 8> MemOpChains2;
1842 // Do not flag preceeding copytoreg stuff together with the following stuff.
1844 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1845 CCValAssign &VA = ArgLocs[i];
1846 if (!VA.isRegLoc()) {
1847 assert(VA.isMemLoc());
1848 SDValue Arg = TheCall->getArg(i);
1849 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1850 // Create frame index.
1851 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1852 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1853 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1854 FIN = DAG.getFrameIndex(FI, getPointerTy());
1856 if (Flags.isByVal()) {
1857 // Copy relative to framepointer.
1858 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1859 if (StackPtr.getNode() == 0)
1860 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1862 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1864 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1867 // Store relative to framepointer.
1868 MemOpChains2.push_back(
1869 DAG.getStore(Chain, dl, Arg, FIN,
1870 PseudoSourceValue::getFixedStack(FI), 0));
1875 if (!MemOpChains2.empty())
1876 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1877 &MemOpChains2[0], MemOpChains2.size());
1879 // Copy arguments to their registers.
1880 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1881 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1882 RegsToPass[i].second, InFlag);
1883 InFlag = Chain.getValue(1);
1887 // Store the return address to the appropriate stack slot.
1888 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1892 // If the callee is a GlobalAddress node (quite common, every direct call is)
1893 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1894 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1895 // We should use extra load for direct calls to dllimported functions in
1897 GlobalValue *GV = G->getGlobal();
1898 if (!GV->hasDLLImportLinkage()) {
1899 unsigned char OpFlags = 0;
1901 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
1902 // external symbols most go through the PLT in PIC mode. If the symbol
1903 // has hidden or protected visibility, or if it is static or local, then
1904 // we don't need to use the PLT - we can directly call it.
1905 if (Subtarget->isTargetELF() &&
1906 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1907 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
1908 OpFlags = X86II::MO_PLT;
1909 } else if (Subtarget->isPICStyleStubAny() &&
1910 (GV->isDeclaration() || GV->isWeakForLinker()) &&
1911 Subtarget->getDarwinVers() < 9) {
1912 // PC-relative references to external symbols should go through $stub,
1913 // unless we're building with the leopard linker or later, which
1914 // automatically synthesizes these stubs.
1915 OpFlags = X86II::MO_DARWIN_STUB;
1918 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
1919 G->getOffset(), OpFlags);
1921 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1922 unsigned char OpFlags = 0;
1924 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
1925 // symbols should go through the PLT.
1926 if (Subtarget->isTargetELF() &&
1927 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1928 OpFlags = X86II::MO_PLT;
1929 } else if (Subtarget->isPICStyleStubAny() &&
1930 Subtarget->getDarwinVers() < 9) {
1931 // PC-relative references to external symbols should go through $stub,
1932 // unless we're building with the leopard linker or later, which
1933 // automatically synthesizes these stubs.
1934 OpFlags = X86II::MO_DARWIN_STUB;
1937 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
1939 } else if (IsTailCall) {
1940 unsigned Opc = Is64Bit ? X86::R11 : X86::EAX;
1942 Chain = DAG.getCopyToReg(Chain, dl,
1943 DAG.getRegister(Opc, getPointerTy()),
1945 Callee = DAG.getRegister(Opc, getPointerTy());
1946 // Add register as live out.
1947 MF.getRegInfo().addLiveOut(Opc);
1950 // Returns a chain & a flag for retval copy to use.
1951 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1952 SmallVector<SDValue, 8> Ops;
1955 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1956 DAG.getIntPtrConstant(0, true), InFlag);
1957 InFlag = Chain.getValue(1);
1959 // Returns a chain & a flag for retval copy to use.
1960 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1964 Ops.push_back(Chain);
1965 Ops.push_back(Callee);
1968 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1970 // Add argument registers to the end of the list so that they are known live
1972 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1973 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1974 RegsToPass[i].second.getValueType()));
1976 // Add an implicit use GOT pointer in EBX.
1977 if (!IsTailCall && Subtarget->isPICStyleGOT())
1978 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1980 // Add an implicit use of AL for x86 vararg functions.
1981 if (Is64Bit && isVarArg)
1982 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1984 if (InFlag.getNode())
1985 Ops.push_back(InFlag);
1988 assert(InFlag.getNode() &&
1989 "Flag must be set. Depend on flag being set in LowerRET");
1990 Chain = DAG.getNode(X86ISD::TAILCALL, dl,
1991 TheCall->getVTList(), &Ops[0], Ops.size());
1993 return SDValue(Chain.getNode(), Op.getResNo());
1996 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
1997 InFlag = Chain.getValue(1);
1999 // Create the CALLSEQ_END node.
2000 unsigned NumBytesForCalleeToPush;
2001 if (IsCalleePop(isVarArg, CC))
2002 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2003 else if (!Is64Bit && CC != CallingConv::Fast && IsStructRet)
2004 // If this is is a call to a struct-return function, the callee
2005 // pops the hidden struct pointer, so we have to push it back.
2006 // This is common for Darwin/X86, Linux & Mingw32 targets.
2007 NumBytesForCalleeToPush = 4;
2009 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2011 // Returns a flag for retval copy to use.
2012 Chain = DAG.getCALLSEQ_END(Chain,
2013 DAG.getIntPtrConstant(NumBytes, true),
2014 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2017 InFlag = Chain.getValue(1);
2019 // Handle result values, copying them out of physregs into vregs that we
2021 return SDValue(LowerCallResult(Chain, InFlag, TheCall, CC, DAG),
2026 //===----------------------------------------------------------------------===//
2027 // Fast Calling Convention (tail call) implementation
2028 //===----------------------------------------------------------------------===//
2030 // Like std call, callee cleans arguments, convention except that ECX is
2031 // reserved for storing the tail called function address. Only 2 registers are
2032 // free for argument passing (inreg). Tail call optimization is performed
2034 // * tailcallopt is enabled
2035 // * caller/callee are fastcc
2036 // On X86_64 architecture with GOT-style position independent code only local
2037 // (within module) calls are supported at the moment.
2038 // To keep the stack aligned according to platform abi the function
2039 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2040 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2041 // If a tail called function callee has more arguments than the caller the
2042 // caller needs to make sure that there is room to move the RETADDR to. This is
2043 // achieved by reserving an area the size of the argument delta right after the
2044 // original REtADDR, but before the saved framepointer or the spilled registers
2045 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2057 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2058 /// for a 16 byte align requirement.
2059 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2060 SelectionDAG& DAG) {
2061 MachineFunction &MF = DAG.getMachineFunction();
2062 const TargetMachine &TM = MF.getTarget();
2063 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2064 unsigned StackAlignment = TFI.getStackAlignment();
2065 uint64_t AlignMask = StackAlignment - 1;
2066 int64_t Offset = StackSize;
2067 uint64_t SlotSize = TD->getPointerSize();
2068 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2069 // Number smaller than 12 so just add the difference.
2070 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2072 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2073 Offset = ((~AlignMask) & Offset) + StackAlignment +
2074 (StackAlignment-SlotSize);
2079 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
2080 /// following the call is a return. A function is eligible if caller/callee
2081 /// calling conventions match, currently only fastcc supports tail calls, and
2082 /// the function CALL is immediatly followed by a RET.
2083 bool X86TargetLowering::IsEligibleForTailCallOptimization(CallSDNode *TheCall,
2085 SelectionDAG& DAG) const {
2086 if (!PerformTailCallOpt)
2089 if (CheckTailCallReturnConstraints(TheCall, Ret)) {
2091 DAG.getMachineFunction().getFunction()->getCallingConv();
2092 unsigned CalleeCC = TheCall->getCallingConv();
2093 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC)
2101 X86TargetLowering::createFastISel(MachineFunction &mf,
2102 MachineModuleInfo *mmo,
2104 DenseMap<const Value *, unsigned> &vm,
2105 DenseMap<const BasicBlock *,
2106 MachineBasicBlock *> &bm,
2107 DenseMap<const AllocaInst *, int> &am
2109 , SmallSet<Instruction*, 8> &cil
2112 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2120 //===----------------------------------------------------------------------===//
2121 // Other Lowering Hooks
2122 //===----------------------------------------------------------------------===//
2125 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2126 MachineFunction &MF = DAG.getMachineFunction();
2127 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2128 int ReturnAddrIndex = FuncInfo->getRAIndex();
2130 if (ReturnAddrIndex == 0) {
2131 // Set up a frame object for the return address.
2132 uint64_t SlotSize = TD->getPointerSize();
2133 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
2134 FuncInfo->setRAIndex(ReturnAddrIndex);
2137 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2141 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2142 /// specific condition code, returning the condition code and the LHS/RHS of the
2143 /// comparison to make.
2144 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2145 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2147 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2148 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2149 // X > -1 -> X == 0, jump !sign.
2150 RHS = DAG.getConstant(0, RHS.getValueType());
2151 return X86::COND_NS;
2152 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2153 // X < 0 -> X == 0, jump on sign.
2155 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2157 RHS = DAG.getConstant(0, RHS.getValueType());
2158 return X86::COND_LE;
2162 switch (SetCCOpcode) {
2163 default: llvm_unreachable("Invalid integer condition!");
2164 case ISD::SETEQ: return X86::COND_E;
2165 case ISD::SETGT: return X86::COND_G;
2166 case ISD::SETGE: return X86::COND_GE;
2167 case ISD::SETLT: return X86::COND_L;
2168 case ISD::SETLE: return X86::COND_LE;
2169 case ISD::SETNE: return X86::COND_NE;
2170 case ISD::SETULT: return X86::COND_B;
2171 case ISD::SETUGT: return X86::COND_A;
2172 case ISD::SETULE: return X86::COND_BE;
2173 case ISD::SETUGE: return X86::COND_AE;
2177 // First determine if it is required or is profitable to flip the operands.
2179 // If LHS is a foldable load, but RHS is not, flip the condition.
2180 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2181 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2182 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2183 std::swap(LHS, RHS);
2186 switch (SetCCOpcode) {
2192 std::swap(LHS, RHS);
2196 // On a floating point condition, the flags are set as follows:
2198 // 0 | 0 | 0 | X > Y
2199 // 0 | 0 | 1 | X < Y
2200 // 1 | 0 | 0 | X == Y
2201 // 1 | 1 | 1 | unordered
2202 switch (SetCCOpcode) {
2203 default: llvm_unreachable("Condcode should be pre-legalized away");
2205 case ISD::SETEQ: return X86::COND_E;
2206 case ISD::SETOLT: // flipped
2208 case ISD::SETGT: return X86::COND_A;
2209 case ISD::SETOLE: // flipped
2211 case ISD::SETGE: return X86::COND_AE;
2212 case ISD::SETUGT: // flipped
2214 case ISD::SETLT: return X86::COND_B;
2215 case ISD::SETUGE: // flipped
2217 case ISD::SETLE: return X86::COND_BE;
2219 case ISD::SETNE: return X86::COND_NE;
2220 case ISD::SETUO: return X86::COND_P;
2221 case ISD::SETO: return X86::COND_NP;
2225 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2226 /// code. Current x86 isa includes the following FP cmov instructions:
2227 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2228 static bool hasFPCMov(unsigned X86CC) {
2244 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2245 /// the specified range (L, H].
2246 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2247 return (Val < 0) || (Val >= Low && Val < Hi);
2250 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2251 /// specified value.
2252 static bool isUndefOrEqual(int Val, int CmpVal) {
2253 if (Val < 0 || Val == CmpVal)
2258 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2259 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2260 /// the second operand.
2261 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2262 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2263 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2264 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2265 return (Mask[0] < 2 && Mask[1] < 2);
2269 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2270 SmallVector<int, 8> M;
2272 return ::isPSHUFDMask(M, N->getValueType(0));
2275 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2276 /// is suitable for input to PSHUFHW.
2277 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2278 if (VT != MVT::v8i16)
2281 // Lower quadword copied in order or undef.
2282 for (int i = 0; i != 4; ++i)
2283 if (Mask[i] >= 0 && Mask[i] != i)
2286 // Upper quadword shuffled.
2287 for (int i = 4; i != 8; ++i)
2288 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2294 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2295 SmallVector<int, 8> M;
2297 return ::isPSHUFHWMask(M, N->getValueType(0));
2300 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2301 /// is suitable for input to PSHUFLW.
2302 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2303 if (VT != MVT::v8i16)
2306 // Upper quadword copied in order.
2307 for (int i = 4; i != 8; ++i)
2308 if (Mask[i] >= 0 && Mask[i] != i)
2311 // Lower quadword shuffled.
2312 for (int i = 0; i != 4; ++i)
2319 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2320 SmallVector<int, 8> M;
2322 return ::isPSHUFLWMask(M, N->getValueType(0));
2325 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2326 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2327 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2328 int NumElems = VT.getVectorNumElements();
2329 if (NumElems != 2 && NumElems != 4)
2332 int Half = NumElems / 2;
2333 for (int i = 0; i < Half; ++i)
2334 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2336 for (int i = Half; i < NumElems; ++i)
2337 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2343 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2344 SmallVector<int, 8> M;
2346 return ::isSHUFPMask(M, N->getValueType(0));
2349 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2350 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2351 /// half elements to come from vector 1 (which would equal the dest.) and
2352 /// the upper half to come from vector 2.
2353 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2354 int NumElems = VT.getVectorNumElements();
2356 if (NumElems != 2 && NumElems != 4)
2359 int Half = NumElems / 2;
2360 for (int i = 0; i < Half; ++i)
2361 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2363 for (int i = Half; i < NumElems; ++i)
2364 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2369 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2370 SmallVector<int, 8> M;
2372 return isCommutedSHUFPMask(M, N->getValueType(0));
2375 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2376 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2377 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2378 if (N->getValueType(0).getVectorNumElements() != 4)
2381 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2382 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2383 isUndefOrEqual(N->getMaskElt(1), 7) &&
2384 isUndefOrEqual(N->getMaskElt(2), 2) &&
2385 isUndefOrEqual(N->getMaskElt(3), 3);
2388 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2389 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2390 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2391 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2393 if (NumElems != 2 && NumElems != 4)
2396 for (unsigned i = 0; i < NumElems/2; ++i)
2397 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2400 for (unsigned i = NumElems/2; i < NumElems; ++i)
2401 if (!isUndefOrEqual(N->getMaskElt(i), i))
2407 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2408 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2410 bool X86::isMOVHPMask(ShuffleVectorSDNode *N) {
2411 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2413 if (NumElems != 2 && NumElems != 4)
2416 for (unsigned i = 0; i < NumElems/2; ++i)
2417 if (!isUndefOrEqual(N->getMaskElt(i), i))
2420 for (unsigned i = 0; i < NumElems/2; ++i)
2421 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2427 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2428 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2430 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2431 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2436 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2437 isUndefOrEqual(N->getMaskElt(1), 3) &&
2438 isUndefOrEqual(N->getMaskElt(2), 2) &&
2439 isUndefOrEqual(N->getMaskElt(3), 3);
2442 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2443 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2444 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, MVT VT,
2445 bool V2IsSplat = false) {
2446 int NumElts = VT.getVectorNumElements();
2447 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2450 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2452 int BitI1 = Mask[i+1];
2453 if (!isUndefOrEqual(BitI, j))
2456 if (!isUndefOrEqual(BitI1, NumElts))
2459 if (!isUndefOrEqual(BitI1, j + NumElts))
2466 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2467 SmallVector<int, 8> M;
2469 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2472 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2473 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2474 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, MVT VT,
2475 bool V2IsSplat = false) {
2476 int NumElts = VT.getVectorNumElements();
2477 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2480 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2482 int BitI1 = Mask[i+1];
2483 if (!isUndefOrEqual(BitI, j + NumElts/2))
2486 if (isUndefOrEqual(BitI1, NumElts))
2489 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2496 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2497 SmallVector<int, 8> M;
2499 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2502 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2503 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2505 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, MVT VT) {
2506 int NumElems = VT.getVectorNumElements();
2507 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2510 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2512 int BitI1 = Mask[i+1];
2513 if (!isUndefOrEqual(BitI, j))
2515 if (!isUndefOrEqual(BitI1, j))
2521 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2522 SmallVector<int, 8> M;
2524 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2527 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2528 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2530 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, MVT VT) {
2531 int NumElems = VT.getVectorNumElements();
2532 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2535 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2537 int BitI1 = Mask[i+1];
2538 if (!isUndefOrEqual(BitI, j))
2540 if (!isUndefOrEqual(BitI1, j))
2546 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2547 SmallVector<int, 8> M;
2549 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2552 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2553 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2554 /// MOVSD, and MOVD, i.e. setting the lowest element.
2555 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2556 if (VT.getVectorElementType().getSizeInBits() < 32)
2559 int NumElts = VT.getVectorNumElements();
2561 if (!isUndefOrEqual(Mask[0], NumElts))
2564 for (int i = 1; i < NumElts; ++i)
2565 if (!isUndefOrEqual(Mask[i], i))
2571 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2572 SmallVector<int, 8> M;
2574 return ::isMOVLMask(M, N->getValueType(0));
2577 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2578 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2579 /// element of vector 2 and the other elements to come from vector 1 in order.
2580 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, MVT VT,
2581 bool V2IsSplat = false, bool V2IsUndef = false) {
2582 int NumOps = VT.getVectorNumElements();
2583 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2586 if (!isUndefOrEqual(Mask[0], 0))
2589 for (int i = 1; i < NumOps; ++i)
2590 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2591 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2592 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2598 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2599 bool V2IsUndef = false) {
2600 SmallVector<int, 8> M;
2602 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2605 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2606 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2607 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2608 if (N->getValueType(0).getVectorNumElements() != 4)
2611 // Expect 1, 1, 3, 3
2612 for (unsigned i = 0; i < 2; ++i) {
2613 int Elt = N->getMaskElt(i);
2614 if (Elt >= 0 && Elt != 1)
2619 for (unsigned i = 2; i < 4; ++i) {
2620 int Elt = N->getMaskElt(i);
2621 if (Elt >= 0 && Elt != 3)
2626 // Don't use movshdup if it can be done with a shufps.
2627 // FIXME: verify that matching u, u, 3, 3 is what we want.
2631 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2632 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2633 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2634 if (N->getValueType(0).getVectorNumElements() != 4)
2637 // Expect 0, 0, 2, 2
2638 for (unsigned i = 0; i < 2; ++i)
2639 if (N->getMaskElt(i) > 0)
2643 for (unsigned i = 2; i < 4; ++i) {
2644 int Elt = N->getMaskElt(i);
2645 if (Elt >= 0 && Elt != 2)
2650 // Don't use movsldup if it can be done with a shufps.
2654 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2655 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2656 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2657 int e = N->getValueType(0).getVectorNumElements() / 2;
2659 for (int i = 0; i < e; ++i)
2660 if (!isUndefOrEqual(N->getMaskElt(i), i))
2662 for (int i = 0; i < e; ++i)
2663 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2668 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2669 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2671 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2672 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2673 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2675 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2677 for (int i = 0; i < NumOperands; ++i) {
2678 int Val = SVOp->getMaskElt(NumOperands-i-1);
2679 if (Val < 0) Val = 0;
2680 if (Val >= NumOperands) Val -= NumOperands;
2682 if (i != NumOperands - 1)
2688 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2689 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2691 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2692 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2694 // 8 nodes, but we only care about the last 4.
2695 for (unsigned i = 7; i >= 4; --i) {
2696 int Val = SVOp->getMaskElt(i);
2705 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2706 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2708 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2709 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2711 // 8 nodes, but we only care about the first 4.
2712 for (int i = 3; i >= 0; --i) {
2713 int Val = SVOp->getMaskElt(i);
2722 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2724 bool X86::isZeroNode(SDValue Elt) {
2725 return ((isa<ConstantSDNode>(Elt) &&
2726 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2727 (isa<ConstantFPSDNode>(Elt) &&
2728 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2731 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
2732 /// their permute mask.
2733 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
2734 SelectionDAG &DAG) {
2735 MVT VT = SVOp->getValueType(0);
2736 unsigned NumElems = VT.getVectorNumElements();
2737 SmallVector<int, 8> MaskVec;
2739 for (unsigned i = 0; i != NumElems; ++i) {
2740 int idx = SVOp->getMaskElt(i);
2742 MaskVec.push_back(idx);
2743 else if (idx < (int)NumElems)
2744 MaskVec.push_back(idx + NumElems);
2746 MaskVec.push_back(idx - NumElems);
2748 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
2749 SVOp->getOperand(0), &MaskVec[0]);
2752 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2753 /// the two vector operands have swapped position.
2754 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, MVT VT) {
2755 unsigned NumElems = VT.getVectorNumElements();
2756 for (unsigned i = 0; i != NumElems; ++i) {
2760 else if (idx < (int)NumElems)
2761 Mask[i] = idx + NumElems;
2763 Mask[i] = idx - NumElems;
2767 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2768 /// match movhlps. The lower half elements should come from upper half of
2769 /// V1 (and in order), and the upper half elements should come from the upper
2770 /// half of V2 (and in order).
2771 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
2772 if (Op->getValueType(0).getVectorNumElements() != 4)
2774 for (unsigned i = 0, e = 2; i != e; ++i)
2775 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
2777 for (unsigned i = 2; i != 4; ++i)
2778 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
2783 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2784 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2786 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2787 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2789 N = N->getOperand(0).getNode();
2790 if (!ISD::isNON_EXTLoad(N))
2793 *LD = cast<LoadSDNode>(N);
2797 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2798 /// match movlp{s|d}. The lower half elements should come from lower half of
2799 /// V1 (and in order), and the upper half elements should come from the upper
2800 /// half of V2 (and in order). And since V1 will become the source of the
2801 /// MOVLP, it must be either a vector load or a scalar load to vector.
2802 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
2803 ShuffleVectorSDNode *Op) {
2804 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2806 // Is V2 is a vector load, don't do this transformation. We will try to use
2807 // load folding shufps op.
2808 if (ISD::isNON_EXTLoad(V2))
2811 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
2813 if (NumElems != 2 && NumElems != 4)
2815 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2816 if (!isUndefOrEqual(Op->getMaskElt(i), i))
2818 for (unsigned i = NumElems/2; i != NumElems; ++i)
2819 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
2824 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2826 static bool isSplatVector(SDNode *N) {
2827 if (N->getOpcode() != ISD::BUILD_VECTOR)
2830 SDValue SplatValue = N->getOperand(0);
2831 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2832 if (N->getOperand(i) != SplatValue)
2837 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2838 /// to an zero vector.
2839 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
2840 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
2841 SDValue V1 = N->getOperand(0);
2842 SDValue V2 = N->getOperand(1);
2843 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2844 for (unsigned i = 0; i != NumElems; ++i) {
2845 int Idx = N->getMaskElt(i);
2846 if (Idx >= (int)NumElems) {
2847 unsigned Opc = V2.getOpcode();
2848 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2850 if (Opc != ISD::BUILD_VECTOR ||
2851 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
2853 } else if (Idx >= 0) {
2854 unsigned Opc = V1.getOpcode();
2855 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2857 if (Opc != ISD::BUILD_VECTOR ||
2858 !X86::isZeroNode(V1.getOperand(Idx)))
2865 /// getZeroVector - Returns a vector of specified type with all zero elements.
2867 static SDValue getZeroVector(MVT VT, bool HasSSE2, SelectionDAG &DAG,
2869 assert(VT.isVector() && "Expected a vector type");
2871 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2872 // type. This ensures they get CSE'd.
2874 if (VT.getSizeInBits() == 64) { // MMX
2875 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2876 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2877 } else if (HasSSE2) { // SSE2
2878 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2879 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2881 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2882 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
2884 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2887 /// getOnesVector - Returns a vector of specified type with all bits set.
2889 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG, DebugLoc dl) {
2890 assert(VT.isVector() && "Expected a vector type");
2892 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2893 // type. This ensures they get CSE'd.
2894 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2896 if (VT.getSizeInBits() == 64) // MMX
2897 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2899 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2900 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2904 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2905 /// that point to V2 points to its first element.
2906 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
2907 MVT VT = SVOp->getValueType(0);
2908 unsigned NumElems = VT.getVectorNumElements();
2910 bool Changed = false;
2911 SmallVector<int, 8> MaskVec;
2912 SVOp->getMask(MaskVec);
2914 for (unsigned i = 0; i != NumElems; ++i) {
2915 if (MaskVec[i] > (int)NumElems) {
2916 MaskVec[i] = NumElems;
2921 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
2922 SVOp->getOperand(1), &MaskVec[0]);
2923 return SDValue(SVOp, 0);
2926 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2927 /// operation of specified width.
2928 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2930 unsigned NumElems = VT.getVectorNumElements();
2931 SmallVector<int, 8> Mask;
2932 Mask.push_back(NumElems);
2933 for (unsigned i = 1; i != NumElems; ++i)
2935 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2938 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
2939 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2941 unsigned NumElems = VT.getVectorNumElements();
2942 SmallVector<int, 8> Mask;
2943 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2945 Mask.push_back(i + NumElems);
2947 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2950 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
2951 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2953 unsigned NumElems = VT.getVectorNumElements();
2954 unsigned Half = NumElems/2;
2955 SmallVector<int, 8> Mask;
2956 for (unsigned i = 0; i != Half; ++i) {
2957 Mask.push_back(i + Half);
2958 Mask.push_back(i + NumElems + Half);
2960 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2963 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
2964 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
2966 if (SV->getValueType(0).getVectorNumElements() <= 4)
2967 return SDValue(SV, 0);
2969 MVT PVT = MVT::v4f32;
2970 MVT VT = SV->getValueType(0);
2971 DebugLoc dl = SV->getDebugLoc();
2972 SDValue V1 = SV->getOperand(0);
2973 int NumElems = VT.getVectorNumElements();
2974 int EltNo = SV->getSplatIndex();
2976 // unpack elements to the correct location
2977 while (NumElems > 4) {
2978 if (EltNo < NumElems/2) {
2979 V1 = getUnpackl(DAG, dl, VT, V1, V1);
2981 V1 = getUnpackh(DAG, dl, VT, V1, V1);
2982 EltNo -= NumElems/2;
2987 // Perform the splat.
2988 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
2989 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
2990 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
2991 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
2994 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
2995 /// vector of zero or undef vector. This produces a shuffle where the low
2996 /// element of V2 is swizzled into the zero/undef vector, landing at element
2997 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
2998 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
2999 bool isZero, bool HasSSE2,
3000 SelectionDAG &DAG) {
3001 MVT VT = V2.getValueType();
3003 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3004 unsigned NumElems = VT.getVectorNumElements();
3005 SmallVector<int, 16> MaskVec;
3006 for (unsigned i = 0; i != NumElems; ++i)
3007 // If this is the insertion idx, put the low elt of V2 here.
3008 MaskVec.push_back(i == Idx ? NumElems : i);
3009 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3012 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3013 /// a shuffle that is zero.
3015 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3016 bool Low, SelectionDAG &DAG) {
3017 unsigned NumZeros = 0;
3018 for (int i = 0; i < NumElems; ++i) {
3019 unsigned Index = Low ? i : NumElems-i-1;
3020 int Idx = SVOp->getMaskElt(Index);
3025 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3026 if (Elt.getNode() && X86::isZeroNode(Elt))
3034 /// isVectorShift - Returns true if the shuffle can be implemented as a
3035 /// logical left or right shift of a vector.
3036 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3037 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3038 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3039 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3042 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3045 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3049 bool SeenV1 = false;
3050 bool SeenV2 = false;
3051 for (int i = NumZeros; i < NumElems; ++i) {
3052 int Val = isLeft ? (i - NumZeros) : i;
3053 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3065 if (SeenV1 && SeenV2)
3068 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3074 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3076 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3077 unsigned NumNonZero, unsigned NumZero,
3078 SelectionDAG &DAG, TargetLowering &TLI) {
3082 DebugLoc dl = Op.getDebugLoc();
3085 for (unsigned i = 0; i < 16; ++i) {
3086 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3087 if (ThisIsNonZero && First) {
3089 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3091 V = DAG.getUNDEF(MVT::v8i16);
3096 SDValue ThisElt(0, 0), LastElt(0, 0);
3097 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3098 if (LastIsNonZero) {
3099 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3100 MVT::i16, Op.getOperand(i-1));
3102 if (ThisIsNonZero) {
3103 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3104 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3105 ThisElt, DAG.getConstant(8, MVT::i8));
3107 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3111 if (ThisElt.getNode())
3112 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3113 DAG.getIntPtrConstant(i/2));
3117 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3120 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3122 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3123 unsigned NumNonZero, unsigned NumZero,
3124 SelectionDAG &DAG, TargetLowering &TLI) {
3128 DebugLoc dl = Op.getDebugLoc();
3131 for (unsigned i = 0; i < 8; ++i) {
3132 bool isNonZero = (NonZeros & (1 << i)) != 0;
3136 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3138 V = DAG.getUNDEF(MVT::v8i16);
3141 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3142 MVT::v8i16, V, Op.getOperand(i),
3143 DAG.getIntPtrConstant(i));
3150 /// getVShift - Return a vector logical shift node.
3152 static SDValue getVShift(bool isLeft, MVT VT, SDValue SrcOp,
3153 unsigned NumBits, SelectionDAG &DAG,
3154 const TargetLowering &TLI, DebugLoc dl) {
3155 bool isMMX = VT.getSizeInBits() == 64;
3156 MVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3157 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3158 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3159 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3160 DAG.getNode(Opc, dl, ShVT, SrcOp,
3161 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3165 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3166 DebugLoc dl = Op.getDebugLoc();
3167 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3168 if (ISD::isBuildVectorAllZeros(Op.getNode())
3169 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3170 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3171 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3172 // eliminated on x86-32 hosts.
3173 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3176 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3177 return getOnesVector(Op.getValueType(), DAG, dl);
3178 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3181 MVT VT = Op.getValueType();
3182 MVT EVT = VT.getVectorElementType();
3183 unsigned EVTBits = EVT.getSizeInBits();
3185 unsigned NumElems = Op.getNumOperands();
3186 unsigned NumZero = 0;
3187 unsigned NumNonZero = 0;
3188 unsigned NonZeros = 0;
3189 bool IsAllConstants = true;
3190 SmallSet<SDValue, 8> Values;
3191 for (unsigned i = 0; i < NumElems; ++i) {
3192 SDValue Elt = Op.getOperand(i);
3193 if (Elt.getOpcode() == ISD::UNDEF)
3196 if (Elt.getOpcode() != ISD::Constant &&
3197 Elt.getOpcode() != ISD::ConstantFP)
3198 IsAllConstants = false;
3199 if (X86::isZeroNode(Elt))
3202 NonZeros |= (1 << i);
3207 if (NumNonZero == 0) {
3208 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3209 return DAG.getUNDEF(VT);
3212 // Special case for single non-zero, non-undef, element.
3213 if (NumNonZero == 1) {
3214 unsigned Idx = CountTrailingZeros_32(NonZeros);
3215 SDValue Item = Op.getOperand(Idx);
3217 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3218 // the value are obviously zero, truncate the value to i32 and do the
3219 // insertion that way. Only do this if the value is non-constant or if the
3220 // value is a constant being inserted into element 0. It is cheaper to do
3221 // a constant pool load than it is to do a movd + shuffle.
3222 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
3223 (!IsAllConstants || Idx == 0)) {
3224 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3225 // Handle MMX and SSE both.
3226 MVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3227 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3229 // Truncate the value (which may itself be a constant) to i32, and
3230 // convert it to a vector with movd (S2V+shuffle to zero extend).
3231 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3232 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3233 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3234 Subtarget->hasSSE2(), DAG);
3236 // Now we have our 32-bit value zero extended in the low element of
3237 // a vector. If Idx != 0, swizzle it into place.
3239 SmallVector<int, 4> Mask;
3240 Mask.push_back(Idx);
3241 for (unsigned i = 1; i != VecElts; ++i)
3243 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3244 DAG.getUNDEF(Item.getValueType()),
3247 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3251 // If we have a constant or non-constant insertion into the low element of
3252 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3253 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3254 // depending on what the source datatype is.
3257 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3258 } else if (EVT == MVT::i32 || EVT == MVT::f32 || EVT == MVT::f64 ||
3259 (EVT == MVT::i64 && Subtarget->is64Bit())) {
3260 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3261 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3262 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3264 } else if (EVT == MVT::i16 || EVT == MVT::i8) {
3265 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3266 MVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3267 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3268 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3269 Subtarget->hasSSE2(), DAG);
3270 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3274 // Is it a vector logical left shift?
3275 if (NumElems == 2 && Idx == 1 &&
3276 X86::isZeroNode(Op.getOperand(0)) &&
3277 !X86::isZeroNode(Op.getOperand(1))) {
3278 unsigned NumBits = VT.getSizeInBits();
3279 return getVShift(true, VT,
3280 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3281 VT, Op.getOperand(1)),
3282 NumBits/2, DAG, *this, dl);
3285 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3288 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3289 // is a non-constant being inserted into an element other than the low one,
3290 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3291 // movd/movss) to move this into the low element, then shuffle it into
3293 if (EVTBits == 32) {
3294 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3296 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3297 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3298 Subtarget->hasSSE2(), DAG);
3299 SmallVector<int, 8> MaskVec;
3300 for (unsigned i = 0; i < NumElems; i++)
3301 MaskVec.push_back(i == Idx ? 0 : 1);
3302 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3306 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3307 if (Values.size() == 1)
3310 // A vector full of immediates; various special cases are already
3311 // handled, so this is best done with a single constant-pool load.
3315 // Let legalizer expand 2-wide build_vectors.
3316 if (EVTBits == 64) {
3317 if (NumNonZero == 1) {
3318 // One half is zero or undef.
3319 unsigned Idx = CountTrailingZeros_32(NonZeros);
3320 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3321 Op.getOperand(Idx));
3322 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3323 Subtarget->hasSSE2(), DAG);
3328 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3329 if (EVTBits == 8 && NumElems == 16) {
3330 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3332 if (V.getNode()) return V;
3335 if (EVTBits == 16 && NumElems == 8) {
3336 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3338 if (V.getNode()) return V;
3341 // If element VT is == 32 bits, turn it into a number of shuffles.
3342 SmallVector<SDValue, 8> V;
3344 if (NumElems == 4 && NumZero > 0) {
3345 for (unsigned i = 0; i < 4; ++i) {
3346 bool isZero = !(NonZeros & (1 << i));
3348 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3350 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3353 for (unsigned i = 0; i < 2; ++i) {
3354 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3357 V[i] = V[i*2]; // Must be a zero vector.
3360 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3363 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3366 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3371 SmallVector<int, 8> MaskVec;
3372 bool Reverse = (NonZeros & 0x3) == 2;
3373 for (unsigned i = 0; i < 2; ++i)
3374 MaskVec.push_back(Reverse ? 1-i : i);
3375 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3376 for (unsigned i = 0; i < 2; ++i)
3377 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3378 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3381 if (Values.size() > 2) {
3382 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3383 // values to be inserted is equal to the number of elements, in which case
3384 // use the unpack code below in the hopes of matching the consecutive elts
3385 // load merge pattern for shuffles.
3386 // FIXME: We could probably just check that here directly.
3387 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3388 getSubtarget()->hasSSE41()) {
3389 V[0] = DAG.getUNDEF(VT);
3390 for (unsigned i = 0; i < NumElems; ++i)
3391 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3392 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3393 Op.getOperand(i), DAG.getIntPtrConstant(i));
3396 // Expand into a number of unpckl*.
3398 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3399 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3400 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3401 for (unsigned i = 0; i < NumElems; ++i)
3402 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3404 while (NumElems != 0) {
3405 for (unsigned i = 0; i < NumElems; ++i)
3406 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3415 // v8i16 shuffles - Prefer shuffles in the following order:
3416 // 1. [all] pshuflw, pshufhw, optional move
3417 // 2. [ssse3] 1 x pshufb
3418 // 3. [ssse3] 2 x pshufb + 1 x por
3419 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3421 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3422 SelectionDAG &DAG, X86TargetLowering &TLI) {
3423 SDValue V1 = SVOp->getOperand(0);
3424 SDValue V2 = SVOp->getOperand(1);
3425 DebugLoc dl = SVOp->getDebugLoc();
3426 SmallVector<int, 8> MaskVals;
3428 // Determine if more than 1 of the words in each of the low and high quadwords
3429 // of the result come from the same quadword of one of the two inputs. Undef
3430 // mask values count as coming from any quadword, for better codegen.
3431 SmallVector<unsigned, 4> LoQuad(4);
3432 SmallVector<unsigned, 4> HiQuad(4);
3433 BitVector InputQuads(4);
3434 for (unsigned i = 0; i < 8; ++i) {
3435 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3436 int EltIdx = SVOp->getMaskElt(i);
3437 MaskVals.push_back(EltIdx);
3446 InputQuads.set(EltIdx / 4);
3449 int BestLoQuad = -1;
3450 unsigned MaxQuad = 1;
3451 for (unsigned i = 0; i < 4; ++i) {
3452 if (LoQuad[i] > MaxQuad) {
3454 MaxQuad = LoQuad[i];
3458 int BestHiQuad = -1;
3460 for (unsigned i = 0; i < 4; ++i) {
3461 if (HiQuad[i] > MaxQuad) {
3463 MaxQuad = HiQuad[i];
3467 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3468 // of the two input vectors, shuffle them into one input vector so only a
3469 // single pshufb instruction is necessary. If There are more than 2 input
3470 // quads, disable the next transformation since it does not help SSSE3.
3471 bool V1Used = InputQuads[0] || InputQuads[1];
3472 bool V2Used = InputQuads[2] || InputQuads[3];
3473 if (TLI.getSubtarget()->hasSSSE3()) {
3474 if (InputQuads.count() == 2 && V1Used && V2Used) {
3475 BestLoQuad = InputQuads.find_first();
3476 BestHiQuad = InputQuads.find_next(BestLoQuad);
3478 if (InputQuads.count() > 2) {
3484 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3485 // the shuffle mask. If a quad is scored as -1, that means that it contains
3486 // words from all 4 input quadwords.
3488 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3489 SmallVector<int, 8> MaskV;
3490 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3491 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3492 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3493 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3494 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3495 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3497 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3498 // source words for the shuffle, to aid later transformations.
3499 bool AllWordsInNewV = true;
3500 bool InOrder[2] = { true, true };
3501 for (unsigned i = 0; i != 8; ++i) {
3502 int idx = MaskVals[i];
3504 InOrder[i/4] = false;
3505 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3507 AllWordsInNewV = false;
3511 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3512 if (AllWordsInNewV) {
3513 for (int i = 0; i != 8; ++i) {
3514 int idx = MaskVals[i];
3517 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3518 if ((idx != i) && idx < 4)
3520 if ((idx != i) && idx > 3)
3529 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3530 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3531 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3532 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3533 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3537 // If we have SSSE3, and all words of the result are from 1 input vector,
3538 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3539 // is present, fall back to case 4.
3540 if (TLI.getSubtarget()->hasSSSE3()) {
3541 SmallVector<SDValue,16> pshufbMask;
3543 // If we have elements from both input vectors, set the high bit of the
3544 // shuffle mask element to zero out elements that come from V2 in the V1
3545 // mask, and elements that come from V1 in the V2 mask, so that the two
3546 // results can be OR'd together.
3547 bool TwoInputs = V1Used && V2Used;
3548 for (unsigned i = 0; i != 8; ++i) {
3549 int EltIdx = MaskVals[i] * 2;
3550 if (TwoInputs && (EltIdx >= 16)) {
3551 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3552 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3555 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3556 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3558 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3559 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3560 DAG.getNode(ISD::BUILD_VECTOR, dl,
3561 MVT::v16i8, &pshufbMask[0], 16));
3563 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3565 // Calculate the shuffle mask for the second input, shuffle it, and
3566 // OR it with the first shuffled input.
3568 for (unsigned i = 0; i != 8; ++i) {
3569 int EltIdx = MaskVals[i] * 2;
3571 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3572 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3575 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3576 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3578 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3579 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3580 DAG.getNode(ISD::BUILD_VECTOR, dl,
3581 MVT::v16i8, &pshufbMask[0], 16));
3582 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3583 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3586 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3587 // and update MaskVals with new element order.
3588 BitVector InOrder(8);
3589 if (BestLoQuad >= 0) {
3590 SmallVector<int, 8> MaskV;
3591 for (int i = 0; i != 4; ++i) {
3592 int idx = MaskVals[i];
3594 MaskV.push_back(-1);
3596 } else if ((idx / 4) == BestLoQuad) {
3597 MaskV.push_back(idx & 3);
3600 MaskV.push_back(-1);
3603 for (unsigned i = 4; i != 8; ++i)
3605 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3609 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
3610 // and update MaskVals with the new element order.
3611 if (BestHiQuad >= 0) {
3612 SmallVector<int, 8> MaskV;
3613 for (unsigned i = 0; i != 4; ++i)
3615 for (unsigned i = 4; i != 8; ++i) {
3616 int idx = MaskVals[i];
3618 MaskV.push_back(-1);
3620 } else if ((idx / 4) == BestHiQuad) {
3621 MaskV.push_back((idx & 3) + 4);
3624 MaskV.push_back(-1);
3627 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3631 // In case BestHi & BestLo were both -1, which means each quadword has a word
3632 // from each of the four input quadwords, calculate the InOrder bitvector now
3633 // before falling through to the insert/extract cleanup.
3634 if (BestLoQuad == -1 && BestHiQuad == -1) {
3636 for (int i = 0; i != 8; ++i)
3637 if (MaskVals[i] < 0 || MaskVals[i] == i)
3641 // The other elements are put in the right place using pextrw and pinsrw.
3642 for (unsigned i = 0; i != 8; ++i) {
3645 int EltIdx = MaskVals[i];
3648 SDValue ExtOp = (EltIdx < 8)
3649 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
3650 DAG.getIntPtrConstant(EltIdx))
3651 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
3652 DAG.getIntPtrConstant(EltIdx - 8));
3653 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3654 DAG.getIntPtrConstant(i));
3659 // v16i8 shuffles - Prefer shuffles in the following order:
3660 // 1. [ssse3] 1 x pshufb
3661 // 2. [ssse3] 2 x pshufb + 1 x por
3662 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
3664 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
3665 SelectionDAG &DAG, X86TargetLowering &TLI) {
3666 SDValue V1 = SVOp->getOperand(0);
3667 SDValue V2 = SVOp->getOperand(1);
3668 DebugLoc dl = SVOp->getDebugLoc();
3669 SmallVector<int, 16> MaskVals;
3670 SVOp->getMask(MaskVals);
3672 // If we have SSSE3, case 1 is generated when all result bytes come from
3673 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
3674 // present, fall back to case 3.
3675 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
3678 for (unsigned i = 0; i < 16; ++i) {
3679 int EltIdx = MaskVals[i];
3688 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
3689 if (TLI.getSubtarget()->hasSSSE3()) {
3690 SmallVector<SDValue,16> pshufbMask;
3692 // If all result elements are from one input vector, then only translate
3693 // undef mask values to 0x80 (zero out result) in the pshufb mask.
3695 // Otherwise, we have elements from both input vectors, and must zero out
3696 // elements that come from V2 in the first mask, and V1 in the second mask
3697 // so that we can OR them together.
3698 bool TwoInputs = !(V1Only || V2Only);
3699 for (unsigned i = 0; i != 16; ++i) {
3700 int EltIdx = MaskVals[i];
3701 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
3702 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3705 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3707 // If all the elements are from V2, assign it to V1 and return after
3708 // building the first pshufb.
3711 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3712 DAG.getNode(ISD::BUILD_VECTOR, dl,
3713 MVT::v16i8, &pshufbMask[0], 16));
3717 // Calculate the shuffle mask for the second input, shuffle it, and
3718 // OR it with the first shuffled input.
3720 for (unsigned i = 0; i != 16; ++i) {
3721 int EltIdx = MaskVals[i];
3723 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3726 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3728 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3729 DAG.getNode(ISD::BUILD_VECTOR, dl,
3730 MVT::v16i8, &pshufbMask[0], 16));
3731 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3734 // No SSSE3 - Calculate in place words and then fix all out of place words
3735 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
3736 // the 16 different words that comprise the two doublequadword input vectors.
3737 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3738 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
3739 SDValue NewV = V2Only ? V2 : V1;
3740 for (int i = 0; i != 8; ++i) {
3741 int Elt0 = MaskVals[i*2];
3742 int Elt1 = MaskVals[i*2+1];
3744 // This word of the result is all undef, skip it.
3745 if (Elt0 < 0 && Elt1 < 0)
3748 // This word of the result is already in the correct place, skip it.
3749 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
3751 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
3754 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
3755 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
3758 // If Elt0 and Elt1 are defined, are consecutive, and can be load
3759 // using a single extract together, load it and store it.
3760 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
3761 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3762 DAG.getIntPtrConstant(Elt1 / 2));
3763 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3764 DAG.getIntPtrConstant(i));
3768 // If Elt1 is defined, extract it from the appropriate source. If the
3769 // source byte is not also odd, shift the extracted word left 8 bits
3770 // otherwise clear the bottom 8 bits if we need to do an or.
3772 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3773 DAG.getIntPtrConstant(Elt1 / 2));
3774 if ((Elt1 & 1) == 0)
3775 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
3776 DAG.getConstant(8, TLI.getShiftAmountTy()));
3778 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
3779 DAG.getConstant(0xFF00, MVT::i16));
3781 // If Elt0 is defined, extract it from the appropriate source. If the
3782 // source byte is not also even, shift the extracted word right 8 bits. If
3783 // Elt1 was also defined, OR the extracted values together before
3784 // inserting them in the result.
3786 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
3787 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
3788 if ((Elt0 & 1) != 0)
3789 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
3790 DAG.getConstant(8, TLI.getShiftAmountTy()));
3792 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
3793 DAG.getConstant(0x00FF, MVT::i16));
3794 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
3797 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3798 DAG.getIntPtrConstant(i));
3800 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
3803 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3804 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3805 /// done when every pair / quad of shuffle mask elements point to elements in
3806 /// the right sequence. e.g.
3807 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3809 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
3811 TargetLowering &TLI, DebugLoc dl) {
3812 MVT VT = SVOp->getValueType(0);
3813 SDValue V1 = SVOp->getOperand(0);
3814 SDValue V2 = SVOp->getOperand(1);
3815 unsigned NumElems = VT.getVectorNumElements();
3816 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3817 MVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3818 MVT MaskEltVT = MaskVT.getVectorElementType();
3820 switch (VT.getSimpleVT()) {
3821 default: assert(false && "Unexpected!");
3822 case MVT::v4f32: NewVT = MVT::v2f64; break;
3823 case MVT::v4i32: NewVT = MVT::v2i64; break;
3824 case MVT::v8i16: NewVT = MVT::v4i32; break;
3825 case MVT::v16i8: NewVT = MVT::v4i32; break;
3828 if (NewWidth == 2) {
3834 int Scale = NumElems / NewWidth;
3835 SmallVector<int, 8> MaskVec;
3836 for (unsigned i = 0; i < NumElems; i += Scale) {
3838 for (int j = 0; j < Scale; ++j) {
3839 int EltIdx = SVOp->getMaskElt(i+j);
3843 StartIdx = EltIdx - (EltIdx % Scale);
3844 if (EltIdx != StartIdx + j)
3848 MaskVec.push_back(-1);
3850 MaskVec.push_back(StartIdx / Scale);
3853 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
3854 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
3855 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
3858 /// getVZextMovL - Return a zero-extending vector move low node.
3860 static SDValue getVZextMovL(MVT VT, MVT OpVT,
3861 SDValue SrcOp, SelectionDAG &DAG,
3862 const X86Subtarget *Subtarget, DebugLoc dl) {
3863 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
3864 LoadSDNode *LD = NULL;
3865 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
3866 LD = dyn_cast<LoadSDNode>(SrcOp);
3868 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
3870 MVT EVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
3871 if ((EVT != MVT::i64 || Subtarget->is64Bit()) &&
3872 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
3873 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
3874 SrcOp.getOperand(0).getOperand(0).getValueType() == EVT) {
3876 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
3877 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3878 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3879 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3887 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3888 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3889 DAG.getNode(ISD::BIT_CONVERT, dl,
3893 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
3896 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3897 SDValue V1 = SVOp->getOperand(0);
3898 SDValue V2 = SVOp->getOperand(1);
3899 DebugLoc dl = SVOp->getDebugLoc();
3900 MVT VT = SVOp->getValueType(0);
3902 SmallVector<std::pair<int, int>, 8> Locs;
3904 SmallVector<int, 8> Mask1(4U, -1);
3905 SmallVector<int, 8> PermMask;
3906 SVOp->getMask(PermMask);
3910 for (unsigned i = 0; i != 4; ++i) {
3911 int Idx = PermMask[i];
3913 Locs[i] = std::make_pair(-1, -1);
3915 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
3917 Locs[i] = std::make_pair(0, NumLo);
3921 Locs[i] = std::make_pair(1, NumHi);
3923 Mask1[2+NumHi] = Idx;
3929 if (NumLo <= 2 && NumHi <= 2) {
3930 // If no more than two elements come from either vector. This can be
3931 // implemented with two shuffles. First shuffle gather the elements.
3932 // The second shuffle, which takes the first shuffle as both of its
3933 // vector operands, put the elements into the right order.
3934 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3936 SmallVector<int, 8> Mask2(4U, -1);
3938 for (unsigned i = 0; i != 4; ++i) {
3939 if (Locs[i].first == -1)
3942 unsigned Idx = (i < 2) ? 0 : 4;
3943 Idx += Locs[i].first * 2 + Locs[i].second;
3948 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
3949 } else if (NumLo == 3 || NumHi == 3) {
3950 // Otherwise, we must have three elements from one vector, call it X, and
3951 // one element from the other, call it Y. First, use a shufps to build an
3952 // intermediate vector with the one element from Y and the element from X
3953 // that will be in the same half in the final destination (the indexes don't
3954 // matter). Then, use a shufps to build the final vector, taking the half
3955 // containing the element from Y from the intermediate, and the other half
3958 // Normalize it so the 3 elements come from V1.
3959 CommuteVectorShuffleMask(PermMask, VT);
3963 // Find the element from V2.
3965 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
3966 int Val = PermMask[HiIndex];
3973 Mask1[0] = PermMask[HiIndex];
3975 Mask1[2] = PermMask[HiIndex^1];
3977 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3980 Mask1[0] = PermMask[0];
3981 Mask1[1] = PermMask[1];
3982 Mask1[2] = HiIndex & 1 ? 6 : 4;
3983 Mask1[3] = HiIndex & 1 ? 4 : 6;
3984 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3986 Mask1[0] = HiIndex & 1 ? 2 : 0;
3987 Mask1[1] = HiIndex & 1 ? 0 : 2;
3988 Mask1[2] = PermMask[2];
3989 Mask1[3] = PermMask[3];
3994 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
3998 // Break it into (shuffle shuffle_hi, shuffle_lo).
4000 SmallVector<int,8> LoMask(4U, -1);
4001 SmallVector<int,8> HiMask(4U, -1);
4003 SmallVector<int,8> *MaskPtr = &LoMask;
4004 unsigned MaskIdx = 0;
4007 for (unsigned i = 0; i != 4; ++i) {
4014 int Idx = PermMask[i];
4016 Locs[i] = std::make_pair(-1, -1);
4017 } else if (Idx < 4) {
4018 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4019 (*MaskPtr)[LoIdx] = Idx;
4022 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4023 (*MaskPtr)[HiIdx] = Idx;
4028 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4029 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4030 SmallVector<int, 8> MaskOps;
4031 for (unsigned i = 0; i != 4; ++i) {
4032 if (Locs[i].first == -1) {
4033 MaskOps.push_back(-1);
4035 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4036 MaskOps.push_back(Idx);
4039 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4043 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4044 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4045 SDValue V1 = Op.getOperand(0);
4046 SDValue V2 = Op.getOperand(1);
4047 MVT VT = Op.getValueType();
4048 DebugLoc dl = Op.getDebugLoc();
4049 unsigned NumElems = VT.getVectorNumElements();
4050 bool isMMX = VT.getSizeInBits() == 64;
4051 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4052 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4053 bool V1IsSplat = false;
4054 bool V2IsSplat = false;
4056 if (isZeroShuffle(SVOp))
4057 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4059 // Promote splats to v4f32.
4060 if (SVOp->isSplat()) {
4061 if (isMMX || NumElems < 4)
4063 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4066 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4068 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4069 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4070 if (NewOp.getNode())
4071 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4072 LowerVECTOR_SHUFFLE(NewOp, DAG));
4073 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4074 // FIXME: Figure out a cleaner way to do this.
4075 // Try to make use of movq to zero out the top part.
4076 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4077 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4078 if (NewOp.getNode()) {
4079 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4080 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4081 DAG, Subtarget, dl);
4083 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4084 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4085 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4086 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4087 DAG, Subtarget, dl);
4091 if (X86::isPSHUFDMask(SVOp))
4094 // Check if this can be converted into a logical shift.
4095 bool isLeft = false;
4098 bool isShift = getSubtarget()->hasSSE2() &&
4099 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4100 if (isShift && ShVal.hasOneUse()) {
4101 // If the shifted value has multiple uses, it may be cheaper to use
4102 // v_set0 + movlhps or movhlps, etc.
4103 MVT EVT = VT.getVectorElementType();
4104 ShAmt *= EVT.getSizeInBits();
4105 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4108 if (X86::isMOVLMask(SVOp)) {
4111 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4112 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4117 // FIXME: fold these into legal mask.
4118 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4119 X86::isMOVSLDUPMask(SVOp) ||
4120 X86::isMOVHLPSMask(SVOp) ||
4121 X86::isMOVHPMask(SVOp) ||
4122 X86::isMOVLPMask(SVOp)))
4125 if (ShouldXformToMOVHLPS(SVOp) ||
4126 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4127 return CommuteVectorShuffle(SVOp, DAG);
4130 // No better options. Use a vshl / vsrl.
4131 MVT EVT = VT.getVectorElementType();
4132 ShAmt *= EVT.getSizeInBits();
4133 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4136 bool Commuted = false;
4137 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4138 // 1,1,1,1 -> v8i16 though.
4139 V1IsSplat = isSplatVector(V1.getNode());
4140 V2IsSplat = isSplatVector(V2.getNode());
4142 // Canonicalize the splat or undef, if present, to be on the RHS.
4143 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4144 Op = CommuteVectorShuffle(SVOp, DAG);
4145 SVOp = cast<ShuffleVectorSDNode>(Op);
4146 V1 = SVOp->getOperand(0);
4147 V2 = SVOp->getOperand(1);
4148 std::swap(V1IsSplat, V2IsSplat);
4149 std::swap(V1IsUndef, V2IsUndef);
4153 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4154 // Shuffling low element of v1 into undef, just return v1.
4157 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4158 // the instruction selector will not match, so get a canonical MOVL with
4159 // swapped operands to undo the commute.
4160 return getMOVL(DAG, dl, VT, V2, V1);
4163 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4164 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4165 X86::isUNPCKLMask(SVOp) ||
4166 X86::isUNPCKHMask(SVOp))
4170 // Normalize mask so all entries that point to V2 points to its first
4171 // element then try to match unpck{h|l} again. If match, return a
4172 // new vector_shuffle with the corrected mask.
4173 SDValue NewMask = NormalizeMask(SVOp, DAG);
4174 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4175 if (NSVOp != SVOp) {
4176 if (X86::isUNPCKLMask(NSVOp, true)) {
4178 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4185 // Commute is back and try unpck* again.
4186 // FIXME: this seems wrong.
4187 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4188 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4189 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4190 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4191 X86::isUNPCKLMask(NewSVOp) ||
4192 X86::isUNPCKHMask(NewSVOp))
4196 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4198 // Normalize the node to match x86 shuffle ops if needed
4199 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4200 return CommuteVectorShuffle(SVOp, DAG);
4202 // Check for legal shuffle and return?
4203 SmallVector<int, 16> PermMask;
4204 SVOp->getMask(PermMask);
4205 if (isShuffleMaskLegal(PermMask, VT))
4208 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4209 if (VT == MVT::v8i16) {
4210 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4211 if (NewOp.getNode())
4215 if (VT == MVT::v16i8) {
4216 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4217 if (NewOp.getNode())
4221 // Handle all 4 wide cases with a number of shuffles except for MMX.
4222 if (NumElems == 4 && !isMMX)
4223 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4229 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4230 SelectionDAG &DAG) {
4231 MVT VT = Op.getValueType();
4232 DebugLoc dl = Op.getDebugLoc();
4233 if (VT.getSizeInBits() == 8) {
4234 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4235 Op.getOperand(0), Op.getOperand(1));
4236 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4237 DAG.getValueType(VT));
4238 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4239 } else if (VT.getSizeInBits() == 16) {
4240 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4241 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4243 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4244 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4245 DAG.getNode(ISD::BIT_CONVERT, dl,
4249 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4250 Op.getOperand(0), Op.getOperand(1));
4251 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4252 DAG.getValueType(VT));
4253 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4254 } else if (VT == MVT::f32) {
4255 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4256 // the result back to FR32 register. It's only worth matching if the
4257 // result has a single use which is a store or a bitcast to i32. And in
4258 // the case of a store, it's not worth it if the index is a constant 0,
4259 // because a MOVSSmr can be used instead, which is smaller and faster.
4260 if (!Op.hasOneUse())
4262 SDNode *User = *Op.getNode()->use_begin();
4263 if ((User->getOpcode() != ISD::STORE ||
4264 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4265 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4266 (User->getOpcode() != ISD::BIT_CONVERT ||
4267 User->getValueType(0) != MVT::i32))
4269 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4270 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4273 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4274 } else if (VT == MVT::i32) {
4275 // ExtractPS works with constant index.
4276 if (isa<ConstantSDNode>(Op.getOperand(1)))
4284 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4285 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4288 if (Subtarget->hasSSE41()) {
4289 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4294 MVT VT = Op.getValueType();
4295 DebugLoc dl = Op.getDebugLoc();
4296 // TODO: handle v16i8.
4297 if (VT.getSizeInBits() == 16) {
4298 SDValue Vec = Op.getOperand(0);
4299 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4301 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4302 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4303 DAG.getNode(ISD::BIT_CONVERT, dl,
4306 // Transform it so it match pextrw which produces a 32-bit result.
4307 MVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT()+1);
4308 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EVT,
4309 Op.getOperand(0), Op.getOperand(1));
4310 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EVT, Extract,
4311 DAG.getValueType(VT));
4312 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4313 } else if (VT.getSizeInBits() == 32) {
4314 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4318 // SHUFPS the element to the lowest double word, then movss.
4319 int Mask[4] = { Idx, -1, -1, -1 };
4320 MVT VVT = Op.getOperand(0).getValueType();
4321 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4322 DAG.getUNDEF(VVT), Mask);
4323 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4324 DAG.getIntPtrConstant(0));
4325 } else if (VT.getSizeInBits() == 64) {
4326 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4327 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4328 // to match extract_elt for f64.
4329 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4333 // UNPCKHPD the element to the lowest double word, then movsd.
4334 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4335 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4336 int Mask[2] = { 1, -1 };
4337 MVT VVT = Op.getOperand(0).getValueType();
4338 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4339 DAG.getUNDEF(VVT), Mask);
4340 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4341 DAG.getIntPtrConstant(0));
4348 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4349 MVT VT = Op.getValueType();
4350 MVT EVT = VT.getVectorElementType();
4351 DebugLoc dl = Op.getDebugLoc();
4353 SDValue N0 = Op.getOperand(0);
4354 SDValue N1 = Op.getOperand(1);
4355 SDValue N2 = Op.getOperand(2);
4357 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4358 isa<ConstantSDNode>(N2)) {
4359 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4361 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4363 if (N1.getValueType() != MVT::i32)
4364 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4365 if (N2.getValueType() != MVT::i32)
4366 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4367 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4368 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4369 // Bits [7:6] of the constant are the source select. This will always be
4370 // zero here. The DAG Combiner may combine an extract_elt index into these
4371 // bits. For example (insert (extract, 3), 2) could be matched by putting
4372 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4373 // Bits [5:4] of the constant are the destination select. This is the
4374 // value of the incoming immediate.
4375 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4376 // combine either bitwise AND or insert of float 0.0 to set these bits.
4377 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4378 // Create this as a scalar to vector..
4379 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4380 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4381 } else if (EVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4382 // PINSR* works with constant index.
4389 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4390 MVT VT = Op.getValueType();
4391 MVT EVT = VT.getVectorElementType();
4393 if (Subtarget->hasSSE41())
4394 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4399 DebugLoc dl = Op.getDebugLoc();
4400 SDValue N0 = Op.getOperand(0);
4401 SDValue N1 = Op.getOperand(1);
4402 SDValue N2 = Op.getOperand(2);
4404 if (EVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4405 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4406 // as its second argument.
4407 if (N1.getValueType() != MVT::i32)
4408 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4409 if (N2.getValueType() != MVT::i32)
4410 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4411 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4417 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4418 DebugLoc dl = Op.getDebugLoc();
4419 if (Op.getValueType() == MVT::v2f32)
4420 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4421 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4422 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4423 Op.getOperand(0))));
4425 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4426 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4428 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4429 MVT VT = MVT::v2i32;
4430 switch (Op.getValueType().getSimpleVT()) {
4437 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4438 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4441 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4442 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4443 // one of the above mentioned nodes. It has to be wrapped because otherwise
4444 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4445 // be used to form addressing mode. These wrapped nodes will be selected
4448 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4449 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4451 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4453 unsigned char OpFlag = 0;
4454 unsigned WrapperKind = X86ISD::Wrapper;
4456 if (Subtarget->isPICStyleRIPRel() &&
4457 getTargetMachine().getCodeModel() == CodeModel::Small)
4458 WrapperKind = X86ISD::WrapperRIP;
4459 else if (Subtarget->isPICStyleGOT())
4460 OpFlag = X86II::MO_GOTOFF;
4461 else if (Subtarget->isPICStyleStubPIC())
4462 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4464 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4466 CP->getOffset(), OpFlag);
4467 DebugLoc DL = CP->getDebugLoc();
4468 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4469 // With PIC, the address is actually $g + Offset.
4471 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4472 DAG.getNode(X86ISD::GlobalBaseReg,
4473 DebugLoc::getUnknownLoc(), getPointerTy()),
4480 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4481 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4483 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4485 unsigned char OpFlag = 0;
4486 unsigned WrapperKind = X86ISD::Wrapper;
4488 if (Subtarget->isPICStyleRIPRel() &&
4489 getTargetMachine().getCodeModel() == CodeModel::Small)
4490 WrapperKind = X86ISD::WrapperRIP;
4491 else if (Subtarget->isPICStyleGOT())
4492 OpFlag = X86II::MO_GOTOFF;
4493 else if (Subtarget->isPICStyleStubPIC())
4494 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4496 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4498 DebugLoc DL = JT->getDebugLoc();
4499 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4501 // With PIC, the address is actually $g + Offset.
4503 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4504 DAG.getNode(X86ISD::GlobalBaseReg,
4505 DebugLoc::getUnknownLoc(), getPointerTy()),
4513 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4514 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4516 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4518 unsigned char OpFlag = 0;
4519 unsigned WrapperKind = X86ISD::Wrapper;
4520 if (Subtarget->isPICStyleRIPRel() &&
4521 getTargetMachine().getCodeModel() == CodeModel::Small)
4522 WrapperKind = X86ISD::WrapperRIP;
4523 else if (Subtarget->isPICStyleGOT())
4524 OpFlag = X86II::MO_GOTOFF;
4525 else if (Subtarget->isPICStyleStubPIC())
4526 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4528 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4530 DebugLoc DL = Op.getDebugLoc();
4531 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4534 // With PIC, the address is actually $g + Offset.
4535 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4536 !Subtarget->is64Bit()) {
4537 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4538 DAG.getNode(X86ISD::GlobalBaseReg,
4539 DebugLoc::getUnknownLoc(),
4548 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4550 SelectionDAG &DAG) const {
4551 // Create the TargetGlobalAddress node, folding in the constant
4552 // offset if it is legal.
4553 unsigned char OpFlags =
4554 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
4556 if (OpFlags == X86II::MO_NO_FLAG && isInt32(Offset)) {
4557 // A direct static reference to a global.
4558 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4561 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
4564 if (Subtarget->isPICStyleRIPRel() &&
4565 getTargetMachine().getCodeModel() == CodeModel::Small)
4566 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4568 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4570 // With PIC, the address is actually $g + Offset.
4571 if (isGlobalRelativeToPICBase(OpFlags)) {
4572 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4573 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4577 // For globals that require a load from a stub to get the address, emit the
4579 if (isGlobalStubReference(OpFlags))
4580 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4581 PseudoSourceValue::getGOT(), 0);
4583 // If there was a non-zero offset that we didn't fold, create an explicit
4586 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
4587 DAG.getConstant(Offset, getPointerTy()));
4593 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4594 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4595 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4596 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
4600 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
4601 SDValue *InFlag, const MVT PtrVT, unsigned ReturnReg,
4602 unsigned char OperandFlags) {
4603 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4604 DebugLoc dl = GA->getDebugLoc();
4605 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4606 GA->getValueType(0),
4610 SDValue Ops[] = { Chain, TGA, *InFlag };
4611 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
4613 SDValue Ops[] = { Chain, TGA };
4614 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
4616 SDValue Flag = Chain.getValue(1);
4617 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
4620 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4622 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4625 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4626 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
4627 DAG.getNode(X86ISD::GlobalBaseReg,
4628 DebugLoc::getUnknownLoc(),
4630 InFlag = Chain.getValue(1);
4632 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
4635 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4637 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4639 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
4640 X86::RAX, X86II::MO_TLSGD);
4643 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4644 // "local exec" model.
4645 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4646 const MVT PtrVT, TLSModel::Model model,
4648 DebugLoc dl = GA->getDebugLoc();
4649 // Get the Thread Pointer
4650 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
4651 DebugLoc::getUnknownLoc(), PtrVT,
4652 DAG.getRegister(is64Bit? X86::FS : X86::GS,
4655 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
4658 unsigned char OperandFlags = 0;
4659 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
4661 unsigned WrapperKind = X86ISD::Wrapper;
4662 if (model == TLSModel::LocalExec) {
4663 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
4664 } else if (is64Bit) {
4665 assert(model == TLSModel::InitialExec);
4666 OperandFlags = X86II::MO_GOTTPOFF;
4667 WrapperKind = X86ISD::WrapperRIP;
4669 assert(model == TLSModel::InitialExec);
4670 OperandFlags = X86II::MO_INDNTPOFF;
4673 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4675 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
4676 GA->getOffset(), OperandFlags);
4677 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
4679 if (model == TLSModel::InitialExec)
4680 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
4681 PseudoSourceValue::getGOT(), 0);
4683 // The address of the thread local variable is the add of the thread
4684 // pointer with the offset of the variable.
4685 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
4689 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4690 // TODO: implement the "local dynamic" model
4691 // TODO: implement the "initial exec"model for pic executables
4692 assert(Subtarget->isTargetELF() &&
4693 "TLS not implemented for non-ELF targets");
4694 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4695 const GlobalValue *GV = GA->getGlobal();
4697 // If GV is an alias then use the aliasee for determining
4698 // thread-localness.
4699 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
4700 GV = GA->resolveAliasedGlobal(false);
4702 TLSModel::Model model = getTLSModel(GV,
4703 getTargetMachine().getRelocationModel());
4706 case TLSModel::GeneralDynamic:
4707 case TLSModel::LocalDynamic: // not implemented
4708 if (Subtarget->is64Bit())
4709 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4710 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4712 case TLSModel::InitialExec:
4713 case TLSModel::LocalExec:
4714 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
4715 Subtarget->is64Bit());
4718 llvm_unreachable("Unreachable");
4723 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4724 /// take a 2 x i32 value to shift plus a shift amount.
4725 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4726 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4727 MVT VT = Op.getValueType();
4728 unsigned VTBits = VT.getSizeInBits();
4729 DebugLoc dl = Op.getDebugLoc();
4730 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4731 SDValue ShOpLo = Op.getOperand(0);
4732 SDValue ShOpHi = Op.getOperand(1);
4733 SDValue ShAmt = Op.getOperand(2);
4734 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
4735 DAG.getConstant(VTBits - 1, MVT::i8))
4736 : DAG.getConstant(0, VT);
4739 if (Op.getOpcode() == ISD::SHL_PARTS) {
4740 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
4741 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4743 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
4744 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
4747 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
4748 DAG.getConstant(VTBits, MVT::i8));
4749 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
4750 AndNode, DAG.getConstant(0, MVT::i8));
4753 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4754 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4755 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4757 if (Op.getOpcode() == ISD::SHL_PARTS) {
4758 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4759 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4761 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4762 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4765 SDValue Ops[2] = { Lo, Hi };
4766 return DAG.getMergeValues(Ops, 2, dl);
4769 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4770 MVT SrcVT = Op.getOperand(0).getValueType();
4772 if (SrcVT.isVector()) {
4773 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
4779 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4780 "Unknown SINT_TO_FP to lower!");
4782 // These are really Legal; return the operand so the caller accepts it as
4784 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4786 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
4787 Subtarget->is64Bit()) {
4791 DebugLoc dl = Op.getDebugLoc();
4792 unsigned Size = SrcVT.getSizeInBits()/8;
4793 MachineFunction &MF = DAG.getMachineFunction();
4794 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4795 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4796 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
4798 PseudoSourceValue::getFixedStack(SSFI), 0);
4799 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
4802 SDValue X86TargetLowering::BuildFILD(SDValue Op, MVT SrcVT, SDValue Chain,
4804 SelectionDAG &DAG) {
4806 DebugLoc dl = Op.getDebugLoc();
4808 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4810 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4812 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4813 SmallVector<SDValue, 8> Ops;
4814 Ops.push_back(Chain);
4815 Ops.push_back(StackSlot);
4816 Ops.push_back(DAG.getValueType(SrcVT));
4817 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
4818 Tys, &Ops[0], Ops.size());
4821 Chain = Result.getValue(1);
4822 SDValue InFlag = Result.getValue(2);
4824 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4825 // shouldn't be necessary except that RFP cannot be live across
4826 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4827 MachineFunction &MF = DAG.getMachineFunction();
4828 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4829 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4830 Tys = DAG.getVTList(MVT::Other);
4831 SmallVector<SDValue, 8> Ops;
4832 Ops.push_back(Chain);
4833 Ops.push_back(Result);
4834 Ops.push_back(StackSlot);
4835 Ops.push_back(DAG.getValueType(Op.getValueType()));
4836 Ops.push_back(InFlag);
4837 Chain = DAG.getNode(X86ISD::FST, dl, Tys, &Ops[0], Ops.size());
4838 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
4839 PseudoSourceValue::getFixedStack(SSFI), 0);
4845 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
4846 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
4847 // This algorithm is not obvious. Here it is in C code, more or less:
4849 double uint64_to_double( uint32_t hi, uint32_t lo ) {
4850 static const __m128i exp = { 0x4330000045300000ULL, 0 };
4851 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
4853 // Copy ints to xmm registers.
4854 __m128i xh = _mm_cvtsi32_si128( hi );
4855 __m128i xl = _mm_cvtsi32_si128( lo );
4857 // Combine into low half of a single xmm register.
4858 __m128i x = _mm_unpacklo_epi32( xh, xl );
4862 // Merge in appropriate exponents to give the integer bits the right
4864 x = _mm_unpacklo_epi32( x, exp );
4866 // Subtract away the biases to deal with the IEEE-754 double precision
4868 d = _mm_sub_pd( (__m128d) x, bias );
4870 // All conversions up to here are exact. The correctly rounded result is
4871 // calculated using the current rounding mode using the following
4873 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
4874 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
4875 // store doesn't really need to be here (except
4876 // maybe to zero the other double)
4881 DebugLoc dl = Op.getDebugLoc();
4882 LLVMContext *Context = DAG.getContext();
4884 // Build some magic constants.
4885 std::vector<Constant*> CV0;
4886 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
4887 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
4888 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
4889 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
4890 Constant *C0 = ConstantVector::get(CV0);
4891 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
4893 std::vector<Constant*> CV1;
4895 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
4897 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
4898 Constant *C1 = ConstantVector::get(CV1);
4899 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
4901 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4902 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4904 DAG.getIntPtrConstant(1)));
4905 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4906 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4908 DAG.getIntPtrConstant(0)));
4909 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
4910 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
4911 PseudoSourceValue::getConstantPool(), 0,
4913 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
4914 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
4915 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
4916 PseudoSourceValue::getConstantPool(), 0,
4918 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
4920 // Add the halves; easiest way is to swap them into another reg first.
4921 int ShufMask[2] = { 1, -1 };
4922 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
4923 DAG.getUNDEF(MVT::v2f64), ShufMask);
4924 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
4925 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
4926 DAG.getIntPtrConstant(0));
4929 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
4930 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
4931 DebugLoc dl = Op.getDebugLoc();
4932 // FP constant to bias correct the final result.
4933 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
4936 // Load the 32-bit value into an XMM register.
4937 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4938 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4940 DAG.getIntPtrConstant(0)));
4942 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
4943 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
4944 DAG.getIntPtrConstant(0));
4946 // Or the load with the bias.
4947 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
4948 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4949 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4951 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4952 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4953 MVT::v2f64, Bias)));
4954 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
4955 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
4956 DAG.getIntPtrConstant(0));
4958 // Subtract the bias.
4959 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
4961 // Handle final rounding.
4962 MVT DestVT = Op.getValueType();
4964 if (DestVT.bitsLT(MVT::f64)) {
4965 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
4966 DAG.getIntPtrConstant(0));
4967 } else if (DestVT.bitsGT(MVT::f64)) {
4968 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
4971 // Handle final rounding.
4975 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4976 SDValue N0 = Op.getOperand(0);
4977 DebugLoc dl = Op.getDebugLoc();
4979 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
4980 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
4981 // the optimization here.
4982 if (DAG.SignBitIsZero(N0))
4983 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
4985 MVT SrcVT = N0.getValueType();
4986 if (SrcVT == MVT::i64) {
4987 // We only handle SSE2 f64 target here; caller can expand the rest.
4988 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
4991 return LowerUINT_TO_FP_i64(Op, DAG);
4992 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
4993 return LowerUINT_TO_FP_i32(Op, DAG);
4996 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
4998 // Make a 64-bit buffer, and use it to build an FILD.
4999 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5000 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5001 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5002 getPointerTy(), StackSlot, WordOff);
5003 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5004 StackSlot, NULL, 0);
5005 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5006 OffsetSlot, NULL, 0);
5007 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5010 std::pair<SDValue,SDValue> X86TargetLowering::
5011 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5012 DebugLoc dl = Op.getDebugLoc();
5014 MVT DstTy = Op.getValueType();
5017 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5021 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5022 DstTy.getSimpleVT() >= MVT::i16 &&
5023 "Unknown FP_TO_SINT to lower!");
5025 // These are really Legal.
5026 if (DstTy == MVT::i32 &&
5027 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5028 return std::make_pair(SDValue(), SDValue());
5029 if (Subtarget->is64Bit() &&
5030 DstTy == MVT::i64 &&
5031 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5032 return std::make_pair(SDValue(), SDValue());
5034 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5036 MachineFunction &MF = DAG.getMachineFunction();
5037 unsigned MemSize = DstTy.getSizeInBits()/8;
5038 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5039 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5042 switch (DstTy.getSimpleVT()) {
5043 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5044 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5045 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5046 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5049 SDValue Chain = DAG.getEntryNode();
5050 SDValue Value = Op.getOperand(0);
5051 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5052 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5053 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5054 PseudoSourceValue::getFixedStack(SSFI), 0);
5055 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5057 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5059 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5060 Chain = Value.getValue(1);
5061 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5062 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5065 // Build the FP_TO_INT*_IN_MEM
5066 SDValue Ops[] = { Chain, Value, StackSlot };
5067 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5069 return std::make_pair(FIST, StackSlot);
5072 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5073 if (Op.getValueType().isVector()) {
5074 if (Op.getValueType() == MVT::v2i32 &&
5075 Op.getOperand(0).getValueType() == MVT::v2f64) {
5081 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5082 SDValue FIST = Vals.first, StackSlot = Vals.second;
5083 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5084 if (FIST.getNode() == 0) return Op;
5087 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5088 FIST, StackSlot, NULL, 0);
5091 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5092 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5093 SDValue FIST = Vals.first, StackSlot = Vals.second;
5094 assert(FIST.getNode() && "Unexpected failure");
5097 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5098 FIST, StackSlot, NULL, 0);
5101 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5102 LLVMContext *Context = DAG.getContext();
5103 DebugLoc dl = Op.getDebugLoc();
5104 MVT VT = Op.getValueType();
5107 EltVT = VT.getVectorElementType();
5108 std::vector<Constant*> CV;
5109 if (EltVT == MVT::f64) {
5110 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5114 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5120 Constant *C = ConstantVector::get(CV);
5121 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5122 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5123 PseudoSourceValue::getConstantPool(), 0,
5125 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5128 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5129 LLVMContext *Context = DAG.getContext();
5130 DebugLoc dl = Op.getDebugLoc();
5131 MVT VT = Op.getValueType();
5133 unsigned EltNum = 1;
5134 if (VT.isVector()) {
5135 EltVT = VT.getVectorElementType();
5136 EltNum = VT.getVectorNumElements();
5138 std::vector<Constant*> CV;
5139 if (EltVT == MVT::f64) {
5140 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5144 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5150 Constant *C = ConstantVector::get(CV);
5151 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5152 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5153 PseudoSourceValue::getConstantPool(), 0,
5155 if (VT.isVector()) {
5156 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5157 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5158 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5160 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5162 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5166 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5167 LLVMContext *Context = DAG.getContext();
5168 SDValue Op0 = Op.getOperand(0);
5169 SDValue Op1 = Op.getOperand(1);
5170 DebugLoc dl = Op.getDebugLoc();
5171 MVT VT = Op.getValueType();
5172 MVT SrcVT = Op1.getValueType();
5174 // If second operand is smaller, extend it first.
5175 if (SrcVT.bitsLT(VT)) {
5176 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5179 // And if it is bigger, shrink it first.
5180 if (SrcVT.bitsGT(VT)) {
5181 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5185 // At this point the operands and the result should have the same
5186 // type, and that won't be f80 since that is not custom lowered.
5188 // First get the sign bit of second operand.
5189 std::vector<Constant*> CV;
5190 if (SrcVT == MVT::f64) {
5191 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5192 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5194 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5195 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5196 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5197 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5199 Constant *C = ConstantVector::get(CV);
5200 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5201 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5202 PseudoSourceValue::getConstantPool(), 0,
5204 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5206 // Shift sign bit right or left if the two operands have different types.
5207 if (SrcVT.bitsGT(VT)) {
5208 // Op0 is MVT::f32, Op1 is MVT::f64.
5209 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5210 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5211 DAG.getConstant(32, MVT::i32));
5212 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5213 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5214 DAG.getIntPtrConstant(0));
5217 // Clear first operand sign bit.
5219 if (VT == MVT::f64) {
5220 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5221 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5223 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5224 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5225 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5226 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5228 C = ConstantVector::get(CV);
5229 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5230 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5231 PseudoSourceValue::getConstantPool(), 0,
5233 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5235 // Or the value with the sign bit.
5236 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5239 /// Emit nodes that will be selected as "test Op0,Op0", or something
5241 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5242 SelectionDAG &DAG) {
5243 DebugLoc dl = Op.getDebugLoc();
5245 // CF and OF aren't always set the way we want. Determine which
5246 // of these we need.
5247 bool NeedCF = false;
5248 bool NeedOF = false;
5250 case X86::COND_A: case X86::COND_AE:
5251 case X86::COND_B: case X86::COND_BE:
5254 case X86::COND_G: case X86::COND_GE:
5255 case X86::COND_L: case X86::COND_LE:
5256 case X86::COND_O: case X86::COND_NO:
5262 // See if we can use the EFLAGS value from the operand instead of
5263 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5264 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5265 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5266 unsigned Opcode = 0;
5267 unsigned NumOperands = 0;
5268 switch (Op.getNode()->getOpcode()) {
5270 // Due to an isel shortcoming, be conservative if this add is likely to
5271 // be selected as part of a load-modify-store instruction. When the root
5272 // node in a match is a store, isel doesn't know how to remap non-chain
5273 // non-flag uses of other nodes in the match, such as the ADD in this
5274 // case. This leads to the ADD being left around and reselected, with
5275 // the result being two adds in the output.
5276 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5277 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5278 if (UI->getOpcode() == ISD::STORE)
5280 if (ConstantSDNode *C =
5281 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5282 // An add of one will be selected as an INC.
5283 if (C->getAPIntValue() == 1) {
5284 Opcode = X86ISD::INC;
5288 // An add of negative one (subtract of one) will be selected as a DEC.
5289 if (C->getAPIntValue().isAllOnesValue()) {
5290 Opcode = X86ISD::DEC;
5295 // Otherwise use a regular EFLAGS-setting add.
5296 Opcode = X86ISD::ADD;
5300 // Due to the ISEL shortcoming noted above, be conservative if this sub is
5301 // likely to be selected as part of a load-modify-store instruction.
5302 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5303 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5304 if (UI->getOpcode() == ISD::STORE)
5306 // Otherwise use a regular EFLAGS-setting sub.
5307 Opcode = X86ISD::SUB;
5314 return SDValue(Op.getNode(), 1);
5320 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5321 SmallVector<SDValue, 4> Ops;
5322 for (unsigned i = 0; i != NumOperands; ++i)
5323 Ops.push_back(Op.getOperand(i));
5324 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5325 DAG.ReplaceAllUsesWith(Op, New);
5326 return SDValue(New.getNode(), 1);
5330 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5331 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5332 DAG.getConstant(0, Op.getValueType()));
5335 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5337 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5338 SelectionDAG &DAG) {
5339 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5340 if (C->getAPIntValue() == 0)
5341 return EmitTest(Op0, X86CC, DAG);
5343 DebugLoc dl = Op0.getDebugLoc();
5344 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5347 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5348 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5349 SDValue Op0 = Op.getOperand(0);
5350 SDValue Op1 = Op.getOperand(1);
5351 DebugLoc dl = Op.getDebugLoc();
5352 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5354 // Lower (X & (1 << N)) == 0 to BT(X, N).
5355 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5356 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5357 if (Op0.getOpcode() == ISD::AND &&
5359 Op1.getOpcode() == ISD::Constant &&
5360 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5361 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5363 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5364 if (ConstantSDNode *Op010C =
5365 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5366 if (Op010C->getZExtValue() == 1) {
5367 LHS = Op0.getOperand(0);
5368 RHS = Op0.getOperand(1).getOperand(1);
5370 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5371 if (ConstantSDNode *Op000C =
5372 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5373 if (Op000C->getZExtValue() == 1) {
5374 LHS = Op0.getOperand(1);
5375 RHS = Op0.getOperand(0).getOperand(1);
5377 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5378 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5379 SDValue AndLHS = Op0.getOperand(0);
5380 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5381 LHS = AndLHS.getOperand(0);
5382 RHS = AndLHS.getOperand(1);
5386 if (LHS.getNode()) {
5387 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5388 // instruction. Since the shift amount is in-range-or-undefined, we know
5389 // that doing a bittest on the i16 value is ok. We extend to i32 because
5390 // the encoding for the i16 version is larger than the i32 version.
5391 if (LHS.getValueType() == MVT::i8)
5392 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5394 // If the operand types disagree, extend the shift amount to match. Since
5395 // BT ignores high bits (like shifts) we can use anyextend.
5396 if (LHS.getValueType() != RHS.getValueType())
5397 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5399 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5400 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5401 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5402 DAG.getConstant(Cond, MVT::i8), BT);
5406 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5407 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5409 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5410 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5411 DAG.getConstant(X86CC, MVT::i8), Cond);
5414 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5416 SDValue Op0 = Op.getOperand(0);
5417 SDValue Op1 = Op.getOperand(1);
5418 SDValue CC = Op.getOperand(2);
5419 MVT VT = Op.getValueType();
5420 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5421 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5422 DebugLoc dl = Op.getDebugLoc();
5426 MVT VT0 = Op0.getValueType();
5427 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5428 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5431 switch (SetCCOpcode) {
5434 case ISD::SETEQ: SSECC = 0; break;
5436 case ISD::SETGT: Swap = true; // Fallthrough
5438 case ISD::SETOLT: SSECC = 1; break;
5440 case ISD::SETGE: Swap = true; // Fallthrough
5442 case ISD::SETOLE: SSECC = 2; break;
5443 case ISD::SETUO: SSECC = 3; break;
5445 case ISD::SETNE: SSECC = 4; break;
5446 case ISD::SETULE: Swap = true;
5447 case ISD::SETUGE: SSECC = 5; break;
5448 case ISD::SETULT: Swap = true;
5449 case ISD::SETUGT: SSECC = 6; break;
5450 case ISD::SETO: SSECC = 7; break;
5453 std::swap(Op0, Op1);
5455 // In the two special cases we can't handle, emit two comparisons.
5457 if (SetCCOpcode == ISD::SETUEQ) {
5459 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5460 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5461 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5463 else if (SetCCOpcode == ISD::SETONE) {
5465 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5466 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5467 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5469 llvm_unreachable("Illegal FP comparison");
5471 // Handle all other FP comparisons here.
5472 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5475 // We are handling one of the integer comparisons here. Since SSE only has
5476 // GT and EQ comparisons for integer, swapping operands and multiple
5477 // operations may be required for some comparisons.
5478 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5479 bool Swap = false, Invert = false, FlipSigns = false;
5481 switch (VT.getSimpleVT()) {
5484 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5486 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5488 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5489 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5492 switch (SetCCOpcode) {
5494 case ISD::SETNE: Invert = true;
5495 case ISD::SETEQ: Opc = EQOpc; break;
5496 case ISD::SETLT: Swap = true;
5497 case ISD::SETGT: Opc = GTOpc; break;
5498 case ISD::SETGE: Swap = true;
5499 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5500 case ISD::SETULT: Swap = true;
5501 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5502 case ISD::SETUGE: Swap = true;
5503 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5506 std::swap(Op0, Op1);
5508 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5509 // bits of the inputs before performing those operations.
5511 MVT EltVT = VT.getVectorElementType();
5512 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5514 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5515 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5517 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5518 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5521 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5523 // If the logical-not of the result is required, perform that now.
5525 Result = DAG.getNOT(dl, Result, VT);
5530 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5531 static bool isX86LogicalCmp(SDValue Op) {
5532 unsigned Opc = Op.getNode()->getOpcode();
5533 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
5535 if (Op.getResNo() == 1 &&
5536 (Opc == X86ISD::ADD ||
5537 Opc == X86ISD::SUB ||
5538 Opc == X86ISD::SMUL ||
5539 Opc == X86ISD::UMUL ||
5540 Opc == X86ISD::INC ||
5541 Opc == X86ISD::DEC))
5547 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5548 bool addTest = true;
5549 SDValue Cond = Op.getOperand(0);
5550 DebugLoc dl = Op.getDebugLoc();
5553 if (Cond.getOpcode() == ISD::SETCC)
5554 Cond = LowerSETCC(Cond, DAG);
5556 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5557 // setting operand in place of the X86ISD::SETCC.
5558 if (Cond.getOpcode() == X86ISD::SETCC) {
5559 CC = Cond.getOperand(0);
5561 SDValue Cmp = Cond.getOperand(1);
5562 unsigned Opc = Cmp.getOpcode();
5563 MVT VT = Op.getValueType();
5565 bool IllegalFPCMov = false;
5566 if (VT.isFloatingPoint() && !VT.isVector() &&
5567 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5568 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5570 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
5571 Opc == X86ISD::BT) { // FIXME
5578 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5579 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5582 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
5583 SmallVector<SDValue, 4> Ops;
5584 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5585 // condition is true.
5586 Ops.push_back(Op.getOperand(2));
5587 Ops.push_back(Op.getOperand(1));
5589 Ops.push_back(Cond);
5590 return DAG.getNode(X86ISD::CMOV, dl, VTs, &Ops[0], Ops.size());
5593 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
5594 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
5595 // from the AND / OR.
5596 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
5597 Opc = Op.getOpcode();
5598 if (Opc != ISD::OR && Opc != ISD::AND)
5600 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5601 Op.getOperand(0).hasOneUse() &&
5602 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
5603 Op.getOperand(1).hasOneUse());
5606 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
5607 // 1 and that the SETCC node has a single use.
5608 static bool isXor1OfSetCC(SDValue Op) {
5609 if (Op.getOpcode() != ISD::XOR)
5611 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5612 if (N1C && N1C->getAPIntValue() == 1) {
5613 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5614 Op.getOperand(0).hasOneUse();
5619 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5620 bool addTest = true;
5621 SDValue Chain = Op.getOperand(0);
5622 SDValue Cond = Op.getOperand(1);
5623 SDValue Dest = Op.getOperand(2);
5624 DebugLoc dl = Op.getDebugLoc();
5627 if (Cond.getOpcode() == ISD::SETCC)
5628 Cond = LowerSETCC(Cond, DAG);
5630 // FIXME: LowerXALUO doesn't handle these!!
5631 else if (Cond.getOpcode() == X86ISD::ADD ||
5632 Cond.getOpcode() == X86ISD::SUB ||
5633 Cond.getOpcode() == X86ISD::SMUL ||
5634 Cond.getOpcode() == X86ISD::UMUL)
5635 Cond = LowerXALUO(Cond, DAG);
5638 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5639 // setting operand in place of the X86ISD::SETCC.
5640 if (Cond.getOpcode() == X86ISD::SETCC) {
5641 CC = Cond.getOperand(0);
5643 SDValue Cmp = Cond.getOperand(1);
5644 unsigned Opc = Cmp.getOpcode();
5645 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
5646 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
5650 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
5654 // These can only come from an arithmetic instruction with overflow,
5655 // e.g. SADDO, UADDO.
5656 Cond = Cond.getNode()->getOperand(1);
5663 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
5664 SDValue Cmp = Cond.getOperand(0).getOperand(1);
5665 if (CondOpc == ISD::OR) {
5666 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
5667 // two branches instead of an explicit OR instruction with a
5669 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5670 isX86LogicalCmp(Cmp)) {
5671 CC = Cond.getOperand(0).getOperand(0);
5672 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5673 Chain, Dest, CC, Cmp);
5674 CC = Cond.getOperand(1).getOperand(0);
5678 } else { // ISD::AND
5679 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
5680 // two branches instead of an explicit AND instruction with a
5681 // separate test. However, we only do this if this block doesn't
5682 // have a fall-through edge, because this requires an explicit
5683 // jmp when the condition is false.
5684 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5685 isX86LogicalCmp(Cmp) &&
5686 Op.getNode()->hasOneUse()) {
5687 X86::CondCode CCode =
5688 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5689 CCode = X86::GetOppositeBranchCondition(CCode);
5690 CC = DAG.getConstant(CCode, MVT::i8);
5691 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
5692 // Look for an unconditional branch following this conditional branch.
5693 // We need this because we need to reverse the successors in order
5694 // to implement FCMP_OEQ.
5695 if (User.getOpcode() == ISD::BR) {
5696 SDValue FalseBB = User.getOperand(1);
5698 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
5699 assert(NewBR == User);
5702 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5703 Chain, Dest, CC, Cmp);
5704 X86::CondCode CCode =
5705 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
5706 CCode = X86::GetOppositeBranchCondition(CCode);
5707 CC = DAG.getConstant(CCode, MVT::i8);
5713 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
5714 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
5715 // It should be transformed during dag combiner except when the condition
5716 // is set by a arithmetics with overflow node.
5717 X86::CondCode CCode =
5718 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5719 CCode = X86::GetOppositeBranchCondition(CCode);
5720 CC = DAG.getConstant(CCode, MVT::i8);
5721 Cond = Cond.getOperand(0).getOperand(1);
5727 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5728 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5730 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5731 Chain, Dest, CC, Cond);
5735 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5736 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5737 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5738 // that the guard pages used by the OS virtual memory manager are allocated in
5739 // correct sequence.
5741 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5742 SelectionDAG &DAG) {
5743 assert(Subtarget->isTargetCygMing() &&
5744 "This should be used only on Cygwin/Mingw targets");
5745 DebugLoc dl = Op.getDebugLoc();
5748 SDValue Chain = Op.getOperand(0);
5749 SDValue Size = Op.getOperand(1);
5750 // FIXME: Ensure alignment here
5754 MVT IntPtr = getPointerTy();
5755 MVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5757 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
5759 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
5760 Flag = Chain.getValue(1);
5762 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5763 SDValue Ops[] = { Chain,
5764 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5765 DAG.getRegister(X86::EAX, IntPtr),
5766 DAG.getRegister(X86StackPtr, SPTy),
5768 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
5769 Flag = Chain.getValue(1);
5771 Chain = DAG.getCALLSEQ_END(Chain,
5772 DAG.getIntPtrConstant(0, true),
5773 DAG.getIntPtrConstant(0, true),
5776 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
5778 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5779 return DAG.getMergeValues(Ops1, 2, dl);
5783 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
5785 SDValue Dst, SDValue Src,
5786 SDValue Size, unsigned Align,
5788 uint64_t DstSVOff) {
5789 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5791 // If not DWORD aligned or size is more than the threshold, call the library.
5792 // The libc version is likely to be faster for these cases. It can use the
5793 // address value and run time information about the CPU.
5794 if ((Align & 3) != 0 ||
5796 ConstantSize->getZExtValue() >
5797 getSubtarget()->getMaxInlineSizeThreshold()) {
5798 SDValue InFlag(0, 0);
5800 // Check to see if there is a specialized entry-point for memory zeroing.
5801 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5803 if (const char *bzeroEntry = V &&
5804 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5805 MVT IntPtr = getPointerTy();
5806 const Type *IntPtrTy = TD->getIntPtrType();
5807 TargetLowering::ArgListTy Args;
5808 TargetLowering::ArgListEntry Entry;
5810 Entry.Ty = IntPtrTy;
5811 Args.push_back(Entry);
5813 Args.push_back(Entry);
5814 std::pair<SDValue,SDValue> CallResult =
5815 LowerCallTo(Chain, Type::VoidTy, false, false, false, false,
5816 0, CallingConv::C, false,
5817 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
5818 return CallResult.second;
5821 // Otherwise have the target-independent code call memset.
5825 uint64_t SizeVal = ConstantSize->getZExtValue();
5826 SDValue InFlag(0, 0);
5829 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5830 unsigned BytesLeft = 0;
5831 bool TwoRepStos = false;
5834 uint64_t Val = ValC->getZExtValue() & 255;
5836 // If the value is a constant, then we can potentially use larger sets.
5837 switch (Align & 3) {
5838 case 2: // WORD aligned
5841 Val = (Val << 8) | Val;
5843 case 0: // DWORD aligned
5846 Val = (Val << 8) | Val;
5847 Val = (Val << 16) | Val;
5848 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
5851 Val = (Val << 32) | Val;
5854 default: // Byte aligned
5857 Count = DAG.getIntPtrConstant(SizeVal);
5861 if (AVT.bitsGT(MVT::i8)) {
5862 unsigned UBytes = AVT.getSizeInBits() / 8;
5863 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
5864 BytesLeft = SizeVal % UBytes;
5867 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
5869 InFlag = Chain.getValue(1);
5872 Count = DAG.getIntPtrConstant(SizeVal);
5873 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
5874 InFlag = Chain.getValue(1);
5877 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5880 InFlag = Chain.getValue(1);
5881 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5884 InFlag = Chain.getValue(1);
5886 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5887 SmallVector<SDValue, 8> Ops;
5888 Ops.push_back(Chain);
5889 Ops.push_back(DAG.getValueType(AVT));
5890 Ops.push_back(InFlag);
5891 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5894 InFlag = Chain.getValue(1);
5896 MVT CVT = Count.getValueType();
5897 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
5898 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
5899 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
5902 InFlag = Chain.getValue(1);
5903 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5905 Ops.push_back(Chain);
5906 Ops.push_back(DAG.getValueType(MVT::i8));
5907 Ops.push_back(InFlag);
5908 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5909 } else if (BytesLeft) {
5910 // Handle the last 1 - 7 bytes.
5911 unsigned Offset = SizeVal - BytesLeft;
5912 MVT AddrVT = Dst.getValueType();
5913 MVT SizeVT = Size.getValueType();
5915 Chain = DAG.getMemset(Chain, dl,
5916 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
5917 DAG.getConstant(Offset, AddrVT)),
5919 DAG.getConstant(BytesLeft, SizeVT),
5920 Align, DstSV, DstSVOff + Offset);
5923 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
5928 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
5929 SDValue Chain, SDValue Dst, SDValue Src,
5930 SDValue Size, unsigned Align,
5932 const Value *DstSV, uint64_t DstSVOff,
5933 const Value *SrcSV, uint64_t SrcSVOff) {
5934 // This requires the copy size to be a constant, preferrably
5935 // within a subtarget-specific limit.
5936 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5939 uint64_t SizeVal = ConstantSize->getZExtValue();
5940 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
5943 /// If not DWORD aligned, call the library.
5944 if ((Align & 3) != 0)
5949 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
5952 unsigned UBytes = AVT.getSizeInBits() / 8;
5953 unsigned CountVal = SizeVal / UBytes;
5954 SDValue Count = DAG.getIntPtrConstant(CountVal);
5955 unsigned BytesLeft = SizeVal % UBytes;
5957 SDValue InFlag(0, 0);
5958 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5961 InFlag = Chain.getValue(1);
5962 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5965 InFlag = Chain.getValue(1);
5966 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
5969 InFlag = Chain.getValue(1);
5971 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5972 SmallVector<SDValue, 8> Ops;
5973 Ops.push_back(Chain);
5974 Ops.push_back(DAG.getValueType(AVT));
5975 Ops.push_back(InFlag);
5976 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, &Ops[0], Ops.size());
5978 SmallVector<SDValue, 4> Results;
5979 Results.push_back(RepMovs);
5981 // Handle the last 1 - 7 bytes.
5982 unsigned Offset = SizeVal - BytesLeft;
5983 MVT DstVT = Dst.getValueType();
5984 MVT SrcVT = Src.getValueType();
5985 MVT SizeVT = Size.getValueType();
5986 Results.push_back(DAG.getMemcpy(Chain, dl,
5987 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
5988 DAG.getConstant(Offset, DstVT)),
5989 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
5990 DAG.getConstant(Offset, SrcVT)),
5991 DAG.getConstant(BytesLeft, SizeVT),
5992 Align, AlwaysInline,
5993 DstSV, DstSVOff + Offset,
5994 SrcSV, SrcSVOff + Offset));
5997 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
5998 &Results[0], Results.size());
6001 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6002 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6003 DebugLoc dl = Op.getDebugLoc();
6005 if (!Subtarget->is64Bit()) {
6006 // vastart just stores the address of the VarArgsFrameIndex slot into the
6007 // memory location argument.
6008 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6009 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6013 // gp_offset (0 - 6 * 8)
6014 // fp_offset (48 - 48 + 8 * 16)
6015 // overflow_arg_area (point to parameters coming in memory).
6017 SmallVector<SDValue, 8> MemOps;
6018 SDValue FIN = Op.getOperand(1);
6020 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6021 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6023 MemOps.push_back(Store);
6026 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6027 FIN, DAG.getIntPtrConstant(4));
6028 Store = DAG.getStore(Op.getOperand(0), dl,
6029 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6031 MemOps.push_back(Store);
6033 // Store ptr to overflow_arg_area
6034 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6035 FIN, DAG.getIntPtrConstant(4));
6036 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6037 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6038 MemOps.push_back(Store);
6040 // Store ptr to reg_save_area.
6041 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6042 FIN, DAG.getIntPtrConstant(8));
6043 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6044 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6045 MemOps.push_back(Store);
6046 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6047 &MemOps[0], MemOps.size());
6050 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6051 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6052 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6053 SDValue Chain = Op.getOperand(0);
6054 SDValue SrcPtr = Op.getOperand(1);
6055 SDValue SrcSV = Op.getOperand(2);
6057 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6061 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6062 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6063 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6064 SDValue Chain = Op.getOperand(0);
6065 SDValue DstPtr = Op.getOperand(1);
6066 SDValue SrcPtr = Op.getOperand(2);
6067 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6068 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6069 DebugLoc dl = Op.getDebugLoc();
6071 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6072 DAG.getIntPtrConstant(24), 8, false,
6073 DstSV, 0, SrcSV, 0);
6077 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6078 DebugLoc dl = Op.getDebugLoc();
6079 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6081 default: return SDValue(); // Don't custom lower most intrinsics.
6082 // Comparison intrinsics.
6083 case Intrinsic::x86_sse_comieq_ss:
6084 case Intrinsic::x86_sse_comilt_ss:
6085 case Intrinsic::x86_sse_comile_ss:
6086 case Intrinsic::x86_sse_comigt_ss:
6087 case Intrinsic::x86_sse_comige_ss:
6088 case Intrinsic::x86_sse_comineq_ss:
6089 case Intrinsic::x86_sse_ucomieq_ss:
6090 case Intrinsic::x86_sse_ucomilt_ss:
6091 case Intrinsic::x86_sse_ucomile_ss:
6092 case Intrinsic::x86_sse_ucomigt_ss:
6093 case Intrinsic::x86_sse_ucomige_ss:
6094 case Intrinsic::x86_sse_ucomineq_ss:
6095 case Intrinsic::x86_sse2_comieq_sd:
6096 case Intrinsic::x86_sse2_comilt_sd:
6097 case Intrinsic::x86_sse2_comile_sd:
6098 case Intrinsic::x86_sse2_comigt_sd:
6099 case Intrinsic::x86_sse2_comige_sd:
6100 case Intrinsic::x86_sse2_comineq_sd:
6101 case Intrinsic::x86_sse2_ucomieq_sd:
6102 case Intrinsic::x86_sse2_ucomilt_sd:
6103 case Intrinsic::x86_sse2_ucomile_sd:
6104 case Intrinsic::x86_sse2_ucomigt_sd:
6105 case Intrinsic::x86_sse2_ucomige_sd:
6106 case Intrinsic::x86_sse2_ucomineq_sd: {
6108 ISD::CondCode CC = ISD::SETCC_INVALID;
6111 case Intrinsic::x86_sse_comieq_ss:
6112 case Intrinsic::x86_sse2_comieq_sd:
6116 case Intrinsic::x86_sse_comilt_ss:
6117 case Intrinsic::x86_sse2_comilt_sd:
6121 case Intrinsic::x86_sse_comile_ss:
6122 case Intrinsic::x86_sse2_comile_sd:
6126 case Intrinsic::x86_sse_comigt_ss:
6127 case Intrinsic::x86_sse2_comigt_sd:
6131 case Intrinsic::x86_sse_comige_ss:
6132 case Intrinsic::x86_sse2_comige_sd:
6136 case Intrinsic::x86_sse_comineq_ss:
6137 case Intrinsic::x86_sse2_comineq_sd:
6141 case Intrinsic::x86_sse_ucomieq_ss:
6142 case Intrinsic::x86_sse2_ucomieq_sd:
6143 Opc = X86ISD::UCOMI;
6146 case Intrinsic::x86_sse_ucomilt_ss:
6147 case Intrinsic::x86_sse2_ucomilt_sd:
6148 Opc = X86ISD::UCOMI;
6151 case Intrinsic::x86_sse_ucomile_ss:
6152 case Intrinsic::x86_sse2_ucomile_sd:
6153 Opc = X86ISD::UCOMI;
6156 case Intrinsic::x86_sse_ucomigt_ss:
6157 case Intrinsic::x86_sse2_ucomigt_sd:
6158 Opc = X86ISD::UCOMI;
6161 case Intrinsic::x86_sse_ucomige_ss:
6162 case Intrinsic::x86_sse2_ucomige_sd:
6163 Opc = X86ISD::UCOMI;
6166 case Intrinsic::x86_sse_ucomineq_ss:
6167 case Intrinsic::x86_sse2_ucomineq_sd:
6168 Opc = X86ISD::UCOMI;
6173 SDValue LHS = Op.getOperand(1);
6174 SDValue RHS = Op.getOperand(2);
6175 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6176 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6177 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6178 DAG.getConstant(X86CC, MVT::i8), Cond);
6179 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6181 // ptest intrinsics. The intrinsic these come from are designed to return
6182 // an integer value, not just an instruction so lower it to the ptest
6183 // pattern and a setcc for the result.
6184 case Intrinsic::x86_sse41_ptestz:
6185 case Intrinsic::x86_sse41_ptestc:
6186 case Intrinsic::x86_sse41_ptestnzc:{
6189 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6190 case Intrinsic::x86_sse41_ptestz:
6192 X86CC = X86::COND_E;
6194 case Intrinsic::x86_sse41_ptestc:
6196 X86CC = X86::COND_B;
6198 case Intrinsic::x86_sse41_ptestnzc:
6200 X86CC = X86::COND_A;
6204 SDValue LHS = Op.getOperand(1);
6205 SDValue RHS = Op.getOperand(2);
6206 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6207 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6208 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6209 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6212 // Fix vector shift instructions where the last operand is a non-immediate
6214 case Intrinsic::x86_sse2_pslli_w:
6215 case Intrinsic::x86_sse2_pslli_d:
6216 case Intrinsic::x86_sse2_pslli_q:
6217 case Intrinsic::x86_sse2_psrli_w:
6218 case Intrinsic::x86_sse2_psrli_d:
6219 case Intrinsic::x86_sse2_psrli_q:
6220 case Intrinsic::x86_sse2_psrai_w:
6221 case Intrinsic::x86_sse2_psrai_d:
6222 case Intrinsic::x86_mmx_pslli_w:
6223 case Intrinsic::x86_mmx_pslli_d:
6224 case Intrinsic::x86_mmx_pslli_q:
6225 case Intrinsic::x86_mmx_psrli_w:
6226 case Intrinsic::x86_mmx_psrli_d:
6227 case Intrinsic::x86_mmx_psrli_q:
6228 case Intrinsic::x86_mmx_psrai_w:
6229 case Intrinsic::x86_mmx_psrai_d: {
6230 SDValue ShAmt = Op.getOperand(2);
6231 if (isa<ConstantSDNode>(ShAmt))
6234 unsigned NewIntNo = 0;
6235 MVT ShAmtVT = MVT::v4i32;
6237 case Intrinsic::x86_sse2_pslli_w:
6238 NewIntNo = Intrinsic::x86_sse2_psll_w;
6240 case Intrinsic::x86_sse2_pslli_d:
6241 NewIntNo = Intrinsic::x86_sse2_psll_d;
6243 case Intrinsic::x86_sse2_pslli_q:
6244 NewIntNo = Intrinsic::x86_sse2_psll_q;
6246 case Intrinsic::x86_sse2_psrli_w:
6247 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6249 case Intrinsic::x86_sse2_psrli_d:
6250 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6252 case Intrinsic::x86_sse2_psrli_q:
6253 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6255 case Intrinsic::x86_sse2_psrai_w:
6256 NewIntNo = Intrinsic::x86_sse2_psra_w;
6258 case Intrinsic::x86_sse2_psrai_d:
6259 NewIntNo = Intrinsic::x86_sse2_psra_d;
6262 ShAmtVT = MVT::v2i32;
6264 case Intrinsic::x86_mmx_pslli_w:
6265 NewIntNo = Intrinsic::x86_mmx_psll_w;
6267 case Intrinsic::x86_mmx_pslli_d:
6268 NewIntNo = Intrinsic::x86_mmx_psll_d;
6270 case Intrinsic::x86_mmx_pslli_q:
6271 NewIntNo = Intrinsic::x86_mmx_psll_q;
6273 case Intrinsic::x86_mmx_psrli_w:
6274 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6276 case Intrinsic::x86_mmx_psrli_d:
6277 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6279 case Intrinsic::x86_mmx_psrli_q:
6280 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6282 case Intrinsic::x86_mmx_psrai_w:
6283 NewIntNo = Intrinsic::x86_mmx_psra_w;
6285 case Intrinsic::x86_mmx_psrai_d:
6286 NewIntNo = Intrinsic::x86_mmx_psra_d;
6288 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6293 MVT VT = Op.getValueType();
6294 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6295 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, ShAmtVT, ShAmt));
6296 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6297 DAG.getConstant(NewIntNo, MVT::i32),
6298 Op.getOperand(1), ShAmt);
6303 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6304 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6305 DebugLoc dl = Op.getDebugLoc();
6308 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6310 DAG.getConstant(TD->getPointerSize(),
6311 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6312 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6313 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6318 // Just load the return address.
6319 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6320 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6321 RetAddrFI, NULL, 0);
6324 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6325 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6326 MFI->setFrameAddressIsTaken(true);
6327 MVT VT = Op.getValueType();
6328 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6329 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6330 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6331 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6333 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6337 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6338 SelectionDAG &DAG) {
6339 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6342 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6344 MachineFunction &MF = DAG.getMachineFunction();
6345 SDValue Chain = Op.getOperand(0);
6346 SDValue Offset = Op.getOperand(1);
6347 SDValue Handler = Op.getOperand(2);
6348 DebugLoc dl = Op.getDebugLoc();
6350 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6352 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6354 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6355 DAG.getIntPtrConstant(-TD->getPointerSize()));
6356 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6357 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6358 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6359 MF.getRegInfo().addLiveOut(StoreAddrReg);
6361 return DAG.getNode(X86ISD::EH_RETURN, dl,
6363 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6366 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6367 SelectionDAG &DAG) {
6368 SDValue Root = Op.getOperand(0);
6369 SDValue Trmp = Op.getOperand(1); // trampoline
6370 SDValue FPtr = Op.getOperand(2); // nested function
6371 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6372 DebugLoc dl = Op.getDebugLoc();
6374 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6376 const X86InstrInfo *TII =
6377 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6379 if (Subtarget->is64Bit()) {
6380 SDValue OutChains[6];
6382 // Large code-model.
6384 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6385 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6387 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6388 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6390 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6392 // Load the pointer to the nested function into R11.
6393 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6394 SDValue Addr = Trmp;
6395 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6398 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6399 DAG.getConstant(2, MVT::i64));
6400 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6402 // Load the 'nest' parameter value into R10.
6403 // R10 is specified in X86CallingConv.td
6404 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6405 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6406 DAG.getConstant(10, MVT::i64));
6407 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6408 Addr, TrmpAddr, 10);
6410 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6411 DAG.getConstant(12, MVT::i64));
6412 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6414 // Jump to the nested function.
6415 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6416 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6417 DAG.getConstant(20, MVT::i64));
6418 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6419 Addr, TrmpAddr, 20);
6421 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6422 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6423 DAG.getConstant(22, MVT::i64));
6424 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6428 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6429 return DAG.getMergeValues(Ops, 2, dl);
6431 const Function *Func =
6432 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6433 unsigned CC = Func->getCallingConv();
6438 llvm_unreachable("Unsupported calling convention");
6439 case CallingConv::C:
6440 case CallingConv::X86_StdCall: {
6441 // Pass 'nest' parameter in ECX.
6442 // Must be kept in sync with X86CallingConv.td
6445 // Check that ECX wasn't needed by an 'inreg' parameter.
6446 const FunctionType *FTy = Func->getFunctionType();
6447 const AttrListPtr &Attrs = Func->getAttributes();
6449 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6450 unsigned InRegCount = 0;
6453 for (FunctionType::param_iterator I = FTy->param_begin(),
6454 E = FTy->param_end(); I != E; ++I, ++Idx)
6455 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6456 // FIXME: should only count parameters that are lowered to integers.
6457 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6459 if (InRegCount > 2) {
6460 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
6465 case CallingConv::X86_FastCall:
6466 case CallingConv::Fast:
6467 // Pass 'nest' parameter in EAX.
6468 // Must be kept in sync with X86CallingConv.td
6473 SDValue OutChains[4];
6476 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6477 DAG.getConstant(10, MVT::i32));
6478 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
6480 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6481 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6482 OutChains[0] = DAG.getStore(Root, dl,
6483 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6486 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6487 DAG.getConstant(1, MVT::i32));
6488 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
6490 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6491 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6492 DAG.getConstant(5, MVT::i32));
6493 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
6494 TrmpAddr, 5, false, 1);
6496 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6497 DAG.getConstant(6, MVT::i32));
6498 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
6501 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
6502 return DAG.getMergeValues(Ops, 2, dl);
6506 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6508 The rounding mode is in bits 11:10 of FPSR, and has the following
6515 FLT_ROUNDS, on the other hand, expects the following:
6522 To perform the conversion, we do:
6523 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6526 MachineFunction &MF = DAG.getMachineFunction();
6527 const TargetMachine &TM = MF.getTarget();
6528 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6529 unsigned StackAlignment = TFI.getStackAlignment();
6530 MVT VT = Op.getValueType();
6531 DebugLoc dl = Op.getDebugLoc();
6533 // Save FP Control Word to stack slot
6534 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
6535 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6537 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
6538 DAG.getEntryNode(), StackSlot);
6540 // Load FP Control Word from stack slot
6541 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
6543 // Transform as necessary
6545 DAG.getNode(ISD::SRL, dl, MVT::i16,
6546 DAG.getNode(ISD::AND, dl, MVT::i16,
6547 CWD, DAG.getConstant(0x800, MVT::i16)),
6548 DAG.getConstant(11, MVT::i8));
6550 DAG.getNode(ISD::SRL, dl, MVT::i16,
6551 DAG.getNode(ISD::AND, dl, MVT::i16,
6552 CWD, DAG.getConstant(0x400, MVT::i16)),
6553 DAG.getConstant(9, MVT::i8));
6556 DAG.getNode(ISD::AND, dl, MVT::i16,
6557 DAG.getNode(ISD::ADD, dl, MVT::i16,
6558 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
6559 DAG.getConstant(1, MVT::i16)),
6560 DAG.getConstant(3, MVT::i16));
6563 return DAG.getNode((VT.getSizeInBits() < 16 ?
6564 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
6567 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
6568 MVT VT = Op.getValueType();
6570 unsigned NumBits = VT.getSizeInBits();
6571 DebugLoc dl = Op.getDebugLoc();
6573 Op = Op.getOperand(0);
6574 if (VT == MVT::i8) {
6575 // Zero extend to i32 since there is not an i8 bsr.
6577 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6580 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
6581 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6582 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
6584 // If src is zero (i.e. bsr sets ZF), returns NumBits.
6585 SmallVector<SDValue, 4> Ops;
6587 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
6588 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6589 Ops.push_back(Op.getValue(1));
6590 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6592 // Finally xor with NumBits-1.
6593 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
6596 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6600 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
6601 MVT VT = Op.getValueType();
6603 unsigned NumBits = VT.getSizeInBits();
6604 DebugLoc dl = Op.getDebugLoc();
6606 Op = Op.getOperand(0);
6607 if (VT == MVT::i8) {
6609 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6612 // Issue a bsf (scan bits forward) which also sets EFLAGS.
6613 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6614 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
6616 // If src is zero (i.e. bsf sets ZF), returns NumBits.
6617 SmallVector<SDValue, 4> Ops;
6619 Ops.push_back(DAG.getConstant(NumBits, OpVT));
6620 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6621 Ops.push_back(Op.getValue(1));
6622 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6625 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6629 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
6630 MVT VT = Op.getValueType();
6631 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
6632 DebugLoc dl = Op.getDebugLoc();
6634 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
6635 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
6636 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
6637 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
6638 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
6640 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
6641 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
6642 // return AloBlo + AloBhi + AhiBlo;
6644 SDValue A = Op.getOperand(0);
6645 SDValue B = Op.getOperand(1);
6647 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6648 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6649 A, DAG.getConstant(32, MVT::i32));
6650 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6651 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6652 B, DAG.getConstant(32, MVT::i32));
6653 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6654 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6656 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6657 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6659 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6660 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6662 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6663 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6664 AloBhi, DAG.getConstant(32, MVT::i32));
6665 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6666 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6667 AhiBlo, DAG.getConstant(32, MVT::i32));
6668 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
6669 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
6674 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
6675 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
6676 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
6677 // looks for this combo and may remove the "setcc" instruction if the "setcc"
6678 // has only one use.
6679 SDNode *N = Op.getNode();
6680 SDValue LHS = N->getOperand(0);
6681 SDValue RHS = N->getOperand(1);
6682 unsigned BaseOp = 0;
6684 DebugLoc dl = Op.getDebugLoc();
6686 switch (Op.getOpcode()) {
6687 default: llvm_unreachable("Unknown ovf instruction!");
6689 // A subtract of one will be selected as a INC. Note that INC doesn't
6690 // set CF, so we can't do this for UADDO.
6691 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6692 if (C->getAPIntValue() == 1) {
6693 BaseOp = X86ISD::INC;
6697 BaseOp = X86ISD::ADD;
6701 BaseOp = X86ISD::ADD;
6705 // A subtract of one will be selected as a DEC. Note that DEC doesn't
6706 // set CF, so we can't do this for USUBO.
6707 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6708 if (C->getAPIntValue() == 1) {
6709 BaseOp = X86ISD::DEC;
6713 BaseOp = X86ISD::SUB;
6717 BaseOp = X86ISD::SUB;
6721 BaseOp = X86ISD::SMUL;
6725 BaseOp = X86ISD::UMUL;
6730 // Also sets EFLAGS.
6731 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
6732 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
6735 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
6736 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
6738 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
6742 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
6743 MVT T = Op.getValueType();
6744 DebugLoc dl = Op.getDebugLoc();
6747 switch(T.getSimpleVT()) {
6749 assert(false && "Invalid value type!");
6750 case MVT::i8: Reg = X86::AL; size = 1; break;
6751 case MVT::i16: Reg = X86::AX; size = 2; break;
6752 case MVT::i32: Reg = X86::EAX; size = 4; break;
6754 assert(Subtarget->is64Bit() && "Node not type legal!");
6755 Reg = X86::RAX; size = 8;
6758 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
6759 Op.getOperand(2), SDValue());
6760 SDValue Ops[] = { cpIn.getValue(0),
6763 DAG.getTargetConstant(size, MVT::i8),
6765 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6766 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
6768 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
6772 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
6773 SelectionDAG &DAG) {
6774 assert(Subtarget->is64Bit() && "Result not type legalized?");
6775 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6776 SDValue TheChain = Op.getOperand(0);
6777 DebugLoc dl = Op.getDebugLoc();
6778 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6779 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
6780 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
6782 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
6783 DAG.getConstant(32, MVT::i8));
6785 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
6788 return DAG.getMergeValues(Ops, 2, dl);
6791 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
6792 SDNode *Node = Op.getNode();
6793 DebugLoc dl = Node->getDebugLoc();
6794 MVT T = Node->getValueType(0);
6795 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
6796 DAG.getConstant(0, T), Node->getOperand(2));
6797 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
6798 cast<AtomicSDNode>(Node)->getMemoryVT(),
6799 Node->getOperand(0),
6800 Node->getOperand(1), negOp,
6801 cast<AtomicSDNode>(Node)->getSrcValue(),
6802 cast<AtomicSDNode>(Node)->getAlignment());
6805 /// LowerOperation - Provide custom lowering hooks for some operations.
6807 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
6808 switch (Op.getOpcode()) {
6809 default: llvm_unreachable("Should not custom lower this!");
6810 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
6811 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
6812 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
6813 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6814 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6815 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6816 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
6817 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6818 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
6819 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6820 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
6821 case ISD::SHL_PARTS:
6822 case ISD::SRA_PARTS:
6823 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
6824 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
6825 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
6826 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
6827 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
6828 case ISD::FABS: return LowerFABS(Op, DAG);
6829 case ISD::FNEG: return LowerFNEG(Op, DAG);
6830 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6831 case ISD::SETCC: return LowerSETCC(Op, DAG);
6832 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
6833 case ISD::SELECT: return LowerSELECT(Op, DAG);
6834 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
6835 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
6836 case ISD::CALL: return LowerCALL(Op, DAG);
6837 case ISD::RET: return LowerRET(Op, DAG);
6838 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
6839 case ISD::VASTART: return LowerVASTART(Op, DAG);
6840 case ISD::VAARG: return LowerVAARG(Op, DAG);
6841 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
6842 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
6843 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6844 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6845 case ISD::FRAME_TO_ARGS_OFFSET:
6846 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6847 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6848 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6849 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6850 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6851 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6852 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6853 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
6859 case ISD::UMULO: return LowerXALUO(Op, DAG);
6860 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
6864 void X86TargetLowering::
6865 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
6866 SelectionDAG &DAG, unsigned NewOp) {
6867 MVT T = Node->getValueType(0);
6868 DebugLoc dl = Node->getDebugLoc();
6869 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
6871 SDValue Chain = Node->getOperand(0);
6872 SDValue In1 = Node->getOperand(1);
6873 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6874 Node->getOperand(2), DAG.getIntPtrConstant(0));
6875 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6876 Node->getOperand(2), DAG.getIntPtrConstant(1));
6877 // This is a generalized SDNode, not an AtomicSDNode, so it doesn't
6878 // have a MemOperand. Pass the info through as a normal operand.
6879 SDValue LSI = DAG.getMemOperand(cast<MemSDNode>(Node)->getMemOperand());
6880 SDValue Ops[] = { Chain, In1, In2L, In2H, LSI };
6881 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
6882 SDValue Result = DAG.getNode(NewOp, dl, Tys, Ops, 5);
6883 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
6884 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6885 Results.push_back(Result.getValue(2));
6888 /// ReplaceNodeResults - Replace a node with an illegal result type
6889 /// with a new node built out of custom code.
6890 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
6891 SmallVectorImpl<SDValue>&Results,
6892 SelectionDAG &DAG) {
6893 DebugLoc dl = N->getDebugLoc();
6894 switch (N->getOpcode()) {
6896 assert(false && "Do not know how to custom type legalize this operation!");
6898 case ISD::FP_TO_SINT: {
6899 std::pair<SDValue,SDValue> Vals =
6900 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
6901 SDValue FIST = Vals.first, StackSlot = Vals.second;
6902 if (FIST.getNode() != 0) {
6903 MVT VT = N->getValueType(0);
6904 // Return a load from the stack slot.
6905 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
6909 case ISD::READCYCLECOUNTER: {
6910 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6911 SDValue TheChain = N->getOperand(0);
6912 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6913 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
6915 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
6917 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
6918 SDValue Ops[] = { eax, edx };
6919 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
6920 Results.push_back(edx.getValue(1));
6923 case ISD::ATOMIC_CMP_SWAP: {
6924 MVT T = N->getValueType(0);
6925 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
6926 SDValue cpInL, cpInH;
6927 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
6928 DAG.getConstant(0, MVT::i32));
6929 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
6930 DAG.getConstant(1, MVT::i32));
6931 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
6932 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
6934 SDValue swapInL, swapInH;
6935 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
6936 DAG.getConstant(0, MVT::i32));
6937 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
6938 DAG.getConstant(1, MVT::i32));
6939 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
6941 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
6942 swapInL.getValue(1));
6943 SDValue Ops[] = { swapInH.getValue(0),
6945 swapInH.getValue(1) };
6946 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6947 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
6948 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
6949 MVT::i32, Result.getValue(1));
6950 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
6951 MVT::i32, cpOutL.getValue(2));
6952 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
6953 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6954 Results.push_back(cpOutH.getValue(1));
6957 case ISD::ATOMIC_LOAD_ADD:
6958 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
6960 case ISD::ATOMIC_LOAD_AND:
6961 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
6963 case ISD::ATOMIC_LOAD_NAND:
6964 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
6966 case ISD::ATOMIC_LOAD_OR:
6967 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
6969 case ISD::ATOMIC_LOAD_SUB:
6970 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
6972 case ISD::ATOMIC_LOAD_XOR:
6973 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
6975 case ISD::ATOMIC_SWAP:
6976 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
6981 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
6983 default: return NULL;
6984 case X86ISD::BSF: return "X86ISD::BSF";
6985 case X86ISD::BSR: return "X86ISD::BSR";
6986 case X86ISD::SHLD: return "X86ISD::SHLD";
6987 case X86ISD::SHRD: return "X86ISD::SHRD";
6988 case X86ISD::FAND: return "X86ISD::FAND";
6989 case X86ISD::FOR: return "X86ISD::FOR";
6990 case X86ISD::FXOR: return "X86ISD::FXOR";
6991 case X86ISD::FSRL: return "X86ISD::FSRL";
6992 case X86ISD::FILD: return "X86ISD::FILD";
6993 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
6994 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
6995 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
6996 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
6997 case X86ISD::FLD: return "X86ISD::FLD";
6998 case X86ISD::FST: return "X86ISD::FST";
6999 case X86ISD::CALL: return "X86ISD::CALL";
7000 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
7001 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7002 case X86ISD::BT: return "X86ISD::BT";
7003 case X86ISD::CMP: return "X86ISD::CMP";
7004 case X86ISD::COMI: return "X86ISD::COMI";
7005 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7006 case X86ISD::SETCC: return "X86ISD::SETCC";
7007 case X86ISD::CMOV: return "X86ISD::CMOV";
7008 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7009 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7010 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7011 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7012 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7013 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7014 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7015 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7016 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7017 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7018 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7019 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7020 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7021 case X86ISD::FMAX: return "X86ISD::FMAX";
7022 case X86ISD::FMIN: return "X86ISD::FMIN";
7023 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7024 case X86ISD::FRCP: return "X86ISD::FRCP";
7025 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7026 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7027 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7028 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7029 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7030 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7031 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7032 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7033 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7034 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7035 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7036 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7037 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7038 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7039 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7040 case X86ISD::VSHL: return "X86ISD::VSHL";
7041 case X86ISD::VSRL: return "X86ISD::VSRL";
7042 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7043 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7044 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7045 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7046 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7047 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7048 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7049 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7050 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7051 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7052 case X86ISD::ADD: return "X86ISD::ADD";
7053 case X86ISD::SUB: return "X86ISD::SUB";
7054 case X86ISD::SMUL: return "X86ISD::SMUL";
7055 case X86ISD::UMUL: return "X86ISD::UMUL";
7056 case X86ISD::INC: return "X86ISD::INC";
7057 case X86ISD::DEC: return "X86ISD::DEC";
7058 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7059 case X86ISD::PTEST: return "X86ISD::PTEST";
7063 // isLegalAddressingMode - Return true if the addressing mode represented
7064 // by AM is legal for this target, for a load/store of the specified type.
7065 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7066 const Type *Ty) const {
7067 // X86 supports extremely general addressing modes.
7069 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7070 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
7075 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7077 // If a reference to this global requires an extra load, we can't fold it.
7078 if (isGlobalStubReference(GVFlags))
7081 // If BaseGV requires a register for the PIC base, we cannot also have a
7082 // BaseReg specified.
7083 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7086 // X86-64 only supports addr of globals in small code model.
7087 if (Subtarget->is64Bit()) {
7088 if (getTargetMachine().getCodeModel() != CodeModel::Small)
7090 // If lower 4G is not available, then we must use rip-relative addressing.
7091 if (AM.BaseOffs || AM.Scale > 1)
7102 // These scales always work.
7107 // These scales are formed with basereg+scalereg. Only accept if there is
7112 default: // Other stuff never works.
7120 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7121 if (!Ty1->isInteger() || !Ty2->isInteger())
7123 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7124 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7125 if (NumBits1 <= NumBits2)
7127 return Subtarget->is64Bit() || NumBits1 < 64;
7130 bool X86TargetLowering::isTruncateFree(MVT VT1, MVT VT2) const {
7131 if (!VT1.isInteger() || !VT2.isInteger())
7133 unsigned NumBits1 = VT1.getSizeInBits();
7134 unsigned NumBits2 = VT2.getSizeInBits();
7135 if (NumBits1 <= NumBits2)
7137 return Subtarget->is64Bit() || NumBits1 < 64;
7140 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7141 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7142 return Ty1 == Type::Int32Ty && Ty2 == Type::Int64Ty && Subtarget->is64Bit();
7145 bool X86TargetLowering::isZExtFree(MVT VT1, MVT VT2) const {
7146 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7147 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7150 bool X86TargetLowering::isNarrowingProfitable(MVT VT1, MVT VT2) const {
7151 // i16 instructions are longer (0x66 prefix) and potentially slower.
7152 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7155 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7156 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7157 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7158 /// are assumed to be legal.
7160 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7162 // Only do shuffles on 128-bit vector types for now.
7163 if (VT.getSizeInBits() == 64)
7166 // FIXME: pshufb, blends, palignr, shifts.
7167 return (VT.getVectorNumElements() == 2 ||
7168 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7169 isMOVLMask(M, VT) ||
7170 isSHUFPMask(M, VT) ||
7171 isPSHUFDMask(M, VT) ||
7172 isPSHUFHWMask(M, VT) ||
7173 isPSHUFLWMask(M, VT) ||
7174 isUNPCKLMask(M, VT) ||
7175 isUNPCKHMask(M, VT) ||
7176 isUNPCKL_v_undef_Mask(M, VT) ||
7177 isUNPCKH_v_undef_Mask(M, VT));
7181 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7183 unsigned NumElts = VT.getVectorNumElements();
7184 // FIXME: This collection of masks seems suspect.
7187 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7188 return (isMOVLMask(Mask, VT) ||
7189 isCommutedMOVLMask(Mask, VT, true) ||
7190 isSHUFPMask(Mask, VT) ||
7191 isCommutedSHUFPMask(Mask, VT));
7196 //===----------------------------------------------------------------------===//
7197 // X86 Scheduler Hooks
7198 //===----------------------------------------------------------------------===//
7200 // private utility function
7202 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7203 MachineBasicBlock *MBB,
7211 TargetRegisterClass *RC,
7212 bool invSrc) const {
7213 // For the atomic bitwise operator, we generate
7216 // ld t1 = [bitinstr.addr]
7217 // op t2 = t1, [bitinstr.val]
7219 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7221 // fallthrough -->nextMBB
7222 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7223 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7224 MachineFunction::iterator MBBIter = MBB;
7227 /// First build the CFG
7228 MachineFunction *F = MBB->getParent();
7229 MachineBasicBlock *thisMBB = MBB;
7230 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7231 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7232 F->insert(MBBIter, newMBB);
7233 F->insert(MBBIter, nextMBB);
7235 // Move all successors to thisMBB to nextMBB
7236 nextMBB->transferSuccessors(thisMBB);
7238 // Update thisMBB to fall through to newMBB
7239 thisMBB->addSuccessor(newMBB);
7241 // newMBB jumps to itself and fall through to nextMBB
7242 newMBB->addSuccessor(nextMBB);
7243 newMBB->addSuccessor(newMBB);
7245 // Insert instructions into newMBB based on incoming instruction
7246 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7247 "unexpected number of operands");
7248 DebugLoc dl = bInstr->getDebugLoc();
7249 MachineOperand& destOper = bInstr->getOperand(0);
7250 MachineOperand* argOpers[2 + X86AddrNumOperands];
7251 int numArgs = bInstr->getNumOperands() - 1;
7252 for (int i=0; i < numArgs; ++i)
7253 argOpers[i] = &bInstr->getOperand(i+1);
7255 // x86 address has 4 operands: base, index, scale, and displacement
7256 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7257 int valArgIndx = lastAddrIndx + 1;
7259 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7260 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7261 for (int i=0; i <= lastAddrIndx; ++i)
7262 (*MIB).addOperand(*argOpers[i]);
7264 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7266 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7271 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7272 assert((argOpers[valArgIndx]->isReg() ||
7273 argOpers[valArgIndx]->isImm()) &&
7275 if (argOpers[valArgIndx]->isReg())
7276 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7278 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7280 (*MIB).addOperand(*argOpers[valArgIndx]);
7282 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7285 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7286 for (int i=0; i <= lastAddrIndx; ++i)
7287 (*MIB).addOperand(*argOpers[i]);
7289 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7290 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7292 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7296 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7298 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7302 // private utility function: 64 bit atomics on 32 bit host.
7304 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7305 MachineBasicBlock *MBB,
7310 bool invSrc) const {
7311 // For the atomic bitwise operator, we generate
7312 // thisMBB (instructions are in pairs, except cmpxchg8b)
7313 // ld t1,t2 = [bitinstr.addr]
7315 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7316 // op t5, t6 <- out1, out2, [bitinstr.val]
7317 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7318 // mov ECX, EBX <- t5, t6
7319 // mov EAX, EDX <- t1, t2
7320 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7321 // mov t3, t4 <- EAX, EDX
7323 // result in out1, out2
7324 // fallthrough -->nextMBB
7326 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7327 const unsigned LoadOpc = X86::MOV32rm;
7328 const unsigned copyOpc = X86::MOV32rr;
7329 const unsigned NotOpc = X86::NOT32r;
7330 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7331 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7332 MachineFunction::iterator MBBIter = MBB;
7335 /// First build the CFG
7336 MachineFunction *F = MBB->getParent();
7337 MachineBasicBlock *thisMBB = MBB;
7338 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7339 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7340 F->insert(MBBIter, newMBB);
7341 F->insert(MBBIter, nextMBB);
7343 // Move all successors to thisMBB to nextMBB
7344 nextMBB->transferSuccessors(thisMBB);
7346 // Update thisMBB to fall through to newMBB
7347 thisMBB->addSuccessor(newMBB);
7349 // newMBB jumps to itself and fall through to nextMBB
7350 newMBB->addSuccessor(nextMBB);
7351 newMBB->addSuccessor(newMBB);
7353 DebugLoc dl = bInstr->getDebugLoc();
7354 // Insert instructions into newMBB based on incoming instruction
7355 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7356 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7357 "unexpected number of operands");
7358 MachineOperand& dest1Oper = bInstr->getOperand(0);
7359 MachineOperand& dest2Oper = bInstr->getOperand(1);
7360 MachineOperand* argOpers[2 + X86AddrNumOperands];
7361 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7362 argOpers[i] = &bInstr->getOperand(i+2);
7364 // x86 address has 4 operands: base, index, scale, and displacement
7365 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7367 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7368 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7369 for (int i=0; i <= lastAddrIndx; ++i)
7370 (*MIB).addOperand(*argOpers[i]);
7371 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7372 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7373 // add 4 to displacement.
7374 for (int i=0; i <= lastAddrIndx-2; ++i)
7375 (*MIB).addOperand(*argOpers[i]);
7376 MachineOperand newOp3 = *(argOpers[3]);
7378 newOp3.setImm(newOp3.getImm()+4);
7380 newOp3.setOffset(newOp3.getOffset()+4);
7381 (*MIB).addOperand(newOp3);
7382 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7384 // t3/4 are defined later, at the bottom of the loop
7385 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7386 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7387 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7388 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7389 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7390 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7392 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
7393 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
7395 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt1).addReg(t1);
7396 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt2).addReg(t2);
7402 int valArgIndx = lastAddrIndx + 1;
7403 assert((argOpers[valArgIndx]->isReg() ||
7404 argOpers[valArgIndx]->isImm()) &&
7406 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7407 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7408 if (argOpers[valArgIndx]->isReg())
7409 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7411 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7412 if (regOpcL != X86::MOV32rr)
7414 (*MIB).addOperand(*argOpers[valArgIndx]);
7415 assert(argOpers[valArgIndx + 1]->isReg() ==
7416 argOpers[valArgIndx]->isReg());
7417 assert(argOpers[valArgIndx + 1]->isImm() ==
7418 argOpers[valArgIndx]->isImm());
7419 if (argOpers[valArgIndx + 1]->isReg())
7420 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7422 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7423 if (regOpcH != X86::MOV32rr)
7425 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
7427 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
7429 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
7432 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
7434 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
7437 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
7438 for (int i=0; i <= lastAddrIndx; ++i)
7439 (*MIB).addOperand(*argOpers[i]);
7441 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7442 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7444 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
7445 MIB.addReg(X86::EAX);
7446 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
7447 MIB.addReg(X86::EDX);
7450 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7452 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7456 // private utility function
7458 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
7459 MachineBasicBlock *MBB,
7460 unsigned cmovOpc) const {
7461 // For the atomic min/max operator, we generate
7464 // ld t1 = [min/max.addr]
7465 // mov t2 = [min/max.val]
7467 // cmov[cond] t2 = t1
7469 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7471 // fallthrough -->nextMBB
7473 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7474 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7475 MachineFunction::iterator MBBIter = MBB;
7478 /// First build the CFG
7479 MachineFunction *F = MBB->getParent();
7480 MachineBasicBlock *thisMBB = MBB;
7481 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7482 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7483 F->insert(MBBIter, newMBB);
7484 F->insert(MBBIter, nextMBB);
7486 // Move all successors to thisMBB to nextMBB
7487 nextMBB->transferSuccessors(thisMBB);
7489 // Update thisMBB to fall through to newMBB
7490 thisMBB->addSuccessor(newMBB);
7492 // newMBB jumps to newMBB and fall through to nextMBB
7493 newMBB->addSuccessor(nextMBB);
7494 newMBB->addSuccessor(newMBB);
7496 DebugLoc dl = mInstr->getDebugLoc();
7497 // Insert instructions into newMBB based on incoming instruction
7498 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7499 "unexpected number of operands");
7500 MachineOperand& destOper = mInstr->getOperand(0);
7501 MachineOperand* argOpers[2 + X86AddrNumOperands];
7502 int numArgs = mInstr->getNumOperands() - 1;
7503 for (int i=0; i < numArgs; ++i)
7504 argOpers[i] = &mInstr->getOperand(i+1);
7506 // x86 address has 4 operands: base, index, scale, and displacement
7507 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7508 int valArgIndx = lastAddrIndx + 1;
7510 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7511 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
7512 for (int i=0; i <= lastAddrIndx; ++i)
7513 (*MIB).addOperand(*argOpers[i]);
7515 // We only support register and immediate values
7516 assert((argOpers[valArgIndx]->isReg() ||
7517 argOpers[valArgIndx]->isImm()) &&
7520 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7521 if (argOpers[valArgIndx]->isReg())
7522 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7524 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7525 (*MIB).addOperand(*argOpers[valArgIndx]);
7527 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
7530 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
7535 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7536 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
7540 // Cmp and exchange if none has modified the memory location
7541 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
7542 for (int i=0; i <= lastAddrIndx; ++i)
7543 (*MIB).addOperand(*argOpers[i]);
7545 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7546 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
7548 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
7549 MIB.addReg(X86::EAX);
7552 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7554 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
7560 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7561 MachineBasicBlock *BB) const {
7562 DebugLoc dl = MI->getDebugLoc();
7563 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7564 switch (MI->getOpcode()) {
7565 default: assert(false && "Unexpected instr type to insert");
7566 case X86::CMOV_V1I64:
7567 case X86::CMOV_FR32:
7568 case X86::CMOV_FR64:
7569 case X86::CMOV_V4F32:
7570 case X86::CMOV_V2F64:
7571 case X86::CMOV_V2I64: {
7572 // To "insert" a SELECT_CC instruction, we actually have to insert the
7573 // diamond control-flow pattern. The incoming instruction knows the
7574 // destination vreg to set, the condition code register to branch on, the
7575 // true/false values to select between, and a branch opcode to use.
7576 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7577 MachineFunction::iterator It = BB;
7583 // cmpTY ccX, r1, r2
7585 // fallthrough --> copy0MBB
7586 MachineBasicBlock *thisMBB = BB;
7587 MachineFunction *F = BB->getParent();
7588 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7589 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7591 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
7592 BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
7593 F->insert(It, copy0MBB);
7594 F->insert(It, sinkMBB);
7595 // Update machine-CFG edges by transferring all successors of the current
7596 // block to the new block which will contain the Phi node for the select.
7597 sinkMBB->transferSuccessors(BB);
7599 // Add the true and fallthrough blocks as its successors.
7600 BB->addSuccessor(copy0MBB);
7601 BB->addSuccessor(sinkMBB);
7604 // %FalseValue = ...
7605 // # fallthrough to sinkMBB
7608 // Update machine-CFG edges
7609 BB->addSuccessor(sinkMBB);
7612 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7615 BuildMI(BB, dl, TII->get(X86::PHI), MI->getOperand(0).getReg())
7616 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7617 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7619 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7623 case X86::FP32_TO_INT16_IN_MEM:
7624 case X86::FP32_TO_INT32_IN_MEM:
7625 case X86::FP32_TO_INT64_IN_MEM:
7626 case X86::FP64_TO_INT16_IN_MEM:
7627 case X86::FP64_TO_INT32_IN_MEM:
7628 case X86::FP64_TO_INT64_IN_MEM:
7629 case X86::FP80_TO_INT16_IN_MEM:
7630 case X86::FP80_TO_INT32_IN_MEM:
7631 case X86::FP80_TO_INT64_IN_MEM: {
7632 // Change the floating point control register to use "round towards zero"
7633 // mode when truncating to an integer value.
7634 MachineFunction *F = BB->getParent();
7635 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
7636 addFrameReference(BuildMI(BB, dl, TII->get(X86::FNSTCW16m)), CWFrameIdx);
7638 // Load the old value of the high byte of the control word...
7640 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
7641 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16rm), OldCW),
7644 // Set the high part to be round to zero...
7645 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mi)), CWFrameIdx)
7648 // Reload the modified control word now...
7649 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7651 // Restore the memory image of control word to original value
7652 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mr)), CWFrameIdx)
7655 // Get the X86 opcode to use.
7657 switch (MI->getOpcode()) {
7658 default: llvm_unreachable("illegal opcode!");
7659 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
7660 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
7661 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
7662 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
7663 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
7664 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
7665 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
7666 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
7667 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
7671 MachineOperand &Op = MI->getOperand(0);
7673 AM.BaseType = X86AddressMode::RegBase;
7674 AM.Base.Reg = Op.getReg();
7676 AM.BaseType = X86AddressMode::FrameIndexBase;
7677 AM.Base.FrameIndex = Op.getIndex();
7679 Op = MI->getOperand(1);
7681 AM.Scale = Op.getImm();
7682 Op = MI->getOperand(2);
7684 AM.IndexReg = Op.getImm();
7685 Op = MI->getOperand(3);
7686 if (Op.isGlobal()) {
7687 AM.GV = Op.getGlobal();
7689 AM.Disp = Op.getImm();
7691 addFullAddress(BuildMI(BB, dl, TII->get(Opc)), AM)
7692 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
7694 // Reload the original control word now.
7695 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7697 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7700 case X86::ATOMAND32:
7701 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7702 X86::AND32ri, X86::MOV32rm,
7703 X86::LCMPXCHG32, X86::MOV32rr,
7704 X86::NOT32r, X86::EAX,
7705 X86::GR32RegisterClass);
7707 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
7708 X86::OR32ri, X86::MOV32rm,
7709 X86::LCMPXCHG32, X86::MOV32rr,
7710 X86::NOT32r, X86::EAX,
7711 X86::GR32RegisterClass);
7712 case X86::ATOMXOR32:
7713 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
7714 X86::XOR32ri, X86::MOV32rm,
7715 X86::LCMPXCHG32, X86::MOV32rr,
7716 X86::NOT32r, X86::EAX,
7717 X86::GR32RegisterClass);
7718 case X86::ATOMNAND32:
7719 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7720 X86::AND32ri, X86::MOV32rm,
7721 X86::LCMPXCHG32, X86::MOV32rr,
7722 X86::NOT32r, X86::EAX,
7723 X86::GR32RegisterClass, true);
7724 case X86::ATOMMIN32:
7725 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
7726 case X86::ATOMMAX32:
7727 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
7728 case X86::ATOMUMIN32:
7729 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
7730 case X86::ATOMUMAX32:
7731 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
7733 case X86::ATOMAND16:
7734 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7735 X86::AND16ri, X86::MOV16rm,
7736 X86::LCMPXCHG16, X86::MOV16rr,
7737 X86::NOT16r, X86::AX,
7738 X86::GR16RegisterClass);
7740 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
7741 X86::OR16ri, X86::MOV16rm,
7742 X86::LCMPXCHG16, X86::MOV16rr,
7743 X86::NOT16r, X86::AX,
7744 X86::GR16RegisterClass);
7745 case X86::ATOMXOR16:
7746 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
7747 X86::XOR16ri, X86::MOV16rm,
7748 X86::LCMPXCHG16, X86::MOV16rr,
7749 X86::NOT16r, X86::AX,
7750 X86::GR16RegisterClass);
7751 case X86::ATOMNAND16:
7752 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7753 X86::AND16ri, X86::MOV16rm,
7754 X86::LCMPXCHG16, X86::MOV16rr,
7755 X86::NOT16r, X86::AX,
7756 X86::GR16RegisterClass, true);
7757 case X86::ATOMMIN16:
7758 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
7759 case X86::ATOMMAX16:
7760 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
7761 case X86::ATOMUMIN16:
7762 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
7763 case X86::ATOMUMAX16:
7764 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
7767 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7768 X86::AND8ri, X86::MOV8rm,
7769 X86::LCMPXCHG8, X86::MOV8rr,
7770 X86::NOT8r, X86::AL,
7771 X86::GR8RegisterClass);
7773 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
7774 X86::OR8ri, X86::MOV8rm,
7775 X86::LCMPXCHG8, X86::MOV8rr,
7776 X86::NOT8r, X86::AL,
7777 X86::GR8RegisterClass);
7779 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
7780 X86::XOR8ri, X86::MOV8rm,
7781 X86::LCMPXCHG8, X86::MOV8rr,
7782 X86::NOT8r, X86::AL,
7783 X86::GR8RegisterClass);
7784 case X86::ATOMNAND8:
7785 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7786 X86::AND8ri, X86::MOV8rm,
7787 X86::LCMPXCHG8, X86::MOV8rr,
7788 X86::NOT8r, X86::AL,
7789 X86::GR8RegisterClass, true);
7790 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
7791 // This group is for 64-bit host.
7792 case X86::ATOMAND64:
7793 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7794 X86::AND64ri32, X86::MOV64rm,
7795 X86::LCMPXCHG64, X86::MOV64rr,
7796 X86::NOT64r, X86::RAX,
7797 X86::GR64RegisterClass);
7799 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
7800 X86::OR64ri32, X86::MOV64rm,
7801 X86::LCMPXCHG64, X86::MOV64rr,
7802 X86::NOT64r, X86::RAX,
7803 X86::GR64RegisterClass);
7804 case X86::ATOMXOR64:
7805 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
7806 X86::XOR64ri32, X86::MOV64rm,
7807 X86::LCMPXCHG64, X86::MOV64rr,
7808 X86::NOT64r, X86::RAX,
7809 X86::GR64RegisterClass);
7810 case X86::ATOMNAND64:
7811 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7812 X86::AND64ri32, X86::MOV64rm,
7813 X86::LCMPXCHG64, X86::MOV64rr,
7814 X86::NOT64r, X86::RAX,
7815 X86::GR64RegisterClass, true);
7816 case X86::ATOMMIN64:
7817 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
7818 case X86::ATOMMAX64:
7819 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
7820 case X86::ATOMUMIN64:
7821 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
7822 case X86::ATOMUMAX64:
7823 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
7825 // This group does 64-bit operations on a 32-bit host.
7826 case X86::ATOMAND6432:
7827 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7828 X86::AND32rr, X86::AND32rr,
7829 X86::AND32ri, X86::AND32ri,
7831 case X86::ATOMOR6432:
7832 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7833 X86::OR32rr, X86::OR32rr,
7834 X86::OR32ri, X86::OR32ri,
7836 case X86::ATOMXOR6432:
7837 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7838 X86::XOR32rr, X86::XOR32rr,
7839 X86::XOR32ri, X86::XOR32ri,
7841 case X86::ATOMNAND6432:
7842 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7843 X86::AND32rr, X86::AND32rr,
7844 X86::AND32ri, X86::AND32ri,
7846 case X86::ATOMADD6432:
7847 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7848 X86::ADD32rr, X86::ADC32rr,
7849 X86::ADD32ri, X86::ADC32ri,
7851 case X86::ATOMSUB6432:
7852 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7853 X86::SUB32rr, X86::SBB32rr,
7854 X86::SUB32ri, X86::SBB32ri,
7856 case X86::ATOMSWAP6432:
7857 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7858 X86::MOV32rr, X86::MOV32rr,
7859 X86::MOV32ri, X86::MOV32ri,
7864 //===----------------------------------------------------------------------===//
7865 // X86 Optimization Hooks
7866 //===----------------------------------------------------------------------===//
7868 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
7872 const SelectionDAG &DAG,
7873 unsigned Depth) const {
7874 unsigned Opc = Op.getOpcode();
7875 assert((Opc >= ISD::BUILTIN_OP_END ||
7876 Opc == ISD::INTRINSIC_WO_CHAIN ||
7877 Opc == ISD::INTRINSIC_W_CHAIN ||
7878 Opc == ISD::INTRINSIC_VOID) &&
7879 "Should use MaskedValueIsZero if you don't know whether Op"
7880 " is a target node!");
7882 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
7891 // These nodes' second result is a boolean.
7892 if (Op.getResNo() == 0)
7896 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
7897 Mask.getBitWidth() - 1);
7902 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
7903 /// node is a GlobalAddress + offset.
7904 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
7905 GlobalValue* &GA, int64_t &Offset) const{
7906 if (N->getOpcode() == X86ISD::Wrapper) {
7907 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
7908 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
7909 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
7913 return TargetLowering::isGAPlusOffset(N, GA, Offset);
7916 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
7917 const TargetLowering &TLI) {
7920 if (TLI.isGAPlusOffset(Base, GV, Offset))
7921 return (GV->getAlignment() >= N && (Offset % N) == 0);
7922 // DAG combine handles the stack object case.
7926 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
7927 MVT EVT, LoadSDNode *&LDBase,
7928 unsigned &LastLoadedElt,
7929 SelectionDAG &DAG, MachineFrameInfo *MFI,
7930 const TargetLowering &TLI) {
7932 LastLoadedElt = -1U;
7933 for (unsigned i = 0; i < NumElems; ++i) {
7934 if (N->getMaskElt(i) < 0) {
7940 SDValue Elt = DAG.getShuffleScalarElt(N, i);
7941 if (!Elt.getNode() ||
7942 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
7945 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
7947 LDBase = cast<LoadSDNode>(Elt.getNode());
7951 if (Elt.getOpcode() == ISD::UNDEF)
7954 LoadSDNode *LD = cast<LoadSDNode>(Elt);
7955 if (!TLI.isConsecutiveLoad(LD, LDBase, EVT.getSizeInBits()/8, i, MFI))
7962 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
7963 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
7964 /// if the load addresses are consecutive, non-overlapping, and in the right
7965 /// order. In the case of v2i64, it will see if it can rewrite the
7966 /// shuffle to be an appropriate build vector so it can take advantage of
7967 // performBuildVectorCombine.
7968 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
7969 const TargetLowering &TLI) {
7970 DebugLoc dl = N->getDebugLoc();
7971 MVT VT = N->getValueType(0);
7972 MVT EVT = VT.getVectorElementType();
7973 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7974 unsigned NumElems = VT.getVectorNumElements();
7976 if (VT.getSizeInBits() != 128)
7979 // Try to combine a vector_shuffle into a 128-bit load.
7980 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7981 LoadSDNode *LD = NULL;
7982 unsigned LastLoadedElt;
7983 if (!EltsFromConsecutiveLoads(SVN, NumElems, EVT, LD, LastLoadedElt, DAG,
7987 if (LastLoadedElt == NumElems - 1) {
7988 if (isBaseAlignmentOfN(16, LD->getBasePtr().getNode(), TLI))
7989 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
7990 LD->getSrcValue(), LD->getSrcValueOffset(),
7992 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
7993 LD->getSrcValue(), LD->getSrcValueOffset(),
7994 LD->isVolatile(), LD->getAlignment());
7995 } else if (NumElems == 4 && LastLoadedElt == 1) {
7996 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
7997 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
7998 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
7999 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8004 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8005 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8006 const X86Subtarget *Subtarget) {
8007 DebugLoc DL = N->getDebugLoc();
8008 SDValue Cond = N->getOperand(0);
8009 // Get the LHS/RHS of the select.
8010 SDValue LHS = N->getOperand(1);
8011 SDValue RHS = N->getOperand(2);
8013 // If we have SSE[12] support, try to form min/max nodes.
8014 if (Subtarget->hasSSE2() &&
8015 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8016 Cond.getOpcode() == ISD::SETCC) {
8017 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8019 unsigned Opcode = 0;
8020 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8023 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
8026 if (!UnsafeFPMath) break;
8028 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
8030 Opcode = X86ISD::FMIN;
8033 case ISD::SETOGT: // (X > Y) ? X : Y -> max
8036 if (!UnsafeFPMath) break;
8038 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
8040 Opcode = X86ISD::FMAX;
8043 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8046 case ISD::SETOGT: // (X > Y) ? Y : X -> min
8049 if (!UnsafeFPMath) break;
8051 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
8053 Opcode = X86ISD::FMIN;
8056 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
8059 if (!UnsafeFPMath) break;
8061 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
8063 Opcode = X86ISD::FMAX;
8069 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8072 // If this is a select between two integer constants, try to do some
8074 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8075 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8076 // Don't do this for crazy integer types.
8077 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8078 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8079 // so that TrueC (the true value) is larger than FalseC.
8080 bool NeedsCondInvert = false;
8082 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8083 // Efficiently invertible.
8084 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8085 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8086 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8087 NeedsCondInvert = true;
8088 std::swap(TrueC, FalseC);
8091 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8092 if (FalseC->getAPIntValue() == 0 &&
8093 TrueC->getAPIntValue().isPowerOf2()) {
8094 if (NeedsCondInvert) // Invert the condition if needed.
8095 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8096 DAG.getConstant(1, Cond.getValueType()));
8098 // Zero extend the condition if needed.
8099 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8101 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8102 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8103 DAG.getConstant(ShAmt, MVT::i8));
8106 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8107 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8108 if (NeedsCondInvert) // Invert the condition if needed.
8109 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8110 DAG.getConstant(1, Cond.getValueType()));
8112 // Zero extend the condition if needed.
8113 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8114 FalseC->getValueType(0), Cond);
8115 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8116 SDValue(FalseC, 0));
8119 // Optimize cases that will turn into an LEA instruction. This requires
8120 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8121 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8122 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8123 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8125 bool isFastMultiplier = false;
8127 switch ((unsigned char)Diff) {
8129 case 1: // result = add base, cond
8130 case 2: // result = lea base( , cond*2)
8131 case 3: // result = lea base(cond, cond*2)
8132 case 4: // result = lea base( , cond*4)
8133 case 5: // result = lea base(cond, cond*4)
8134 case 8: // result = lea base( , cond*8)
8135 case 9: // result = lea base(cond, cond*8)
8136 isFastMultiplier = true;
8141 if (isFastMultiplier) {
8142 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8143 if (NeedsCondInvert) // Invert the condition if needed.
8144 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8145 DAG.getConstant(1, Cond.getValueType()));
8147 // Zero extend the condition if needed.
8148 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8150 // Scale the condition by the difference.
8152 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8153 DAG.getConstant(Diff, Cond.getValueType()));
8155 // Add the base if non-zero.
8156 if (FalseC->getAPIntValue() != 0)
8157 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8158 SDValue(FalseC, 0));
8168 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
8169 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8170 TargetLowering::DAGCombinerInfo &DCI) {
8171 DebugLoc DL = N->getDebugLoc();
8173 // If the flag operand isn't dead, don't touch this CMOV.
8174 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8177 // If this is a select between two integer constants, try to do some
8178 // optimizations. Note that the operands are ordered the opposite of SELECT
8180 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
8181 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8182 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
8183 // larger than FalseC (the false value).
8184 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
8186 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
8187 CC = X86::GetOppositeBranchCondition(CC);
8188 std::swap(TrueC, FalseC);
8191 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
8192 // This is efficient for any integer data type (including i8/i16) and
8194 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
8195 SDValue Cond = N->getOperand(3);
8196 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8197 DAG.getConstant(CC, MVT::i8), Cond);
8199 // Zero extend the condition if needed.
8200 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
8202 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8203 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
8204 DAG.getConstant(ShAmt, MVT::i8));
8205 if (N->getNumValues() == 2) // Dead flag value?
8206 return DCI.CombineTo(N, Cond, SDValue());
8210 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
8211 // for any integer data type, including i8/i16.
8212 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8213 SDValue Cond = N->getOperand(3);
8214 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8215 DAG.getConstant(CC, MVT::i8), Cond);
8217 // Zero extend the condition if needed.
8218 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8219 FalseC->getValueType(0), Cond);
8220 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8221 SDValue(FalseC, 0));
8223 if (N->getNumValues() == 2) // Dead flag value?
8224 return DCI.CombineTo(N, Cond, SDValue());
8228 // Optimize cases that will turn into an LEA instruction. This requires
8229 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8230 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8231 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8232 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8234 bool isFastMultiplier = false;
8236 switch ((unsigned char)Diff) {
8238 case 1: // result = add base, cond
8239 case 2: // result = lea base( , cond*2)
8240 case 3: // result = lea base(cond, cond*2)
8241 case 4: // result = lea base( , cond*4)
8242 case 5: // result = lea base(cond, cond*4)
8243 case 8: // result = lea base( , cond*8)
8244 case 9: // result = lea base(cond, cond*8)
8245 isFastMultiplier = true;
8250 if (isFastMultiplier) {
8251 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8252 SDValue Cond = N->getOperand(3);
8253 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8254 DAG.getConstant(CC, MVT::i8), Cond);
8255 // Zero extend the condition if needed.
8256 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8258 // Scale the condition by the difference.
8260 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8261 DAG.getConstant(Diff, Cond.getValueType()));
8263 // Add the base if non-zero.
8264 if (FalseC->getAPIntValue() != 0)
8265 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8266 SDValue(FalseC, 0));
8267 if (N->getNumValues() == 2) // Dead flag value?
8268 return DCI.CombineTo(N, Cond, SDValue());
8278 /// PerformMulCombine - Optimize a single multiply with constant into two
8279 /// in order to implement it with two cheaper instructions, e.g.
8280 /// LEA + SHL, LEA + LEA.
8281 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
8282 TargetLowering::DAGCombinerInfo &DCI) {
8283 if (DAG.getMachineFunction().
8284 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
8287 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8290 MVT VT = N->getValueType(0);
8294 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8297 uint64_t MulAmt = C->getZExtValue();
8298 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
8301 uint64_t MulAmt1 = 0;
8302 uint64_t MulAmt2 = 0;
8303 if ((MulAmt % 9) == 0) {
8305 MulAmt2 = MulAmt / 9;
8306 } else if ((MulAmt % 5) == 0) {
8308 MulAmt2 = MulAmt / 5;
8309 } else if ((MulAmt % 3) == 0) {
8311 MulAmt2 = MulAmt / 3;
8314 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
8315 DebugLoc DL = N->getDebugLoc();
8317 if (isPowerOf2_64(MulAmt2) &&
8318 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
8319 // If second multiplifer is pow2, issue it first. We want the multiply by
8320 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
8322 std::swap(MulAmt1, MulAmt2);
8325 if (isPowerOf2_64(MulAmt1))
8326 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
8327 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
8329 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
8330 DAG.getConstant(MulAmt1, VT));
8332 if (isPowerOf2_64(MulAmt2))
8333 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
8334 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
8336 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
8337 DAG.getConstant(MulAmt2, VT));
8339 // Do not add new nodes to DAG combiner worklist.
8340 DCI.CombineTo(N, NewMul, false);
8346 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
8348 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
8349 const X86Subtarget *Subtarget) {
8350 // On X86 with SSE2 support, we can transform this to a vector shift if
8351 // all elements are shifted by the same amount. We can't do this in legalize
8352 // because the a constant vector is typically transformed to a constant pool
8353 // so we have no knowledge of the shift amount.
8354 if (!Subtarget->hasSSE2())
8357 MVT VT = N->getValueType(0);
8358 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
8361 SDValue ShAmtOp = N->getOperand(1);
8362 MVT EltVT = VT.getVectorElementType();
8363 DebugLoc DL = N->getDebugLoc();
8365 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
8366 unsigned NumElts = VT.getVectorNumElements();
8368 for (; i != NumElts; ++i) {
8369 SDValue Arg = ShAmtOp.getOperand(i);
8370 if (Arg.getOpcode() == ISD::UNDEF) continue;
8374 for (; i != NumElts; ++i) {
8375 SDValue Arg = ShAmtOp.getOperand(i);
8376 if (Arg.getOpcode() == ISD::UNDEF) continue;
8377 if (Arg != BaseShAmt) {
8381 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
8382 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
8383 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
8384 DAG.getIntPtrConstant(0));
8388 if (EltVT.bitsGT(MVT::i32))
8389 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
8390 else if (EltVT.bitsLT(MVT::i32))
8391 BaseShAmt = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, BaseShAmt);
8393 // The shift amount is identical so we can do a vector shift.
8394 SDValue ValOp = N->getOperand(0);
8395 switch (N->getOpcode()) {
8397 llvm_unreachable("Unknown shift opcode!");
8400 if (VT == MVT::v2i64)
8401 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8402 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8404 if (VT == MVT::v4i32)
8405 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8406 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8408 if (VT == MVT::v8i16)
8409 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8410 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8414 if (VT == MVT::v4i32)
8415 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8416 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
8418 if (VT == MVT::v8i16)
8419 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8420 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
8424 if (VT == MVT::v2i64)
8425 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8426 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8428 if (VT == MVT::v4i32)
8429 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8430 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
8432 if (VT == MVT::v8i16)
8433 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8434 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
8441 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
8442 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
8443 const X86Subtarget *Subtarget) {
8444 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
8445 // the FP state in cases where an emms may be missing.
8446 // A preferable solution to the general problem is to figure out the right
8447 // places to insert EMMS. This qualifies as a quick hack.
8449 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
8450 StoreSDNode *St = cast<StoreSDNode>(N);
8451 MVT VT = St->getValue().getValueType();
8452 if (VT.getSizeInBits() != 64)
8455 const Function *F = DAG.getMachineFunction().getFunction();
8456 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
8457 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
8458 && Subtarget->hasSSE2();
8459 if ((VT.isVector() ||
8460 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
8461 isa<LoadSDNode>(St->getValue()) &&
8462 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
8463 St->getChain().hasOneUse() && !St->isVolatile()) {
8464 SDNode* LdVal = St->getValue().getNode();
8466 int TokenFactorIndex = -1;
8467 SmallVector<SDValue, 8> Ops;
8468 SDNode* ChainVal = St->getChain().getNode();
8469 // Must be a store of a load. We currently handle two cases: the load
8470 // is a direct child, and it's under an intervening TokenFactor. It is
8471 // possible to dig deeper under nested TokenFactors.
8472 if (ChainVal == LdVal)
8473 Ld = cast<LoadSDNode>(St->getChain());
8474 else if (St->getValue().hasOneUse() &&
8475 ChainVal->getOpcode() == ISD::TokenFactor) {
8476 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
8477 if (ChainVal->getOperand(i).getNode() == LdVal) {
8478 TokenFactorIndex = i;
8479 Ld = cast<LoadSDNode>(St->getValue());
8481 Ops.push_back(ChainVal->getOperand(i));
8485 if (!Ld || !ISD::isNormalLoad(Ld))
8488 // If this is not the MMX case, i.e. we are just turning i64 load/store
8489 // into f64 load/store, avoid the transformation if there are multiple
8490 // uses of the loaded value.
8491 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
8494 DebugLoc LdDL = Ld->getDebugLoc();
8495 DebugLoc StDL = N->getDebugLoc();
8496 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
8497 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
8499 if (Subtarget->is64Bit() || F64IsLegal) {
8500 MVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
8501 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
8502 Ld->getBasePtr(), Ld->getSrcValue(),
8503 Ld->getSrcValueOffset(), Ld->isVolatile(),
8504 Ld->getAlignment());
8505 SDValue NewChain = NewLd.getValue(1);
8506 if (TokenFactorIndex != -1) {
8507 Ops.push_back(NewChain);
8508 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8511 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
8512 St->getSrcValue(), St->getSrcValueOffset(),
8513 St->isVolatile(), St->getAlignment());
8516 // Otherwise, lower to two pairs of 32-bit loads / stores.
8517 SDValue LoAddr = Ld->getBasePtr();
8518 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
8519 DAG.getConstant(4, MVT::i32));
8521 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
8522 Ld->getSrcValue(), Ld->getSrcValueOffset(),
8523 Ld->isVolatile(), Ld->getAlignment());
8524 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
8525 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
8527 MinAlign(Ld->getAlignment(), 4));
8529 SDValue NewChain = LoLd.getValue(1);
8530 if (TokenFactorIndex != -1) {
8531 Ops.push_back(LoLd);
8532 Ops.push_back(HiLd);
8533 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8537 LoAddr = St->getBasePtr();
8538 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
8539 DAG.getConstant(4, MVT::i32));
8541 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
8542 St->getSrcValue(), St->getSrcValueOffset(),
8543 St->isVolatile(), St->getAlignment());
8544 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
8546 St->getSrcValueOffset() + 4,
8548 MinAlign(St->getAlignment(), 4));
8549 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
8554 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
8555 /// X86ISD::FXOR nodes.
8556 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
8557 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
8558 // F[X]OR(0.0, x) -> x
8559 // F[X]OR(x, 0.0) -> x
8560 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8561 if (C->getValueAPF().isPosZero())
8562 return N->getOperand(1);
8563 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8564 if (C->getValueAPF().isPosZero())
8565 return N->getOperand(0);
8569 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
8570 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
8571 // FAND(0.0, x) -> 0.0
8572 // FAND(x, 0.0) -> 0.0
8573 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8574 if (C->getValueAPF().isPosZero())
8575 return N->getOperand(0);
8576 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8577 if (C->getValueAPF().isPosZero())
8578 return N->getOperand(1);
8582 static SDValue PerformBTCombine(SDNode *N,
8584 TargetLowering::DAGCombinerInfo &DCI) {
8585 // BT ignores high bits in the bit index operand.
8586 SDValue Op1 = N->getOperand(1);
8587 if (Op1.hasOneUse()) {
8588 unsigned BitWidth = Op1.getValueSizeInBits();
8589 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
8590 APInt KnownZero, KnownOne;
8591 TargetLowering::TargetLoweringOpt TLO(DAG);
8592 TargetLowering &TLI = DAG.getTargetLoweringInfo();
8593 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
8594 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
8595 DCI.CommitTargetLoweringOpt(TLO);
8600 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
8601 SDValue Op = N->getOperand(0);
8602 if (Op.getOpcode() == ISD::BIT_CONVERT)
8603 Op = Op.getOperand(0);
8604 MVT VT = N->getValueType(0), OpVT = Op.getValueType();
8605 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
8606 VT.getVectorElementType().getSizeInBits() ==
8607 OpVT.getVectorElementType().getSizeInBits()) {
8608 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
8613 // On X86 and X86-64, atomic operations are lowered to locked instructions.
8614 // Locked instructions, in turn, have implicit fence semantics (all memory
8615 // operations are flushed before issuing the locked instruction, and the
8616 // are not buffered), so we can fold away the common pattern of
8617 // fence-atomic-fence.
8618 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
8619 SDValue atomic = N->getOperand(0);
8620 switch (atomic.getOpcode()) {
8621 case ISD::ATOMIC_CMP_SWAP:
8622 case ISD::ATOMIC_SWAP:
8623 case ISD::ATOMIC_LOAD_ADD:
8624 case ISD::ATOMIC_LOAD_SUB:
8625 case ISD::ATOMIC_LOAD_AND:
8626 case ISD::ATOMIC_LOAD_OR:
8627 case ISD::ATOMIC_LOAD_XOR:
8628 case ISD::ATOMIC_LOAD_NAND:
8629 case ISD::ATOMIC_LOAD_MIN:
8630 case ISD::ATOMIC_LOAD_MAX:
8631 case ISD::ATOMIC_LOAD_UMIN:
8632 case ISD::ATOMIC_LOAD_UMAX:
8638 SDValue fence = atomic.getOperand(0);
8639 if (fence.getOpcode() != ISD::MEMBARRIER)
8642 switch (atomic.getOpcode()) {
8643 case ISD::ATOMIC_CMP_SWAP:
8644 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
8645 atomic.getOperand(1), atomic.getOperand(2),
8646 atomic.getOperand(3));
8647 case ISD::ATOMIC_SWAP:
8648 case ISD::ATOMIC_LOAD_ADD:
8649 case ISD::ATOMIC_LOAD_SUB:
8650 case ISD::ATOMIC_LOAD_AND:
8651 case ISD::ATOMIC_LOAD_OR:
8652 case ISD::ATOMIC_LOAD_XOR:
8653 case ISD::ATOMIC_LOAD_NAND:
8654 case ISD::ATOMIC_LOAD_MIN:
8655 case ISD::ATOMIC_LOAD_MAX:
8656 case ISD::ATOMIC_LOAD_UMIN:
8657 case ISD::ATOMIC_LOAD_UMAX:
8658 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
8659 atomic.getOperand(1), atomic.getOperand(2));
8665 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
8666 DAGCombinerInfo &DCI) const {
8667 SelectionDAG &DAG = DCI.DAG;
8668 switch (N->getOpcode()) {
8670 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
8671 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
8672 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
8673 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
8676 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
8677 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
8679 case X86ISD::FOR: return PerformFORCombine(N, DAG);
8680 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
8681 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
8682 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
8683 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
8689 //===----------------------------------------------------------------------===//
8690 // X86 Inline Assembly Support
8691 //===----------------------------------------------------------------------===//
8693 static bool LowerToBSwap(CallInst *CI) {
8694 // FIXME: this should verify that we are targetting a 486 or better. If not,
8695 // we will turn this bswap into something that will be lowered to logical ops
8696 // instead of emitting the bswap asm. For now, we don't support 486 or lower
8697 // so don't worry about this.
8699 // Verify this is a simple bswap.
8700 if (CI->getNumOperands() != 2 ||
8701 CI->getType() != CI->getOperand(1)->getType() ||
8702 !CI->getType()->isInteger())
8705 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
8706 if (!Ty || Ty->getBitWidth() % 16 != 0)
8709 // Okay, we can do this xform, do so now.
8710 const Type *Tys[] = { Ty };
8711 Module *M = CI->getParent()->getParent()->getParent();
8712 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
8714 Value *Op = CI->getOperand(1);
8715 Op = CallInst::Create(Int, Op, CI->getName(), CI);
8717 CI->replaceAllUsesWith(Op);
8718 CI->eraseFromParent();
8722 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
8723 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
8724 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
8726 std::string AsmStr = IA->getAsmString();
8728 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
8729 std::vector<std::string> AsmPieces;
8730 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
8732 switch (AsmPieces.size()) {
8733 default: return false;
8735 AsmStr = AsmPieces[0];
8737 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
8740 if (AsmPieces.size() == 2 &&
8741 (AsmPieces[0] == "bswap" ||
8742 AsmPieces[0] == "bswapq" ||
8743 AsmPieces[0] == "bswapl") &&
8744 (AsmPieces[1] == "$0" ||
8745 AsmPieces[1] == "${0:q}")) {
8746 // No need to check constraints, nothing other than the equivalent of
8747 // "=r,0" would be valid here.
8748 return LowerToBSwap(CI);
8750 // rorw $$8, ${0:w} --> llvm.bswap.i16
8751 if (CI->getType() == Type::Int16Ty &&
8752 AsmPieces.size() == 3 &&
8753 AsmPieces[0] == "rorw" &&
8754 AsmPieces[1] == "$$8," &&
8755 AsmPieces[2] == "${0:w}" &&
8756 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
8757 return LowerToBSwap(CI);
8761 if (CI->getType() == Type::Int64Ty && Constraints.size() >= 2 &&
8762 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
8763 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
8764 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
8765 std::vector<std::string> Words;
8766 SplitString(AsmPieces[0], Words, " \t");
8767 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
8769 SplitString(AsmPieces[1], Words, " \t");
8770 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
8772 SplitString(AsmPieces[2], Words, " \t,");
8773 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
8774 Words[2] == "%edx") {
8775 return LowerToBSwap(CI);
8787 /// getConstraintType - Given a constraint letter, return the type of
8788 /// constraint it is for this target.
8789 X86TargetLowering::ConstraintType
8790 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
8791 if (Constraint.size() == 1) {
8792 switch (Constraint[0]) {
8804 return C_RegisterClass;
8812 return TargetLowering::getConstraintType(Constraint);
8815 /// LowerXConstraint - try to replace an X constraint, which matches anything,
8816 /// with another that has more specific requirements based on the type of the
8817 /// corresponding operand.
8818 const char *X86TargetLowering::
8819 LowerXConstraint(MVT ConstraintVT) const {
8820 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
8821 // 'f' like normal targets.
8822 if (ConstraintVT.isFloatingPoint()) {
8823 if (Subtarget->hasSSE2())
8825 if (Subtarget->hasSSE1())
8829 return TargetLowering::LowerXConstraint(ConstraintVT);
8832 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8833 /// vector. If it is invalid, don't add anything to Ops.
8834 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8837 std::vector<SDValue>&Ops,
8838 SelectionDAG &DAG) const {
8839 SDValue Result(0, 0);
8841 switch (Constraint) {
8844 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8845 if (C->getZExtValue() <= 31) {
8846 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8852 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8853 if (C->getZExtValue() <= 63) {
8854 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8860 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8861 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
8862 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8868 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8869 if (C->getZExtValue() <= 255) {
8870 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8876 // 32-bit signed value
8877 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8878 const ConstantInt *CI = C->getConstantIntValue();
8879 if (CI->isValueValidForType(Type::Int32Ty, C->getSExtValue())) {
8880 // Widen to 64 bits here to get it sign extended.
8881 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
8884 // FIXME gcc accepts some relocatable values here too, but only in certain
8885 // memory models; it's complicated.
8890 // 32-bit unsigned value
8891 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8892 const ConstantInt *CI = C->getConstantIntValue();
8893 if (CI->isValueValidForType(Type::Int32Ty, C->getZExtValue())) {
8894 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8898 // FIXME gcc accepts some relocatable values here too, but only in certain
8899 // memory models; it's complicated.
8903 // Literal immediates are always ok.
8904 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
8905 // Widen to 64 bits here to get it sign extended.
8906 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
8910 // If we are in non-pic codegen mode, we allow the address of a global (with
8911 // an optional displacement) to be used with 'i'.
8912 GlobalAddressSDNode *GA = 0;
8915 // Match either (GA), (GA+C), (GA+C1+C2), etc.
8917 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
8918 Offset += GA->getOffset();
8920 } else if (Op.getOpcode() == ISD::ADD) {
8921 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
8922 Offset += C->getZExtValue();
8923 Op = Op.getOperand(0);
8926 } else if (Op.getOpcode() == ISD::SUB) {
8927 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
8928 Offset += -C->getZExtValue();
8929 Op = Op.getOperand(0);
8934 // Otherwise, this isn't something we can handle, reject it.
8938 GlobalValue *GV = GA->getGlobal();
8939 // If we require an extra load to get this address, as in PIC mode, we
8941 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
8942 getTargetMachine())))
8946 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
8948 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
8954 if (Result.getNode()) {
8955 Ops.push_back(Result);
8958 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
8962 std::vector<unsigned> X86TargetLowering::
8963 getRegClassForInlineAsmConstraint(const std::string &Constraint,
8965 if (Constraint.size() == 1) {
8966 // FIXME: not handling fp-stack yet!
8967 switch (Constraint[0]) { // GCC X86 Constraint Letters
8968 default: break; // Unknown constraint letter
8969 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
8970 if (Subtarget->is64Bit()) {
8972 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
8973 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
8974 X86::R10D,X86::R11D,X86::R12D,
8975 X86::R13D,X86::R14D,X86::R15D,
8976 X86::EBP, X86::ESP, 0);
8977 else if (VT == MVT::i16)
8978 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
8979 X86::SI, X86::DI, X86::R8W,X86::R9W,
8980 X86::R10W,X86::R11W,X86::R12W,
8981 X86::R13W,X86::R14W,X86::R15W,
8982 X86::BP, X86::SP, 0);
8983 else if (VT == MVT::i8)
8984 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
8985 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
8986 X86::R10B,X86::R11B,X86::R12B,
8987 X86::R13B,X86::R14B,X86::R15B,
8988 X86::BPL, X86::SPL, 0);
8990 else if (VT == MVT::i64)
8991 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
8992 X86::RSI, X86::RDI, X86::R8, X86::R9,
8993 X86::R10, X86::R11, X86::R12,
8994 X86::R13, X86::R14, X86::R15,
8995 X86::RBP, X86::RSP, 0);
8999 // 32-bit fallthrough
9002 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
9003 else if (VT == MVT::i16)
9004 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
9005 else if (VT == MVT::i8)
9006 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
9007 else if (VT == MVT::i64)
9008 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
9013 return std::vector<unsigned>();
9016 std::pair<unsigned, const TargetRegisterClass*>
9017 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9019 // First, see if this is a constraint that directly corresponds to an LLVM
9021 if (Constraint.size() == 1) {
9022 // GCC Constraint Letters
9023 switch (Constraint[0]) {
9025 case 'r': // GENERAL_REGS
9026 case 'R': // LEGACY_REGS
9027 case 'l': // INDEX_REGS
9029 return std::make_pair(0U, X86::GR8RegisterClass);
9031 return std::make_pair(0U, X86::GR16RegisterClass);
9032 if (VT == MVT::i32 || !Subtarget->is64Bit())
9033 return std::make_pair(0U, X86::GR32RegisterClass);
9034 return std::make_pair(0U, X86::GR64RegisterClass);
9035 case 'f': // FP Stack registers.
9036 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
9037 // value to the correct fpstack register class.
9038 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
9039 return std::make_pair(0U, X86::RFP32RegisterClass);
9040 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
9041 return std::make_pair(0U, X86::RFP64RegisterClass);
9042 return std::make_pair(0U, X86::RFP80RegisterClass);
9043 case 'y': // MMX_REGS if MMX allowed.
9044 if (!Subtarget->hasMMX()) break;
9045 return std::make_pair(0U, X86::VR64RegisterClass);
9046 case 'Y': // SSE_REGS if SSE2 allowed
9047 if (!Subtarget->hasSSE2()) break;
9049 case 'x': // SSE_REGS if SSE1 allowed
9050 if (!Subtarget->hasSSE1()) break;
9052 switch (VT.getSimpleVT()) {
9054 // Scalar SSE types.
9057 return std::make_pair(0U, X86::FR32RegisterClass);
9060 return std::make_pair(0U, X86::FR64RegisterClass);
9068 return std::make_pair(0U, X86::VR128RegisterClass);
9074 // Use the default implementation in TargetLowering to convert the register
9075 // constraint into a member of a register class.
9076 std::pair<unsigned, const TargetRegisterClass*> Res;
9077 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9079 // Not found as a standard register?
9080 if (Res.second == 0) {
9081 // GCC calls "st(0)" just plain "st".
9082 if (StringsEqualNoCase("{st}", Constraint)) {
9083 Res.first = X86::ST0;
9084 Res.second = X86::RFP80RegisterClass;
9086 // 'A' means EAX + EDX.
9087 if (Constraint == "A") {
9088 Res.first = X86::EAX;
9089 Res.second = X86::GR32_ADRegisterClass;
9094 // Otherwise, check to see if this is a register class of the wrong value
9095 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
9096 // turn into {ax},{dx}.
9097 if (Res.second->hasType(VT))
9098 return Res; // Correct type already, nothing to do.
9100 // All of the single-register GCC register classes map their values onto
9101 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
9102 // really want an 8-bit or 32-bit register, map to the appropriate register
9103 // class and return the appropriate register.
9104 if (Res.second == X86::GR16RegisterClass) {
9105 if (VT == MVT::i8) {
9106 unsigned DestReg = 0;
9107 switch (Res.first) {
9109 case X86::AX: DestReg = X86::AL; break;
9110 case X86::DX: DestReg = X86::DL; break;
9111 case X86::CX: DestReg = X86::CL; break;
9112 case X86::BX: DestReg = X86::BL; break;
9115 Res.first = DestReg;
9116 Res.second = X86::GR8RegisterClass;
9118 } else if (VT == MVT::i32) {
9119 unsigned DestReg = 0;
9120 switch (Res.first) {
9122 case X86::AX: DestReg = X86::EAX; break;
9123 case X86::DX: DestReg = X86::EDX; break;
9124 case X86::CX: DestReg = X86::ECX; break;
9125 case X86::BX: DestReg = X86::EBX; break;
9126 case X86::SI: DestReg = X86::ESI; break;
9127 case X86::DI: DestReg = X86::EDI; break;
9128 case X86::BP: DestReg = X86::EBP; break;
9129 case X86::SP: DestReg = X86::ESP; break;
9132 Res.first = DestReg;
9133 Res.second = X86::GR32RegisterClass;
9135 } else if (VT == MVT::i64) {
9136 unsigned DestReg = 0;
9137 switch (Res.first) {
9139 case X86::AX: DestReg = X86::RAX; break;
9140 case X86::DX: DestReg = X86::RDX; break;
9141 case X86::CX: DestReg = X86::RCX; break;
9142 case X86::BX: DestReg = X86::RBX; break;
9143 case X86::SI: DestReg = X86::RSI; break;
9144 case X86::DI: DestReg = X86::RDI; break;
9145 case X86::BP: DestReg = X86::RBP; break;
9146 case X86::SP: DestReg = X86::RSP; break;
9149 Res.first = DestReg;
9150 Res.second = X86::GR64RegisterClass;
9153 } else if (Res.second == X86::FR32RegisterClass ||
9154 Res.second == X86::FR64RegisterClass ||
9155 Res.second == X86::VR128RegisterClass) {
9156 // Handle references to XMM physical registers that got mapped into the
9157 // wrong class. This can happen with constraints like {xmm0} where the
9158 // target independent register mapper will just pick the first match it can
9159 // find, ignoring the required type.
9161 Res.second = X86::FR32RegisterClass;
9162 else if (VT == MVT::f64)
9163 Res.second = X86::FR64RegisterClass;
9164 else if (X86::VR128RegisterClass->hasType(VT))
9165 Res.second = X86::VR128RegisterClass;
9171 //===----------------------------------------------------------------------===//
9172 // X86 Widen vector type
9173 //===----------------------------------------------------------------------===//
9175 /// getWidenVectorType: given a vector type, returns the type to widen
9176 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
9177 /// If there is no vector type that we want to widen to, returns MVT::Other
9178 /// When and where to widen is target dependent based on the cost of
9179 /// scalarizing vs using the wider vector type.
9181 MVT X86TargetLowering::getWidenVectorType(MVT VT) const {
9182 assert(VT.isVector());
9183 if (isTypeLegal(VT))
9186 // TODO: In computeRegisterProperty, we can compute the list of legal vector
9187 // type based on element type. This would speed up our search (though
9188 // it may not be worth it since the size of the list is relatively
9190 MVT EltVT = VT.getVectorElementType();
9191 unsigned NElts = VT.getVectorNumElements();
9193 // On X86, it make sense to widen any vector wider than 1
9197 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
9198 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
9199 MVT SVT = (MVT::SimpleValueType)nVT;
9201 if (isTypeLegal(SVT) &&
9202 SVT.getVectorElementType() == EltVT &&
9203 SVT.getVectorNumElements() > NElts)