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 "X86TargetObjectFile.h"
20 #include "llvm/CallingConv.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/GlobalAlias.h"
24 #include "llvm/GlobalVariable.h"
25 #include "llvm/Function.h"
26 #include "llvm/Instructions.h"
27 #include "llvm/Intrinsics.h"
28 #include "llvm/LLVMContext.h"
29 #include "llvm/ADT/BitVector.h"
30 #include "llvm/ADT/VectorExtras.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/PseudoSourceValue.h"
37 #include "llvm/Support/MathExtras.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/ErrorHandling.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 // Disable16Bit - 16-bit operations typically have a larger encoding than
51 // corresponding 32-bit instructions, and 16-bit code is slow on some
52 // processors. This is an experimental flag to disable 16-bit operations
53 // (which forces them to be Legalized to 32-bit operations).
55 Disable16Bit("disable-16bit", cl::Hidden,
56 cl::desc("Disable use of 16-bit instructions"));
58 // Forward declarations.
59 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
62 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
63 switch (TM.getSubtarget<X86Subtarget>().TargetType) {
64 default: llvm_unreachable("unknown subtarget type");
65 case X86Subtarget::isDarwin:
66 if (TM.getSubtarget<X86Subtarget>().is64Bit())
67 return new X8664_MachoTargetObjectFile();
68 return new X8632_MachoTargetObjectFile();
69 case X86Subtarget::isELF:
70 return new TargetLoweringObjectFileELF();
71 case X86Subtarget::isMingw:
72 case X86Subtarget::isCygwin:
73 case X86Subtarget::isWindows:
74 return new TargetLoweringObjectFileCOFF();
79 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
80 : TargetLowering(TM, createTLOF(TM)) {
81 Subtarget = &TM.getSubtarget<X86Subtarget>();
82 X86ScalarSSEf64 = Subtarget->hasSSE2();
83 X86ScalarSSEf32 = Subtarget->hasSSE1();
84 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
86 RegInfo = TM.getRegisterInfo();
89 // Set up the TargetLowering object.
91 // X86 is weird, it always uses i8 for shift amounts and setcc results.
92 setShiftAmountType(MVT::i8);
93 setBooleanContents(ZeroOrOneBooleanContent);
94 setSchedulingPreference(SchedulingForRegPressure);
95 setStackPointerRegisterToSaveRestore(X86StackPtr);
97 if (Subtarget->isTargetDarwin()) {
98 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
99 setUseUnderscoreSetJmp(false);
100 setUseUnderscoreLongJmp(false);
101 } else if (Subtarget->isTargetMingw()) {
102 // MS runtime is weird: it exports _setjmp, but longjmp!
103 setUseUnderscoreSetJmp(true);
104 setUseUnderscoreLongJmp(false);
106 setUseUnderscoreSetJmp(true);
107 setUseUnderscoreLongJmp(true);
110 // Set up the register classes.
111 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
113 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
114 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
115 if (Subtarget->is64Bit())
116 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
118 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
120 // We don't accept any truncstore of integer registers.
121 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
123 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
124 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
126 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
127 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
128 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
130 // SETOEQ and SETUNE require checking two conditions.
131 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
132 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
133 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
134 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
135 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
136 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
138 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
140 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
141 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
142 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
144 if (Subtarget->is64Bit()) {
145 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
146 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
147 } else if (!UseSoftFloat) {
148 if (X86ScalarSSEf64) {
149 // We have an impenetrably clever algorithm for ui64->double only.
150 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
152 // We have an algorithm for SSE2, and we turn this into a 64-bit
153 // FILD for other targets.
154 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
157 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
159 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
160 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
163 // SSE has no i16 to fp conversion, only i32
164 if (X86ScalarSSEf32) {
165 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
166 // f32 and f64 cases are Legal, f80 case is not
167 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
169 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
170 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
173 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
174 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
177 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
178 // are Legal, f80 is custom lowered.
179 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
180 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
182 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
184 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
185 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
187 if (X86ScalarSSEf32) {
188 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
189 // f32 and f64 cases are Legal, f80 case is not
190 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
192 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
193 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
196 // Handle FP_TO_UINT by promoting the destination to a larger signed
198 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
199 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
200 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
202 if (Subtarget->is64Bit()) {
203 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
204 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
205 } else if (!UseSoftFloat) {
206 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
207 // Expand FP_TO_UINT into a select.
208 // FIXME: We would like to use a Custom expander here eventually to do
209 // the optimal thing for SSE vs. the default expansion in the legalizer.
210 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
212 // With SSE3 we can use fisttpll to convert to a signed i64; without
213 // SSE, we're stuck with a fistpll.
214 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
217 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
218 if (!X86ScalarSSEf64) {
219 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
220 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
223 // Scalar integer divide and remainder are lowered to use operations that
224 // produce two results, to match the available instructions. This exposes
225 // the two-result form to trivial CSE, which is able to combine x/y and x%y
226 // into a single instruction.
228 // Scalar integer multiply-high is also lowered to use two-result
229 // operations, to match the available instructions. However, plain multiply
230 // (low) operations are left as Legal, as there are single-result
231 // instructions for this in x86. Using the two-result multiply instructions
232 // when both high and low results are needed must be arranged by dagcombine.
233 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
234 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
235 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
236 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
237 setOperationAction(ISD::SREM , MVT::i8 , Expand);
238 setOperationAction(ISD::UREM , MVT::i8 , Expand);
239 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
240 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
241 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
242 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
243 setOperationAction(ISD::SREM , MVT::i16 , Expand);
244 setOperationAction(ISD::UREM , MVT::i16 , Expand);
245 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
246 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
247 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
248 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
249 setOperationAction(ISD::SREM , MVT::i32 , Expand);
250 setOperationAction(ISD::UREM , MVT::i32 , Expand);
251 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
252 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
253 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
254 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
255 setOperationAction(ISD::SREM , MVT::i64 , Expand);
256 setOperationAction(ISD::UREM , MVT::i64 , Expand);
258 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
259 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
260 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
261 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
262 if (Subtarget->is64Bit())
263 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
264 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
265 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
266 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
267 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
268 setOperationAction(ISD::FREM , MVT::f32 , Expand);
269 setOperationAction(ISD::FREM , MVT::f64 , Expand);
270 setOperationAction(ISD::FREM , MVT::f80 , Expand);
271 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
273 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
274 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
275 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
276 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
278 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
279 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
281 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
282 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
284 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
285 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
286 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
287 if (Subtarget->is64Bit()) {
288 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
289 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
290 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
293 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
294 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
296 // These should be promoted to a larger select which is supported.
297 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
298 // X86 wants to expand cmov itself.
299 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
301 setOperationAction(ISD::SELECT , MVT::i16 , Expand);
303 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
304 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
305 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
306 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
307 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
308 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
310 setOperationAction(ISD::SETCC , MVT::i16 , Expand);
312 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
313 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
314 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
315 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
316 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
317 if (Subtarget->is64Bit()) {
318 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
319 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
321 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
324 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
325 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
326 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
327 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
328 if (Subtarget->is64Bit())
329 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
330 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
331 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
332 if (Subtarget->is64Bit()) {
333 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
334 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
335 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
336 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
337 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
339 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
340 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
341 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
342 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
343 if (Subtarget->is64Bit()) {
344 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
345 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
346 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
349 if (Subtarget->hasSSE1())
350 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
352 if (!Subtarget->hasSSE2())
353 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
355 // Expand certain atomics
356 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
357 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
358 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
359 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
361 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
362 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
363 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
364 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
366 if (!Subtarget->is64Bit()) {
367 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
368 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
369 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
370 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
371 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
372 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
373 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
376 // FIXME - use subtarget debug flags
377 if (!Subtarget->isTargetDarwin() &&
378 !Subtarget->isTargetELF() &&
379 !Subtarget->isTargetCygMing()) {
380 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
383 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
384 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
385 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
386 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
387 if (Subtarget->is64Bit()) {
388 setExceptionPointerRegister(X86::RAX);
389 setExceptionSelectorRegister(X86::RDX);
391 setExceptionPointerRegister(X86::EAX);
392 setExceptionSelectorRegister(X86::EDX);
394 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
395 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
397 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
399 setOperationAction(ISD::TRAP, MVT::Other, Legal);
401 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
402 setOperationAction(ISD::VASTART , MVT::Other, Custom);
403 setOperationAction(ISD::VAEND , MVT::Other, Expand);
404 if (Subtarget->is64Bit()) {
405 setOperationAction(ISD::VAARG , MVT::Other, Custom);
406 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
408 setOperationAction(ISD::VAARG , MVT::Other, Expand);
409 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
412 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
413 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
414 if (Subtarget->is64Bit())
415 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
416 if (Subtarget->isTargetCygMing())
417 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
419 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
421 if (!UseSoftFloat && X86ScalarSSEf64) {
422 // f32 and f64 use SSE.
423 // Set up the FP register classes.
424 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
425 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
427 // Use ANDPD to simulate FABS.
428 setOperationAction(ISD::FABS , MVT::f64, Custom);
429 setOperationAction(ISD::FABS , MVT::f32, Custom);
431 // Use XORP to simulate FNEG.
432 setOperationAction(ISD::FNEG , MVT::f64, Custom);
433 setOperationAction(ISD::FNEG , MVT::f32, Custom);
435 // Use ANDPD and ORPD to simulate FCOPYSIGN.
436 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
437 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
439 // We don't support sin/cos/fmod
440 setOperationAction(ISD::FSIN , MVT::f64, Expand);
441 setOperationAction(ISD::FCOS , MVT::f64, Expand);
442 setOperationAction(ISD::FSIN , MVT::f32, Expand);
443 setOperationAction(ISD::FCOS , MVT::f32, Expand);
445 // Expand FP immediates into loads from the stack, except for the special
447 addLegalFPImmediate(APFloat(+0.0)); // xorpd
448 addLegalFPImmediate(APFloat(+0.0f)); // xorps
449 } else if (!UseSoftFloat && X86ScalarSSEf32) {
450 // Use SSE for f32, x87 for f64.
451 // Set up the FP register classes.
452 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
453 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
455 // Use ANDPS to simulate FABS.
456 setOperationAction(ISD::FABS , MVT::f32, Custom);
458 // Use XORP to simulate FNEG.
459 setOperationAction(ISD::FNEG , MVT::f32, Custom);
461 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
463 // Use ANDPS and ORPS to simulate FCOPYSIGN.
464 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
465 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
467 // We don't support sin/cos/fmod
468 setOperationAction(ISD::FSIN , MVT::f32, Expand);
469 setOperationAction(ISD::FCOS , MVT::f32, Expand);
471 // Special cases we handle for FP constants.
472 addLegalFPImmediate(APFloat(+0.0f)); // xorps
473 addLegalFPImmediate(APFloat(+0.0)); // FLD0
474 addLegalFPImmediate(APFloat(+1.0)); // FLD1
475 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
476 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
479 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
480 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
482 } else if (!UseSoftFloat) {
483 // f32 and f64 in x87.
484 // Set up the FP register classes.
485 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
486 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
488 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
489 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
490 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
491 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
494 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
495 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
497 addLegalFPImmediate(APFloat(+0.0)); // FLD0
498 addLegalFPImmediate(APFloat(+1.0)); // FLD1
499 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
500 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
501 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
502 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
503 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
504 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
507 // Long double always uses X87.
509 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
510 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
511 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
514 APFloat TmpFlt(+0.0);
515 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
517 addLegalFPImmediate(TmpFlt); // FLD0
519 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
520 APFloat TmpFlt2(+1.0);
521 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
523 addLegalFPImmediate(TmpFlt2); // FLD1
524 TmpFlt2.changeSign();
525 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
529 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
530 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
534 // Always use a library call for pow.
535 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
536 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
537 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
539 setOperationAction(ISD::FLOG, MVT::f80, Expand);
540 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
541 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
542 setOperationAction(ISD::FEXP, MVT::f80, Expand);
543 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
545 // First set operation action for all vector types to either promote
546 // (for widening) or expand (for scalarization). Then we will selectively
547 // turn on ones that can be effectively codegen'd.
548 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
549 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
550 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
565 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
566 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
599 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
601 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
603 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
604 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
605 setTruncStoreAction((MVT::SimpleValueType)VT,
606 (MVT::SimpleValueType)InnerVT, Expand);
607 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
608 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
609 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
612 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
613 // with -msoft-float, disable use of MMX as well.
614 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
615 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
616 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
617 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
618 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
619 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
621 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
622 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
623 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
624 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
626 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
627 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
628 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
629 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
631 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
632 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
634 setOperationAction(ISD::AND, MVT::v8i8, Promote);
635 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
636 setOperationAction(ISD::AND, MVT::v4i16, Promote);
637 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
638 setOperationAction(ISD::AND, MVT::v2i32, Promote);
639 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
640 setOperationAction(ISD::AND, MVT::v1i64, Legal);
642 setOperationAction(ISD::OR, MVT::v8i8, Promote);
643 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
644 setOperationAction(ISD::OR, MVT::v4i16, Promote);
645 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
646 setOperationAction(ISD::OR, MVT::v2i32, Promote);
647 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
648 setOperationAction(ISD::OR, MVT::v1i64, Legal);
650 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
651 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
652 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
653 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
654 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
655 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
656 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
658 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
659 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
660 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
661 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
662 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
663 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
664 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
665 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
666 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
668 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
669 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
670 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
671 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
672 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
674 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
675 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
676 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
677 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
679 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
680 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
681 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
682 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
684 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
686 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
687 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
688 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
689 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
690 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
691 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
692 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
695 if (!UseSoftFloat && Subtarget->hasSSE1()) {
696 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
698 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
699 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
700 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
701 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
702 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
703 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
704 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
705 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
706 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
707 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
708 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
709 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
712 if (!UseSoftFloat && Subtarget->hasSSE2()) {
713 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
715 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
716 // registers cannot be used even for integer operations.
717 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
718 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
719 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
720 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
722 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
723 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
724 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
725 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
726 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
727 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
728 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
729 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
730 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
731 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
732 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
733 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
734 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
735 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
736 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
737 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
739 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
740 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
741 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
742 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
744 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
745 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
746 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
747 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
748 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
750 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
751 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
752 EVT VT = (MVT::SimpleValueType)i;
753 // Do not attempt to custom lower non-power-of-2 vectors
754 if (!isPowerOf2_32(VT.getVectorNumElements()))
756 // Do not attempt to custom lower non-128-bit vectors
757 if (!VT.is128BitVector())
759 setOperationAction(ISD::BUILD_VECTOR,
760 VT.getSimpleVT().SimpleTy, Custom);
761 setOperationAction(ISD::VECTOR_SHUFFLE,
762 VT.getSimpleVT().SimpleTy, Custom);
763 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
764 VT.getSimpleVT().SimpleTy, Custom);
767 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
768 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
769 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
770 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
771 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
772 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
774 if (Subtarget->is64Bit()) {
775 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
776 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
779 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
780 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
781 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
784 // Do not attempt to promote non-128-bit vectors
785 if (!VT.is128BitVector()) {
788 setOperationAction(ISD::AND, SVT, Promote);
789 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
790 setOperationAction(ISD::OR, SVT, Promote);
791 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
792 setOperationAction(ISD::XOR, SVT, Promote);
793 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
794 setOperationAction(ISD::LOAD, SVT, Promote);
795 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
796 setOperationAction(ISD::SELECT, SVT, Promote);
797 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
800 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
802 // Custom lower v2i64 and v2f64 selects.
803 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
804 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
805 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
806 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
808 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
809 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
810 if (!DisableMMX && Subtarget->hasMMX()) {
811 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
812 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
816 if (Subtarget->hasSSE41()) {
817 // FIXME: Do we need to handle scalar-to-vector here?
818 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
820 // i8 and i16 vectors are custom , because the source register and source
821 // source memory operand types are not the same width. f32 vectors are
822 // custom since the immediate controlling the insert encodes additional
824 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
825 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
826 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
827 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
829 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
830 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
831 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
832 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
834 if (Subtarget->is64Bit()) {
835 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
836 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
840 if (Subtarget->hasSSE42()) {
841 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
844 if (!UseSoftFloat && Subtarget->hasAVX()) {
845 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
846 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
847 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
848 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
850 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
851 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
852 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
853 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
854 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
855 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
856 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
857 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
858 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
859 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
860 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
861 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
862 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
863 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
864 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
866 // Operations to consider commented out -v16i16 v32i8
867 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
868 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
869 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
870 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
871 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
872 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
873 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
874 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
875 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
876 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
877 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
878 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
879 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
880 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
882 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
883 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
884 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
885 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
887 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
888 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
889 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
890 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
891 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
893 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
894 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
895 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
896 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
897 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
898 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
901 // Not sure we want to do this since there are no 256-bit integer
904 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
905 // This includes 256-bit vectors
906 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
907 EVT VT = (MVT::SimpleValueType)i;
909 // Do not attempt to custom lower non-power-of-2 vectors
910 if (!isPowerOf2_32(VT.getVectorNumElements()))
913 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
914 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
915 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
918 if (Subtarget->is64Bit()) {
919 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
920 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
925 // Not sure we want to do this since there are no 256-bit integer
928 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
929 // Including 256-bit vectors
930 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
931 EVT VT = (MVT::SimpleValueType)i;
933 if (!VT.is256BitVector()) {
936 setOperationAction(ISD::AND, VT, Promote);
937 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
938 setOperationAction(ISD::OR, VT, Promote);
939 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
940 setOperationAction(ISD::XOR, VT, Promote);
941 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
942 setOperationAction(ISD::LOAD, VT, Promote);
943 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
944 setOperationAction(ISD::SELECT, VT, Promote);
945 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
948 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
952 // We want to custom lower some of our intrinsics.
953 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
955 // Add/Sub/Mul with overflow operations are custom lowered.
956 setOperationAction(ISD::SADDO, MVT::i32, Custom);
957 setOperationAction(ISD::SADDO, MVT::i64, Custom);
958 setOperationAction(ISD::UADDO, MVT::i32, Custom);
959 setOperationAction(ISD::UADDO, MVT::i64, Custom);
960 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
961 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
962 setOperationAction(ISD::USUBO, MVT::i32, Custom);
963 setOperationAction(ISD::USUBO, MVT::i64, Custom);
964 setOperationAction(ISD::SMULO, MVT::i32, Custom);
965 setOperationAction(ISD::SMULO, MVT::i64, Custom);
967 if (!Subtarget->is64Bit()) {
968 // These libcalls are not available in 32-bit.
969 setLibcallName(RTLIB::SHL_I128, 0);
970 setLibcallName(RTLIB::SRL_I128, 0);
971 setLibcallName(RTLIB::SRA_I128, 0);
974 // We have target-specific dag combine patterns for the following nodes:
975 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
976 setTargetDAGCombine(ISD::BUILD_VECTOR);
977 setTargetDAGCombine(ISD::SELECT);
978 setTargetDAGCombine(ISD::SHL);
979 setTargetDAGCombine(ISD::SRA);
980 setTargetDAGCombine(ISD::SRL);
981 setTargetDAGCombine(ISD::STORE);
982 setTargetDAGCombine(ISD::MEMBARRIER);
983 if (Subtarget->is64Bit())
984 setTargetDAGCombine(ISD::MUL);
986 computeRegisterProperties();
988 // Divide and reminder operations have no vector equivalent and can
989 // trap. Do a custom widening for these operations in which we never
990 // generate more divides/remainder than the original vector width.
991 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
992 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
993 if (!isTypeLegal((MVT::SimpleValueType)VT)) {
994 setOperationAction(ISD::SDIV, (MVT::SimpleValueType) VT, Custom);
995 setOperationAction(ISD::UDIV, (MVT::SimpleValueType) VT, Custom);
996 setOperationAction(ISD::SREM, (MVT::SimpleValueType) VT, Custom);
997 setOperationAction(ISD::UREM, (MVT::SimpleValueType) VT, Custom);
1001 // FIXME: These should be based on subtarget info. Plus, the values should
1002 // be smaller when we are in optimizing for size mode.
1003 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1004 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
1005 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1006 setPrefLoopAlignment(16);
1007 benefitFromCodePlacementOpt = true;
1011 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1016 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1017 /// the desired ByVal argument alignment.
1018 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1021 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1022 if (VTy->getBitWidth() == 128)
1024 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1025 unsigned EltAlign = 0;
1026 getMaxByValAlign(ATy->getElementType(), EltAlign);
1027 if (EltAlign > MaxAlign)
1028 MaxAlign = EltAlign;
1029 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1030 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1031 unsigned EltAlign = 0;
1032 getMaxByValAlign(STy->getElementType(i), EltAlign);
1033 if (EltAlign > MaxAlign)
1034 MaxAlign = EltAlign;
1042 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1043 /// function arguments in the caller parameter area. For X86, aggregates
1044 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1045 /// are at 4-byte boundaries.
1046 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1047 if (Subtarget->is64Bit()) {
1048 // Max of 8 and alignment of type.
1049 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1056 if (Subtarget->hasSSE1())
1057 getMaxByValAlign(Ty, Align);
1061 /// getOptimalMemOpType - Returns the target specific optimal type for load
1062 /// and store operations as a result of memset, memcpy, and memmove
1063 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
1066 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1067 bool isSrcConst, bool isSrcStr,
1068 SelectionDAG &DAG) const {
1069 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1070 // linux. This is because the stack realignment code can't handle certain
1071 // cases like PR2962. This should be removed when PR2962 is fixed.
1072 const Function *F = DAG.getMachineFunction().getFunction();
1073 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1074 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1075 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1077 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1080 if (Subtarget->is64Bit() && Size >= 8)
1085 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1087 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1088 SelectionDAG &DAG) const {
1089 if (usesGlobalOffsetTable())
1090 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
1091 if (!Subtarget->is64Bit())
1092 // This doesn't have DebugLoc associated with it, but is not really the
1093 // same as a Register.
1094 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1099 /// getFunctionAlignment - Return the Log2 alignment of this function.
1100 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1101 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1104 //===----------------------------------------------------------------------===//
1105 // Return Value Calling Convention Implementation
1106 //===----------------------------------------------------------------------===//
1108 #include "X86GenCallingConv.inc"
1111 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1112 const SmallVectorImpl<EVT> &OutTys,
1113 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1114 SelectionDAG &DAG) {
1115 SmallVector<CCValAssign, 16> RVLocs;
1116 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1117 RVLocs, *DAG.getContext());
1118 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1122 X86TargetLowering::LowerReturn(SDValue Chain,
1123 CallingConv::ID CallConv, bool isVarArg,
1124 const SmallVectorImpl<ISD::OutputArg> &Outs,
1125 DebugLoc dl, SelectionDAG &DAG) {
1127 SmallVector<CCValAssign, 16> RVLocs;
1128 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1129 RVLocs, *DAG.getContext());
1130 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1132 // If this is the first return lowered for this function, add the regs to the
1133 // liveout set for the function.
1134 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1135 for (unsigned i = 0; i != RVLocs.size(); ++i)
1136 if (RVLocs[i].isRegLoc())
1137 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1142 SmallVector<SDValue, 6> RetOps;
1143 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1144 // Operand #1 = Bytes To Pop
1145 RetOps.push_back(DAG.getTargetConstant(getBytesToPopOnReturn(), MVT::i16));
1147 // Copy the result values into the output registers.
1148 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1149 CCValAssign &VA = RVLocs[i];
1150 assert(VA.isRegLoc() && "Can only return in registers!");
1151 SDValue ValToCopy = Outs[i].Val;
1153 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1154 // the RET instruction and handled by the FP Stackifier.
1155 if (VA.getLocReg() == X86::ST0 ||
1156 VA.getLocReg() == X86::ST1) {
1157 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1158 // change the value to the FP stack register class.
1159 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1160 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1161 RetOps.push_back(ValToCopy);
1162 // Don't emit a copytoreg.
1166 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1167 // which is returned in RAX / RDX.
1168 if (Subtarget->is64Bit()) {
1169 EVT ValVT = ValToCopy.getValueType();
1170 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1171 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1172 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1173 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1177 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1178 Flag = Chain.getValue(1);
1181 // The x86-64 ABI for returning structs by value requires that we copy
1182 // the sret argument into %rax for the return. We saved the argument into
1183 // a virtual register in the entry block, so now we copy the value out
1185 if (Subtarget->is64Bit() &&
1186 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1187 MachineFunction &MF = DAG.getMachineFunction();
1188 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1189 unsigned Reg = FuncInfo->getSRetReturnReg();
1191 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1192 FuncInfo->setSRetReturnReg(Reg);
1194 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1196 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1197 Flag = Chain.getValue(1);
1199 // RAX now acts like a return value.
1200 MF.getRegInfo().addLiveOut(X86::RAX);
1203 RetOps[0] = Chain; // Update chain.
1205 // Add the flag if we have it.
1207 RetOps.push_back(Flag);
1209 return DAG.getNode(X86ISD::RET_FLAG, dl,
1210 MVT::Other, &RetOps[0], RetOps.size());
1213 /// LowerCallResult - Lower the result values of a call into the
1214 /// appropriate copies out of appropriate physical registers.
1217 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1218 CallingConv::ID CallConv, bool isVarArg,
1219 const SmallVectorImpl<ISD::InputArg> &Ins,
1220 DebugLoc dl, SelectionDAG &DAG,
1221 SmallVectorImpl<SDValue> &InVals) {
1223 // Assign locations to each value returned by this call.
1224 SmallVector<CCValAssign, 16> RVLocs;
1225 bool Is64Bit = Subtarget->is64Bit();
1226 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1227 RVLocs, *DAG.getContext());
1228 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1230 // Copy all of the result registers out of their specified physreg.
1231 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1232 CCValAssign &VA = RVLocs[i];
1233 EVT CopyVT = VA.getValVT();
1235 // If this is x86-64, and we disabled SSE, we can't return FP values
1236 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1237 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1238 llvm_report_error("SSE register return with SSE disabled");
1241 // If this is a call to a function that returns an fp value on the floating
1242 // point stack, but where we prefer to use the value in xmm registers, copy
1243 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1244 if ((VA.getLocReg() == X86::ST0 ||
1245 VA.getLocReg() == X86::ST1) &&
1246 isScalarFPTypeInSSEReg(VA.getValVT())) {
1251 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1252 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1253 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1254 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1255 MVT::v2i64, InFlag).getValue(1);
1256 Val = Chain.getValue(0);
1257 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1258 Val, DAG.getConstant(0, MVT::i64));
1260 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1261 MVT::i64, InFlag).getValue(1);
1262 Val = Chain.getValue(0);
1264 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1266 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1267 CopyVT, InFlag).getValue(1);
1268 Val = Chain.getValue(0);
1270 InFlag = Chain.getValue(2);
1272 if (CopyVT != VA.getValVT()) {
1273 // Round the F80 the right size, which also moves to the appropriate xmm
1275 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1276 // This truncation won't change the value.
1277 DAG.getIntPtrConstant(1));
1280 InVals.push_back(Val);
1287 //===----------------------------------------------------------------------===//
1288 // C & StdCall & Fast Calling Convention implementation
1289 //===----------------------------------------------------------------------===//
1290 // StdCall calling convention seems to be standard for many Windows' API
1291 // routines and around. It differs from C calling convention just a little:
1292 // callee should clean up the stack, not caller. Symbols should be also
1293 // decorated in some fancy way :) It doesn't support any vector arguments.
1294 // For info on fast calling convention see Fast Calling Convention (tail call)
1295 // implementation LowerX86_32FastCCCallTo.
1297 /// CallIsStructReturn - Determines whether a call uses struct return
1299 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1303 return Outs[0].Flags.isSRet();
1306 /// ArgsAreStructReturn - Determines whether a function uses struct
1307 /// return semantics.
1309 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1313 return Ins[0].Flags.isSRet();
1316 /// IsCalleePop - Determines whether the callee is required to pop its
1317 /// own arguments. Callee pop is necessary to support tail calls.
1318 bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
1322 switch (CallingConv) {
1325 case CallingConv::X86_StdCall:
1326 return !Subtarget->is64Bit();
1327 case CallingConv::X86_FastCall:
1328 return !Subtarget->is64Bit();
1329 case CallingConv::Fast:
1330 return PerformTailCallOpt;
1334 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1335 /// given CallingConvention value.
1336 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1337 if (Subtarget->is64Bit()) {
1338 if (Subtarget->isTargetWin64())
1339 return CC_X86_Win64_C;
1344 if (CC == CallingConv::X86_FastCall)
1345 return CC_X86_32_FastCall;
1346 else if (CC == CallingConv::Fast)
1347 return CC_X86_32_FastCC;
1352 /// NameDecorationForCallConv - Selects the appropriate decoration to
1353 /// apply to a MachineFunction containing a given calling convention.
1355 X86TargetLowering::NameDecorationForCallConv(CallingConv::ID CallConv) {
1356 if (CallConv == CallingConv::X86_FastCall)
1358 else if (CallConv == CallingConv::X86_StdCall)
1364 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1365 /// by "Src" to address "Dst" with size and alignment information specified by
1366 /// the specific parameter attribute. The copy will be passed as a byval
1367 /// function parameter.
1369 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1370 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1372 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1373 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1374 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1378 X86TargetLowering::LowerMemArgument(SDValue Chain,
1379 CallingConv::ID CallConv,
1380 const SmallVectorImpl<ISD::InputArg> &Ins,
1381 DebugLoc dl, SelectionDAG &DAG,
1382 const CCValAssign &VA,
1383 MachineFrameInfo *MFI,
1386 // Create the nodes corresponding to a load from this parameter slot.
1387 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1388 bool AlwaysUseMutable = (CallConv==CallingConv::Fast) && PerformTailCallOpt;
1389 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1392 // If value is passed by pointer we have address passed instead of the value
1394 if (VA.getLocInfo() == CCValAssign::Indirect)
1395 ValVT = VA.getLocVT();
1397 ValVT = VA.getValVT();
1399 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1400 // changed with more analysis.
1401 // In case of tail call optimization mark all arguments mutable. Since they
1402 // could be overwritten by lowering of arguments in case of a tail call.
1403 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1404 VA.getLocMemOffset(), isImmutable, false);
1405 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1406 if (Flags.isByVal())
1408 return DAG.getLoad(ValVT, dl, Chain, FIN,
1409 PseudoSourceValue::getFixedStack(FI), 0);
1413 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1414 CallingConv::ID CallConv,
1416 const SmallVectorImpl<ISD::InputArg> &Ins,
1419 SmallVectorImpl<SDValue> &InVals) {
1421 MachineFunction &MF = DAG.getMachineFunction();
1422 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1424 const Function* Fn = MF.getFunction();
1425 if (Fn->hasExternalLinkage() &&
1426 Subtarget->isTargetCygMing() &&
1427 Fn->getName() == "main")
1428 FuncInfo->setForceFramePointer(true);
1430 // Decorate the function name.
1431 FuncInfo->setDecorationStyle(NameDecorationForCallConv(CallConv));
1433 MachineFrameInfo *MFI = MF.getFrameInfo();
1434 bool Is64Bit = Subtarget->is64Bit();
1435 bool IsWin64 = Subtarget->isTargetWin64();
1437 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1438 "Var args not supported with calling convention fastcc");
1440 // Assign locations to all of the incoming arguments.
1441 SmallVector<CCValAssign, 16> ArgLocs;
1442 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1443 ArgLocs, *DAG.getContext());
1444 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1446 unsigned LastVal = ~0U;
1448 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1449 CCValAssign &VA = ArgLocs[i];
1450 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1452 assert(VA.getValNo() != LastVal &&
1453 "Don't support value assigned to multiple locs yet");
1454 LastVal = VA.getValNo();
1456 if (VA.isRegLoc()) {
1457 EVT RegVT = VA.getLocVT();
1458 TargetRegisterClass *RC = NULL;
1459 if (RegVT == MVT::i32)
1460 RC = X86::GR32RegisterClass;
1461 else if (Is64Bit && RegVT == MVT::i64)
1462 RC = X86::GR64RegisterClass;
1463 else if (RegVT == MVT::f32)
1464 RC = X86::FR32RegisterClass;
1465 else if (RegVT == MVT::f64)
1466 RC = X86::FR64RegisterClass;
1467 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1468 RC = X86::VR128RegisterClass;
1469 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1470 RC = X86::VR64RegisterClass;
1472 llvm_unreachable("Unknown argument type!");
1474 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1475 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1477 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1478 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1480 if (VA.getLocInfo() == CCValAssign::SExt)
1481 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1482 DAG.getValueType(VA.getValVT()));
1483 else if (VA.getLocInfo() == CCValAssign::ZExt)
1484 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1485 DAG.getValueType(VA.getValVT()));
1486 else if (VA.getLocInfo() == CCValAssign::BCvt)
1487 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1489 if (VA.isExtInLoc()) {
1490 // Handle MMX values passed in XMM regs.
1491 if (RegVT.isVector()) {
1492 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1493 ArgValue, DAG.getConstant(0, MVT::i64));
1494 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1496 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1499 assert(VA.isMemLoc());
1500 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1503 // If value is passed via pointer - do a load.
1504 if (VA.getLocInfo() == CCValAssign::Indirect)
1505 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0);
1507 InVals.push_back(ArgValue);
1510 // The x86-64 ABI for returning structs by value requires that we copy
1511 // the sret argument into %rax for the return. Save the argument into
1512 // a virtual register so that we can access it from the return points.
1513 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1514 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1515 unsigned Reg = FuncInfo->getSRetReturnReg();
1517 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1518 FuncInfo->setSRetReturnReg(Reg);
1520 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1521 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1524 unsigned StackSize = CCInfo.getNextStackOffset();
1525 // align stack specially for tail calls
1526 if (PerformTailCallOpt && CallConv == CallingConv::Fast)
1527 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1529 // If the function takes variable number of arguments, make a frame index for
1530 // the start of the first vararg value... for expansion of llvm.va_start.
1532 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1533 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize, true, false);
1536 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1538 // FIXME: We should really autogenerate these arrays
1539 static const unsigned GPR64ArgRegsWin64[] = {
1540 X86::RCX, X86::RDX, X86::R8, X86::R9
1542 static const unsigned XMMArgRegsWin64[] = {
1543 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1545 static const unsigned GPR64ArgRegs64Bit[] = {
1546 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1548 static const unsigned XMMArgRegs64Bit[] = {
1549 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1550 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1552 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1555 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1556 GPR64ArgRegs = GPR64ArgRegsWin64;
1557 XMMArgRegs = XMMArgRegsWin64;
1559 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1560 GPR64ArgRegs = GPR64ArgRegs64Bit;
1561 XMMArgRegs = XMMArgRegs64Bit;
1563 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1565 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1568 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1569 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1570 "SSE register cannot be used when SSE is disabled!");
1571 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1572 "SSE register cannot be used when SSE is disabled!");
1573 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1574 // Kernel mode asks for SSE to be disabled, so don't push them
1576 TotalNumXMMRegs = 0;
1578 // For X86-64, if there are vararg parameters that are passed via
1579 // registers, then we must store them to their spots on the stack so they
1580 // may be loaded by deferencing the result of va_next.
1581 VarArgsGPOffset = NumIntRegs * 8;
1582 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1583 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1584 TotalNumXMMRegs * 16, 16,
1587 // Store the integer parameter registers.
1588 SmallVector<SDValue, 8> MemOps;
1589 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1590 unsigned Offset = VarArgsGPOffset;
1591 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1592 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1593 DAG.getIntPtrConstant(Offset));
1594 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1595 X86::GR64RegisterClass);
1596 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1598 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1599 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1601 MemOps.push_back(Store);
1605 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1606 // Now store the XMM (fp + vector) parameter registers.
1607 SmallVector<SDValue, 11> SaveXMMOps;
1608 SaveXMMOps.push_back(Chain);
1610 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1611 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1612 SaveXMMOps.push_back(ALVal);
1614 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1615 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1617 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1618 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1619 X86::VR128RegisterClass);
1620 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1621 SaveXMMOps.push_back(Val);
1623 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1625 &SaveXMMOps[0], SaveXMMOps.size()));
1628 if (!MemOps.empty())
1629 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1630 &MemOps[0], MemOps.size());
1634 // Some CCs need callee pop.
1635 if (IsCalleePop(isVarArg, CallConv)) {
1636 BytesToPopOnReturn = StackSize; // Callee pops everything.
1637 BytesCallerReserves = 0;
1639 BytesToPopOnReturn = 0; // Callee pops nothing.
1640 // If this is an sret function, the return should pop the hidden pointer.
1641 if (!Is64Bit && CallConv != CallingConv::Fast && ArgsAreStructReturn(Ins))
1642 BytesToPopOnReturn = 4;
1643 BytesCallerReserves = StackSize;
1647 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1648 if (CallConv == CallingConv::X86_FastCall)
1649 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1652 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1658 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1659 SDValue StackPtr, SDValue Arg,
1660 DebugLoc dl, SelectionDAG &DAG,
1661 const CCValAssign &VA,
1662 ISD::ArgFlagsTy Flags) {
1663 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1664 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1665 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1666 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1667 if (Flags.isByVal()) {
1668 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1670 return DAG.getStore(Chain, dl, Arg, PtrOff,
1671 PseudoSourceValue::getStack(), LocMemOffset);
1674 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1675 /// optimization is performed and it is required.
1677 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1678 SDValue &OutRetAddr,
1684 if (!IsTailCall || FPDiff==0) return Chain;
1686 // Adjust the Return address stack slot.
1687 EVT VT = getPointerTy();
1688 OutRetAddr = getReturnAddressFrameIndex(DAG);
1690 // Load the "old" Return address.
1691 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1692 return SDValue(OutRetAddr.getNode(), 1);
1695 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1696 /// optimization is performed and it is required (FPDiff!=0).
1698 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1699 SDValue Chain, SDValue RetAddrFrIdx,
1700 bool Is64Bit, int FPDiff, DebugLoc dl) {
1701 // Store the return address to the appropriate stack slot.
1702 if (!FPDiff) return Chain;
1703 // Calculate the new stack slot for the return address.
1704 int SlotSize = Is64Bit ? 8 : 4;
1705 int NewReturnAddrFI =
1706 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize,
1708 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1709 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1710 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1711 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1716 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1717 CallingConv::ID CallConv, bool isVarArg,
1719 const SmallVectorImpl<ISD::OutputArg> &Outs,
1720 const SmallVectorImpl<ISD::InputArg> &Ins,
1721 DebugLoc dl, SelectionDAG &DAG,
1722 SmallVectorImpl<SDValue> &InVals) {
1724 MachineFunction &MF = DAG.getMachineFunction();
1725 bool Is64Bit = Subtarget->is64Bit();
1726 bool IsStructRet = CallIsStructReturn(Outs);
1728 assert((!isTailCall ||
1729 (CallConv == CallingConv::Fast && PerformTailCallOpt)) &&
1730 "IsEligibleForTailCallOptimization missed a case!");
1731 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1732 "Var args not supported with calling convention fastcc");
1734 // Analyze operands of the call, assigning locations to each operand.
1735 SmallVector<CCValAssign, 16> ArgLocs;
1736 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1737 ArgLocs, *DAG.getContext());
1738 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1740 // Get a count of how many bytes are to be pushed on the stack.
1741 unsigned NumBytes = CCInfo.getNextStackOffset();
1742 if (PerformTailCallOpt && CallConv == CallingConv::Fast)
1743 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1747 // Lower arguments at fp - stackoffset + fpdiff.
1748 unsigned NumBytesCallerPushed =
1749 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1750 FPDiff = NumBytesCallerPushed - NumBytes;
1752 // Set the delta of movement of the returnaddr stackslot.
1753 // But only set if delta is greater than previous delta.
1754 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1755 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1758 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1760 SDValue RetAddrFrIdx;
1761 // Load return adress for tail calls.
1762 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall, Is64Bit,
1765 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1766 SmallVector<SDValue, 8> MemOpChains;
1769 // Walk the register/memloc assignments, inserting copies/loads. In the case
1770 // of tail call optimization arguments are handle later.
1771 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1772 CCValAssign &VA = ArgLocs[i];
1773 EVT RegVT = VA.getLocVT();
1774 SDValue Arg = Outs[i].Val;
1775 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1776 bool isByVal = Flags.isByVal();
1778 // Promote the value if needed.
1779 switch (VA.getLocInfo()) {
1780 default: llvm_unreachable("Unknown loc info!");
1781 case CCValAssign::Full: break;
1782 case CCValAssign::SExt:
1783 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1785 case CCValAssign::ZExt:
1786 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1788 case CCValAssign::AExt:
1789 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1790 // Special case: passing MMX values in XMM registers.
1791 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1792 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1793 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1795 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1797 case CCValAssign::BCvt:
1798 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1800 case CCValAssign::Indirect: {
1801 // Store the argument.
1802 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1803 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1804 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1805 PseudoSourceValue::getFixedStack(FI), 0);
1811 if (VA.isRegLoc()) {
1812 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1814 if (!isTailCall || (isTailCall && isByVal)) {
1815 assert(VA.isMemLoc());
1816 if (StackPtr.getNode() == 0)
1817 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1819 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1820 dl, DAG, VA, Flags));
1825 if (!MemOpChains.empty())
1826 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1827 &MemOpChains[0], MemOpChains.size());
1829 // Build a sequence of copy-to-reg nodes chained together with token chain
1830 // and flag operands which copy the outgoing args into registers.
1832 // Tail call byval lowering might overwrite argument registers so in case of
1833 // tail call optimization the copies to registers are lowered later.
1835 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1836 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1837 RegsToPass[i].second, InFlag);
1838 InFlag = Chain.getValue(1);
1842 if (Subtarget->isPICStyleGOT()) {
1843 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1846 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1847 DAG.getNode(X86ISD::GlobalBaseReg,
1848 DebugLoc::getUnknownLoc(),
1851 InFlag = Chain.getValue(1);
1853 // If we are tail calling and generating PIC/GOT style code load the
1854 // address of the callee into ECX. The value in ecx is used as target of
1855 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1856 // for tail calls on PIC/GOT architectures. Normally we would just put the
1857 // address of GOT into ebx and then call target@PLT. But for tail calls
1858 // ebx would be restored (since ebx is callee saved) before jumping to the
1861 // Note: The actual moving to ECX is done further down.
1862 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1863 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1864 !G->getGlobal()->hasProtectedVisibility())
1865 Callee = LowerGlobalAddress(Callee, DAG);
1866 else if (isa<ExternalSymbolSDNode>(Callee))
1867 Callee = LowerExternalSymbol(Callee, DAG);
1871 if (Is64Bit && isVarArg) {
1872 // From AMD64 ABI document:
1873 // For calls that may call functions that use varargs or stdargs
1874 // (prototype-less calls or calls to functions containing ellipsis (...) in
1875 // the declaration) %al is used as hidden argument to specify the number
1876 // of SSE registers used. The contents of %al do not need to match exactly
1877 // the number of registers, but must be an ubound on the number of SSE
1878 // registers used and is in the range 0 - 8 inclusive.
1880 // FIXME: Verify this on Win64
1881 // Count the number of XMM registers allocated.
1882 static const unsigned XMMArgRegs[] = {
1883 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1884 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1886 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1887 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1888 && "SSE registers cannot be used when SSE is disabled");
1890 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1891 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1892 InFlag = Chain.getValue(1);
1896 // For tail calls lower the arguments to the 'real' stack slot.
1898 // Force all the incoming stack arguments to be loaded from the stack
1899 // before any new outgoing arguments are stored to the stack, because the
1900 // outgoing stack slots may alias the incoming argument stack slots, and
1901 // the alias isn't otherwise explicit. This is slightly more conservative
1902 // than necessary, because it means that each store effectively depends
1903 // on every argument instead of just those arguments it would clobber.
1904 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1906 SmallVector<SDValue, 8> MemOpChains2;
1909 // Do not flag preceeding copytoreg stuff together with the following stuff.
1911 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1912 CCValAssign &VA = ArgLocs[i];
1913 if (!VA.isRegLoc()) {
1914 assert(VA.isMemLoc());
1915 SDValue Arg = Outs[i].Val;
1916 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1917 // Create frame index.
1918 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1919 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1920 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
1921 FIN = DAG.getFrameIndex(FI, getPointerTy());
1923 if (Flags.isByVal()) {
1924 // Copy relative to framepointer.
1925 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1926 if (StackPtr.getNode() == 0)
1927 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1929 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1931 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
1935 // Store relative to framepointer.
1936 MemOpChains2.push_back(
1937 DAG.getStore(ArgChain, dl, Arg, FIN,
1938 PseudoSourceValue::getFixedStack(FI), 0));
1943 if (!MemOpChains2.empty())
1944 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1945 &MemOpChains2[0], MemOpChains2.size());
1947 // Copy arguments to their registers.
1948 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1949 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1950 RegsToPass[i].second, InFlag);
1951 InFlag = Chain.getValue(1);
1955 // Store the return address to the appropriate stack slot.
1956 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1960 bool WasGlobalOrExternal = false;
1961 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
1962 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
1963 // In the 64-bit large code model, we have to make all calls
1964 // through a register, since the call instruction's 32-bit
1965 // pc-relative offset may not be large enough to hold the whole
1967 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1968 WasGlobalOrExternal = true;
1969 // If the callee is a GlobalAddress node (quite common, every direct call
1970 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
1973 // We should use extra load for direct calls to dllimported functions in
1975 GlobalValue *GV = G->getGlobal();
1976 if (!GV->hasDLLImportLinkage()) {
1977 unsigned char OpFlags = 0;
1979 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
1980 // external symbols most go through the PLT in PIC mode. If the symbol
1981 // has hidden or protected visibility, or if it is static or local, then
1982 // we don't need to use the PLT - we can directly call it.
1983 if (Subtarget->isTargetELF() &&
1984 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1985 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
1986 OpFlags = X86II::MO_PLT;
1987 } else if (Subtarget->isPICStyleStubAny() &&
1988 (GV->isDeclaration() || GV->isWeakForLinker()) &&
1989 Subtarget->getDarwinVers() < 9) {
1990 // PC-relative references to external symbols should go through $stub,
1991 // unless we're building with the leopard linker or later, which
1992 // automatically synthesizes these stubs.
1993 OpFlags = X86II::MO_DARWIN_STUB;
1996 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
1997 G->getOffset(), OpFlags);
1999 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2000 WasGlobalOrExternal = true;
2001 unsigned char OpFlags = 0;
2003 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2004 // symbols should go through the PLT.
2005 if (Subtarget->isTargetELF() &&
2006 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2007 OpFlags = X86II::MO_PLT;
2008 } else if (Subtarget->isPICStyleStubAny() &&
2009 Subtarget->getDarwinVers() < 9) {
2010 // PC-relative references to external symbols should go through $stub,
2011 // unless we're building with the leopard linker or later, which
2012 // automatically synthesizes these stubs.
2013 OpFlags = X86II::MO_DARWIN_STUB;
2016 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2020 if (isTailCall && !WasGlobalOrExternal) {
2021 unsigned Opc = Is64Bit ? X86::R11 : X86::EAX;
2023 Chain = DAG.getCopyToReg(Chain, dl,
2024 DAG.getRegister(Opc, getPointerTy()),
2026 Callee = DAG.getRegister(Opc, getPointerTy());
2027 // Add register as live out.
2028 MF.getRegInfo().addLiveOut(Opc);
2031 // Returns a chain & a flag for retval copy to use.
2032 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2033 SmallVector<SDValue, 8> Ops;
2036 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2037 DAG.getIntPtrConstant(0, true), InFlag);
2038 InFlag = Chain.getValue(1);
2041 Ops.push_back(Chain);
2042 Ops.push_back(Callee);
2045 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2047 // Add argument registers to the end of the list so that they are known live
2049 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2050 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2051 RegsToPass[i].second.getValueType()));
2053 // Add an implicit use GOT pointer in EBX.
2054 if (!isTailCall && Subtarget->isPICStyleGOT())
2055 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2057 // Add an implicit use of AL for x86 vararg functions.
2058 if (Is64Bit && isVarArg)
2059 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2061 if (InFlag.getNode())
2062 Ops.push_back(InFlag);
2065 // If this is the first return lowered for this function, add the regs
2066 // to the liveout set for the function.
2067 if (MF.getRegInfo().liveout_empty()) {
2068 SmallVector<CCValAssign, 16> RVLocs;
2069 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2071 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2072 for (unsigned i = 0; i != RVLocs.size(); ++i)
2073 if (RVLocs[i].isRegLoc())
2074 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2077 assert(((Callee.getOpcode() == ISD::Register &&
2078 (cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
2079 cast<RegisterSDNode>(Callee)->getReg() == X86::R9)) ||
2080 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2081 Callee.getOpcode() == ISD::TargetGlobalAddress) &&
2082 "Expecting an global address, external symbol, or register");
2084 return DAG.getNode(X86ISD::TC_RETURN, dl,
2085 NodeTys, &Ops[0], Ops.size());
2088 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2089 InFlag = Chain.getValue(1);
2091 // Create the CALLSEQ_END node.
2092 unsigned NumBytesForCalleeToPush;
2093 if (IsCalleePop(isVarArg, CallConv))
2094 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2095 else if (!Is64Bit && CallConv != CallingConv::Fast && IsStructRet)
2096 // If this is is a call to a struct-return function, the callee
2097 // pops the hidden struct pointer, so we have to push it back.
2098 // This is common for Darwin/X86, Linux & Mingw32 targets.
2099 NumBytesForCalleeToPush = 4;
2101 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2103 // Returns a flag for retval copy to use.
2104 Chain = DAG.getCALLSEQ_END(Chain,
2105 DAG.getIntPtrConstant(NumBytes, true),
2106 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2109 InFlag = Chain.getValue(1);
2111 // Handle result values, copying them out of physregs into vregs that we
2113 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2114 Ins, dl, DAG, InVals);
2118 //===----------------------------------------------------------------------===//
2119 // Fast Calling Convention (tail call) implementation
2120 //===----------------------------------------------------------------------===//
2122 // Like std call, callee cleans arguments, convention except that ECX is
2123 // reserved for storing the tail called function address. Only 2 registers are
2124 // free for argument passing (inreg). Tail call optimization is performed
2126 // * tailcallopt is enabled
2127 // * caller/callee are fastcc
2128 // On X86_64 architecture with GOT-style position independent code only local
2129 // (within module) calls are supported at the moment.
2130 // To keep the stack aligned according to platform abi the function
2131 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2132 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2133 // If a tail called function callee has more arguments than the caller the
2134 // caller needs to make sure that there is room to move the RETADDR to. This is
2135 // achieved by reserving an area the size of the argument delta right after the
2136 // original REtADDR, but before the saved framepointer or the spilled registers
2137 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2149 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2150 /// for a 16 byte align requirement.
2151 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2152 SelectionDAG& DAG) {
2153 MachineFunction &MF = DAG.getMachineFunction();
2154 const TargetMachine &TM = MF.getTarget();
2155 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2156 unsigned StackAlignment = TFI.getStackAlignment();
2157 uint64_t AlignMask = StackAlignment - 1;
2158 int64_t Offset = StackSize;
2159 uint64_t SlotSize = TD->getPointerSize();
2160 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2161 // Number smaller than 12 so just add the difference.
2162 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2164 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2165 Offset = ((~AlignMask) & Offset) + StackAlignment +
2166 (StackAlignment-SlotSize);
2171 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2172 /// for tail call optimization. Targets which want to do tail call
2173 /// optimization should implement this function.
2175 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2176 CallingConv::ID CalleeCC,
2178 const SmallVectorImpl<ISD::InputArg> &Ins,
2179 SelectionDAG& DAG) const {
2180 MachineFunction &MF = DAG.getMachineFunction();
2181 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2182 return CalleeCC == CallingConv::Fast && CallerCC == CalleeCC;
2186 X86TargetLowering::createFastISel(MachineFunction &mf,
2187 MachineModuleInfo *mmo,
2189 DenseMap<const Value *, unsigned> &vm,
2190 DenseMap<const BasicBlock *,
2191 MachineBasicBlock *> &bm,
2192 DenseMap<const AllocaInst *, int> &am
2194 , SmallSet<Instruction*, 8> &cil
2197 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2205 //===----------------------------------------------------------------------===//
2206 // Other Lowering Hooks
2207 //===----------------------------------------------------------------------===//
2210 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2211 MachineFunction &MF = DAG.getMachineFunction();
2212 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2213 int ReturnAddrIndex = FuncInfo->getRAIndex();
2215 if (ReturnAddrIndex == 0) {
2216 // Set up a frame object for the return address.
2217 uint64_t SlotSize = TD->getPointerSize();
2218 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2220 FuncInfo->setRAIndex(ReturnAddrIndex);
2223 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2227 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2228 bool hasSymbolicDisplacement) {
2229 // Offset should fit into 32 bit immediate field.
2230 if (!isInt32(Offset))
2233 // If we don't have a symbolic displacement - we don't have any extra
2235 if (!hasSymbolicDisplacement)
2238 // FIXME: Some tweaks might be needed for medium code model.
2239 if (M != CodeModel::Small && M != CodeModel::Kernel)
2242 // For small code model we assume that latest object is 16MB before end of 31
2243 // bits boundary. We may also accept pretty large negative constants knowing
2244 // that all objects are in the positive half of address space.
2245 if (M == CodeModel::Small && Offset < 16*1024*1024)
2248 // For kernel code model we know that all object resist in the negative half
2249 // of 32bits address space. We may not accept negative offsets, since they may
2250 // be just off and we may accept pretty large positive ones.
2251 if (M == CodeModel::Kernel && Offset > 0)
2257 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2258 /// specific condition code, returning the condition code and the LHS/RHS of the
2259 /// comparison to make.
2260 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2261 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2263 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2264 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2265 // X > -1 -> X == 0, jump !sign.
2266 RHS = DAG.getConstant(0, RHS.getValueType());
2267 return X86::COND_NS;
2268 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2269 // X < 0 -> X == 0, jump on sign.
2271 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2273 RHS = DAG.getConstant(0, RHS.getValueType());
2274 return X86::COND_LE;
2278 switch (SetCCOpcode) {
2279 default: llvm_unreachable("Invalid integer condition!");
2280 case ISD::SETEQ: return X86::COND_E;
2281 case ISD::SETGT: return X86::COND_G;
2282 case ISD::SETGE: return X86::COND_GE;
2283 case ISD::SETLT: return X86::COND_L;
2284 case ISD::SETLE: return X86::COND_LE;
2285 case ISD::SETNE: return X86::COND_NE;
2286 case ISD::SETULT: return X86::COND_B;
2287 case ISD::SETUGT: return X86::COND_A;
2288 case ISD::SETULE: return X86::COND_BE;
2289 case ISD::SETUGE: return X86::COND_AE;
2293 // First determine if it is required or is profitable to flip the operands.
2295 // If LHS is a foldable load, but RHS is not, flip the condition.
2296 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2297 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2298 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2299 std::swap(LHS, RHS);
2302 switch (SetCCOpcode) {
2308 std::swap(LHS, RHS);
2312 // On a floating point condition, the flags are set as follows:
2314 // 0 | 0 | 0 | X > Y
2315 // 0 | 0 | 1 | X < Y
2316 // 1 | 0 | 0 | X == Y
2317 // 1 | 1 | 1 | unordered
2318 switch (SetCCOpcode) {
2319 default: llvm_unreachable("Condcode should be pre-legalized away");
2321 case ISD::SETEQ: return X86::COND_E;
2322 case ISD::SETOLT: // flipped
2324 case ISD::SETGT: return X86::COND_A;
2325 case ISD::SETOLE: // flipped
2327 case ISD::SETGE: return X86::COND_AE;
2328 case ISD::SETUGT: // flipped
2330 case ISD::SETLT: return X86::COND_B;
2331 case ISD::SETUGE: // flipped
2333 case ISD::SETLE: return X86::COND_BE;
2335 case ISD::SETNE: return X86::COND_NE;
2336 case ISD::SETUO: return X86::COND_P;
2337 case ISD::SETO: return X86::COND_NP;
2339 case ISD::SETUNE: return X86::COND_INVALID;
2343 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2344 /// code. Current x86 isa includes the following FP cmov instructions:
2345 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2346 static bool hasFPCMov(unsigned X86CC) {
2362 /// isFPImmLegal - Returns true if the target can instruction select the
2363 /// specified FP immediate natively. If false, the legalizer will
2364 /// materialize the FP immediate as a load from a constant pool.
2365 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2366 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2367 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2373 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2374 /// the specified range (L, H].
2375 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2376 return (Val < 0) || (Val >= Low && Val < Hi);
2379 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2380 /// specified value.
2381 static bool isUndefOrEqual(int Val, int CmpVal) {
2382 if (Val < 0 || Val == CmpVal)
2387 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2388 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2389 /// the second operand.
2390 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2391 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2392 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2393 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2394 return (Mask[0] < 2 && Mask[1] < 2);
2398 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2399 SmallVector<int, 8> M;
2401 return ::isPSHUFDMask(M, N->getValueType(0));
2404 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2405 /// is suitable for input to PSHUFHW.
2406 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2407 if (VT != MVT::v8i16)
2410 // Lower quadword copied in order or undef.
2411 for (int i = 0; i != 4; ++i)
2412 if (Mask[i] >= 0 && Mask[i] != i)
2415 // Upper quadword shuffled.
2416 for (int i = 4; i != 8; ++i)
2417 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2423 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2424 SmallVector<int, 8> M;
2426 return ::isPSHUFHWMask(M, N->getValueType(0));
2429 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2430 /// is suitable for input to PSHUFLW.
2431 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2432 if (VT != MVT::v8i16)
2435 // Upper quadword copied in order.
2436 for (int i = 4; i != 8; ++i)
2437 if (Mask[i] >= 0 && Mask[i] != i)
2440 // Lower quadword shuffled.
2441 for (int i = 0; i != 4; ++i)
2448 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2449 SmallVector<int, 8> M;
2451 return ::isPSHUFLWMask(M, N->getValueType(0));
2454 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2455 /// is suitable for input to PALIGNR.
2456 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2458 int i, e = VT.getVectorNumElements();
2460 // Do not handle v2i64 / v2f64 shuffles with palignr.
2461 if (e < 4 || !hasSSSE3)
2464 for (i = 0; i != e; ++i)
2468 // All undef, not a palignr.
2472 // Determine if it's ok to perform a palignr with only the LHS, since we
2473 // don't have access to the actual shuffle elements to see if RHS is undef.
2474 bool Unary = Mask[i] < (int)e;
2475 bool NeedsUnary = false;
2477 int s = Mask[i] - i;
2479 // Check the rest of the elements to see if they are consecutive.
2480 for (++i; i != e; ++i) {
2485 Unary = Unary && (m < (int)e);
2486 NeedsUnary = NeedsUnary || (m < s);
2488 if (NeedsUnary && !Unary)
2490 if (Unary && m != ((s+i) & (e-1)))
2492 if (!Unary && m != (s+i))
2498 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2499 SmallVector<int, 8> M;
2501 return ::isPALIGNRMask(M, N->getValueType(0), true);
2504 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2505 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2506 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2507 int NumElems = VT.getVectorNumElements();
2508 if (NumElems != 2 && NumElems != 4)
2511 int Half = NumElems / 2;
2512 for (int i = 0; i < Half; ++i)
2513 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2515 for (int i = Half; i < NumElems; ++i)
2516 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2522 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2523 SmallVector<int, 8> M;
2525 return ::isSHUFPMask(M, N->getValueType(0));
2528 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2529 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2530 /// half elements to come from vector 1 (which would equal the dest.) and
2531 /// the upper half to come from vector 2.
2532 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2533 int NumElems = VT.getVectorNumElements();
2535 if (NumElems != 2 && NumElems != 4)
2538 int Half = NumElems / 2;
2539 for (int i = 0; i < Half; ++i)
2540 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2542 for (int i = Half; i < NumElems; ++i)
2543 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2548 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2549 SmallVector<int, 8> M;
2551 return isCommutedSHUFPMask(M, N->getValueType(0));
2554 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2555 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2556 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2557 if (N->getValueType(0).getVectorNumElements() != 4)
2560 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2561 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2562 isUndefOrEqual(N->getMaskElt(1), 7) &&
2563 isUndefOrEqual(N->getMaskElt(2), 2) &&
2564 isUndefOrEqual(N->getMaskElt(3), 3);
2567 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2568 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2570 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2571 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2576 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2577 isUndefOrEqual(N->getMaskElt(1), 3) &&
2578 isUndefOrEqual(N->getMaskElt(2), 2) &&
2579 isUndefOrEqual(N->getMaskElt(3), 3);
2582 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2583 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2584 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2585 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2587 if (NumElems != 2 && NumElems != 4)
2590 for (unsigned i = 0; i < NumElems/2; ++i)
2591 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2594 for (unsigned i = NumElems/2; i < NumElems; ++i)
2595 if (!isUndefOrEqual(N->getMaskElt(i), i))
2601 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2602 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2603 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2604 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2606 if (NumElems != 2 && NumElems != 4)
2609 for (unsigned i = 0; i < NumElems/2; ++i)
2610 if (!isUndefOrEqual(N->getMaskElt(i), i))
2613 for (unsigned i = 0; i < NumElems/2; ++i)
2614 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2620 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2621 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2622 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2623 bool V2IsSplat = false) {
2624 int NumElts = VT.getVectorNumElements();
2625 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2628 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2630 int BitI1 = Mask[i+1];
2631 if (!isUndefOrEqual(BitI, j))
2634 if (!isUndefOrEqual(BitI1, NumElts))
2637 if (!isUndefOrEqual(BitI1, j + NumElts))
2644 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2645 SmallVector<int, 8> M;
2647 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2650 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2651 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2652 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2653 bool V2IsSplat = false) {
2654 int NumElts = VT.getVectorNumElements();
2655 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2658 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2660 int BitI1 = Mask[i+1];
2661 if (!isUndefOrEqual(BitI, j + NumElts/2))
2664 if (isUndefOrEqual(BitI1, NumElts))
2667 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2674 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2675 SmallVector<int, 8> M;
2677 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2680 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2681 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2683 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2684 int NumElems = VT.getVectorNumElements();
2685 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2688 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2690 int BitI1 = Mask[i+1];
2691 if (!isUndefOrEqual(BitI, j))
2693 if (!isUndefOrEqual(BitI1, j))
2699 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2700 SmallVector<int, 8> M;
2702 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2705 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2706 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2708 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2709 int NumElems = VT.getVectorNumElements();
2710 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2713 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2715 int BitI1 = Mask[i+1];
2716 if (!isUndefOrEqual(BitI, j))
2718 if (!isUndefOrEqual(BitI1, j))
2724 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2725 SmallVector<int, 8> M;
2727 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2730 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2731 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2732 /// MOVSD, and MOVD, i.e. setting the lowest element.
2733 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2734 if (VT.getVectorElementType().getSizeInBits() < 32)
2737 int NumElts = VT.getVectorNumElements();
2739 if (!isUndefOrEqual(Mask[0], NumElts))
2742 for (int i = 1; i < NumElts; ++i)
2743 if (!isUndefOrEqual(Mask[i], i))
2749 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2750 SmallVector<int, 8> M;
2752 return ::isMOVLMask(M, N->getValueType(0));
2755 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2756 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2757 /// element of vector 2 and the other elements to come from vector 1 in order.
2758 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2759 bool V2IsSplat = false, bool V2IsUndef = false) {
2760 int NumOps = VT.getVectorNumElements();
2761 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2764 if (!isUndefOrEqual(Mask[0], 0))
2767 for (int i = 1; i < NumOps; ++i)
2768 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2769 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2770 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2776 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2777 bool V2IsUndef = false) {
2778 SmallVector<int, 8> M;
2780 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2783 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2784 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2785 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2786 if (N->getValueType(0).getVectorNumElements() != 4)
2789 // Expect 1, 1, 3, 3
2790 for (unsigned i = 0; i < 2; ++i) {
2791 int Elt = N->getMaskElt(i);
2792 if (Elt >= 0 && Elt != 1)
2797 for (unsigned i = 2; i < 4; ++i) {
2798 int Elt = N->getMaskElt(i);
2799 if (Elt >= 0 && Elt != 3)
2804 // Don't use movshdup if it can be done with a shufps.
2805 // FIXME: verify that matching u, u, 3, 3 is what we want.
2809 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2810 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2811 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2812 if (N->getValueType(0).getVectorNumElements() != 4)
2815 // Expect 0, 0, 2, 2
2816 for (unsigned i = 0; i < 2; ++i)
2817 if (N->getMaskElt(i) > 0)
2821 for (unsigned i = 2; i < 4; ++i) {
2822 int Elt = N->getMaskElt(i);
2823 if (Elt >= 0 && Elt != 2)
2828 // Don't use movsldup if it can be done with a shufps.
2832 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2833 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2834 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2835 int e = N->getValueType(0).getVectorNumElements() / 2;
2837 for (int i = 0; i < e; ++i)
2838 if (!isUndefOrEqual(N->getMaskElt(i), i))
2840 for (int i = 0; i < e; ++i)
2841 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2846 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2847 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
2848 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2849 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2850 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2852 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2854 for (int i = 0; i < NumOperands; ++i) {
2855 int Val = SVOp->getMaskElt(NumOperands-i-1);
2856 if (Val < 0) Val = 0;
2857 if (Val >= NumOperands) Val -= NumOperands;
2859 if (i != NumOperands - 1)
2865 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2866 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
2867 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2868 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2870 // 8 nodes, but we only care about the last 4.
2871 for (unsigned i = 7; i >= 4; --i) {
2872 int Val = SVOp->getMaskElt(i);
2881 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2882 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
2883 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2884 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2886 // 8 nodes, but we only care about the first 4.
2887 for (int i = 3; i >= 0; --i) {
2888 int Val = SVOp->getMaskElt(i);
2897 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
2898 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
2899 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
2900 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2901 EVT VVT = N->getValueType(0);
2902 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
2906 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
2907 Val = SVOp->getMaskElt(i);
2911 return (Val - i) * EltSize;
2914 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2916 bool X86::isZeroNode(SDValue Elt) {
2917 return ((isa<ConstantSDNode>(Elt) &&
2918 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2919 (isa<ConstantFPSDNode>(Elt) &&
2920 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2923 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
2924 /// their permute mask.
2925 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
2926 SelectionDAG &DAG) {
2927 EVT VT = SVOp->getValueType(0);
2928 unsigned NumElems = VT.getVectorNumElements();
2929 SmallVector<int, 8> MaskVec;
2931 for (unsigned i = 0; i != NumElems; ++i) {
2932 int idx = SVOp->getMaskElt(i);
2934 MaskVec.push_back(idx);
2935 else if (idx < (int)NumElems)
2936 MaskVec.push_back(idx + NumElems);
2938 MaskVec.push_back(idx - NumElems);
2940 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
2941 SVOp->getOperand(0), &MaskVec[0]);
2944 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2945 /// the two vector operands have swapped position.
2946 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
2947 unsigned NumElems = VT.getVectorNumElements();
2948 for (unsigned i = 0; i != NumElems; ++i) {
2952 else if (idx < (int)NumElems)
2953 Mask[i] = idx + NumElems;
2955 Mask[i] = idx - NumElems;
2959 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2960 /// match movhlps. The lower half elements should come from upper half of
2961 /// V1 (and in order), and the upper half elements should come from the upper
2962 /// half of V2 (and in order).
2963 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
2964 if (Op->getValueType(0).getVectorNumElements() != 4)
2966 for (unsigned i = 0, e = 2; i != e; ++i)
2967 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
2969 for (unsigned i = 2; i != 4; ++i)
2970 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
2975 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2976 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2978 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2979 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2981 N = N->getOperand(0).getNode();
2982 if (!ISD::isNON_EXTLoad(N))
2985 *LD = cast<LoadSDNode>(N);
2989 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2990 /// match movlp{s|d}. The lower half elements should come from lower half of
2991 /// V1 (and in order), and the upper half elements should come from the upper
2992 /// half of V2 (and in order). And since V1 will become the source of the
2993 /// MOVLP, it must be either a vector load or a scalar load to vector.
2994 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
2995 ShuffleVectorSDNode *Op) {
2996 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2998 // Is V2 is a vector load, don't do this transformation. We will try to use
2999 // load folding shufps op.
3000 if (ISD::isNON_EXTLoad(V2))
3003 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3005 if (NumElems != 2 && NumElems != 4)
3007 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3008 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3010 for (unsigned i = NumElems/2; i != NumElems; ++i)
3011 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3016 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3018 static bool isSplatVector(SDNode *N) {
3019 if (N->getOpcode() != ISD::BUILD_VECTOR)
3022 SDValue SplatValue = N->getOperand(0);
3023 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3024 if (N->getOperand(i) != SplatValue)
3029 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3030 /// to an zero vector.
3031 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3032 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3033 SDValue V1 = N->getOperand(0);
3034 SDValue V2 = N->getOperand(1);
3035 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3036 for (unsigned i = 0; i != NumElems; ++i) {
3037 int Idx = N->getMaskElt(i);
3038 if (Idx >= (int)NumElems) {
3039 unsigned Opc = V2.getOpcode();
3040 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3042 if (Opc != ISD::BUILD_VECTOR ||
3043 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3045 } else if (Idx >= 0) {
3046 unsigned Opc = V1.getOpcode();
3047 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3049 if (Opc != ISD::BUILD_VECTOR ||
3050 !X86::isZeroNode(V1.getOperand(Idx)))
3057 /// getZeroVector - Returns a vector of specified type with all zero elements.
3059 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3061 assert(VT.isVector() && "Expected a vector type");
3063 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3064 // type. This ensures they get CSE'd.
3066 if (VT.getSizeInBits() == 64) { // MMX
3067 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3068 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3069 } else if (HasSSE2) { // SSE2
3070 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3071 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3073 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3074 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3076 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3079 /// getOnesVector - Returns a vector of specified type with all bits set.
3081 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3082 assert(VT.isVector() && "Expected a vector type");
3084 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3085 // type. This ensures they get CSE'd.
3086 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3088 if (VT.getSizeInBits() == 64) // MMX
3089 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3091 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3092 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3096 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3097 /// that point to V2 points to its first element.
3098 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3099 EVT VT = SVOp->getValueType(0);
3100 unsigned NumElems = VT.getVectorNumElements();
3102 bool Changed = false;
3103 SmallVector<int, 8> MaskVec;
3104 SVOp->getMask(MaskVec);
3106 for (unsigned i = 0; i != NumElems; ++i) {
3107 if (MaskVec[i] > (int)NumElems) {
3108 MaskVec[i] = NumElems;
3113 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3114 SVOp->getOperand(1), &MaskVec[0]);
3115 return SDValue(SVOp, 0);
3118 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3119 /// operation of specified width.
3120 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3122 unsigned NumElems = VT.getVectorNumElements();
3123 SmallVector<int, 8> Mask;
3124 Mask.push_back(NumElems);
3125 for (unsigned i = 1; i != NumElems; ++i)
3127 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3130 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3131 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3133 unsigned NumElems = VT.getVectorNumElements();
3134 SmallVector<int, 8> Mask;
3135 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3137 Mask.push_back(i + NumElems);
3139 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3142 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3143 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3145 unsigned NumElems = VT.getVectorNumElements();
3146 unsigned Half = NumElems/2;
3147 SmallVector<int, 8> Mask;
3148 for (unsigned i = 0; i != Half; ++i) {
3149 Mask.push_back(i + Half);
3150 Mask.push_back(i + NumElems + Half);
3152 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3155 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3156 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3158 if (SV->getValueType(0).getVectorNumElements() <= 4)
3159 return SDValue(SV, 0);
3161 EVT PVT = MVT::v4f32;
3162 EVT VT = SV->getValueType(0);
3163 DebugLoc dl = SV->getDebugLoc();
3164 SDValue V1 = SV->getOperand(0);
3165 int NumElems = VT.getVectorNumElements();
3166 int EltNo = SV->getSplatIndex();
3168 // unpack elements to the correct location
3169 while (NumElems > 4) {
3170 if (EltNo < NumElems/2) {
3171 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3173 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3174 EltNo -= NumElems/2;
3179 // Perform the splat.
3180 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3181 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3182 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3183 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3186 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3187 /// vector of zero or undef vector. This produces a shuffle where the low
3188 /// element of V2 is swizzled into the zero/undef vector, landing at element
3189 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3190 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3191 bool isZero, bool HasSSE2,
3192 SelectionDAG &DAG) {
3193 EVT VT = V2.getValueType();
3195 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3196 unsigned NumElems = VT.getVectorNumElements();
3197 SmallVector<int, 16> MaskVec;
3198 for (unsigned i = 0; i != NumElems; ++i)
3199 // If this is the insertion idx, put the low elt of V2 here.
3200 MaskVec.push_back(i == Idx ? NumElems : i);
3201 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3204 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3205 /// a shuffle that is zero.
3207 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3208 bool Low, SelectionDAG &DAG) {
3209 unsigned NumZeros = 0;
3210 for (int i = 0; i < NumElems; ++i) {
3211 unsigned Index = Low ? i : NumElems-i-1;
3212 int Idx = SVOp->getMaskElt(Index);
3217 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3218 if (Elt.getNode() && X86::isZeroNode(Elt))
3226 /// isVectorShift - Returns true if the shuffle can be implemented as a
3227 /// logical left or right shift of a vector.
3228 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3229 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3230 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3231 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3234 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3237 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3241 bool SeenV1 = false;
3242 bool SeenV2 = false;
3243 for (int i = NumZeros; i < NumElems; ++i) {
3244 int Val = isLeft ? (i - NumZeros) : i;
3245 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3257 if (SeenV1 && SeenV2)
3260 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3266 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3268 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3269 unsigned NumNonZero, unsigned NumZero,
3270 SelectionDAG &DAG, TargetLowering &TLI) {
3274 DebugLoc dl = Op.getDebugLoc();
3277 for (unsigned i = 0; i < 16; ++i) {
3278 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3279 if (ThisIsNonZero && First) {
3281 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3283 V = DAG.getUNDEF(MVT::v8i16);
3288 SDValue ThisElt(0, 0), LastElt(0, 0);
3289 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3290 if (LastIsNonZero) {
3291 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3292 MVT::i16, Op.getOperand(i-1));
3294 if (ThisIsNonZero) {
3295 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3296 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3297 ThisElt, DAG.getConstant(8, MVT::i8));
3299 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3303 if (ThisElt.getNode())
3304 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3305 DAG.getIntPtrConstant(i/2));
3309 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3312 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3314 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3315 unsigned NumNonZero, unsigned NumZero,
3316 SelectionDAG &DAG, TargetLowering &TLI) {
3320 DebugLoc dl = Op.getDebugLoc();
3323 for (unsigned i = 0; i < 8; ++i) {
3324 bool isNonZero = (NonZeros & (1 << i)) != 0;
3328 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3330 V = DAG.getUNDEF(MVT::v8i16);
3333 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3334 MVT::v8i16, V, Op.getOperand(i),
3335 DAG.getIntPtrConstant(i));
3342 /// getVShift - Return a vector logical shift node.
3344 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3345 unsigned NumBits, SelectionDAG &DAG,
3346 const TargetLowering &TLI, DebugLoc dl) {
3347 bool isMMX = VT.getSizeInBits() == 64;
3348 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3349 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3350 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3351 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3352 DAG.getNode(Opc, dl, ShVT, SrcOp,
3353 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3357 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3358 SelectionDAG &DAG) {
3360 // Check if the scalar load can be widened into a vector load. And if
3361 // the address is "base + cst" see if the cst can be "absorbed" into
3362 // the shuffle mask.
3363 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3364 SDValue Ptr = LD->getBasePtr();
3365 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3367 EVT PVT = LD->getValueType(0);
3368 if (PVT != MVT::i32 && PVT != MVT::f32)
3373 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3374 FI = FINode->getIndex();
3376 } else if (Ptr.getOpcode() == ISD::ADD &&
3377 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3378 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3379 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3380 Offset = Ptr.getConstantOperandVal(1);
3381 Ptr = Ptr.getOperand(0);
3386 SDValue Chain = LD->getChain();
3387 // Make sure the stack object alignment is at least 16.
3388 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3389 if (DAG.InferPtrAlignment(Ptr) < 16) {
3390 if (MFI->isFixedObjectIndex(FI)) {
3391 // Can't change the alignment. Reference stack + offset explicitly
3392 // if stack pointer is at least 16-byte aligned.
3393 unsigned StackAlign = Subtarget->getStackAlignment();
3394 if (StackAlign < 16)
3396 Offset = MFI->getObjectOffset(FI) + Offset;
3397 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
3399 Ptr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
3400 DAG.getConstant(Offset & ~15, getPointerTy()));
3403 MFI->setObjectAlignment(FI, 16);
3407 // (Offset % 16) must be multiple of 4. Then address is then
3408 // Ptr + (Offset & ~15).
3411 if ((Offset % 16) & 3)
3413 int64_t StartOffset = Offset & ~15;
3415 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3416 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3418 int EltNo = (Offset - StartOffset) >> 2;
3419 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3420 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3421 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0);
3422 // Canonicalize it to a v4i32 shuffle.
3423 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3424 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3425 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3426 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3433 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3434 DebugLoc dl = Op.getDebugLoc();
3435 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3436 if (ISD::isBuildVectorAllZeros(Op.getNode())
3437 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3438 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3439 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3440 // eliminated on x86-32 hosts.
3441 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3444 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3445 return getOnesVector(Op.getValueType(), DAG, dl);
3446 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3449 EVT VT = Op.getValueType();
3450 EVT ExtVT = VT.getVectorElementType();
3451 unsigned EVTBits = ExtVT.getSizeInBits();
3453 unsigned NumElems = Op.getNumOperands();
3454 unsigned NumZero = 0;
3455 unsigned NumNonZero = 0;
3456 unsigned NonZeros = 0;
3457 bool IsAllConstants = true;
3458 SmallSet<SDValue, 8> Values;
3459 for (unsigned i = 0; i < NumElems; ++i) {
3460 SDValue Elt = Op.getOperand(i);
3461 if (Elt.getOpcode() == ISD::UNDEF)
3464 if (Elt.getOpcode() != ISD::Constant &&
3465 Elt.getOpcode() != ISD::ConstantFP)
3466 IsAllConstants = false;
3467 if (X86::isZeroNode(Elt))
3470 NonZeros |= (1 << i);
3475 if (NumNonZero == 0) {
3476 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3477 return DAG.getUNDEF(VT);
3480 // Special case for single non-zero, non-undef, element.
3481 if (NumNonZero == 1) {
3482 unsigned Idx = CountTrailingZeros_32(NonZeros);
3483 SDValue Item = Op.getOperand(Idx);
3485 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3486 // the value are obviously zero, truncate the value to i32 and do the
3487 // insertion that way. Only do this if the value is non-constant or if the
3488 // value is a constant being inserted into element 0. It is cheaper to do
3489 // a constant pool load than it is to do a movd + shuffle.
3490 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3491 (!IsAllConstants || Idx == 0)) {
3492 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3493 // Handle MMX and SSE both.
3494 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3495 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3497 // Truncate the value (which may itself be a constant) to i32, and
3498 // convert it to a vector with movd (S2V+shuffle to zero extend).
3499 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3500 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3501 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3502 Subtarget->hasSSE2(), DAG);
3504 // Now we have our 32-bit value zero extended in the low element of
3505 // a vector. If Idx != 0, swizzle it into place.
3507 SmallVector<int, 4> Mask;
3508 Mask.push_back(Idx);
3509 for (unsigned i = 1; i != VecElts; ++i)
3511 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3512 DAG.getUNDEF(Item.getValueType()),
3515 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3519 // If we have a constant or non-constant insertion into the low element of
3520 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3521 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3522 // depending on what the source datatype is.
3525 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3526 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3527 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3528 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3529 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3530 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3532 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3533 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3534 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3535 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3536 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3537 Subtarget->hasSSE2(), DAG);
3538 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3542 // Is it a vector logical left shift?
3543 if (NumElems == 2 && Idx == 1 &&
3544 X86::isZeroNode(Op.getOperand(0)) &&
3545 !X86::isZeroNode(Op.getOperand(1))) {
3546 unsigned NumBits = VT.getSizeInBits();
3547 return getVShift(true, VT,
3548 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3549 VT, Op.getOperand(1)),
3550 NumBits/2, DAG, *this, dl);
3553 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3556 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3557 // is a non-constant being inserted into an element other than the low one,
3558 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3559 // movd/movss) to move this into the low element, then shuffle it into
3561 if (EVTBits == 32) {
3562 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3564 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3565 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3566 Subtarget->hasSSE2(), DAG);
3567 SmallVector<int, 8> MaskVec;
3568 for (unsigned i = 0; i < NumElems; i++)
3569 MaskVec.push_back(i == Idx ? 0 : 1);
3570 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3574 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3575 if (Values.size() == 1) {
3576 if (EVTBits == 32) {
3577 // Instead of a shuffle like this:
3578 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3579 // Check if it's possible to issue this instead.
3580 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3581 unsigned Idx = CountTrailingZeros_32(NonZeros);
3582 SDValue Item = Op.getOperand(Idx);
3583 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3584 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3589 // A vector full of immediates; various special cases are already
3590 // handled, so this is best done with a single constant-pool load.
3594 // Let legalizer expand 2-wide build_vectors.
3595 if (EVTBits == 64) {
3596 if (NumNonZero == 1) {
3597 // One half is zero or undef.
3598 unsigned Idx = CountTrailingZeros_32(NonZeros);
3599 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3600 Op.getOperand(Idx));
3601 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3602 Subtarget->hasSSE2(), DAG);
3607 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3608 if (EVTBits == 8 && NumElems == 16) {
3609 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3611 if (V.getNode()) return V;
3614 if (EVTBits == 16 && NumElems == 8) {
3615 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3617 if (V.getNode()) return V;
3620 // If element VT is == 32 bits, turn it into a number of shuffles.
3621 SmallVector<SDValue, 8> V;
3623 if (NumElems == 4 && NumZero > 0) {
3624 for (unsigned i = 0; i < 4; ++i) {
3625 bool isZero = !(NonZeros & (1 << i));
3627 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3629 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3632 for (unsigned i = 0; i < 2; ++i) {
3633 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3636 V[i] = V[i*2]; // Must be a zero vector.
3639 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3642 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3645 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3650 SmallVector<int, 8> MaskVec;
3651 bool Reverse = (NonZeros & 0x3) == 2;
3652 for (unsigned i = 0; i < 2; ++i)
3653 MaskVec.push_back(Reverse ? 1-i : i);
3654 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3655 for (unsigned i = 0; i < 2; ++i)
3656 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3657 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3660 if (Values.size() > 2) {
3661 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3662 // values to be inserted is equal to the number of elements, in which case
3663 // use the unpack code below in the hopes of matching the consecutive elts
3664 // load merge pattern for shuffles.
3665 // FIXME: We could probably just check that here directly.
3666 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3667 getSubtarget()->hasSSE41()) {
3668 V[0] = DAG.getUNDEF(VT);
3669 for (unsigned i = 0; i < NumElems; ++i)
3670 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3671 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3672 Op.getOperand(i), DAG.getIntPtrConstant(i));
3675 // Expand into a number of unpckl*.
3677 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3678 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3679 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3680 for (unsigned i = 0; i < NumElems; ++i)
3681 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3683 while (NumElems != 0) {
3684 for (unsigned i = 0; i < NumElems; ++i)
3685 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3694 // v8i16 shuffles - Prefer shuffles in the following order:
3695 // 1. [all] pshuflw, pshufhw, optional move
3696 // 2. [ssse3] 1 x pshufb
3697 // 3. [ssse3] 2 x pshufb + 1 x por
3698 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3700 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3701 SelectionDAG &DAG, X86TargetLowering &TLI) {
3702 SDValue V1 = SVOp->getOperand(0);
3703 SDValue V2 = SVOp->getOperand(1);
3704 DebugLoc dl = SVOp->getDebugLoc();
3705 SmallVector<int, 8> MaskVals;
3707 // Determine if more than 1 of the words in each of the low and high quadwords
3708 // of the result come from the same quadword of one of the two inputs. Undef
3709 // mask values count as coming from any quadword, for better codegen.
3710 SmallVector<unsigned, 4> LoQuad(4);
3711 SmallVector<unsigned, 4> HiQuad(4);
3712 BitVector InputQuads(4);
3713 for (unsigned i = 0; i < 8; ++i) {
3714 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3715 int EltIdx = SVOp->getMaskElt(i);
3716 MaskVals.push_back(EltIdx);
3725 InputQuads.set(EltIdx / 4);
3728 int BestLoQuad = -1;
3729 unsigned MaxQuad = 1;
3730 for (unsigned i = 0; i < 4; ++i) {
3731 if (LoQuad[i] > MaxQuad) {
3733 MaxQuad = LoQuad[i];
3737 int BestHiQuad = -1;
3739 for (unsigned i = 0; i < 4; ++i) {
3740 if (HiQuad[i] > MaxQuad) {
3742 MaxQuad = HiQuad[i];
3746 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3747 // of the two input vectors, shuffle them into one input vector so only a
3748 // single pshufb instruction is necessary. If There are more than 2 input
3749 // quads, disable the next transformation since it does not help SSSE3.
3750 bool V1Used = InputQuads[0] || InputQuads[1];
3751 bool V2Used = InputQuads[2] || InputQuads[3];
3752 if (TLI.getSubtarget()->hasSSSE3()) {
3753 if (InputQuads.count() == 2 && V1Used && V2Used) {
3754 BestLoQuad = InputQuads.find_first();
3755 BestHiQuad = InputQuads.find_next(BestLoQuad);
3757 if (InputQuads.count() > 2) {
3763 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3764 // the shuffle mask. If a quad is scored as -1, that means that it contains
3765 // words from all 4 input quadwords.
3767 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3768 SmallVector<int, 8> MaskV;
3769 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3770 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3771 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3772 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3773 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3774 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3776 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3777 // source words for the shuffle, to aid later transformations.
3778 bool AllWordsInNewV = true;
3779 bool InOrder[2] = { true, true };
3780 for (unsigned i = 0; i != 8; ++i) {
3781 int idx = MaskVals[i];
3783 InOrder[i/4] = false;
3784 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3786 AllWordsInNewV = false;
3790 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3791 if (AllWordsInNewV) {
3792 for (int i = 0; i != 8; ++i) {
3793 int idx = MaskVals[i];
3796 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3797 if ((idx != i) && idx < 4)
3799 if ((idx != i) && idx > 3)
3808 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3809 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3810 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3811 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3812 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3816 // If we have SSSE3, and all words of the result are from 1 input vector,
3817 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3818 // is present, fall back to case 4.
3819 if (TLI.getSubtarget()->hasSSSE3()) {
3820 SmallVector<SDValue,16> pshufbMask;
3822 // If we have elements from both input vectors, set the high bit of the
3823 // shuffle mask element to zero out elements that come from V2 in the V1
3824 // mask, and elements that come from V1 in the V2 mask, so that the two
3825 // results can be OR'd together.
3826 bool TwoInputs = V1Used && V2Used;
3827 for (unsigned i = 0; i != 8; ++i) {
3828 int EltIdx = MaskVals[i] * 2;
3829 if (TwoInputs && (EltIdx >= 16)) {
3830 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3831 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3834 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3835 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3837 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3838 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3839 DAG.getNode(ISD::BUILD_VECTOR, dl,
3840 MVT::v16i8, &pshufbMask[0], 16));
3842 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3844 // Calculate the shuffle mask for the second input, shuffle it, and
3845 // OR it with the first shuffled input.
3847 for (unsigned i = 0; i != 8; ++i) {
3848 int EltIdx = MaskVals[i] * 2;
3850 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3851 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3854 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3855 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3857 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3858 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3859 DAG.getNode(ISD::BUILD_VECTOR, dl,
3860 MVT::v16i8, &pshufbMask[0], 16));
3861 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3862 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3865 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3866 // and update MaskVals with new element order.
3867 BitVector InOrder(8);
3868 if (BestLoQuad >= 0) {
3869 SmallVector<int, 8> MaskV;
3870 for (int i = 0; i != 4; ++i) {
3871 int idx = MaskVals[i];
3873 MaskV.push_back(-1);
3875 } else if ((idx / 4) == BestLoQuad) {
3876 MaskV.push_back(idx & 3);
3879 MaskV.push_back(-1);
3882 for (unsigned i = 4; i != 8; ++i)
3884 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3888 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
3889 // and update MaskVals with the new element order.
3890 if (BestHiQuad >= 0) {
3891 SmallVector<int, 8> MaskV;
3892 for (unsigned i = 0; i != 4; ++i)
3894 for (unsigned i = 4; i != 8; ++i) {
3895 int idx = MaskVals[i];
3897 MaskV.push_back(-1);
3899 } else if ((idx / 4) == BestHiQuad) {
3900 MaskV.push_back((idx & 3) + 4);
3903 MaskV.push_back(-1);
3906 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3910 // In case BestHi & BestLo were both -1, which means each quadword has a word
3911 // from each of the four input quadwords, calculate the InOrder bitvector now
3912 // before falling through to the insert/extract cleanup.
3913 if (BestLoQuad == -1 && BestHiQuad == -1) {
3915 for (int i = 0; i != 8; ++i)
3916 if (MaskVals[i] < 0 || MaskVals[i] == i)
3920 // The other elements are put in the right place using pextrw and pinsrw.
3921 for (unsigned i = 0; i != 8; ++i) {
3924 int EltIdx = MaskVals[i];
3927 SDValue ExtOp = (EltIdx < 8)
3928 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
3929 DAG.getIntPtrConstant(EltIdx))
3930 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
3931 DAG.getIntPtrConstant(EltIdx - 8));
3932 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3933 DAG.getIntPtrConstant(i));
3938 // v16i8 shuffles - Prefer shuffles in the following order:
3939 // 1. [ssse3] 1 x pshufb
3940 // 2. [ssse3] 2 x pshufb + 1 x por
3941 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
3943 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
3944 SelectionDAG &DAG, X86TargetLowering &TLI) {
3945 SDValue V1 = SVOp->getOperand(0);
3946 SDValue V2 = SVOp->getOperand(1);
3947 DebugLoc dl = SVOp->getDebugLoc();
3948 SmallVector<int, 16> MaskVals;
3949 SVOp->getMask(MaskVals);
3951 // If we have SSSE3, case 1 is generated when all result bytes come from
3952 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
3953 // present, fall back to case 3.
3954 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
3957 for (unsigned i = 0; i < 16; ++i) {
3958 int EltIdx = MaskVals[i];
3967 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
3968 if (TLI.getSubtarget()->hasSSSE3()) {
3969 SmallVector<SDValue,16> pshufbMask;
3971 // If all result elements are from one input vector, then only translate
3972 // undef mask values to 0x80 (zero out result) in the pshufb mask.
3974 // Otherwise, we have elements from both input vectors, and must zero out
3975 // elements that come from V2 in the first mask, and V1 in the second mask
3976 // so that we can OR them together.
3977 bool TwoInputs = !(V1Only || V2Only);
3978 for (unsigned i = 0; i != 16; ++i) {
3979 int EltIdx = MaskVals[i];
3980 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
3981 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3984 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3986 // If all the elements are from V2, assign it to V1 and return after
3987 // building the first pshufb.
3990 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3991 DAG.getNode(ISD::BUILD_VECTOR, dl,
3992 MVT::v16i8, &pshufbMask[0], 16));
3996 // Calculate the shuffle mask for the second input, shuffle it, and
3997 // OR it with the first shuffled input.
3999 for (unsigned i = 0; i != 16; ++i) {
4000 int EltIdx = MaskVals[i];
4002 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4005 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4007 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4008 DAG.getNode(ISD::BUILD_VECTOR, dl,
4009 MVT::v16i8, &pshufbMask[0], 16));
4010 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4013 // No SSSE3 - Calculate in place words and then fix all out of place words
4014 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4015 // the 16 different words that comprise the two doublequadword input vectors.
4016 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4017 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4018 SDValue NewV = V2Only ? V2 : V1;
4019 for (int i = 0; i != 8; ++i) {
4020 int Elt0 = MaskVals[i*2];
4021 int Elt1 = MaskVals[i*2+1];
4023 // This word of the result is all undef, skip it.
4024 if (Elt0 < 0 && Elt1 < 0)
4027 // This word of the result is already in the correct place, skip it.
4028 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4030 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4033 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4034 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4037 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4038 // using a single extract together, load it and store it.
4039 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4040 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4041 DAG.getIntPtrConstant(Elt1 / 2));
4042 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4043 DAG.getIntPtrConstant(i));
4047 // If Elt1 is defined, extract it from the appropriate source. If the
4048 // source byte is not also odd, shift the extracted word left 8 bits
4049 // otherwise clear the bottom 8 bits if we need to do an or.
4051 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4052 DAG.getIntPtrConstant(Elt1 / 2));
4053 if ((Elt1 & 1) == 0)
4054 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4055 DAG.getConstant(8, TLI.getShiftAmountTy()));
4057 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4058 DAG.getConstant(0xFF00, MVT::i16));
4060 // If Elt0 is defined, extract it from the appropriate source. If the
4061 // source byte is not also even, shift the extracted word right 8 bits. If
4062 // Elt1 was also defined, OR the extracted values together before
4063 // inserting them in the result.
4065 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4066 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4067 if ((Elt0 & 1) != 0)
4068 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4069 DAG.getConstant(8, TLI.getShiftAmountTy()));
4071 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4072 DAG.getConstant(0x00FF, MVT::i16));
4073 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4076 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4077 DAG.getIntPtrConstant(i));
4079 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4082 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4083 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4084 /// done when every pair / quad of shuffle mask elements point to elements in
4085 /// the right sequence. e.g.
4086 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4088 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4090 TargetLowering &TLI, DebugLoc dl) {
4091 EVT VT = SVOp->getValueType(0);
4092 SDValue V1 = SVOp->getOperand(0);
4093 SDValue V2 = SVOp->getOperand(1);
4094 unsigned NumElems = VT.getVectorNumElements();
4095 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4096 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4097 EVT MaskEltVT = MaskVT.getVectorElementType();
4099 switch (VT.getSimpleVT().SimpleTy) {
4100 default: assert(false && "Unexpected!");
4101 case MVT::v4f32: NewVT = MVT::v2f64; break;
4102 case MVT::v4i32: NewVT = MVT::v2i64; break;
4103 case MVT::v8i16: NewVT = MVT::v4i32; break;
4104 case MVT::v16i8: NewVT = MVT::v4i32; break;
4107 if (NewWidth == 2) {
4113 int Scale = NumElems / NewWidth;
4114 SmallVector<int, 8> MaskVec;
4115 for (unsigned i = 0; i < NumElems; i += Scale) {
4117 for (int j = 0; j < Scale; ++j) {
4118 int EltIdx = SVOp->getMaskElt(i+j);
4122 StartIdx = EltIdx - (EltIdx % Scale);
4123 if (EltIdx != StartIdx + j)
4127 MaskVec.push_back(-1);
4129 MaskVec.push_back(StartIdx / Scale);
4132 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4133 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4134 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4137 /// getVZextMovL - Return a zero-extending vector move low node.
4139 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4140 SDValue SrcOp, SelectionDAG &DAG,
4141 const X86Subtarget *Subtarget, DebugLoc dl) {
4142 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4143 LoadSDNode *LD = NULL;
4144 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4145 LD = dyn_cast<LoadSDNode>(SrcOp);
4147 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4149 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4150 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4151 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4152 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4153 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4155 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4156 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4157 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4158 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4166 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4167 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4168 DAG.getNode(ISD::BIT_CONVERT, dl,
4172 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4175 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4176 SDValue V1 = SVOp->getOperand(0);
4177 SDValue V2 = SVOp->getOperand(1);
4178 DebugLoc dl = SVOp->getDebugLoc();
4179 EVT VT = SVOp->getValueType(0);
4181 SmallVector<std::pair<int, int>, 8> Locs;
4183 SmallVector<int, 8> Mask1(4U, -1);
4184 SmallVector<int, 8> PermMask;
4185 SVOp->getMask(PermMask);
4189 for (unsigned i = 0; i != 4; ++i) {
4190 int Idx = PermMask[i];
4192 Locs[i] = std::make_pair(-1, -1);
4194 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4196 Locs[i] = std::make_pair(0, NumLo);
4200 Locs[i] = std::make_pair(1, NumHi);
4202 Mask1[2+NumHi] = Idx;
4208 if (NumLo <= 2 && NumHi <= 2) {
4209 // If no more than two elements come from either vector. This can be
4210 // implemented with two shuffles. First shuffle gather the elements.
4211 // The second shuffle, which takes the first shuffle as both of its
4212 // vector operands, put the elements into the right order.
4213 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4215 SmallVector<int, 8> Mask2(4U, -1);
4217 for (unsigned i = 0; i != 4; ++i) {
4218 if (Locs[i].first == -1)
4221 unsigned Idx = (i < 2) ? 0 : 4;
4222 Idx += Locs[i].first * 2 + Locs[i].second;
4227 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4228 } else if (NumLo == 3 || NumHi == 3) {
4229 // Otherwise, we must have three elements from one vector, call it X, and
4230 // one element from the other, call it Y. First, use a shufps to build an
4231 // intermediate vector with the one element from Y and the element from X
4232 // that will be in the same half in the final destination (the indexes don't
4233 // matter). Then, use a shufps to build the final vector, taking the half
4234 // containing the element from Y from the intermediate, and the other half
4237 // Normalize it so the 3 elements come from V1.
4238 CommuteVectorShuffleMask(PermMask, VT);
4242 // Find the element from V2.
4244 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4245 int Val = PermMask[HiIndex];
4252 Mask1[0] = PermMask[HiIndex];
4254 Mask1[2] = PermMask[HiIndex^1];
4256 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4259 Mask1[0] = PermMask[0];
4260 Mask1[1] = PermMask[1];
4261 Mask1[2] = HiIndex & 1 ? 6 : 4;
4262 Mask1[3] = HiIndex & 1 ? 4 : 6;
4263 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4265 Mask1[0] = HiIndex & 1 ? 2 : 0;
4266 Mask1[1] = HiIndex & 1 ? 0 : 2;
4267 Mask1[2] = PermMask[2];
4268 Mask1[3] = PermMask[3];
4273 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4277 // Break it into (shuffle shuffle_hi, shuffle_lo).
4279 SmallVector<int,8> LoMask(4U, -1);
4280 SmallVector<int,8> HiMask(4U, -1);
4282 SmallVector<int,8> *MaskPtr = &LoMask;
4283 unsigned MaskIdx = 0;
4286 for (unsigned i = 0; i != 4; ++i) {
4293 int Idx = PermMask[i];
4295 Locs[i] = std::make_pair(-1, -1);
4296 } else if (Idx < 4) {
4297 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4298 (*MaskPtr)[LoIdx] = Idx;
4301 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4302 (*MaskPtr)[HiIdx] = Idx;
4307 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4308 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4309 SmallVector<int, 8> MaskOps;
4310 for (unsigned i = 0; i != 4; ++i) {
4311 if (Locs[i].first == -1) {
4312 MaskOps.push_back(-1);
4314 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4315 MaskOps.push_back(Idx);
4318 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4322 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4323 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4324 SDValue V1 = Op.getOperand(0);
4325 SDValue V2 = Op.getOperand(1);
4326 EVT VT = Op.getValueType();
4327 DebugLoc dl = Op.getDebugLoc();
4328 unsigned NumElems = VT.getVectorNumElements();
4329 bool isMMX = VT.getSizeInBits() == 64;
4330 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4331 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4332 bool V1IsSplat = false;
4333 bool V2IsSplat = false;
4335 if (isZeroShuffle(SVOp))
4336 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4338 // Promote splats to v4f32.
4339 if (SVOp->isSplat()) {
4340 if (isMMX || NumElems < 4)
4342 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4345 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4347 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4348 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4349 if (NewOp.getNode())
4350 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4351 LowerVECTOR_SHUFFLE(NewOp, DAG));
4352 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4353 // FIXME: Figure out a cleaner way to do this.
4354 // Try to make use of movq to zero out the top part.
4355 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4356 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4357 if (NewOp.getNode()) {
4358 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4359 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4360 DAG, Subtarget, dl);
4362 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4363 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4364 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4365 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4366 DAG, Subtarget, dl);
4370 if (X86::isPSHUFDMask(SVOp))
4373 // Check if this can be converted into a logical shift.
4374 bool isLeft = false;
4377 bool isShift = getSubtarget()->hasSSE2() &&
4378 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4379 if (isShift && ShVal.hasOneUse()) {
4380 // If the shifted value has multiple uses, it may be cheaper to use
4381 // v_set0 + movlhps or movhlps, etc.
4382 EVT EltVT = VT.getVectorElementType();
4383 ShAmt *= EltVT.getSizeInBits();
4384 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4387 if (X86::isMOVLMask(SVOp)) {
4390 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4391 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4396 // FIXME: fold these into legal mask.
4397 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4398 X86::isMOVSLDUPMask(SVOp) ||
4399 X86::isMOVHLPSMask(SVOp) ||
4400 X86::isMOVLHPSMask(SVOp) ||
4401 X86::isMOVLPMask(SVOp)))
4404 if (ShouldXformToMOVHLPS(SVOp) ||
4405 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4406 return CommuteVectorShuffle(SVOp, DAG);
4409 // No better options. Use a vshl / vsrl.
4410 EVT EltVT = VT.getVectorElementType();
4411 ShAmt *= EltVT.getSizeInBits();
4412 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4415 bool Commuted = false;
4416 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4417 // 1,1,1,1 -> v8i16 though.
4418 V1IsSplat = isSplatVector(V1.getNode());
4419 V2IsSplat = isSplatVector(V2.getNode());
4421 // Canonicalize the splat or undef, if present, to be on the RHS.
4422 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4423 Op = CommuteVectorShuffle(SVOp, DAG);
4424 SVOp = cast<ShuffleVectorSDNode>(Op);
4425 V1 = SVOp->getOperand(0);
4426 V2 = SVOp->getOperand(1);
4427 std::swap(V1IsSplat, V2IsSplat);
4428 std::swap(V1IsUndef, V2IsUndef);
4432 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4433 // Shuffling low element of v1 into undef, just return v1.
4436 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4437 // the instruction selector will not match, so get a canonical MOVL with
4438 // swapped operands to undo the commute.
4439 return getMOVL(DAG, dl, VT, V2, V1);
4442 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4443 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4444 X86::isUNPCKLMask(SVOp) ||
4445 X86::isUNPCKHMask(SVOp))
4449 // Normalize mask so all entries that point to V2 points to its first
4450 // element then try to match unpck{h|l} again. If match, return a
4451 // new vector_shuffle with the corrected mask.
4452 SDValue NewMask = NormalizeMask(SVOp, DAG);
4453 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4454 if (NSVOp != SVOp) {
4455 if (X86::isUNPCKLMask(NSVOp, true)) {
4457 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4464 // Commute is back and try unpck* again.
4465 // FIXME: this seems wrong.
4466 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4467 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4468 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4469 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4470 X86::isUNPCKLMask(NewSVOp) ||
4471 X86::isUNPCKHMask(NewSVOp))
4475 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4477 // Normalize the node to match x86 shuffle ops if needed
4478 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4479 return CommuteVectorShuffle(SVOp, DAG);
4481 // Check for legal shuffle and return?
4482 SmallVector<int, 16> PermMask;
4483 SVOp->getMask(PermMask);
4484 if (isShuffleMaskLegal(PermMask, VT))
4487 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4488 if (VT == MVT::v8i16) {
4489 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4490 if (NewOp.getNode())
4494 if (VT == MVT::v16i8) {
4495 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4496 if (NewOp.getNode())
4500 // Handle all 4 wide cases with a number of shuffles except for MMX.
4501 if (NumElems == 4 && !isMMX)
4502 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4508 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4509 SelectionDAG &DAG) {
4510 EVT VT = Op.getValueType();
4511 DebugLoc dl = Op.getDebugLoc();
4512 if (VT.getSizeInBits() == 8) {
4513 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4514 Op.getOperand(0), Op.getOperand(1));
4515 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4516 DAG.getValueType(VT));
4517 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4518 } else if (VT.getSizeInBits() == 16) {
4519 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4520 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4522 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4523 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4524 DAG.getNode(ISD::BIT_CONVERT, dl,
4528 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4529 Op.getOperand(0), Op.getOperand(1));
4530 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4531 DAG.getValueType(VT));
4532 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4533 } else if (VT == MVT::f32) {
4534 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4535 // the result back to FR32 register. It's only worth matching if the
4536 // result has a single use which is a store or a bitcast to i32. And in
4537 // the case of a store, it's not worth it if the index is a constant 0,
4538 // because a MOVSSmr can be used instead, which is smaller and faster.
4539 if (!Op.hasOneUse())
4541 SDNode *User = *Op.getNode()->use_begin();
4542 if ((User->getOpcode() != ISD::STORE ||
4543 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4544 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4545 (User->getOpcode() != ISD::BIT_CONVERT ||
4546 User->getValueType(0) != MVT::i32))
4548 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4549 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4552 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4553 } else if (VT == MVT::i32) {
4554 // ExtractPS works with constant index.
4555 if (isa<ConstantSDNode>(Op.getOperand(1)))
4563 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4564 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4567 if (Subtarget->hasSSE41()) {
4568 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4573 EVT VT = Op.getValueType();
4574 DebugLoc dl = Op.getDebugLoc();
4575 // TODO: handle v16i8.
4576 if (VT.getSizeInBits() == 16) {
4577 SDValue Vec = Op.getOperand(0);
4578 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4580 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4581 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4582 DAG.getNode(ISD::BIT_CONVERT, dl,
4585 // Transform it so it match pextrw which produces a 32-bit result.
4586 EVT EltVT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy+1);
4587 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4588 Op.getOperand(0), Op.getOperand(1));
4589 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4590 DAG.getValueType(VT));
4591 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4592 } else if (VT.getSizeInBits() == 32) {
4593 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4597 // SHUFPS the element to the lowest double word, then movss.
4598 int Mask[4] = { Idx, -1, -1, -1 };
4599 EVT VVT = Op.getOperand(0).getValueType();
4600 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4601 DAG.getUNDEF(VVT), Mask);
4602 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4603 DAG.getIntPtrConstant(0));
4604 } else if (VT.getSizeInBits() == 64) {
4605 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4606 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4607 // to match extract_elt for f64.
4608 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4612 // UNPCKHPD the element to the lowest double word, then movsd.
4613 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4614 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4615 int Mask[2] = { 1, -1 };
4616 EVT VVT = Op.getOperand(0).getValueType();
4617 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4618 DAG.getUNDEF(VVT), Mask);
4619 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4620 DAG.getIntPtrConstant(0));
4627 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4628 EVT VT = Op.getValueType();
4629 EVT EltVT = VT.getVectorElementType();
4630 DebugLoc dl = Op.getDebugLoc();
4632 SDValue N0 = Op.getOperand(0);
4633 SDValue N1 = Op.getOperand(1);
4634 SDValue N2 = Op.getOperand(2);
4636 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4637 isa<ConstantSDNode>(N2)) {
4638 unsigned Opc = (EltVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4640 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4642 if (N1.getValueType() != MVT::i32)
4643 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4644 if (N2.getValueType() != MVT::i32)
4645 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4646 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4647 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4648 // Bits [7:6] of the constant are the source select. This will always be
4649 // zero here. The DAG Combiner may combine an extract_elt index into these
4650 // bits. For example (insert (extract, 3), 2) could be matched by putting
4651 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4652 // Bits [5:4] of the constant are the destination select. This is the
4653 // value of the incoming immediate.
4654 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4655 // combine either bitwise AND or insert of float 0.0 to set these bits.
4656 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4657 // Create this as a scalar to vector..
4658 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4659 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4660 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4661 // PINSR* works with constant index.
4668 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4669 EVT VT = Op.getValueType();
4670 EVT EltVT = VT.getVectorElementType();
4672 if (Subtarget->hasSSE41())
4673 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4675 if (EltVT == MVT::i8)
4678 DebugLoc dl = Op.getDebugLoc();
4679 SDValue N0 = Op.getOperand(0);
4680 SDValue N1 = Op.getOperand(1);
4681 SDValue N2 = Op.getOperand(2);
4683 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4684 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4685 // as its second argument.
4686 if (N1.getValueType() != MVT::i32)
4687 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4688 if (N2.getValueType() != MVT::i32)
4689 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4690 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4696 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4697 DebugLoc dl = Op.getDebugLoc();
4698 if (Op.getValueType() == MVT::v2f32)
4699 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4700 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4701 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4702 Op.getOperand(0))));
4704 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4705 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4707 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4708 EVT VT = MVT::v2i32;
4709 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4716 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4717 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4720 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4721 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4722 // one of the above mentioned nodes. It has to be wrapped because otherwise
4723 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4724 // be used to form addressing mode. These wrapped nodes will be selected
4727 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4728 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4730 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4732 unsigned char OpFlag = 0;
4733 unsigned WrapperKind = X86ISD::Wrapper;
4734 CodeModel::Model M = getTargetMachine().getCodeModel();
4736 if (Subtarget->isPICStyleRIPRel() &&
4737 (M == CodeModel::Small || M == CodeModel::Kernel))
4738 WrapperKind = X86ISD::WrapperRIP;
4739 else if (Subtarget->isPICStyleGOT())
4740 OpFlag = X86II::MO_GOTOFF;
4741 else if (Subtarget->isPICStyleStubPIC())
4742 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4744 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4746 CP->getOffset(), OpFlag);
4747 DebugLoc DL = CP->getDebugLoc();
4748 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4749 // With PIC, the address is actually $g + Offset.
4751 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4752 DAG.getNode(X86ISD::GlobalBaseReg,
4753 DebugLoc::getUnknownLoc(), getPointerTy()),
4760 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4761 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4763 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4765 unsigned char OpFlag = 0;
4766 unsigned WrapperKind = X86ISD::Wrapper;
4767 CodeModel::Model M = getTargetMachine().getCodeModel();
4769 if (Subtarget->isPICStyleRIPRel() &&
4770 (M == CodeModel::Small || M == CodeModel::Kernel))
4771 WrapperKind = X86ISD::WrapperRIP;
4772 else if (Subtarget->isPICStyleGOT())
4773 OpFlag = X86II::MO_GOTOFF;
4774 else if (Subtarget->isPICStyleStubPIC())
4775 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4777 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4779 DebugLoc DL = JT->getDebugLoc();
4780 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4782 // With PIC, the address is actually $g + Offset.
4784 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4785 DAG.getNode(X86ISD::GlobalBaseReg,
4786 DebugLoc::getUnknownLoc(), getPointerTy()),
4794 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4795 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4797 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4799 unsigned char OpFlag = 0;
4800 unsigned WrapperKind = X86ISD::Wrapper;
4801 CodeModel::Model M = getTargetMachine().getCodeModel();
4803 if (Subtarget->isPICStyleRIPRel() &&
4804 (M == CodeModel::Small || M == CodeModel::Kernel))
4805 WrapperKind = X86ISD::WrapperRIP;
4806 else if (Subtarget->isPICStyleGOT())
4807 OpFlag = X86II::MO_GOTOFF;
4808 else if (Subtarget->isPICStyleStubPIC())
4809 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4811 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4813 DebugLoc DL = Op.getDebugLoc();
4814 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4817 // With PIC, the address is actually $g + Offset.
4818 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4819 !Subtarget->is64Bit()) {
4820 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4821 DAG.getNode(X86ISD::GlobalBaseReg,
4822 DebugLoc::getUnknownLoc(),
4831 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
4832 // Create the TargetBlockAddressAddress node.
4833 unsigned char OpFlags =
4834 Subtarget->ClassifyBlockAddressReference();
4835 CodeModel::Model M = getTargetMachine().getCodeModel();
4836 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
4837 DebugLoc dl = Op.getDebugLoc();
4838 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
4839 /*isTarget=*/true, OpFlags);
4841 if (Subtarget->isPICStyleRIPRel() &&
4842 (M == CodeModel::Small || M == CodeModel::Kernel))
4843 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4845 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4847 // With PIC, the address is actually $g + Offset.
4848 if (isGlobalRelativeToPICBase(OpFlags)) {
4849 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4850 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4858 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4860 SelectionDAG &DAG) const {
4861 // Create the TargetGlobalAddress node, folding in the constant
4862 // offset if it is legal.
4863 unsigned char OpFlags =
4864 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
4865 CodeModel::Model M = getTargetMachine().getCodeModel();
4867 if (OpFlags == X86II::MO_NO_FLAG &&
4868 X86::isOffsetSuitableForCodeModel(Offset, M)) {
4869 // A direct static reference to a global.
4870 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4873 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
4876 if (Subtarget->isPICStyleRIPRel() &&
4877 (M == CodeModel::Small || M == CodeModel::Kernel))
4878 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4880 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4882 // With PIC, the address is actually $g + Offset.
4883 if (isGlobalRelativeToPICBase(OpFlags)) {
4884 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4885 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4889 // For globals that require a load from a stub to get the address, emit the
4891 if (isGlobalStubReference(OpFlags))
4892 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4893 PseudoSourceValue::getGOT(), 0);
4895 // If there was a non-zero offset that we didn't fold, create an explicit
4898 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
4899 DAG.getConstant(Offset, getPointerTy()));
4905 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4906 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4907 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4908 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
4912 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
4913 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
4914 unsigned char OperandFlags) {
4915 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4916 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4917 DebugLoc dl = GA->getDebugLoc();
4918 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4919 GA->getValueType(0),
4923 SDValue Ops[] = { Chain, TGA, *InFlag };
4924 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
4926 SDValue Ops[] = { Chain, TGA };
4927 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
4930 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
4931 MFI->setHasCalls(true);
4933 SDValue Flag = Chain.getValue(1);
4934 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
4937 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4939 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4942 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4943 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
4944 DAG.getNode(X86ISD::GlobalBaseReg,
4945 DebugLoc::getUnknownLoc(),
4947 InFlag = Chain.getValue(1);
4949 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
4952 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4954 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4956 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
4957 X86::RAX, X86II::MO_TLSGD);
4960 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4961 // "local exec" model.
4962 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4963 const EVT PtrVT, TLSModel::Model model,
4965 DebugLoc dl = GA->getDebugLoc();
4966 // Get the Thread Pointer
4967 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
4968 DebugLoc::getUnknownLoc(), PtrVT,
4969 DAG.getRegister(is64Bit? X86::FS : X86::GS,
4972 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
4975 unsigned char OperandFlags = 0;
4976 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
4978 unsigned WrapperKind = X86ISD::Wrapper;
4979 if (model == TLSModel::LocalExec) {
4980 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
4981 } else if (is64Bit) {
4982 assert(model == TLSModel::InitialExec);
4983 OperandFlags = X86II::MO_GOTTPOFF;
4984 WrapperKind = X86ISD::WrapperRIP;
4986 assert(model == TLSModel::InitialExec);
4987 OperandFlags = X86II::MO_INDNTPOFF;
4990 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4992 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
4993 GA->getOffset(), OperandFlags);
4994 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
4996 if (model == TLSModel::InitialExec)
4997 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
4998 PseudoSourceValue::getGOT(), 0);
5000 // The address of the thread local variable is the add of the thread
5001 // pointer with the offset of the variable.
5002 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5006 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
5007 // TODO: implement the "local dynamic" model
5008 // TODO: implement the "initial exec"model for pic executables
5009 assert(Subtarget->isTargetELF() &&
5010 "TLS not implemented for non-ELF targets");
5011 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5012 const GlobalValue *GV = GA->getGlobal();
5014 // If GV is an alias then use the aliasee for determining
5015 // thread-localness.
5016 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5017 GV = GA->resolveAliasedGlobal(false);
5019 TLSModel::Model model = getTLSModel(GV,
5020 getTargetMachine().getRelocationModel());
5023 case TLSModel::GeneralDynamic:
5024 case TLSModel::LocalDynamic: // not implemented
5025 if (Subtarget->is64Bit())
5026 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5027 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5029 case TLSModel::InitialExec:
5030 case TLSModel::LocalExec:
5031 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5032 Subtarget->is64Bit());
5035 llvm_unreachable("Unreachable");
5040 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5041 /// take a 2 x i32 value to shift plus a shift amount.
5042 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
5043 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5044 EVT VT = Op.getValueType();
5045 unsigned VTBits = VT.getSizeInBits();
5046 DebugLoc dl = Op.getDebugLoc();
5047 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5048 SDValue ShOpLo = Op.getOperand(0);
5049 SDValue ShOpHi = Op.getOperand(1);
5050 SDValue ShAmt = Op.getOperand(2);
5051 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5052 DAG.getConstant(VTBits - 1, MVT::i8))
5053 : DAG.getConstant(0, VT);
5056 if (Op.getOpcode() == ISD::SHL_PARTS) {
5057 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5058 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5060 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5061 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5064 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5065 DAG.getConstant(VTBits, MVT::i8));
5066 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
5067 AndNode, DAG.getConstant(0, MVT::i8));
5070 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5071 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5072 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5074 if (Op.getOpcode() == ISD::SHL_PARTS) {
5075 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5076 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5078 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5079 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5082 SDValue Ops[2] = { Lo, Hi };
5083 return DAG.getMergeValues(Ops, 2, dl);
5086 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5087 EVT SrcVT = Op.getOperand(0).getValueType();
5089 if (SrcVT.isVector()) {
5090 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5096 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5097 "Unknown SINT_TO_FP to lower!");
5099 // These are really Legal; return the operand so the caller accepts it as
5101 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5103 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5104 Subtarget->is64Bit()) {
5108 DebugLoc dl = Op.getDebugLoc();
5109 unsigned Size = SrcVT.getSizeInBits()/8;
5110 MachineFunction &MF = DAG.getMachineFunction();
5111 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5112 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5113 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5115 PseudoSourceValue::getFixedStack(SSFI), 0);
5116 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5119 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5121 SelectionDAG &DAG) {
5123 DebugLoc dl = Op.getDebugLoc();
5125 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5127 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5129 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5130 SmallVector<SDValue, 8> Ops;
5131 Ops.push_back(Chain);
5132 Ops.push_back(StackSlot);
5133 Ops.push_back(DAG.getValueType(SrcVT));
5134 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5135 Tys, &Ops[0], Ops.size());
5138 Chain = Result.getValue(1);
5139 SDValue InFlag = Result.getValue(2);
5141 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5142 // shouldn't be necessary except that RFP cannot be live across
5143 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5144 MachineFunction &MF = DAG.getMachineFunction();
5145 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5146 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5147 Tys = DAG.getVTList(MVT::Other);
5148 SmallVector<SDValue, 8> Ops;
5149 Ops.push_back(Chain);
5150 Ops.push_back(Result);
5151 Ops.push_back(StackSlot);
5152 Ops.push_back(DAG.getValueType(Op.getValueType()));
5153 Ops.push_back(InFlag);
5154 Chain = DAG.getNode(X86ISD::FST, dl, Tys, &Ops[0], Ops.size());
5155 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5156 PseudoSourceValue::getFixedStack(SSFI), 0);
5162 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5163 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5164 // This algorithm is not obvious. Here it is in C code, more or less:
5166 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5167 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5168 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5170 // Copy ints to xmm registers.
5171 __m128i xh = _mm_cvtsi32_si128( hi );
5172 __m128i xl = _mm_cvtsi32_si128( lo );
5174 // Combine into low half of a single xmm register.
5175 __m128i x = _mm_unpacklo_epi32( xh, xl );
5179 // Merge in appropriate exponents to give the integer bits the right
5181 x = _mm_unpacklo_epi32( x, exp );
5183 // Subtract away the biases to deal with the IEEE-754 double precision
5185 d = _mm_sub_pd( (__m128d) x, bias );
5187 // All conversions up to here are exact. The correctly rounded result is
5188 // calculated using the current rounding mode using the following
5190 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5191 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5192 // store doesn't really need to be here (except
5193 // maybe to zero the other double)
5198 DebugLoc dl = Op.getDebugLoc();
5199 LLVMContext *Context = DAG.getContext();
5201 // Build some magic constants.
5202 std::vector<Constant*> CV0;
5203 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5204 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5205 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5206 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5207 Constant *C0 = ConstantVector::get(CV0);
5208 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5210 std::vector<Constant*> CV1;
5212 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5214 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5215 Constant *C1 = ConstantVector::get(CV1);
5216 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5218 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5219 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5221 DAG.getIntPtrConstant(1)));
5222 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5223 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5225 DAG.getIntPtrConstant(0)));
5226 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5227 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5228 PseudoSourceValue::getConstantPool(), 0,
5230 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5231 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5232 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5233 PseudoSourceValue::getConstantPool(), 0,
5235 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5237 // Add the halves; easiest way is to swap them into another reg first.
5238 int ShufMask[2] = { 1, -1 };
5239 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5240 DAG.getUNDEF(MVT::v2f64), ShufMask);
5241 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5242 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5243 DAG.getIntPtrConstant(0));
5246 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5247 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5248 DebugLoc dl = Op.getDebugLoc();
5249 // FP constant to bias correct the final result.
5250 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5253 // Load the 32-bit value into an XMM register.
5254 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5255 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5257 DAG.getIntPtrConstant(0)));
5259 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5260 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5261 DAG.getIntPtrConstant(0));
5263 // Or the load with the bias.
5264 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5265 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5266 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5268 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5269 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5270 MVT::v2f64, Bias)));
5271 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5272 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5273 DAG.getIntPtrConstant(0));
5275 // Subtract the bias.
5276 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5278 // Handle final rounding.
5279 EVT DestVT = Op.getValueType();
5281 if (DestVT.bitsLT(MVT::f64)) {
5282 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5283 DAG.getIntPtrConstant(0));
5284 } else if (DestVT.bitsGT(MVT::f64)) {
5285 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5288 // Handle final rounding.
5292 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5293 SDValue N0 = Op.getOperand(0);
5294 DebugLoc dl = Op.getDebugLoc();
5296 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5297 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5298 // the optimization here.
5299 if (DAG.SignBitIsZero(N0))
5300 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5302 EVT SrcVT = N0.getValueType();
5303 if (SrcVT == MVT::i64) {
5304 // We only handle SSE2 f64 target here; caller can expand the rest.
5305 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5308 return LowerUINT_TO_FP_i64(Op, DAG);
5309 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5310 return LowerUINT_TO_FP_i32(Op, DAG);
5313 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5315 // Make a 64-bit buffer, and use it to build an FILD.
5316 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5317 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5318 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5319 getPointerTy(), StackSlot, WordOff);
5320 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5321 StackSlot, NULL, 0);
5322 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5323 OffsetSlot, NULL, 0);
5324 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5327 std::pair<SDValue,SDValue> X86TargetLowering::
5328 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5329 DebugLoc dl = Op.getDebugLoc();
5331 EVT DstTy = Op.getValueType();
5334 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5338 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5339 DstTy.getSimpleVT() >= MVT::i16 &&
5340 "Unknown FP_TO_SINT to lower!");
5342 // These are really Legal.
5343 if (DstTy == MVT::i32 &&
5344 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5345 return std::make_pair(SDValue(), SDValue());
5346 if (Subtarget->is64Bit() &&
5347 DstTy == MVT::i64 &&
5348 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5349 return std::make_pair(SDValue(), SDValue());
5351 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5353 MachineFunction &MF = DAG.getMachineFunction();
5354 unsigned MemSize = DstTy.getSizeInBits()/8;
5355 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5356 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5359 switch (DstTy.getSimpleVT().SimpleTy) {
5360 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5361 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5362 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5363 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5366 SDValue Chain = DAG.getEntryNode();
5367 SDValue Value = Op.getOperand(0);
5368 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5369 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5370 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5371 PseudoSourceValue::getFixedStack(SSFI), 0);
5372 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5374 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5376 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5377 Chain = Value.getValue(1);
5378 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5379 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5382 // Build the FP_TO_INT*_IN_MEM
5383 SDValue Ops[] = { Chain, Value, StackSlot };
5384 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5386 return std::make_pair(FIST, StackSlot);
5389 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5390 if (Op.getValueType().isVector()) {
5391 if (Op.getValueType() == MVT::v2i32 &&
5392 Op.getOperand(0).getValueType() == MVT::v2f64) {
5398 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5399 SDValue FIST = Vals.first, StackSlot = Vals.second;
5400 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5401 if (FIST.getNode() == 0) return Op;
5404 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5405 FIST, StackSlot, NULL, 0);
5408 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5409 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5410 SDValue FIST = Vals.first, StackSlot = Vals.second;
5411 assert(FIST.getNode() && "Unexpected failure");
5414 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5415 FIST, StackSlot, NULL, 0);
5418 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5419 LLVMContext *Context = DAG.getContext();
5420 DebugLoc dl = Op.getDebugLoc();
5421 EVT VT = Op.getValueType();
5424 EltVT = VT.getVectorElementType();
5425 std::vector<Constant*> CV;
5426 if (EltVT == MVT::f64) {
5427 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5431 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5437 Constant *C = ConstantVector::get(CV);
5438 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5439 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5440 PseudoSourceValue::getConstantPool(), 0,
5442 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5445 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5446 LLVMContext *Context = DAG.getContext();
5447 DebugLoc dl = Op.getDebugLoc();
5448 EVT VT = Op.getValueType();
5451 EltVT = VT.getVectorElementType();
5452 std::vector<Constant*> CV;
5453 if (EltVT == MVT::f64) {
5454 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5458 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5464 Constant *C = ConstantVector::get(CV);
5465 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5466 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5467 PseudoSourceValue::getConstantPool(), 0,
5469 if (VT.isVector()) {
5470 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5471 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5472 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5474 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5476 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5480 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5481 LLVMContext *Context = DAG.getContext();
5482 SDValue Op0 = Op.getOperand(0);
5483 SDValue Op1 = Op.getOperand(1);
5484 DebugLoc dl = Op.getDebugLoc();
5485 EVT VT = Op.getValueType();
5486 EVT SrcVT = Op1.getValueType();
5488 // If second operand is smaller, extend it first.
5489 if (SrcVT.bitsLT(VT)) {
5490 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5493 // And if it is bigger, shrink it first.
5494 if (SrcVT.bitsGT(VT)) {
5495 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5499 // At this point the operands and the result should have the same
5500 // type, and that won't be f80 since that is not custom lowered.
5502 // First get the sign bit of second operand.
5503 std::vector<Constant*> CV;
5504 if (SrcVT == MVT::f64) {
5505 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5506 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5508 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5509 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5510 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5511 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5513 Constant *C = ConstantVector::get(CV);
5514 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5515 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5516 PseudoSourceValue::getConstantPool(), 0,
5518 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5520 // Shift sign bit right or left if the two operands have different types.
5521 if (SrcVT.bitsGT(VT)) {
5522 // Op0 is MVT::f32, Op1 is MVT::f64.
5523 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5524 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5525 DAG.getConstant(32, MVT::i32));
5526 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5527 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5528 DAG.getIntPtrConstant(0));
5531 // Clear first operand sign bit.
5533 if (VT == MVT::f64) {
5534 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5535 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5537 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5538 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5539 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5540 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5542 C = ConstantVector::get(CV);
5543 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5544 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5545 PseudoSourceValue::getConstantPool(), 0,
5547 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5549 // Or the value with the sign bit.
5550 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5553 /// Emit nodes that will be selected as "test Op0,Op0", or something
5555 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5556 SelectionDAG &DAG) {
5557 DebugLoc dl = Op.getDebugLoc();
5559 // CF and OF aren't always set the way we want. Determine which
5560 // of these we need.
5561 bool NeedCF = false;
5562 bool NeedOF = false;
5564 case X86::COND_A: case X86::COND_AE:
5565 case X86::COND_B: case X86::COND_BE:
5568 case X86::COND_G: case X86::COND_GE:
5569 case X86::COND_L: case X86::COND_LE:
5570 case X86::COND_O: case X86::COND_NO:
5576 // See if we can use the EFLAGS value from the operand instead of
5577 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5578 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5579 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5580 unsigned Opcode = 0;
5581 unsigned NumOperands = 0;
5582 switch (Op.getNode()->getOpcode()) {
5584 // Due to an isel shortcoming, be conservative if this add is likely to
5585 // be selected as part of a load-modify-store instruction. When the root
5586 // node in a match is a store, isel doesn't know how to remap non-chain
5587 // non-flag uses of other nodes in the match, such as the ADD in this
5588 // case. This leads to the ADD being left around and reselected, with
5589 // the result being two adds in the output.
5590 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5591 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5592 if (UI->getOpcode() == ISD::STORE)
5594 if (ConstantSDNode *C =
5595 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5596 // An add of one will be selected as an INC.
5597 if (C->getAPIntValue() == 1) {
5598 Opcode = X86ISD::INC;
5602 // An add of negative one (subtract of one) will be selected as a DEC.
5603 if (C->getAPIntValue().isAllOnesValue()) {
5604 Opcode = X86ISD::DEC;
5609 // Otherwise use a regular EFLAGS-setting add.
5610 Opcode = X86ISD::ADD;
5614 // If the primary and result isn't used, don't bother using X86ISD::AND,
5615 // because a TEST instruction will be better.
5616 bool NonFlagUse = false;
5617 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5618 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5619 if (UI->getOpcode() != ISD::BRCOND &&
5620 UI->getOpcode() != ISD::SELECT &&
5621 UI->getOpcode() != ISD::SETCC) {
5632 // Due to the ISEL shortcoming noted above, be conservative if this op is
5633 // likely to be selected as part of a load-modify-store instruction.
5634 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5635 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5636 if (UI->getOpcode() == ISD::STORE)
5638 // Otherwise use a regular EFLAGS-setting instruction.
5639 switch (Op.getNode()->getOpcode()) {
5640 case ISD::SUB: Opcode = X86ISD::SUB; break;
5641 case ISD::OR: Opcode = X86ISD::OR; break;
5642 case ISD::XOR: Opcode = X86ISD::XOR; break;
5643 case ISD::AND: Opcode = X86ISD::AND; break;
5644 default: llvm_unreachable("unexpected operator!");
5655 return SDValue(Op.getNode(), 1);
5661 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5662 SmallVector<SDValue, 4> Ops;
5663 for (unsigned i = 0; i != NumOperands; ++i)
5664 Ops.push_back(Op.getOperand(i));
5665 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5666 DAG.ReplaceAllUsesWith(Op, New);
5667 return SDValue(New.getNode(), 1);
5671 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5672 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5673 DAG.getConstant(0, Op.getValueType()));
5676 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5678 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5679 SelectionDAG &DAG) {
5680 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5681 if (C->getAPIntValue() == 0)
5682 return EmitTest(Op0, X86CC, DAG);
5684 DebugLoc dl = Op0.getDebugLoc();
5685 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5688 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5689 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5690 SDValue Op0 = Op.getOperand(0);
5691 SDValue Op1 = Op.getOperand(1);
5692 DebugLoc dl = Op.getDebugLoc();
5693 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5695 // Lower (X & (1 << N)) == 0 to BT(X, N).
5696 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5697 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5698 if (Op0.getOpcode() == ISD::AND &&
5700 Op1.getOpcode() == ISD::Constant &&
5701 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5702 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5704 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5705 if (ConstantSDNode *Op010C =
5706 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5707 if (Op010C->getZExtValue() == 1) {
5708 LHS = Op0.getOperand(0);
5709 RHS = Op0.getOperand(1).getOperand(1);
5711 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5712 if (ConstantSDNode *Op000C =
5713 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5714 if (Op000C->getZExtValue() == 1) {
5715 LHS = Op0.getOperand(1);
5716 RHS = Op0.getOperand(0).getOperand(1);
5718 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5719 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5720 SDValue AndLHS = Op0.getOperand(0);
5721 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5722 LHS = AndLHS.getOperand(0);
5723 RHS = AndLHS.getOperand(1);
5727 if (LHS.getNode()) {
5728 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5729 // instruction. Since the shift amount is in-range-or-undefined, we know
5730 // that doing a bittest on the i16 value is ok. We extend to i32 because
5731 // the encoding for the i16 version is larger than the i32 version.
5732 if (LHS.getValueType() == MVT::i8)
5733 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5735 // If the operand types disagree, extend the shift amount to match. Since
5736 // BT ignores high bits (like shifts) we can use anyextend.
5737 if (LHS.getValueType() != RHS.getValueType())
5738 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5740 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5741 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5742 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5743 DAG.getConstant(Cond, MVT::i8), BT);
5747 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5748 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5749 if (X86CC == X86::COND_INVALID)
5752 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5754 // Use sbb x, x to materialize carry bit into a GPR.
5755 // FIXME: Temporarily disabled since it breaks self-hosting. It's apparently
5756 // miscompiling ARMISelDAGToDAG.cpp.
5757 if (0 && !isFP && X86CC == X86::COND_B) {
5758 return DAG.getNode(ISD::AND, dl, MVT::i8,
5759 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
5760 DAG.getConstant(X86CC, MVT::i8), Cond),
5761 DAG.getConstant(1, MVT::i8));
5764 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5765 DAG.getConstant(X86CC, MVT::i8), Cond);
5768 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5770 SDValue Op0 = Op.getOperand(0);
5771 SDValue Op1 = Op.getOperand(1);
5772 SDValue CC = Op.getOperand(2);
5773 EVT VT = Op.getValueType();
5774 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5775 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5776 DebugLoc dl = Op.getDebugLoc();
5780 EVT VT0 = Op0.getValueType();
5781 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5782 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5785 switch (SetCCOpcode) {
5788 case ISD::SETEQ: SSECC = 0; break;
5790 case ISD::SETGT: Swap = true; // Fallthrough
5792 case ISD::SETOLT: SSECC = 1; break;
5794 case ISD::SETGE: Swap = true; // Fallthrough
5796 case ISD::SETOLE: SSECC = 2; break;
5797 case ISD::SETUO: SSECC = 3; break;
5799 case ISD::SETNE: SSECC = 4; break;
5800 case ISD::SETULE: Swap = true;
5801 case ISD::SETUGE: SSECC = 5; break;
5802 case ISD::SETULT: Swap = true;
5803 case ISD::SETUGT: SSECC = 6; break;
5804 case ISD::SETO: SSECC = 7; break;
5807 std::swap(Op0, Op1);
5809 // In the two special cases we can't handle, emit two comparisons.
5811 if (SetCCOpcode == ISD::SETUEQ) {
5813 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5814 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5815 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5817 else if (SetCCOpcode == ISD::SETONE) {
5819 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5820 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5821 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5823 llvm_unreachable("Illegal FP comparison");
5825 // Handle all other FP comparisons here.
5826 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5829 // We are handling one of the integer comparisons here. Since SSE only has
5830 // GT and EQ comparisons for integer, swapping operands and multiple
5831 // operations may be required for some comparisons.
5832 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5833 bool Swap = false, Invert = false, FlipSigns = false;
5835 switch (VT.getSimpleVT().SimpleTy) {
5838 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5840 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5842 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5843 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5846 switch (SetCCOpcode) {
5848 case ISD::SETNE: Invert = true;
5849 case ISD::SETEQ: Opc = EQOpc; break;
5850 case ISD::SETLT: Swap = true;
5851 case ISD::SETGT: Opc = GTOpc; break;
5852 case ISD::SETGE: Swap = true;
5853 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5854 case ISD::SETULT: Swap = true;
5855 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5856 case ISD::SETUGE: Swap = true;
5857 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5860 std::swap(Op0, Op1);
5862 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5863 // bits of the inputs before performing those operations.
5865 EVT EltVT = VT.getVectorElementType();
5866 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5868 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5869 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5871 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5872 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5875 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5877 // If the logical-not of the result is required, perform that now.
5879 Result = DAG.getNOT(dl, Result, VT);
5884 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5885 static bool isX86LogicalCmp(SDValue Op) {
5886 unsigned Opc = Op.getNode()->getOpcode();
5887 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
5889 if (Op.getResNo() == 1 &&
5890 (Opc == X86ISD::ADD ||
5891 Opc == X86ISD::SUB ||
5892 Opc == X86ISD::SMUL ||
5893 Opc == X86ISD::UMUL ||
5894 Opc == X86ISD::INC ||
5895 Opc == X86ISD::DEC ||
5896 Opc == X86ISD::OR ||
5897 Opc == X86ISD::XOR ||
5898 Opc == X86ISD::AND))
5904 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5905 bool addTest = true;
5906 SDValue Cond = Op.getOperand(0);
5907 DebugLoc dl = Op.getDebugLoc();
5910 if (Cond.getOpcode() == ISD::SETCC) {
5911 SDValue NewCond = LowerSETCC(Cond, DAG);
5912 if (NewCond.getNode())
5916 // Look pass (and (setcc_carry (cmp ...)), 1).
5917 if (Cond.getOpcode() == ISD::AND &&
5918 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
5919 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
5920 if (C && C->getAPIntValue() == 1)
5921 Cond = Cond.getOperand(0);
5924 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5925 // setting operand in place of the X86ISD::SETCC.
5926 if (Cond.getOpcode() == X86ISD::SETCC ||
5927 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
5928 CC = Cond.getOperand(0);
5930 SDValue Cmp = Cond.getOperand(1);
5931 unsigned Opc = Cmp.getOpcode();
5932 EVT VT = Op.getValueType();
5934 bool IllegalFPCMov = false;
5935 if (VT.isFloatingPoint() && !VT.isVector() &&
5936 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5937 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5939 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
5940 Opc == X86ISD::BT) { // FIXME
5947 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5948 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5951 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
5952 SmallVector<SDValue, 4> Ops;
5953 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5954 // condition is true.
5955 Ops.push_back(Op.getOperand(2));
5956 Ops.push_back(Op.getOperand(1));
5958 Ops.push_back(Cond);
5959 return DAG.getNode(X86ISD::CMOV, dl, VTs, &Ops[0], Ops.size());
5962 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
5963 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
5964 // from the AND / OR.
5965 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
5966 Opc = Op.getOpcode();
5967 if (Opc != ISD::OR && Opc != ISD::AND)
5969 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5970 Op.getOperand(0).hasOneUse() &&
5971 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
5972 Op.getOperand(1).hasOneUse());
5975 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
5976 // 1 and that the SETCC node has a single use.
5977 static bool isXor1OfSetCC(SDValue Op) {
5978 if (Op.getOpcode() != ISD::XOR)
5980 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5981 if (N1C && N1C->getAPIntValue() == 1) {
5982 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5983 Op.getOperand(0).hasOneUse();
5988 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5989 bool addTest = true;
5990 SDValue Chain = Op.getOperand(0);
5991 SDValue Cond = Op.getOperand(1);
5992 SDValue Dest = Op.getOperand(2);
5993 DebugLoc dl = Op.getDebugLoc();
5996 if (Cond.getOpcode() == ISD::SETCC) {
5997 SDValue NewCond = LowerSETCC(Cond, DAG);
5998 if (NewCond.getNode())
6002 // FIXME: LowerXALUO doesn't handle these!!
6003 else if (Cond.getOpcode() == X86ISD::ADD ||
6004 Cond.getOpcode() == X86ISD::SUB ||
6005 Cond.getOpcode() == X86ISD::SMUL ||
6006 Cond.getOpcode() == X86ISD::UMUL)
6007 Cond = LowerXALUO(Cond, DAG);
6010 // Look pass (and (setcc_carry (cmp ...)), 1).
6011 if (Cond.getOpcode() == ISD::AND &&
6012 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6013 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6014 if (C && C->getAPIntValue() == 1)
6015 Cond = Cond.getOperand(0);
6018 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6019 // setting operand in place of the X86ISD::SETCC.
6020 if (Cond.getOpcode() == X86ISD::SETCC ||
6021 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6022 CC = Cond.getOperand(0);
6024 SDValue Cmp = Cond.getOperand(1);
6025 unsigned Opc = Cmp.getOpcode();
6026 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6027 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6031 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6035 // These can only come from an arithmetic instruction with overflow,
6036 // e.g. SADDO, UADDO.
6037 Cond = Cond.getNode()->getOperand(1);
6044 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6045 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6046 if (CondOpc == ISD::OR) {
6047 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6048 // two branches instead of an explicit OR instruction with a
6050 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6051 isX86LogicalCmp(Cmp)) {
6052 CC = Cond.getOperand(0).getOperand(0);
6053 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6054 Chain, Dest, CC, Cmp);
6055 CC = Cond.getOperand(1).getOperand(0);
6059 } else { // ISD::AND
6060 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6061 // two branches instead of an explicit AND instruction with a
6062 // separate test. However, we only do this if this block doesn't
6063 // have a fall-through edge, because this requires an explicit
6064 // jmp when the condition is false.
6065 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6066 isX86LogicalCmp(Cmp) &&
6067 Op.getNode()->hasOneUse()) {
6068 X86::CondCode CCode =
6069 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6070 CCode = X86::GetOppositeBranchCondition(CCode);
6071 CC = DAG.getConstant(CCode, MVT::i8);
6072 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6073 // Look for an unconditional branch following this conditional branch.
6074 // We need this because we need to reverse the successors in order
6075 // to implement FCMP_OEQ.
6076 if (User.getOpcode() == ISD::BR) {
6077 SDValue FalseBB = User.getOperand(1);
6079 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6080 assert(NewBR == User);
6083 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6084 Chain, Dest, CC, Cmp);
6085 X86::CondCode CCode =
6086 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6087 CCode = X86::GetOppositeBranchCondition(CCode);
6088 CC = DAG.getConstant(CCode, MVT::i8);
6094 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6095 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6096 // It should be transformed during dag combiner except when the condition
6097 // is set by a arithmetics with overflow node.
6098 X86::CondCode CCode =
6099 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6100 CCode = X86::GetOppositeBranchCondition(CCode);
6101 CC = DAG.getConstant(CCode, MVT::i8);
6102 Cond = Cond.getOperand(0).getOperand(1);
6108 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6109 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6111 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6112 Chain, Dest, CC, Cond);
6116 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6117 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6118 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6119 // that the guard pages used by the OS virtual memory manager are allocated in
6120 // correct sequence.
6122 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6123 SelectionDAG &DAG) {
6124 assert(Subtarget->isTargetCygMing() &&
6125 "This should be used only on Cygwin/Mingw targets");
6126 DebugLoc dl = Op.getDebugLoc();
6129 SDValue Chain = Op.getOperand(0);
6130 SDValue Size = Op.getOperand(1);
6131 // FIXME: Ensure alignment here
6135 EVT IntPtr = getPointerTy();
6136 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6138 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
6140 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6141 Flag = Chain.getValue(1);
6143 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6144 SDValue Ops[] = { Chain,
6145 DAG.getTargetExternalSymbol("_alloca", IntPtr),
6146 DAG.getRegister(X86::EAX, IntPtr),
6147 DAG.getRegister(X86StackPtr, SPTy),
6149 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
6150 Flag = Chain.getValue(1);
6152 Chain = DAG.getCALLSEQ_END(Chain,
6153 DAG.getIntPtrConstant(0, true),
6154 DAG.getIntPtrConstant(0, true),
6157 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6159 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6160 return DAG.getMergeValues(Ops1, 2, dl);
6164 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6166 SDValue Dst, SDValue Src,
6167 SDValue Size, unsigned Align,
6169 uint64_t DstSVOff) {
6170 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6172 // If not DWORD aligned or size is more than the threshold, call the library.
6173 // The libc version is likely to be faster for these cases. It can use the
6174 // address value and run time information about the CPU.
6175 if ((Align & 3) != 0 ||
6177 ConstantSize->getZExtValue() >
6178 getSubtarget()->getMaxInlineSizeThreshold()) {
6179 SDValue InFlag(0, 0);
6181 // Check to see if there is a specialized entry-point for memory zeroing.
6182 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6184 if (const char *bzeroEntry = V &&
6185 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6186 EVT IntPtr = getPointerTy();
6187 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6188 TargetLowering::ArgListTy Args;
6189 TargetLowering::ArgListEntry Entry;
6191 Entry.Ty = IntPtrTy;
6192 Args.push_back(Entry);
6194 Args.push_back(Entry);
6195 std::pair<SDValue,SDValue> CallResult =
6196 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6197 false, false, false, false,
6198 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6199 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
6200 return CallResult.second;
6203 // Otherwise have the target-independent code call memset.
6207 uint64_t SizeVal = ConstantSize->getZExtValue();
6208 SDValue InFlag(0, 0);
6211 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6212 unsigned BytesLeft = 0;
6213 bool TwoRepStos = false;
6216 uint64_t Val = ValC->getZExtValue() & 255;
6218 // If the value is a constant, then we can potentially use larger sets.
6219 switch (Align & 3) {
6220 case 2: // WORD aligned
6223 Val = (Val << 8) | Val;
6225 case 0: // DWORD aligned
6228 Val = (Val << 8) | Val;
6229 Val = (Val << 16) | Val;
6230 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6233 Val = (Val << 32) | Val;
6236 default: // Byte aligned
6239 Count = DAG.getIntPtrConstant(SizeVal);
6243 if (AVT.bitsGT(MVT::i8)) {
6244 unsigned UBytes = AVT.getSizeInBits() / 8;
6245 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6246 BytesLeft = SizeVal % UBytes;
6249 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6251 InFlag = Chain.getValue(1);
6254 Count = DAG.getIntPtrConstant(SizeVal);
6255 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6256 InFlag = Chain.getValue(1);
6259 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6262 InFlag = Chain.getValue(1);
6263 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6266 InFlag = Chain.getValue(1);
6268 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6269 SmallVector<SDValue, 8> Ops;
6270 Ops.push_back(Chain);
6271 Ops.push_back(DAG.getValueType(AVT));
6272 Ops.push_back(InFlag);
6273 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
6276 InFlag = Chain.getValue(1);
6278 EVT CVT = Count.getValueType();
6279 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6280 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6281 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6284 InFlag = Chain.getValue(1);
6285 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6287 Ops.push_back(Chain);
6288 Ops.push_back(DAG.getValueType(MVT::i8));
6289 Ops.push_back(InFlag);
6290 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
6291 } else if (BytesLeft) {
6292 // Handle the last 1 - 7 bytes.
6293 unsigned Offset = SizeVal - BytesLeft;
6294 EVT AddrVT = Dst.getValueType();
6295 EVT SizeVT = Size.getValueType();
6297 Chain = DAG.getMemset(Chain, dl,
6298 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6299 DAG.getConstant(Offset, AddrVT)),
6301 DAG.getConstant(BytesLeft, SizeVT),
6302 Align, DstSV, DstSVOff + Offset);
6305 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6310 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6311 SDValue Chain, SDValue Dst, SDValue Src,
6312 SDValue Size, unsigned Align,
6314 const Value *DstSV, uint64_t DstSVOff,
6315 const Value *SrcSV, uint64_t SrcSVOff) {
6316 // This requires the copy size to be a constant, preferrably
6317 // within a subtarget-specific limit.
6318 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6321 uint64_t SizeVal = ConstantSize->getZExtValue();
6322 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6325 /// If not DWORD aligned, call the library.
6326 if ((Align & 3) != 0)
6331 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6334 unsigned UBytes = AVT.getSizeInBits() / 8;
6335 unsigned CountVal = SizeVal / UBytes;
6336 SDValue Count = DAG.getIntPtrConstant(CountVal);
6337 unsigned BytesLeft = SizeVal % UBytes;
6339 SDValue InFlag(0, 0);
6340 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6343 InFlag = Chain.getValue(1);
6344 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6347 InFlag = Chain.getValue(1);
6348 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6351 InFlag = Chain.getValue(1);
6353 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6354 SmallVector<SDValue, 8> Ops;
6355 Ops.push_back(Chain);
6356 Ops.push_back(DAG.getValueType(AVT));
6357 Ops.push_back(InFlag);
6358 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, &Ops[0], Ops.size());
6360 SmallVector<SDValue, 4> Results;
6361 Results.push_back(RepMovs);
6363 // Handle the last 1 - 7 bytes.
6364 unsigned Offset = SizeVal - BytesLeft;
6365 EVT DstVT = Dst.getValueType();
6366 EVT SrcVT = Src.getValueType();
6367 EVT SizeVT = Size.getValueType();
6368 Results.push_back(DAG.getMemcpy(Chain, dl,
6369 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6370 DAG.getConstant(Offset, DstVT)),
6371 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6372 DAG.getConstant(Offset, SrcVT)),
6373 DAG.getConstant(BytesLeft, SizeVT),
6374 Align, AlwaysInline,
6375 DstSV, DstSVOff + Offset,
6376 SrcSV, SrcSVOff + Offset));
6379 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6380 &Results[0], Results.size());
6383 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6384 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6385 DebugLoc dl = Op.getDebugLoc();
6387 if (!Subtarget->is64Bit()) {
6388 // vastart just stores the address of the VarArgsFrameIndex slot into the
6389 // memory location argument.
6390 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6391 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6395 // gp_offset (0 - 6 * 8)
6396 // fp_offset (48 - 48 + 8 * 16)
6397 // overflow_arg_area (point to parameters coming in memory).
6399 SmallVector<SDValue, 8> MemOps;
6400 SDValue FIN = Op.getOperand(1);
6402 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6403 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6405 MemOps.push_back(Store);
6408 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6409 FIN, DAG.getIntPtrConstant(4));
6410 Store = DAG.getStore(Op.getOperand(0), dl,
6411 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6413 MemOps.push_back(Store);
6415 // Store ptr to overflow_arg_area
6416 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6417 FIN, DAG.getIntPtrConstant(4));
6418 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6419 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6420 MemOps.push_back(Store);
6422 // Store ptr to reg_save_area.
6423 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6424 FIN, DAG.getIntPtrConstant(8));
6425 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6426 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6427 MemOps.push_back(Store);
6428 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6429 &MemOps[0], MemOps.size());
6432 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6433 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6434 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6435 SDValue Chain = Op.getOperand(0);
6436 SDValue SrcPtr = Op.getOperand(1);
6437 SDValue SrcSV = Op.getOperand(2);
6439 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6443 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6444 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6445 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6446 SDValue Chain = Op.getOperand(0);
6447 SDValue DstPtr = Op.getOperand(1);
6448 SDValue SrcPtr = Op.getOperand(2);
6449 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6450 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6451 DebugLoc dl = Op.getDebugLoc();
6453 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6454 DAG.getIntPtrConstant(24), 8, false,
6455 DstSV, 0, SrcSV, 0);
6459 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6460 DebugLoc dl = Op.getDebugLoc();
6461 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6463 default: return SDValue(); // Don't custom lower most intrinsics.
6464 // Comparison intrinsics.
6465 case Intrinsic::x86_sse_comieq_ss:
6466 case Intrinsic::x86_sse_comilt_ss:
6467 case Intrinsic::x86_sse_comile_ss:
6468 case Intrinsic::x86_sse_comigt_ss:
6469 case Intrinsic::x86_sse_comige_ss:
6470 case Intrinsic::x86_sse_comineq_ss:
6471 case Intrinsic::x86_sse_ucomieq_ss:
6472 case Intrinsic::x86_sse_ucomilt_ss:
6473 case Intrinsic::x86_sse_ucomile_ss:
6474 case Intrinsic::x86_sse_ucomigt_ss:
6475 case Intrinsic::x86_sse_ucomige_ss:
6476 case Intrinsic::x86_sse_ucomineq_ss:
6477 case Intrinsic::x86_sse2_comieq_sd:
6478 case Intrinsic::x86_sse2_comilt_sd:
6479 case Intrinsic::x86_sse2_comile_sd:
6480 case Intrinsic::x86_sse2_comigt_sd:
6481 case Intrinsic::x86_sse2_comige_sd:
6482 case Intrinsic::x86_sse2_comineq_sd:
6483 case Intrinsic::x86_sse2_ucomieq_sd:
6484 case Intrinsic::x86_sse2_ucomilt_sd:
6485 case Intrinsic::x86_sse2_ucomile_sd:
6486 case Intrinsic::x86_sse2_ucomigt_sd:
6487 case Intrinsic::x86_sse2_ucomige_sd:
6488 case Intrinsic::x86_sse2_ucomineq_sd: {
6490 ISD::CondCode CC = ISD::SETCC_INVALID;
6493 case Intrinsic::x86_sse_comieq_ss:
6494 case Intrinsic::x86_sse2_comieq_sd:
6498 case Intrinsic::x86_sse_comilt_ss:
6499 case Intrinsic::x86_sse2_comilt_sd:
6503 case Intrinsic::x86_sse_comile_ss:
6504 case Intrinsic::x86_sse2_comile_sd:
6508 case Intrinsic::x86_sse_comigt_ss:
6509 case Intrinsic::x86_sse2_comigt_sd:
6513 case Intrinsic::x86_sse_comige_ss:
6514 case Intrinsic::x86_sse2_comige_sd:
6518 case Intrinsic::x86_sse_comineq_ss:
6519 case Intrinsic::x86_sse2_comineq_sd:
6523 case Intrinsic::x86_sse_ucomieq_ss:
6524 case Intrinsic::x86_sse2_ucomieq_sd:
6525 Opc = X86ISD::UCOMI;
6528 case Intrinsic::x86_sse_ucomilt_ss:
6529 case Intrinsic::x86_sse2_ucomilt_sd:
6530 Opc = X86ISD::UCOMI;
6533 case Intrinsic::x86_sse_ucomile_ss:
6534 case Intrinsic::x86_sse2_ucomile_sd:
6535 Opc = X86ISD::UCOMI;
6538 case Intrinsic::x86_sse_ucomigt_ss:
6539 case Intrinsic::x86_sse2_ucomigt_sd:
6540 Opc = X86ISD::UCOMI;
6543 case Intrinsic::x86_sse_ucomige_ss:
6544 case Intrinsic::x86_sse2_ucomige_sd:
6545 Opc = X86ISD::UCOMI;
6548 case Intrinsic::x86_sse_ucomineq_ss:
6549 case Intrinsic::x86_sse2_ucomineq_sd:
6550 Opc = X86ISD::UCOMI;
6555 SDValue LHS = Op.getOperand(1);
6556 SDValue RHS = Op.getOperand(2);
6557 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6558 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6559 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6560 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6561 DAG.getConstant(X86CC, MVT::i8), Cond);
6562 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6564 // ptest intrinsics. The intrinsic these come from are designed to return
6565 // an integer value, not just an instruction so lower it to the ptest
6566 // pattern and a setcc for the result.
6567 case Intrinsic::x86_sse41_ptestz:
6568 case Intrinsic::x86_sse41_ptestc:
6569 case Intrinsic::x86_sse41_ptestnzc:{
6572 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6573 case Intrinsic::x86_sse41_ptestz:
6575 X86CC = X86::COND_E;
6577 case Intrinsic::x86_sse41_ptestc:
6579 X86CC = X86::COND_B;
6581 case Intrinsic::x86_sse41_ptestnzc:
6583 X86CC = X86::COND_A;
6587 SDValue LHS = Op.getOperand(1);
6588 SDValue RHS = Op.getOperand(2);
6589 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6590 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6591 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6592 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6595 // Fix vector shift instructions where the last operand is a non-immediate
6597 case Intrinsic::x86_sse2_pslli_w:
6598 case Intrinsic::x86_sse2_pslli_d:
6599 case Intrinsic::x86_sse2_pslli_q:
6600 case Intrinsic::x86_sse2_psrli_w:
6601 case Intrinsic::x86_sse2_psrli_d:
6602 case Intrinsic::x86_sse2_psrli_q:
6603 case Intrinsic::x86_sse2_psrai_w:
6604 case Intrinsic::x86_sse2_psrai_d:
6605 case Intrinsic::x86_mmx_pslli_w:
6606 case Intrinsic::x86_mmx_pslli_d:
6607 case Intrinsic::x86_mmx_pslli_q:
6608 case Intrinsic::x86_mmx_psrli_w:
6609 case Intrinsic::x86_mmx_psrli_d:
6610 case Intrinsic::x86_mmx_psrli_q:
6611 case Intrinsic::x86_mmx_psrai_w:
6612 case Intrinsic::x86_mmx_psrai_d: {
6613 SDValue ShAmt = Op.getOperand(2);
6614 if (isa<ConstantSDNode>(ShAmt))
6617 unsigned NewIntNo = 0;
6618 EVT ShAmtVT = MVT::v4i32;
6620 case Intrinsic::x86_sse2_pslli_w:
6621 NewIntNo = Intrinsic::x86_sse2_psll_w;
6623 case Intrinsic::x86_sse2_pslli_d:
6624 NewIntNo = Intrinsic::x86_sse2_psll_d;
6626 case Intrinsic::x86_sse2_pslli_q:
6627 NewIntNo = Intrinsic::x86_sse2_psll_q;
6629 case Intrinsic::x86_sse2_psrli_w:
6630 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6632 case Intrinsic::x86_sse2_psrli_d:
6633 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6635 case Intrinsic::x86_sse2_psrli_q:
6636 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6638 case Intrinsic::x86_sse2_psrai_w:
6639 NewIntNo = Intrinsic::x86_sse2_psra_w;
6641 case Intrinsic::x86_sse2_psrai_d:
6642 NewIntNo = Intrinsic::x86_sse2_psra_d;
6645 ShAmtVT = MVT::v2i32;
6647 case Intrinsic::x86_mmx_pslli_w:
6648 NewIntNo = Intrinsic::x86_mmx_psll_w;
6650 case Intrinsic::x86_mmx_pslli_d:
6651 NewIntNo = Intrinsic::x86_mmx_psll_d;
6653 case Intrinsic::x86_mmx_pslli_q:
6654 NewIntNo = Intrinsic::x86_mmx_psll_q;
6656 case Intrinsic::x86_mmx_psrli_w:
6657 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6659 case Intrinsic::x86_mmx_psrli_d:
6660 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6662 case Intrinsic::x86_mmx_psrli_q:
6663 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6665 case Intrinsic::x86_mmx_psrai_w:
6666 NewIntNo = Intrinsic::x86_mmx_psra_w;
6668 case Intrinsic::x86_mmx_psrai_d:
6669 NewIntNo = Intrinsic::x86_mmx_psra_d;
6671 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6677 // The vector shift intrinsics with scalars uses 32b shift amounts but
6678 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
6682 ShOps[1] = DAG.getConstant(0, MVT::i32);
6683 if (ShAmtVT == MVT::v4i32) {
6684 ShOps[2] = DAG.getUNDEF(MVT::i32);
6685 ShOps[3] = DAG.getUNDEF(MVT::i32);
6686 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
6688 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
6691 EVT VT = Op.getValueType();
6692 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
6693 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6694 DAG.getConstant(NewIntNo, MVT::i32),
6695 Op.getOperand(1), ShAmt);
6700 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6701 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6702 DebugLoc dl = Op.getDebugLoc();
6705 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6707 DAG.getConstant(TD->getPointerSize(),
6708 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6709 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6710 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6715 // Just load the return address.
6716 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6717 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6718 RetAddrFI, NULL, 0);
6721 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6722 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6723 MFI->setFrameAddressIsTaken(true);
6724 EVT VT = Op.getValueType();
6725 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6726 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6727 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6728 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6730 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6734 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6735 SelectionDAG &DAG) {
6736 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6739 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6741 MachineFunction &MF = DAG.getMachineFunction();
6742 SDValue Chain = Op.getOperand(0);
6743 SDValue Offset = Op.getOperand(1);
6744 SDValue Handler = Op.getOperand(2);
6745 DebugLoc dl = Op.getDebugLoc();
6747 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6749 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6751 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6752 DAG.getIntPtrConstant(-TD->getPointerSize()));
6753 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6754 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6755 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6756 MF.getRegInfo().addLiveOut(StoreAddrReg);
6758 return DAG.getNode(X86ISD::EH_RETURN, dl,
6760 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6763 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6764 SelectionDAG &DAG) {
6765 SDValue Root = Op.getOperand(0);
6766 SDValue Trmp = Op.getOperand(1); // trampoline
6767 SDValue FPtr = Op.getOperand(2); // nested function
6768 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6769 DebugLoc dl = Op.getDebugLoc();
6771 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6773 const X86InstrInfo *TII =
6774 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6776 if (Subtarget->is64Bit()) {
6777 SDValue OutChains[6];
6779 // Large code-model.
6781 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6782 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6784 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6785 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6787 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6789 // Load the pointer to the nested function into R11.
6790 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6791 SDValue Addr = Trmp;
6792 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6795 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6796 DAG.getConstant(2, MVT::i64));
6797 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6799 // Load the 'nest' parameter value into R10.
6800 // R10 is specified in X86CallingConv.td
6801 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6802 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6803 DAG.getConstant(10, MVT::i64));
6804 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6805 Addr, TrmpAddr, 10);
6807 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6808 DAG.getConstant(12, MVT::i64));
6809 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6811 // Jump to the nested function.
6812 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6813 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6814 DAG.getConstant(20, MVT::i64));
6815 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6816 Addr, TrmpAddr, 20);
6818 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6819 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6820 DAG.getConstant(22, MVT::i64));
6821 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6825 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6826 return DAG.getMergeValues(Ops, 2, dl);
6828 const Function *Func =
6829 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6830 CallingConv::ID CC = Func->getCallingConv();
6835 llvm_unreachable("Unsupported calling convention");
6836 case CallingConv::C:
6837 case CallingConv::X86_StdCall: {
6838 // Pass 'nest' parameter in ECX.
6839 // Must be kept in sync with X86CallingConv.td
6842 // Check that ECX wasn't needed by an 'inreg' parameter.
6843 const FunctionType *FTy = Func->getFunctionType();
6844 const AttrListPtr &Attrs = Func->getAttributes();
6846 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6847 unsigned InRegCount = 0;
6850 for (FunctionType::param_iterator I = FTy->param_begin(),
6851 E = FTy->param_end(); I != E; ++I, ++Idx)
6852 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6853 // FIXME: should only count parameters that are lowered to integers.
6854 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6856 if (InRegCount > 2) {
6857 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
6862 case CallingConv::X86_FastCall:
6863 case CallingConv::Fast:
6864 // Pass 'nest' parameter in EAX.
6865 // Must be kept in sync with X86CallingConv.td
6870 SDValue OutChains[4];
6873 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6874 DAG.getConstant(10, MVT::i32));
6875 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
6877 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6878 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6879 OutChains[0] = DAG.getStore(Root, dl,
6880 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6883 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6884 DAG.getConstant(1, MVT::i32));
6885 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
6887 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6888 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6889 DAG.getConstant(5, MVT::i32));
6890 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
6891 TrmpAddr, 5, false, 1);
6893 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6894 DAG.getConstant(6, MVT::i32));
6895 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
6898 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
6899 return DAG.getMergeValues(Ops, 2, dl);
6903 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6905 The rounding mode is in bits 11:10 of FPSR, and has the following
6912 FLT_ROUNDS, on the other hand, expects the following:
6919 To perform the conversion, we do:
6920 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6923 MachineFunction &MF = DAG.getMachineFunction();
6924 const TargetMachine &TM = MF.getTarget();
6925 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6926 unsigned StackAlignment = TFI.getStackAlignment();
6927 EVT VT = Op.getValueType();
6928 DebugLoc dl = Op.getDebugLoc();
6930 // Save FP Control Word to stack slot
6931 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
6932 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6934 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
6935 DAG.getEntryNode(), StackSlot);
6937 // Load FP Control Word from stack slot
6938 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
6940 // Transform as necessary
6942 DAG.getNode(ISD::SRL, dl, MVT::i16,
6943 DAG.getNode(ISD::AND, dl, MVT::i16,
6944 CWD, DAG.getConstant(0x800, MVT::i16)),
6945 DAG.getConstant(11, MVT::i8));
6947 DAG.getNode(ISD::SRL, dl, MVT::i16,
6948 DAG.getNode(ISD::AND, dl, MVT::i16,
6949 CWD, DAG.getConstant(0x400, MVT::i16)),
6950 DAG.getConstant(9, MVT::i8));
6953 DAG.getNode(ISD::AND, dl, MVT::i16,
6954 DAG.getNode(ISD::ADD, dl, MVT::i16,
6955 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
6956 DAG.getConstant(1, MVT::i16)),
6957 DAG.getConstant(3, MVT::i16));
6960 return DAG.getNode((VT.getSizeInBits() < 16 ?
6961 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
6964 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
6965 EVT VT = Op.getValueType();
6967 unsigned NumBits = VT.getSizeInBits();
6968 DebugLoc dl = Op.getDebugLoc();
6970 Op = Op.getOperand(0);
6971 if (VT == MVT::i8) {
6972 // Zero extend to i32 since there is not an i8 bsr.
6974 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6977 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
6978 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6979 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
6981 // If src is zero (i.e. bsr sets ZF), returns NumBits.
6982 SmallVector<SDValue, 4> Ops;
6984 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
6985 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6986 Ops.push_back(Op.getValue(1));
6987 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6989 // Finally xor with NumBits-1.
6990 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
6993 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6997 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
6998 EVT VT = Op.getValueType();
7000 unsigned NumBits = VT.getSizeInBits();
7001 DebugLoc dl = Op.getDebugLoc();
7003 Op = Op.getOperand(0);
7004 if (VT == MVT::i8) {
7006 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7009 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7010 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7011 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7013 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7014 SmallVector<SDValue, 4> Ops;
7016 Ops.push_back(DAG.getConstant(NumBits, OpVT));
7017 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
7018 Ops.push_back(Op.getValue(1));
7019 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
7022 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7026 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
7027 EVT VT = Op.getValueType();
7028 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7029 DebugLoc dl = Op.getDebugLoc();
7031 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7032 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7033 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7034 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7035 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7037 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7038 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7039 // return AloBlo + AloBhi + AhiBlo;
7041 SDValue A = Op.getOperand(0);
7042 SDValue B = Op.getOperand(1);
7044 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7045 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7046 A, DAG.getConstant(32, MVT::i32));
7047 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7048 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7049 B, DAG.getConstant(32, MVT::i32));
7050 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7051 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7053 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7054 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7056 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7057 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7059 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7060 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7061 AloBhi, DAG.getConstant(32, MVT::i32));
7062 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7063 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7064 AhiBlo, DAG.getConstant(32, MVT::i32));
7065 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7066 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7071 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
7072 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7073 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7074 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7075 // has only one use.
7076 SDNode *N = Op.getNode();
7077 SDValue LHS = N->getOperand(0);
7078 SDValue RHS = N->getOperand(1);
7079 unsigned BaseOp = 0;
7081 DebugLoc dl = Op.getDebugLoc();
7083 switch (Op.getOpcode()) {
7084 default: llvm_unreachable("Unknown ovf instruction!");
7086 // A subtract of one will be selected as a INC. Note that INC doesn't
7087 // set CF, so we can't do this for UADDO.
7088 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7089 if (C->getAPIntValue() == 1) {
7090 BaseOp = X86ISD::INC;
7094 BaseOp = X86ISD::ADD;
7098 BaseOp = X86ISD::ADD;
7102 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7103 // set CF, so we can't do this for USUBO.
7104 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7105 if (C->getAPIntValue() == 1) {
7106 BaseOp = X86ISD::DEC;
7110 BaseOp = X86ISD::SUB;
7114 BaseOp = X86ISD::SUB;
7118 BaseOp = X86ISD::SMUL;
7122 BaseOp = X86ISD::UMUL;
7127 // Also sets EFLAGS.
7128 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7129 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7132 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7133 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7135 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7139 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
7140 EVT T = Op.getValueType();
7141 DebugLoc dl = Op.getDebugLoc();
7144 switch(T.getSimpleVT().SimpleTy) {
7146 assert(false && "Invalid value type!");
7147 case MVT::i8: Reg = X86::AL; size = 1; break;
7148 case MVT::i16: Reg = X86::AX; size = 2; break;
7149 case MVT::i32: Reg = X86::EAX; size = 4; break;
7151 assert(Subtarget->is64Bit() && "Node not type legal!");
7152 Reg = X86::RAX; size = 8;
7155 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7156 Op.getOperand(2), SDValue());
7157 SDValue Ops[] = { cpIn.getValue(0),
7160 DAG.getTargetConstant(size, MVT::i8),
7162 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7163 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7165 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7169 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7170 SelectionDAG &DAG) {
7171 assert(Subtarget->is64Bit() && "Result not type legalized?");
7172 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7173 SDValue TheChain = Op.getOperand(0);
7174 DebugLoc dl = Op.getDebugLoc();
7175 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7176 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7177 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7179 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7180 DAG.getConstant(32, MVT::i8));
7182 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7185 return DAG.getMergeValues(Ops, 2, dl);
7188 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7189 SDNode *Node = Op.getNode();
7190 DebugLoc dl = Node->getDebugLoc();
7191 EVT T = Node->getValueType(0);
7192 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7193 DAG.getConstant(0, T), Node->getOperand(2));
7194 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7195 cast<AtomicSDNode>(Node)->getMemoryVT(),
7196 Node->getOperand(0),
7197 Node->getOperand(1), negOp,
7198 cast<AtomicSDNode>(Node)->getSrcValue(),
7199 cast<AtomicSDNode>(Node)->getAlignment());
7202 /// LowerOperation - Provide custom lowering hooks for some operations.
7204 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7205 switch (Op.getOpcode()) {
7206 default: llvm_unreachable("Should not custom lower this!");
7207 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7208 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7209 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7210 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7211 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7212 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7213 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7214 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7215 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7216 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7217 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7218 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7219 case ISD::SHL_PARTS:
7220 case ISD::SRA_PARTS:
7221 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7222 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7223 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7224 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7225 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7226 case ISD::FABS: return LowerFABS(Op, DAG);
7227 case ISD::FNEG: return LowerFNEG(Op, DAG);
7228 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7229 case ISD::SETCC: return LowerSETCC(Op, DAG);
7230 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7231 case ISD::SELECT: return LowerSELECT(Op, DAG);
7232 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7233 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7234 case ISD::VASTART: return LowerVASTART(Op, DAG);
7235 case ISD::VAARG: return LowerVAARG(Op, DAG);
7236 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7237 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7238 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7239 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7240 case ISD::FRAME_TO_ARGS_OFFSET:
7241 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7242 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7243 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7244 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7245 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7246 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7247 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7248 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7254 case ISD::UMULO: return LowerXALUO(Op, DAG);
7255 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7259 void X86TargetLowering::
7260 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7261 SelectionDAG &DAG, unsigned NewOp) {
7262 EVT T = Node->getValueType(0);
7263 DebugLoc dl = Node->getDebugLoc();
7264 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7266 SDValue Chain = Node->getOperand(0);
7267 SDValue In1 = Node->getOperand(1);
7268 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7269 Node->getOperand(2), DAG.getIntPtrConstant(0));
7270 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7271 Node->getOperand(2), DAG.getIntPtrConstant(1));
7272 SDValue Ops[] = { Chain, In1, In2L, In2H };
7273 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7275 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7276 cast<MemSDNode>(Node)->getMemOperand());
7277 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7278 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7279 Results.push_back(Result.getValue(2));
7282 /// ReplaceNodeResults - Replace a node with an illegal result type
7283 /// with a new node built out of custom code.
7284 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7285 SmallVectorImpl<SDValue>&Results,
7286 SelectionDAG &DAG) {
7287 DebugLoc dl = N->getDebugLoc();
7288 switch (N->getOpcode()) {
7290 assert(false && "Do not know how to custom type legalize this operation!");
7292 case ISD::FP_TO_SINT: {
7293 std::pair<SDValue,SDValue> Vals =
7294 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7295 SDValue FIST = Vals.first, StackSlot = Vals.second;
7296 if (FIST.getNode() != 0) {
7297 EVT VT = N->getValueType(0);
7298 // Return a load from the stack slot.
7299 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
7303 case ISD::READCYCLECOUNTER: {
7304 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7305 SDValue TheChain = N->getOperand(0);
7306 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7307 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7309 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7311 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7312 SDValue Ops[] = { eax, edx };
7313 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7314 Results.push_back(edx.getValue(1));
7321 EVT WidenVT = getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
7322 Results.push_back(DAG.UnrollVectorOp(N, WidenVT.getVectorNumElements()));
7325 case ISD::ATOMIC_CMP_SWAP: {
7326 EVT T = N->getValueType(0);
7327 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7328 SDValue cpInL, cpInH;
7329 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7330 DAG.getConstant(0, MVT::i32));
7331 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7332 DAG.getConstant(1, MVT::i32));
7333 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7334 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7336 SDValue swapInL, swapInH;
7337 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7338 DAG.getConstant(0, MVT::i32));
7339 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7340 DAG.getConstant(1, MVT::i32));
7341 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7343 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7344 swapInL.getValue(1));
7345 SDValue Ops[] = { swapInH.getValue(0),
7347 swapInH.getValue(1) };
7348 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7349 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7350 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7351 MVT::i32, Result.getValue(1));
7352 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7353 MVT::i32, cpOutL.getValue(2));
7354 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7355 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7356 Results.push_back(cpOutH.getValue(1));
7359 case ISD::ATOMIC_LOAD_ADD:
7360 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7362 case ISD::ATOMIC_LOAD_AND:
7363 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7365 case ISD::ATOMIC_LOAD_NAND:
7366 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7368 case ISD::ATOMIC_LOAD_OR:
7369 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7371 case ISD::ATOMIC_LOAD_SUB:
7372 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7374 case ISD::ATOMIC_LOAD_XOR:
7375 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7377 case ISD::ATOMIC_SWAP:
7378 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7383 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7385 default: return NULL;
7386 case X86ISD::BSF: return "X86ISD::BSF";
7387 case X86ISD::BSR: return "X86ISD::BSR";
7388 case X86ISD::SHLD: return "X86ISD::SHLD";
7389 case X86ISD::SHRD: return "X86ISD::SHRD";
7390 case X86ISD::FAND: return "X86ISD::FAND";
7391 case X86ISD::FOR: return "X86ISD::FOR";
7392 case X86ISD::FXOR: return "X86ISD::FXOR";
7393 case X86ISD::FSRL: return "X86ISD::FSRL";
7394 case X86ISD::FILD: return "X86ISD::FILD";
7395 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7396 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7397 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7398 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7399 case X86ISD::FLD: return "X86ISD::FLD";
7400 case X86ISD::FST: return "X86ISD::FST";
7401 case X86ISD::CALL: return "X86ISD::CALL";
7402 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7403 case X86ISD::BT: return "X86ISD::BT";
7404 case X86ISD::CMP: return "X86ISD::CMP";
7405 case X86ISD::COMI: return "X86ISD::COMI";
7406 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7407 case X86ISD::SETCC: return "X86ISD::SETCC";
7408 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7409 case X86ISD::CMOV: return "X86ISD::CMOV";
7410 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7411 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7412 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7413 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7414 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7415 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7416 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7417 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7418 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7419 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7420 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7421 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7422 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7423 case X86ISD::FMAX: return "X86ISD::FMAX";
7424 case X86ISD::FMIN: return "X86ISD::FMIN";
7425 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7426 case X86ISD::FRCP: return "X86ISD::FRCP";
7427 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7428 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7429 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7430 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7431 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7432 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7433 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7434 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7435 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7436 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7437 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7438 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7439 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7440 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7441 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7442 case X86ISD::VSHL: return "X86ISD::VSHL";
7443 case X86ISD::VSRL: return "X86ISD::VSRL";
7444 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7445 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7446 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7447 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7448 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7449 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7450 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7451 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7452 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7453 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7454 case X86ISD::ADD: return "X86ISD::ADD";
7455 case X86ISD::SUB: return "X86ISD::SUB";
7456 case X86ISD::SMUL: return "X86ISD::SMUL";
7457 case X86ISD::UMUL: return "X86ISD::UMUL";
7458 case X86ISD::INC: return "X86ISD::INC";
7459 case X86ISD::DEC: return "X86ISD::DEC";
7460 case X86ISD::OR: return "X86ISD::OR";
7461 case X86ISD::XOR: return "X86ISD::XOR";
7462 case X86ISD::AND: return "X86ISD::AND";
7463 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7464 case X86ISD::PTEST: return "X86ISD::PTEST";
7465 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7469 // isLegalAddressingMode - Return true if the addressing mode represented
7470 // by AM is legal for this target, for a load/store of the specified type.
7471 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7472 const Type *Ty) const {
7473 // X86 supports extremely general addressing modes.
7474 CodeModel::Model M = getTargetMachine().getCodeModel();
7476 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7477 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7482 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7484 // If a reference to this global requires an extra load, we can't fold it.
7485 if (isGlobalStubReference(GVFlags))
7488 // If BaseGV requires a register for the PIC base, we cannot also have a
7489 // BaseReg specified.
7490 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7493 // If lower 4G is not available, then we must use rip-relative addressing.
7494 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7504 // These scales always work.
7509 // These scales are formed with basereg+scalereg. Only accept if there is
7514 default: // Other stuff never works.
7522 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7523 if (!Ty1->isInteger() || !Ty2->isInteger())
7525 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7526 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7527 if (NumBits1 <= NumBits2)
7529 return Subtarget->is64Bit() || NumBits1 < 64;
7532 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7533 if (!VT1.isInteger() || !VT2.isInteger())
7535 unsigned NumBits1 = VT1.getSizeInBits();
7536 unsigned NumBits2 = VT2.getSizeInBits();
7537 if (NumBits1 <= NumBits2)
7539 return Subtarget->is64Bit() || NumBits1 < 64;
7542 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7543 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7544 return Ty1 == Type::getInt32Ty(Ty1->getContext()) &&
7545 Ty2 == Type::getInt64Ty(Ty1->getContext()) && Subtarget->is64Bit();
7548 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7549 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7550 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7553 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7554 // i16 instructions are longer (0x66 prefix) and potentially slower.
7555 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7558 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7559 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7560 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7561 /// are assumed to be legal.
7563 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7565 // Only do shuffles on 128-bit vector types for now.
7566 if (VT.getSizeInBits() == 64)
7569 // FIXME: pshufb, blends, shifts.
7570 return (VT.getVectorNumElements() == 2 ||
7571 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7572 isMOVLMask(M, VT) ||
7573 isSHUFPMask(M, VT) ||
7574 isPSHUFDMask(M, VT) ||
7575 isPSHUFHWMask(M, VT) ||
7576 isPSHUFLWMask(M, VT) ||
7577 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7578 isUNPCKLMask(M, VT) ||
7579 isUNPCKHMask(M, VT) ||
7580 isUNPCKL_v_undef_Mask(M, VT) ||
7581 isUNPCKH_v_undef_Mask(M, VT));
7585 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7587 unsigned NumElts = VT.getVectorNumElements();
7588 // FIXME: This collection of masks seems suspect.
7591 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7592 return (isMOVLMask(Mask, VT) ||
7593 isCommutedMOVLMask(Mask, VT, true) ||
7594 isSHUFPMask(Mask, VT) ||
7595 isCommutedSHUFPMask(Mask, VT));
7600 //===----------------------------------------------------------------------===//
7601 // X86 Scheduler Hooks
7602 //===----------------------------------------------------------------------===//
7604 // private utility function
7606 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7607 MachineBasicBlock *MBB,
7615 TargetRegisterClass *RC,
7616 bool invSrc) const {
7617 // For the atomic bitwise operator, we generate
7620 // ld t1 = [bitinstr.addr]
7621 // op t2 = t1, [bitinstr.val]
7623 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7625 // fallthrough -->nextMBB
7626 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7627 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7628 MachineFunction::iterator MBBIter = MBB;
7631 /// First build the CFG
7632 MachineFunction *F = MBB->getParent();
7633 MachineBasicBlock *thisMBB = MBB;
7634 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7635 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7636 F->insert(MBBIter, newMBB);
7637 F->insert(MBBIter, nextMBB);
7639 // Move all successors to thisMBB to nextMBB
7640 nextMBB->transferSuccessors(thisMBB);
7642 // Update thisMBB to fall through to newMBB
7643 thisMBB->addSuccessor(newMBB);
7645 // newMBB jumps to itself and fall through to nextMBB
7646 newMBB->addSuccessor(nextMBB);
7647 newMBB->addSuccessor(newMBB);
7649 // Insert instructions into newMBB based on incoming instruction
7650 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7651 "unexpected number of operands");
7652 DebugLoc dl = bInstr->getDebugLoc();
7653 MachineOperand& destOper = bInstr->getOperand(0);
7654 MachineOperand* argOpers[2 + X86AddrNumOperands];
7655 int numArgs = bInstr->getNumOperands() - 1;
7656 for (int i=0; i < numArgs; ++i)
7657 argOpers[i] = &bInstr->getOperand(i+1);
7659 // x86 address has 4 operands: base, index, scale, and displacement
7660 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7661 int valArgIndx = lastAddrIndx + 1;
7663 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7664 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7665 for (int i=0; i <= lastAddrIndx; ++i)
7666 (*MIB).addOperand(*argOpers[i]);
7668 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7670 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7675 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7676 assert((argOpers[valArgIndx]->isReg() ||
7677 argOpers[valArgIndx]->isImm()) &&
7679 if (argOpers[valArgIndx]->isReg())
7680 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7682 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7684 (*MIB).addOperand(*argOpers[valArgIndx]);
7686 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7689 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7690 for (int i=0; i <= lastAddrIndx; ++i)
7691 (*MIB).addOperand(*argOpers[i]);
7693 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7694 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7695 bInstr->memoperands_end());
7697 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7701 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7703 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7707 // private utility function: 64 bit atomics on 32 bit host.
7709 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7710 MachineBasicBlock *MBB,
7715 bool invSrc) const {
7716 // For the atomic bitwise operator, we generate
7717 // thisMBB (instructions are in pairs, except cmpxchg8b)
7718 // ld t1,t2 = [bitinstr.addr]
7720 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7721 // op t5, t6 <- out1, out2, [bitinstr.val]
7722 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7723 // mov ECX, EBX <- t5, t6
7724 // mov EAX, EDX <- t1, t2
7725 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7726 // mov t3, t4 <- EAX, EDX
7728 // result in out1, out2
7729 // fallthrough -->nextMBB
7731 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7732 const unsigned LoadOpc = X86::MOV32rm;
7733 const unsigned copyOpc = X86::MOV32rr;
7734 const unsigned NotOpc = X86::NOT32r;
7735 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7736 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7737 MachineFunction::iterator MBBIter = MBB;
7740 /// First build the CFG
7741 MachineFunction *F = MBB->getParent();
7742 MachineBasicBlock *thisMBB = MBB;
7743 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7744 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7745 F->insert(MBBIter, newMBB);
7746 F->insert(MBBIter, nextMBB);
7748 // Move all successors to thisMBB to nextMBB
7749 nextMBB->transferSuccessors(thisMBB);
7751 // Update thisMBB to fall through to newMBB
7752 thisMBB->addSuccessor(newMBB);
7754 // newMBB jumps to itself and fall through to nextMBB
7755 newMBB->addSuccessor(nextMBB);
7756 newMBB->addSuccessor(newMBB);
7758 DebugLoc dl = bInstr->getDebugLoc();
7759 // Insert instructions into newMBB based on incoming instruction
7760 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7761 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7762 "unexpected number of operands");
7763 MachineOperand& dest1Oper = bInstr->getOperand(0);
7764 MachineOperand& dest2Oper = bInstr->getOperand(1);
7765 MachineOperand* argOpers[2 + X86AddrNumOperands];
7766 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7767 argOpers[i] = &bInstr->getOperand(i+2);
7769 // x86 address has 4 operands: base, index, scale, and displacement
7770 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7772 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7773 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7774 for (int i=0; i <= lastAddrIndx; ++i)
7775 (*MIB).addOperand(*argOpers[i]);
7776 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7777 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7778 // add 4 to displacement.
7779 for (int i=0; i <= lastAddrIndx-2; ++i)
7780 (*MIB).addOperand(*argOpers[i]);
7781 MachineOperand newOp3 = *(argOpers[3]);
7783 newOp3.setImm(newOp3.getImm()+4);
7785 newOp3.setOffset(newOp3.getOffset()+4);
7786 (*MIB).addOperand(newOp3);
7787 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7789 // t3/4 are defined later, at the bottom of the loop
7790 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7791 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7792 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7793 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7794 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7795 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7797 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
7798 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
7800 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt1).addReg(t1);
7801 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt2).addReg(t2);
7807 int valArgIndx = lastAddrIndx + 1;
7808 assert((argOpers[valArgIndx]->isReg() ||
7809 argOpers[valArgIndx]->isImm()) &&
7811 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7812 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7813 if (argOpers[valArgIndx]->isReg())
7814 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7816 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7817 if (regOpcL != X86::MOV32rr)
7819 (*MIB).addOperand(*argOpers[valArgIndx]);
7820 assert(argOpers[valArgIndx + 1]->isReg() ==
7821 argOpers[valArgIndx]->isReg());
7822 assert(argOpers[valArgIndx + 1]->isImm() ==
7823 argOpers[valArgIndx]->isImm());
7824 if (argOpers[valArgIndx + 1]->isReg())
7825 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7827 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7828 if (regOpcH != X86::MOV32rr)
7830 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
7832 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
7834 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
7837 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
7839 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
7842 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
7843 for (int i=0; i <= lastAddrIndx; ++i)
7844 (*MIB).addOperand(*argOpers[i]);
7846 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7847 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7848 bInstr->memoperands_end());
7850 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
7851 MIB.addReg(X86::EAX);
7852 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
7853 MIB.addReg(X86::EDX);
7856 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7858 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7862 // private utility function
7864 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
7865 MachineBasicBlock *MBB,
7866 unsigned cmovOpc) const {
7867 // For the atomic min/max operator, we generate
7870 // ld t1 = [min/max.addr]
7871 // mov t2 = [min/max.val]
7873 // cmov[cond] t2 = t1
7875 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7877 // fallthrough -->nextMBB
7879 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7880 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7881 MachineFunction::iterator MBBIter = MBB;
7884 /// First build the CFG
7885 MachineFunction *F = MBB->getParent();
7886 MachineBasicBlock *thisMBB = MBB;
7887 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7888 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7889 F->insert(MBBIter, newMBB);
7890 F->insert(MBBIter, nextMBB);
7892 // Move all successors of thisMBB to nextMBB
7893 nextMBB->transferSuccessors(thisMBB);
7895 // Update thisMBB to fall through to newMBB
7896 thisMBB->addSuccessor(newMBB);
7898 // newMBB jumps to newMBB and fall through to nextMBB
7899 newMBB->addSuccessor(nextMBB);
7900 newMBB->addSuccessor(newMBB);
7902 DebugLoc dl = mInstr->getDebugLoc();
7903 // Insert instructions into newMBB based on incoming instruction
7904 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7905 "unexpected number of operands");
7906 MachineOperand& destOper = mInstr->getOperand(0);
7907 MachineOperand* argOpers[2 + X86AddrNumOperands];
7908 int numArgs = mInstr->getNumOperands() - 1;
7909 for (int i=0; i < numArgs; ++i)
7910 argOpers[i] = &mInstr->getOperand(i+1);
7912 // x86 address has 4 operands: base, index, scale, and displacement
7913 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7914 int valArgIndx = lastAddrIndx + 1;
7916 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7917 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
7918 for (int i=0; i <= lastAddrIndx; ++i)
7919 (*MIB).addOperand(*argOpers[i]);
7921 // We only support register and immediate values
7922 assert((argOpers[valArgIndx]->isReg() ||
7923 argOpers[valArgIndx]->isImm()) &&
7926 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7927 if (argOpers[valArgIndx]->isReg())
7928 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7930 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7931 (*MIB).addOperand(*argOpers[valArgIndx]);
7933 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
7936 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
7941 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7942 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
7946 // Cmp and exchange if none has modified the memory location
7947 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
7948 for (int i=0; i <= lastAddrIndx; ++i)
7949 (*MIB).addOperand(*argOpers[i]);
7951 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7952 (*MIB).setMemRefs(mInstr->memoperands_begin(),
7953 mInstr->memoperands_end());
7955 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
7956 MIB.addReg(X86::EAX);
7959 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7961 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
7965 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
7966 // all of this code can be replaced with that in the .td file.
7968 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
7969 unsigned numArgs, bool memArg) const {
7971 MachineFunction *F = BB->getParent();
7972 DebugLoc dl = MI->getDebugLoc();
7973 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7977 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
7979 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
7981 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
7983 for (unsigned i = 0; i < numArgs; ++i) {
7984 MachineOperand &Op = MI->getOperand(i+1);
7986 if (!(Op.isReg() && Op.isImplicit()))
7990 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
7993 F->DeleteMachineInstr(MI);
7999 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8001 MachineBasicBlock *MBB) const {
8002 // Emit code to save XMM registers to the stack. The ABI says that the
8003 // number of registers to save is given in %al, so it's theoretically
8004 // possible to do an indirect jump trick to avoid saving all of them,
8005 // however this code takes a simpler approach and just executes all
8006 // of the stores if %al is non-zero. It's less code, and it's probably
8007 // easier on the hardware branch predictor, and stores aren't all that
8008 // expensive anyway.
8010 // Create the new basic blocks. One block contains all the XMM stores,
8011 // and one block is the final destination regardless of whether any
8012 // stores were performed.
8013 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8014 MachineFunction *F = MBB->getParent();
8015 MachineFunction::iterator MBBIter = MBB;
8017 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8018 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8019 F->insert(MBBIter, XMMSaveMBB);
8020 F->insert(MBBIter, EndMBB);
8023 // Move any original successors of MBB to the end block.
8024 EndMBB->transferSuccessors(MBB);
8025 // The original block will now fall through to the XMM save block.
8026 MBB->addSuccessor(XMMSaveMBB);
8027 // The XMMSaveMBB will fall through to the end block.
8028 XMMSaveMBB->addSuccessor(EndMBB);
8030 // Now add the instructions.
8031 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8032 DebugLoc DL = MI->getDebugLoc();
8034 unsigned CountReg = MI->getOperand(0).getReg();
8035 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8036 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8038 if (!Subtarget->isTargetWin64()) {
8039 // If %al is 0, branch around the XMM save block.
8040 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8041 BuildMI(MBB, DL, TII->get(X86::JE)).addMBB(EndMBB);
8042 MBB->addSuccessor(EndMBB);
8045 // In the XMM save block, save all the XMM argument registers.
8046 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8047 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8048 MachineMemOperand *MMO =
8049 F->getMachineMemOperand(
8050 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8051 MachineMemOperand::MOStore, Offset,
8052 /*Size=*/16, /*Align=*/16);
8053 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8054 .addFrameIndex(RegSaveFrameIndex)
8055 .addImm(/*Scale=*/1)
8056 .addReg(/*IndexReg=*/0)
8057 .addImm(/*Disp=*/Offset)
8058 .addReg(/*Segment=*/0)
8059 .addReg(MI->getOperand(i).getReg())
8060 .addMemOperand(MMO);
8063 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8069 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8070 MachineBasicBlock *BB,
8071 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8072 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8073 DebugLoc DL = MI->getDebugLoc();
8075 // To "insert" a SELECT_CC instruction, we actually have to insert the
8076 // diamond control-flow pattern. The incoming instruction knows the
8077 // destination vreg to set, the condition code register to branch on, the
8078 // true/false values to select between, and a branch opcode to use.
8079 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8080 MachineFunction::iterator It = BB;
8086 // cmpTY ccX, r1, r2
8088 // fallthrough --> copy0MBB
8089 MachineBasicBlock *thisMBB = BB;
8090 MachineFunction *F = BB->getParent();
8091 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8092 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8094 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8095 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8096 F->insert(It, copy0MBB);
8097 F->insert(It, sinkMBB);
8098 // Update machine-CFG edges by first adding all successors of the current
8099 // block to the new block which will contain the Phi node for the select.
8100 // Also inform sdisel of the edge changes.
8101 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8102 E = BB->succ_end(); I != E; ++I) {
8103 EM->insert(std::make_pair(*I, sinkMBB));
8104 sinkMBB->addSuccessor(*I);
8106 // Next, remove all successors of the current block, and add the true
8107 // and fallthrough blocks as its successors.
8108 while (!BB->succ_empty())
8109 BB->removeSuccessor(BB->succ_begin());
8110 // Add the true and fallthrough blocks as its successors.
8111 BB->addSuccessor(copy0MBB);
8112 BB->addSuccessor(sinkMBB);
8115 // %FalseValue = ...
8116 // # fallthrough to sinkMBB
8119 // Update machine-CFG edges
8120 BB->addSuccessor(sinkMBB);
8123 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8126 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8127 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8128 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8130 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8136 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8137 MachineBasicBlock *BB,
8138 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8139 switch (MI->getOpcode()) {
8140 default: assert(false && "Unexpected instr type to insert");
8142 case X86::CMOV_V1I64:
8143 case X86::CMOV_FR32:
8144 case X86::CMOV_FR64:
8145 case X86::CMOV_V4F32:
8146 case X86::CMOV_V2F64:
8147 case X86::CMOV_V2I64:
8148 return EmitLoweredSelect(MI, BB, EM);
8150 case X86::FP32_TO_INT16_IN_MEM:
8151 case X86::FP32_TO_INT32_IN_MEM:
8152 case X86::FP32_TO_INT64_IN_MEM:
8153 case X86::FP64_TO_INT16_IN_MEM:
8154 case X86::FP64_TO_INT32_IN_MEM:
8155 case X86::FP64_TO_INT64_IN_MEM:
8156 case X86::FP80_TO_INT16_IN_MEM:
8157 case X86::FP80_TO_INT32_IN_MEM:
8158 case X86::FP80_TO_INT64_IN_MEM: {
8159 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8160 DebugLoc DL = MI->getDebugLoc();
8162 // Change the floating point control register to use "round towards zero"
8163 // mode when truncating to an integer value.
8164 MachineFunction *F = BB->getParent();
8165 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8166 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8168 // Load the old value of the high byte of the control word...
8170 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8171 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8174 // Set the high part to be round to zero...
8175 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8178 // Reload the modified control word now...
8179 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8181 // Restore the memory image of control word to original value
8182 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8185 // Get the X86 opcode to use.
8187 switch (MI->getOpcode()) {
8188 default: llvm_unreachable("illegal opcode!");
8189 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8190 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8191 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8192 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8193 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8194 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8195 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8196 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8197 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8201 MachineOperand &Op = MI->getOperand(0);
8203 AM.BaseType = X86AddressMode::RegBase;
8204 AM.Base.Reg = Op.getReg();
8206 AM.BaseType = X86AddressMode::FrameIndexBase;
8207 AM.Base.FrameIndex = Op.getIndex();
8209 Op = MI->getOperand(1);
8211 AM.Scale = Op.getImm();
8212 Op = MI->getOperand(2);
8214 AM.IndexReg = Op.getImm();
8215 Op = MI->getOperand(3);
8216 if (Op.isGlobal()) {
8217 AM.GV = Op.getGlobal();
8219 AM.Disp = Op.getImm();
8221 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8222 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8224 // Reload the original control word now.
8225 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8227 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8230 // String/text processing lowering.
8231 case X86::PCMPISTRM128REG:
8232 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8233 case X86::PCMPISTRM128MEM:
8234 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8235 case X86::PCMPESTRM128REG:
8236 return EmitPCMP(MI, BB, 5, false /* in mem */);
8237 case X86::PCMPESTRM128MEM:
8238 return EmitPCMP(MI, BB, 5, true /* in mem */);
8241 case X86::ATOMAND32:
8242 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8243 X86::AND32ri, X86::MOV32rm,
8244 X86::LCMPXCHG32, X86::MOV32rr,
8245 X86::NOT32r, X86::EAX,
8246 X86::GR32RegisterClass);
8248 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8249 X86::OR32ri, X86::MOV32rm,
8250 X86::LCMPXCHG32, X86::MOV32rr,
8251 X86::NOT32r, X86::EAX,
8252 X86::GR32RegisterClass);
8253 case X86::ATOMXOR32:
8254 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8255 X86::XOR32ri, X86::MOV32rm,
8256 X86::LCMPXCHG32, X86::MOV32rr,
8257 X86::NOT32r, X86::EAX,
8258 X86::GR32RegisterClass);
8259 case X86::ATOMNAND32:
8260 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8261 X86::AND32ri, X86::MOV32rm,
8262 X86::LCMPXCHG32, X86::MOV32rr,
8263 X86::NOT32r, X86::EAX,
8264 X86::GR32RegisterClass, true);
8265 case X86::ATOMMIN32:
8266 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8267 case X86::ATOMMAX32:
8268 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8269 case X86::ATOMUMIN32:
8270 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8271 case X86::ATOMUMAX32:
8272 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8274 case X86::ATOMAND16:
8275 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8276 X86::AND16ri, X86::MOV16rm,
8277 X86::LCMPXCHG16, X86::MOV16rr,
8278 X86::NOT16r, X86::AX,
8279 X86::GR16RegisterClass);
8281 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8282 X86::OR16ri, X86::MOV16rm,
8283 X86::LCMPXCHG16, X86::MOV16rr,
8284 X86::NOT16r, X86::AX,
8285 X86::GR16RegisterClass);
8286 case X86::ATOMXOR16:
8287 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8288 X86::XOR16ri, X86::MOV16rm,
8289 X86::LCMPXCHG16, X86::MOV16rr,
8290 X86::NOT16r, X86::AX,
8291 X86::GR16RegisterClass);
8292 case X86::ATOMNAND16:
8293 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8294 X86::AND16ri, X86::MOV16rm,
8295 X86::LCMPXCHG16, X86::MOV16rr,
8296 X86::NOT16r, X86::AX,
8297 X86::GR16RegisterClass, true);
8298 case X86::ATOMMIN16:
8299 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8300 case X86::ATOMMAX16:
8301 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8302 case X86::ATOMUMIN16:
8303 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8304 case X86::ATOMUMAX16:
8305 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8308 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8309 X86::AND8ri, X86::MOV8rm,
8310 X86::LCMPXCHG8, X86::MOV8rr,
8311 X86::NOT8r, X86::AL,
8312 X86::GR8RegisterClass);
8314 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8315 X86::OR8ri, X86::MOV8rm,
8316 X86::LCMPXCHG8, X86::MOV8rr,
8317 X86::NOT8r, X86::AL,
8318 X86::GR8RegisterClass);
8320 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8321 X86::XOR8ri, X86::MOV8rm,
8322 X86::LCMPXCHG8, X86::MOV8rr,
8323 X86::NOT8r, X86::AL,
8324 X86::GR8RegisterClass);
8325 case X86::ATOMNAND8:
8326 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8327 X86::AND8ri, X86::MOV8rm,
8328 X86::LCMPXCHG8, X86::MOV8rr,
8329 X86::NOT8r, X86::AL,
8330 X86::GR8RegisterClass, true);
8331 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8332 // This group is for 64-bit host.
8333 case X86::ATOMAND64:
8334 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8335 X86::AND64ri32, X86::MOV64rm,
8336 X86::LCMPXCHG64, X86::MOV64rr,
8337 X86::NOT64r, X86::RAX,
8338 X86::GR64RegisterClass);
8340 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8341 X86::OR64ri32, X86::MOV64rm,
8342 X86::LCMPXCHG64, X86::MOV64rr,
8343 X86::NOT64r, X86::RAX,
8344 X86::GR64RegisterClass);
8345 case X86::ATOMXOR64:
8346 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8347 X86::XOR64ri32, X86::MOV64rm,
8348 X86::LCMPXCHG64, X86::MOV64rr,
8349 X86::NOT64r, X86::RAX,
8350 X86::GR64RegisterClass);
8351 case X86::ATOMNAND64:
8352 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8353 X86::AND64ri32, X86::MOV64rm,
8354 X86::LCMPXCHG64, X86::MOV64rr,
8355 X86::NOT64r, X86::RAX,
8356 X86::GR64RegisterClass, true);
8357 case X86::ATOMMIN64:
8358 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8359 case X86::ATOMMAX64:
8360 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8361 case X86::ATOMUMIN64:
8362 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8363 case X86::ATOMUMAX64:
8364 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8366 // This group does 64-bit operations on a 32-bit host.
8367 case X86::ATOMAND6432:
8368 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8369 X86::AND32rr, X86::AND32rr,
8370 X86::AND32ri, X86::AND32ri,
8372 case X86::ATOMOR6432:
8373 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8374 X86::OR32rr, X86::OR32rr,
8375 X86::OR32ri, X86::OR32ri,
8377 case X86::ATOMXOR6432:
8378 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8379 X86::XOR32rr, X86::XOR32rr,
8380 X86::XOR32ri, X86::XOR32ri,
8382 case X86::ATOMNAND6432:
8383 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8384 X86::AND32rr, X86::AND32rr,
8385 X86::AND32ri, X86::AND32ri,
8387 case X86::ATOMADD6432:
8388 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8389 X86::ADD32rr, X86::ADC32rr,
8390 X86::ADD32ri, X86::ADC32ri,
8392 case X86::ATOMSUB6432:
8393 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8394 X86::SUB32rr, X86::SBB32rr,
8395 X86::SUB32ri, X86::SBB32ri,
8397 case X86::ATOMSWAP6432:
8398 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8399 X86::MOV32rr, X86::MOV32rr,
8400 X86::MOV32ri, X86::MOV32ri,
8402 case X86::VASTART_SAVE_XMM_REGS:
8403 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8407 //===----------------------------------------------------------------------===//
8408 // X86 Optimization Hooks
8409 //===----------------------------------------------------------------------===//
8411 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8415 const SelectionDAG &DAG,
8416 unsigned Depth) const {
8417 unsigned Opc = Op.getOpcode();
8418 assert((Opc >= ISD::BUILTIN_OP_END ||
8419 Opc == ISD::INTRINSIC_WO_CHAIN ||
8420 Opc == ISD::INTRINSIC_W_CHAIN ||
8421 Opc == ISD::INTRINSIC_VOID) &&
8422 "Should use MaskedValueIsZero if you don't know whether Op"
8423 " is a target node!");
8425 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8437 // These nodes' second result is a boolean.
8438 if (Op.getResNo() == 0)
8442 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8443 Mask.getBitWidth() - 1);
8448 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8449 /// node is a GlobalAddress + offset.
8450 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8451 GlobalValue* &GA, int64_t &Offset) const{
8452 if (N->getOpcode() == X86ISD::Wrapper) {
8453 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8454 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8455 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8459 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8462 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8463 EVT EltVT, LoadSDNode *&LDBase,
8464 unsigned &LastLoadedElt,
8465 SelectionDAG &DAG, MachineFrameInfo *MFI,
8466 const TargetLowering &TLI) {
8468 LastLoadedElt = -1U;
8469 for (unsigned i = 0; i < NumElems; ++i) {
8470 if (N->getMaskElt(i) < 0) {
8476 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8477 if (!Elt.getNode() ||
8478 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8481 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8483 LDBase = cast<LoadSDNode>(Elt.getNode());
8487 if (Elt.getOpcode() == ISD::UNDEF)
8490 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8491 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
8498 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8499 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8500 /// if the load addresses are consecutive, non-overlapping, and in the right
8501 /// order. In the case of v2i64, it will see if it can rewrite the
8502 /// shuffle to be an appropriate build vector so it can take advantage of
8503 // performBuildVectorCombine.
8504 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8505 const TargetLowering &TLI) {
8506 DebugLoc dl = N->getDebugLoc();
8507 EVT VT = N->getValueType(0);
8508 EVT EltVT = VT.getVectorElementType();
8509 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8510 unsigned NumElems = VT.getVectorNumElements();
8512 if (VT.getSizeInBits() != 128)
8515 // Try to combine a vector_shuffle into a 128-bit load.
8516 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8517 LoadSDNode *LD = NULL;
8518 unsigned LastLoadedElt;
8519 if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
8523 if (LastLoadedElt == NumElems - 1) {
8524 if (DAG.InferPtrAlignment(LD->getBasePtr()) >= 16)
8525 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8526 LD->getSrcValue(), LD->getSrcValueOffset(),
8528 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8529 LD->getSrcValue(), LD->getSrcValueOffset(),
8530 LD->isVolatile(), LD->getAlignment());
8531 } else if (NumElems == 4 && LastLoadedElt == 1) {
8532 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8533 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8534 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8535 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8540 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8541 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8542 const X86Subtarget *Subtarget) {
8543 DebugLoc DL = N->getDebugLoc();
8544 SDValue Cond = N->getOperand(0);
8545 // Get the LHS/RHS of the select.
8546 SDValue LHS = N->getOperand(1);
8547 SDValue RHS = N->getOperand(2);
8549 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8550 // instructions have the peculiarity that if either operand is a NaN,
8551 // they chose what we call the RHS operand (and as such are not symmetric).
8552 // It happens that this matches the semantics of the common C idiom
8553 // x<y?x:y and related forms, so we can recognize these cases.
8554 if (Subtarget->hasSSE2() &&
8555 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8556 Cond.getOpcode() == ISD::SETCC) {
8557 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8559 unsigned Opcode = 0;
8560 // Check for x CC y ? x : y.
8561 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8565 // This can be a min if we can prove that at least one of the operands
8567 if (!FiniteOnlyFPMath()) {
8568 if (DAG.isKnownNeverNaN(RHS)) {
8569 // Put the potential NaN in the RHS so that SSE will preserve it.
8570 std::swap(LHS, RHS);
8571 } else if (!DAG.isKnownNeverNaN(LHS))
8574 Opcode = X86ISD::FMIN;
8577 // This can be a min if we can prove that at least one of the operands
8579 if (!FiniteOnlyFPMath()) {
8580 if (DAG.isKnownNeverNaN(LHS)) {
8581 // Put the potential NaN in the RHS so that SSE will preserve it.
8582 std::swap(LHS, RHS);
8583 } else if (!DAG.isKnownNeverNaN(RHS))
8586 Opcode = X86ISD::FMIN;
8589 // This can be a min, but if either operand is a NaN we need it to
8590 // preserve the original LHS.
8591 std::swap(LHS, RHS);
8595 Opcode = X86ISD::FMIN;
8599 // This can be a max if we can prove that at least one of the operands
8601 if (!FiniteOnlyFPMath()) {
8602 if (DAG.isKnownNeverNaN(LHS)) {
8603 // Put the potential NaN in the RHS so that SSE will preserve it.
8604 std::swap(LHS, RHS);
8605 } else if (!DAG.isKnownNeverNaN(RHS))
8608 Opcode = X86ISD::FMAX;
8611 // This can be a max if we can prove that at least one of the operands
8613 if (!FiniteOnlyFPMath()) {
8614 if (DAG.isKnownNeverNaN(RHS)) {
8615 // Put the potential NaN in the RHS so that SSE will preserve it.
8616 std::swap(LHS, RHS);
8617 } else if (!DAG.isKnownNeverNaN(LHS))
8620 Opcode = X86ISD::FMAX;
8623 // This can be a max, but if either operand is a NaN we need it to
8624 // preserve the original LHS.
8625 std::swap(LHS, RHS);
8629 Opcode = X86ISD::FMAX;
8632 // Check for x CC y ? y : x -- a min/max with reversed arms.
8633 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8637 // This can be a min if we can prove that at least one of the operands
8639 if (!FiniteOnlyFPMath()) {
8640 if (DAG.isKnownNeverNaN(RHS)) {
8641 // Put the potential NaN in the RHS so that SSE will preserve it.
8642 std::swap(LHS, RHS);
8643 } else if (!DAG.isKnownNeverNaN(LHS))
8646 Opcode = X86ISD::FMIN;
8649 // This can be a min if we can prove that at least one of the operands
8651 if (!FiniteOnlyFPMath()) {
8652 if (DAG.isKnownNeverNaN(LHS)) {
8653 // Put the potential NaN in the RHS so that SSE will preserve it.
8654 std::swap(LHS, RHS);
8655 } else if (!DAG.isKnownNeverNaN(RHS))
8658 Opcode = X86ISD::FMIN;
8661 // This can be a min, but if either operand is a NaN we need it to
8662 // preserve the original LHS.
8663 std::swap(LHS, RHS);
8667 Opcode = X86ISD::FMIN;
8671 // This can be a max if we can prove that at least one of the operands
8673 if (!FiniteOnlyFPMath()) {
8674 if (DAG.isKnownNeverNaN(LHS)) {
8675 // Put the potential NaN in the RHS so that SSE will preserve it.
8676 std::swap(LHS, RHS);
8677 } else if (!DAG.isKnownNeverNaN(RHS))
8680 Opcode = X86ISD::FMAX;
8683 // This can be a max if we can prove that at least one of the operands
8685 if (!FiniteOnlyFPMath()) {
8686 if (DAG.isKnownNeverNaN(RHS)) {
8687 // Put the potential NaN in the RHS so that SSE will preserve it.
8688 std::swap(LHS, RHS);
8689 } else if (!DAG.isKnownNeverNaN(LHS))
8692 Opcode = X86ISD::FMAX;
8695 // This can be a max, but if either operand is a NaN we need it to
8696 // preserve the original LHS.
8697 std::swap(LHS, RHS);
8701 Opcode = X86ISD::FMAX;
8707 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8710 // If this is a select between two integer constants, try to do some
8712 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8713 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8714 // Don't do this for crazy integer types.
8715 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8716 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8717 // so that TrueC (the true value) is larger than FalseC.
8718 bool NeedsCondInvert = false;
8720 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8721 // Efficiently invertible.
8722 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8723 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8724 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8725 NeedsCondInvert = true;
8726 std::swap(TrueC, FalseC);
8729 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8730 if (FalseC->getAPIntValue() == 0 &&
8731 TrueC->getAPIntValue().isPowerOf2()) {
8732 if (NeedsCondInvert) // Invert the condition if needed.
8733 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8734 DAG.getConstant(1, Cond.getValueType()));
8736 // Zero extend the condition if needed.
8737 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8739 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8740 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8741 DAG.getConstant(ShAmt, MVT::i8));
8744 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8745 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8746 if (NeedsCondInvert) // Invert the condition if needed.
8747 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8748 DAG.getConstant(1, Cond.getValueType()));
8750 // Zero extend the condition if needed.
8751 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8752 FalseC->getValueType(0), Cond);
8753 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8754 SDValue(FalseC, 0));
8757 // Optimize cases that will turn into an LEA instruction. This requires
8758 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8759 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8760 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8761 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8763 bool isFastMultiplier = false;
8765 switch ((unsigned char)Diff) {
8767 case 1: // result = add base, cond
8768 case 2: // result = lea base( , cond*2)
8769 case 3: // result = lea base(cond, cond*2)
8770 case 4: // result = lea base( , cond*4)
8771 case 5: // result = lea base(cond, cond*4)
8772 case 8: // result = lea base( , cond*8)
8773 case 9: // result = lea base(cond, cond*8)
8774 isFastMultiplier = true;
8779 if (isFastMultiplier) {
8780 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8781 if (NeedsCondInvert) // Invert the condition if needed.
8782 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8783 DAG.getConstant(1, Cond.getValueType()));
8785 // Zero extend the condition if needed.
8786 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8788 // Scale the condition by the difference.
8790 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8791 DAG.getConstant(Diff, Cond.getValueType()));
8793 // Add the base if non-zero.
8794 if (FalseC->getAPIntValue() != 0)
8795 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8796 SDValue(FalseC, 0));
8806 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
8807 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8808 TargetLowering::DAGCombinerInfo &DCI) {
8809 DebugLoc DL = N->getDebugLoc();
8811 // If the flag operand isn't dead, don't touch this CMOV.
8812 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8815 // If this is a select between two integer constants, try to do some
8816 // optimizations. Note that the operands are ordered the opposite of SELECT
8818 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
8819 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8820 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
8821 // larger than FalseC (the false value).
8822 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
8824 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
8825 CC = X86::GetOppositeBranchCondition(CC);
8826 std::swap(TrueC, FalseC);
8829 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
8830 // This is efficient for any integer data type (including i8/i16) and
8832 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
8833 SDValue Cond = N->getOperand(3);
8834 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8835 DAG.getConstant(CC, MVT::i8), Cond);
8837 // Zero extend the condition if needed.
8838 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
8840 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8841 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
8842 DAG.getConstant(ShAmt, MVT::i8));
8843 if (N->getNumValues() == 2) // Dead flag value?
8844 return DCI.CombineTo(N, Cond, SDValue());
8848 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
8849 // for any integer data type, including i8/i16.
8850 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8851 SDValue Cond = N->getOperand(3);
8852 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8853 DAG.getConstant(CC, MVT::i8), Cond);
8855 // Zero extend the condition if needed.
8856 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8857 FalseC->getValueType(0), Cond);
8858 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8859 SDValue(FalseC, 0));
8861 if (N->getNumValues() == 2) // Dead flag value?
8862 return DCI.CombineTo(N, Cond, SDValue());
8866 // Optimize cases that will turn into an LEA instruction. This requires
8867 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8868 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8869 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8870 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8872 bool isFastMultiplier = false;
8874 switch ((unsigned char)Diff) {
8876 case 1: // result = add base, cond
8877 case 2: // result = lea base( , cond*2)
8878 case 3: // result = lea base(cond, cond*2)
8879 case 4: // result = lea base( , cond*4)
8880 case 5: // result = lea base(cond, cond*4)
8881 case 8: // result = lea base( , cond*8)
8882 case 9: // result = lea base(cond, cond*8)
8883 isFastMultiplier = true;
8888 if (isFastMultiplier) {
8889 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8890 SDValue Cond = N->getOperand(3);
8891 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8892 DAG.getConstant(CC, MVT::i8), Cond);
8893 // Zero extend the condition if needed.
8894 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8896 // Scale the condition by the difference.
8898 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8899 DAG.getConstant(Diff, Cond.getValueType()));
8901 // Add the base if non-zero.
8902 if (FalseC->getAPIntValue() != 0)
8903 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8904 SDValue(FalseC, 0));
8905 if (N->getNumValues() == 2) // Dead flag value?
8906 return DCI.CombineTo(N, Cond, SDValue());
8916 /// PerformMulCombine - Optimize a single multiply with constant into two
8917 /// in order to implement it with two cheaper instructions, e.g.
8918 /// LEA + SHL, LEA + LEA.
8919 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
8920 TargetLowering::DAGCombinerInfo &DCI) {
8921 if (DAG.getMachineFunction().
8922 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
8925 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8928 EVT VT = N->getValueType(0);
8932 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8935 uint64_t MulAmt = C->getZExtValue();
8936 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
8939 uint64_t MulAmt1 = 0;
8940 uint64_t MulAmt2 = 0;
8941 if ((MulAmt % 9) == 0) {
8943 MulAmt2 = MulAmt / 9;
8944 } else if ((MulAmt % 5) == 0) {
8946 MulAmt2 = MulAmt / 5;
8947 } else if ((MulAmt % 3) == 0) {
8949 MulAmt2 = MulAmt / 3;
8952 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
8953 DebugLoc DL = N->getDebugLoc();
8955 if (isPowerOf2_64(MulAmt2) &&
8956 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
8957 // If second multiplifer is pow2, issue it first. We want the multiply by
8958 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
8960 std::swap(MulAmt1, MulAmt2);
8963 if (isPowerOf2_64(MulAmt1))
8964 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
8965 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
8967 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
8968 DAG.getConstant(MulAmt1, VT));
8970 if (isPowerOf2_64(MulAmt2))
8971 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
8972 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
8974 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
8975 DAG.getConstant(MulAmt2, VT));
8977 // Do not add new nodes to DAG combiner worklist.
8978 DCI.CombineTo(N, NewMul, false);
8983 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
8984 SDValue N0 = N->getOperand(0);
8985 SDValue N1 = N->getOperand(1);
8986 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
8987 EVT VT = N0.getValueType();
8989 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
8990 // since the result of setcc_c is all zero's or all ones.
8991 if (N1C && N0.getOpcode() == ISD::AND &&
8992 N0.getOperand(1).getOpcode() == ISD::Constant) {
8993 SDValue N00 = N0.getOperand(0);
8994 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
8995 ((N00.getOpcode() == ISD::ANY_EXTEND ||
8996 N00.getOpcode() == ISD::ZERO_EXTEND) &&
8997 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
8998 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
8999 APInt ShAmt = N1C->getAPIntValue();
9000 Mask = Mask.shl(ShAmt);
9002 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9003 N00, DAG.getConstant(Mask, VT));
9010 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9012 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9013 const X86Subtarget *Subtarget) {
9014 EVT VT = N->getValueType(0);
9015 if (!VT.isVector() && VT.isInteger() &&
9016 N->getOpcode() == ISD::SHL)
9017 return PerformSHLCombine(N, DAG);
9019 // On X86 with SSE2 support, we can transform this to a vector shift if
9020 // all elements are shifted by the same amount. We can't do this in legalize
9021 // because the a constant vector is typically transformed to a constant pool
9022 // so we have no knowledge of the shift amount.
9023 if (!Subtarget->hasSSE2())
9026 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9029 SDValue ShAmtOp = N->getOperand(1);
9030 EVT EltVT = VT.getVectorElementType();
9031 DebugLoc DL = N->getDebugLoc();
9032 SDValue BaseShAmt = SDValue();
9033 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9034 unsigned NumElts = VT.getVectorNumElements();
9036 for (; i != NumElts; ++i) {
9037 SDValue Arg = ShAmtOp.getOperand(i);
9038 if (Arg.getOpcode() == ISD::UNDEF) continue;
9042 for (; i != NumElts; ++i) {
9043 SDValue Arg = ShAmtOp.getOperand(i);
9044 if (Arg.getOpcode() == ISD::UNDEF) continue;
9045 if (Arg != BaseShAmt) {
9049 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9050 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9051 SDValue InVec = ShAmtOp.getOperand(0);
9052 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9053 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9055 for (; i != NumElts; ++i) {
9056 SDValue Arg = InVec.getOperand(i);
9057 if (Arg.getOpcode() == ISD::UNDEF) continue;
9061 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9062 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9063 unsigned SplatIdx = cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9064 if (C->getZExtValue() == SplatIdx)
9065 BaseShAmt = InVec.getOperand(1);
9068 if (BaseShAmt.getNode() == 0)
9069 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9070 DAG.getIntPtrConstant(0));
9074 // The shift amount is an i32.
9075 if (EltVT.bitsGT(MVT::i32))
9076 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9077 else if (EltVT.bitsLT(MVT::i32))
9078 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9080 // The shift amount is identical so we can do a vector shift.
9081 SDValue ValOp = N->getOperand(0);
9082 switch (N->getOpcode()) {
9084 llvm_unreachable("Unknown shift opcode!");
9087 if (VT == MVT::v2i64)
9088 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9089 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9091 if (VT == MVT::v4i32)
9092 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9093 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9095 if (VT == MVT::v8i16)
9096 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9097 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9101 if (VT == MVT::v4i32)
9102 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9103 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9105 if (VT == MVT::v8i16)
9106 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9107 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9111 if (VT == MVT::v2i64)
9112 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9113 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9115 if (VT == MVT::v4i32)
9116 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9117 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9119 if (VT == MVT::v8i16)
9120 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9121 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9128 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9129 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9130 const X86Subtarget *Subtarget) {
9131 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9132 // the FP state in cases where an emms may be missing.
9133 // A preferable solution to the general problem is to figure out the right
9134 // places to insert EMMS. This qualifies as a quick hack.
9136 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9137 StoreSDNode *St = cast<StoreSDNode>(N);
9138 EVT VT = St->getValue().getValueType();
9139 if (VT.getSizeInBits() != 64)
9142 const Function *F = DAG.getMachineFunction().getFunction();
9143 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9144 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9145 && Subtarget->hasSSE2();
9146 if ((VT.isVector() ||
9147 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9148 isa<LoadSDNode>(St->getValue()) &&
9149 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9150 St->getChain().hasOneUse() && !St->isVolatile()) {
9151 SDNode* LdVal = St->getValue().getNode();
9153 int TokenFactorIndex = -1;
9154 SmallVector<SDValue, 8> Ops;
9155 SDNode* ChainVal = St->getChain().getNode();
9156 // Must be a store of a load. We currently handle two cases: the load
9157 // is a direct child, and it's under an intervening TokenFactor. It is
9158 // possible to dig deeper under nested TokenFactors.
9159 if (ChainVal == LdVal)
9160 Ld = cast<LoadSDNode>(St->getChain());
9161 else if (St->getValue().hasOneUse() &&
9162 ChainVal->getOpcode() == ISD::TokenFactor) {
9163 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9164 if (ChainVal->getOperand(i).getNode() == LdVal) {
9165 TokenFactorIndex = i;
9166 Ld = cast<LoadSDNode>(St->getValue());
9168 Ops.push_back(ChainVal->getOperand(i));
9172 if (!Ld || !ISD::isNormalLoad(Ld))
9175 // If this is not the MMX case, i.e. we are just turning i64 load/store
9176 // into f64 load/store, avoid the transformation if there are multiple
9177 // uses of the loaded value.
9178 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9181 DebugLoc LdDL = Ld->getDebugLoc();
9182 DebugLoc StDL = N->getDebugLoc();
9183 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9184 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9186 if (Subtarget->is64Bit() || F64IsLegal) {
9187 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9188 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9189 Ld->getBasePtr(), Ld->getSrcValue(),
9190 Ld->getSrcValueOffset(), Ld->isVolatile(),
9191 Ld->getAlignment());
9192 SDValue NewChain = NewLd.getValue(1);
9193 if (TokenFactorIndex != -1) {
9194 Ops.push_back(NewChain);
9195 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9198 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9199 St->getSrcValue(), St->getSrcValueOffset(),
9200 St->isVolatile(), St->getAlignment());
9203 // Otherwise, lower to two pairs of 32-bit loads / stores.
9204 SDValue LoAddr = Ld->getBasePtr();
9205 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9206 DAG.getConstant(4, MVT::i32));
9208 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9209 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9210 Ld->isVolatile(), Ld->getAlignment());
9211 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9212 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9214 MinAlign(Ld->getAlignment(), 4));
9216 SDValue NewChain = LoLd.getValue(1);
9217 if (TokenFactorIndex != -1) {
9218 Ops.push_back(LoLd);
9219 Ops.push_back(HiLd);
9220 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9224 LoAddr = St->getBasePtr();
9225 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9226 DAG.getConstant(4, MVT::i32));
9228 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9229 St->getSrcValue(), St->getSrcValueOffset(),
9230 St->isVolatile(), St->getAlignment());
9231 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9233 St->getSrcValueOffset() + 4,
9235 MinAlign(St->getAlignment(), 4));
9236 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9241 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9242 /// X86ISD::FXOR nodes.
9243 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9244 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9245 // F[X]OR(0.0, x) -> x
9246 // F[X]OR(x, 0.0) -> x
9247 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9248 if (C->getValueAPF().isPosZero())
9249 return N->getOperand(1);
9250 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9251 if (C->getValueAPF().isPosZero())
9252 return N->getOperand(0);
9256 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9257 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9258 // FAND(0.0, x) -> 0.0
9259 // FAND(x, 0.0) -> 0.0
9260 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9261 if (C->getValueAPF().isPosZero())
9262 return N->getOperand(0);
9263 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9264 if (C->getValueAPF().isPosZero())
9265 return N->getOperand(1);
9269 static SDValue PerformBTCombine(SDNode *N,
9271 TargetLowering::DAGCombinerInfo &DCI) {
9272 // BT ignores high bits in the bit index operand.
9273 SDValue Op1 = N->getOperand(1);
9274 if (Op1.hasOneUse()) {
9275 unsigned BitWidth = Op1.getValueSizeInBits();
9276 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9277 APInt KnownZero, KnownOne;
9278 TargetLowering::TargetLoweringOpt TLO(DAG);
9279 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9280 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9281 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9282 DCI.CommitTargetLoweringOpt(TLO);
9287 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9288 SDValue Op = N->getOperand(0);
9289 if (Op.getOpcode() == ISD::BIT_CONVERT)
9290 Op = Op.getOperand(0);
9291 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9292 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9293 VT.getVectorElementType().getSizeInBits() ==
9294 OpVT.getVectorElementType().getSizeInBits()) {
9295 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9300 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9301 // Locked instructions, in turn, have implicit fence semantics (all memory
9302 // operations are flushed before issuing the locked instruction, and the
9303 // are not buffered), so we can fold away the common pattern of
9304 // fence-atomic-fence.
9305 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9306 SDValue atomic = N->getOperand(0);
9307 switch (atomic.getOpcode()) {
9308 case ISD::ATOMIC_CMP_SWAP:
9309 case ISD::ATOMIC_SWAP:
9310 case ISD::ATOMIC_LOAD_ADD:
9311 case ISD::ATOMIC_LOAD_SUB:
9312 case ISD::ATOMIC_LOAD_AND:
9313 case ISD::ATOMIC_LOAD_OR:
9314 case ISD::ATOMIC_LOAD_XOR:
9315 case ISD::ATOMIC_LOAD_NAND:
9316 case ISD::ATOMIC_LOAD_MIN:
9317 case ISD::ATOMIC_LOAD_MAX:
9318 case ISD::ATOMIC_LOAD_UMIN:
9319 case ISD::ATOMIC_LOAD_UMAX:
9325 SDValue fence = atomic.getOperand(0);
9326 if (fence.getOpcode() != ISD::MEMBARRIER)
9329 switch (atomic.getOpcode()) {
9330 case ISD::ATOMIC_CMP_SWAP:
9331 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9332 atomic.getOperand(1), atomic.getOperand(2),
9333 atomic.getOperand(3));
9334 case ISD::ATOMIC_SWAP:
9335 case ISD::ATOMIC_LOAD_ADD:
9336 case ISD::ATOMIC_LOAD_SUB:
9337 case ISD::ATOMIC_LOAD_AND:
9338 case ISD::ATOMIC_LOAD_OR:
9339 case ISD::ATOMIC_LOAD_XOR:
9340 case ISD::ATOMIC_LOAD_NAND:
9341 case ISD::ATOMIC_LOAD_MIN:
9342 case ISD::ATOMIC_LOAD_MAX:
9343 case ISD::ATOMIC_LOAD_UMIN:
9344 case ISD::ATOMIC_LOAD_UMAX:
9345 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9346 atomic.getOperand(1), atomic.getOperand(2));
9352 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9353 DAGCombinerInfo &DCI) const {
9354 SelectionDAG &DAG = DCI.DAG;
9355 switch (N->getOpcode()) {
9357 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9358 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9359 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9360 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9363 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9364 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9366 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9367 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9368 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9369 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9370 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9376 //===----------------------------------------------------------------------===//
9377 // X86 Inline Assembly Support
9378 //===----------------------------------------------------------------------===//
9380 static bool LowerToBSwap(CallInst *CI) {
9381 // FIXME: this should verify that we are targetting a 486 or better. If not,
9382 // we will turn this bswap into something that will be lowered to logical ops
9383 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9384 // so don't worry about this.
9386 // Verify this is a simple bswap.
9387 if (CI->getNumOperands() != 2 ||
9388 CI->getType() != CI->getOperand(1)->getType() ||
9389 !CI->getType()->isInteger())
9392 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9393 if (!Ty || Ty->getBitWidth() % 16 != 0)
9396 // Okay, we can do this xform, do so now.
9397 const Type *Tys[] = { Ty };
9398 Module *M = CI->getParent()->getParent()->getParent();
9399 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9401 Value *Op = CI->getOperand(1);
9402 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9404 CI->replaceAllUsesWith(Op);
9405 CI->eraseFromParent();
9409 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9410 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9411 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9413 std::string AsmStr = IA->getAsmString();
9415 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9416 std::vector<std::string> AsmPieces;
9417 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9419 switch (AsmPieces.size()) {
9420 default: return false;
9422 AsmStr = AsmPieces[0];
9424 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9427 if (AsmPieces.size() == 2 &&
9428 (AsmPieces[0] == "bswap" ||
9429 AsmPieces[0] == "bswapq" ||
9430 AsmPieces[0] == "bswapl") &&
9431 (AsmPieces[1] == "$0" ||
9432 AsmPieces[1] == "${0:q}")) {
9433 // No need to check constraints, nothing other than the equivalent of
9434 // "=r,0" would be valid here.
9435 return LowerToBSwap(CI);
9437 // rorw $$8, ${0:w} --> llvm.bswap.i16
9438 if (CI->getType() == Type::getInt16Ty(CI->getContext()) &&
9439 AsmPieces.size() == 3 &&
9440 AsmPieces[0] == "rorw" &&
9441 AsmPieces[1] == "$$8," &&
9442 AsmPieces[2] == "${0:w}" &&
9443 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
9444 return LowerToBSwap(CI);
9448 if (CI->getType() == Type::getInt64Ty(CI->getContext()) &&
9449 Constraints.size() >= 2 &&
9450 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9451 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9452 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9453 std::vector<std::string> Words;
9454 SplitString(AsmPieces[0], Words, " \t");
9455 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9457 SplitString(AsmPieces[1], Words, " \t");
9458 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9460 SplitString(AsmPieces[2], Words, " \t,");
9461 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9462 Words[2] == "%edx") {
9463 return LowerToBSwap(CI);
9475 /// getConstraintType - Given a constraint letter, return the type of
9476 /// constraint it is for this target.
9477 X86TargetLowering::ConstraintType
9478 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
9479 if (Constraint.size() == 1) {
9480 switch (Constraint[0]) {
9492 return C_RegisterClass;
9500 return TargetLowering::getConstraintType(Constraint);
9503 /// LowerXConstraint - try to replace an X constraint, which matches anything,
9504 /// with another that has more specific requirements based on the type of the
9505 /// corresponding operand.
9506 const char *X86TargetLowering::
9507 LowerXConstraint(EVT ConstraintVT) const {
9508 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
9509 // 'f' like normal targets.
9510 if (ConstraintVT.isFloatingPoint()) {
9511 if (Subtarget->hasSSE2())
9513 if (Subtarget->hasSSE1())
9517 return TargetLowering::LowerXConstraint(ConstraintVT);
9520 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9521 /// vector. If it is invalid, don't add anything to Ops.
9522 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9525 std::vector<SDValue>&Ops,
9526 SelectionDAG &DAG) const {
9527 SDValue Result(0, 0);
9529 switch (Constraint) {
9532 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9533 if (C->getZExtValue() <= 31) {
9534 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9540 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9541 if (C->getZExtValue() <= 63) {
9542 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9548 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9549 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9550 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9556 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9557 if (C->getZExtValue() <= 255) {
9558 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9564 // 32-bit signed value
9565 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9566 const ConstantInt *CI = C->getConstantIntValue();
9567 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9568 C->getSExtValue())) {
9569 // Widen to 64 bits here to get it sign extended.
9570 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9573 // FIXME gcc accepts some relocatable values here too, but only in certain
9574 // memory models; it's complicated.
9579 // 32-bit unsigned value
9580 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9581 const ConstantInt *CI = C->getConstantIntValue();
9582 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9583 C->getZExtValue())) {
9584 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9588 // FIXME gcc accepts some relocatable values here too, but only in certain
9589 // memory models; it's complicated.
9593 // Literal immediates are always ok.
9594 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
9595 // Widen to 64 bits here to get it sign extended.
9596 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
9600 // If we are in non-pic codegen mode, we allow the address of a global (with
9601 // an optional displacement) to be used with 'i'.
9602 GlobalAddressSDNode *GA = 0;
9605 // Match either (GA), (GA+C), (GA+C1+C2), etc.
9607 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
9608 Offset += GA->getOffset();
9610 } else if (Op.getOpcode() == ISD::ADD) {
9611 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9612 Offset += C->getZExtValue();
9613 Op = Op.getOperand(0);
9616 } else if (Op.getOpcode() == ISD::SUB) {
9617 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9618 Offset += -C->getZExtValue();
9619 Op = Op.getOperand(0);
9624 // Otherwise, this isn't something we can handle, reject it.
9628 GlobalValue *GV = GA->getGlobal();
9629 // If we require an extra load to get this address, as in PIC mode, we
9631 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
9632 getTargetMachine())))
9636 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
9638 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
9644 if (Result.getNode()) {
9645 Ops.push_back(Result);
9648 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
9652 std::vector<unsigned> X86TargetLowering::
9653 getRegClassForInlineAsmConstraint(const std::string &Constraint,
9655 if (Constraint.size() == 1) {
9656 // FIXME: not handling fp-stack yet!
9657 switch (Constraint[0]) { // GCC X86 Constraint Letters
9658 default: break; // Unknown constraint letter
9659 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
9660 if (Subtarget->is64Bit()) {
9662 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
9663 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
9664 X86::R10D,X86::R11D,X86::R12D,
9665 X86::R13D,X86::R14D,X86::R15D,
9666 X86::EBP, X86::ESP, 0);
9667 else if (VT == MVT::i16)
9668 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
9669 X86::SI, X86::DI, X86::R8W,X86::R9W,
9670 X86::R10W,X86::R11W,X86::R12W,
9671 X86::R13W,X86::R14W,X86::R15W,
9672 X86::BP, X86::SP, 0);
9673 else if (VT == MVT::i8)
9674 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
9675 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
9676 X86::R10B,X86::R11B,X86::R12B,
9677 X86::R13B,X86::R14B,X86::R15B,
9678 X86::BPL, X86::SPL, 0);
9680 else if (VT == MVT::i64)
9681 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
9682 X86::RSI, X86::RDI, X86::R8, X86::R9,
9683 X86::R10, X86::R11, X86::R12,
9684 X86::R13, X86::R14, X86::R15,
9685 X86::RBP, X86::RSP, 0);
9689 // 32-bit fallthrough
9692 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
9693 else if (VT == MVT::i16)
9694 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
9695 else if (VT == MVT::i8)
9696 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
9697 else if (VT == MVT::i64)
9698 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
9703 return std::vector<unsigned>();
9706 std::pair<unsigned, const TargetRegisterClass*>
9707 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9709 // First, see if this is a constraint that directly corresponds to an LLVM
9711 if (Constraint.size() == 1) {
9712 // GCC Constraint Letters
9713 switch (Constraint[0]) {
9715 case 'r': // GENERAL_REGS
9716 case 'l': // INDEX_REGS
9718 return std::make_pair(0U, X86::GR8RegisterClass);
9720 return std::make_pair(0U, X86::GR16RegisterClass);
9721 if (VT == MVT::i32 || !Subtarget->is64Bit())
9722 return std::make_pair(0U, X86::GR32RegisterClass);
9723 return std::make_pair(0U, X86::GR64RegisterClass);
9724 case 'R': // LEGACY_REGS
9726 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
9728 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
9729 if (VT == MVT::i32 || !Subtarget->is64Bit())
9730 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
9731 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
9732 case 'f': // FP Stack registers.
9733 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
9734 // value to the correct fpstack register class.
9735 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
9736 return std::make_pair(0U, X86::RFP32RegisterClass);
9737 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
9738 return std::make_pair(0U, X86::RFP64RegisterClass);
9739 return std::make_pair(0U, X86::RFP80RegisterClass);
9740 case 'y': // MMX_REGS if MMX allowed.
9741 if (!Subtarget->hasMMX()) break;
9742 return std::make_pair(0U, X86::VR64RegisterClass);
9743 case 'Y': // SSE_REGS if SSE2 allowed
9744 if (!Subtarget->hasSSE2()) break;
9746 case 'x': // SSE_REGS if SSE1 allowed
9747 if (!Subtarget->hasSSE1()) break;
9749 switch (VT.getSimpleVT().SimpleTy) {
9751 // Scalar SSE types.
9754 return std::make_pair(0U, X86::FR32RegisterClass);
9757 return std::make_pair(0U, X86::FR64RegisterClass);
9765 return std::make_pair(0U, X86::VR128RegisterClass);
9771 // Use the default implementation in TargetLowering to convert the register
9772 // constraint into a member of a register class.
9773 std::pair<unsigned, const TargetRegisterClass*> Res;
9774 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9776 // Not found as a standard register?
9777 if (Res.second == 0) {
9778 // Map st(0) -> st(7) -> ST0
9779 if (Constraint.size() == 7 && Constraint[0] == '{' &&
9780 tolower(Constraint[1]) == 's' &&
9781 tolower(Constraint[2]) == 't' &&
9782 Constraint[3] == '(' &&
9783 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
9784 Constraint[5] == ')' &&
9785 Constraint[6] == '}') {
9787 Res.first = X86::ST0+Constraint[4]-'0';
9788 Res.second = X86::RFP80RegisterClass;
9792 // GCC allows "st(0)" to be called just plain "st".
9793 if (StringRef("{st}").equals_lower(Constraint)) {
9794 Res.first = X86::ST0;
9795 Res.second = X86::RFP80RegisterClass;
9800 if (StringRef("{flags}").equals_lower(Constraint)) {
9801 Res.first = X86::EFLAGS;
9802 Res.second = X86::CCRRegisterClass;
9806 // 'A' means EAX + EDX.
9807 if (Constraint == "A") {
9808 Res.first = X86::EAX;
9809 Res.second = X86::GR32_ADRegisterClass;
9815 // Otherwise, check to see if this is a register class of the wrong value
9816 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
9817 // turn into {ax},{dx}.
9818 if (Res.second->hasType(VT))
9819 return Res; // Correct type already, nothing to do.
9821 // All of the single-register GCC register classes map their values onto
9822 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
9823 // really want an 8-bit or 32-bit register, map to the appropriate register
9824 // class and return the appropriate register.
9825 if (Res.second == X86::GR16RegisterClass) {
9826 if (VT == MVT::i8) {
9827 unsigned DestReg = 0;
9828 switch (Res.first) {
9830 case X86::AX: DestReg = X86::AL; break;
9831 case X86::DX: DestReg = X86::DL; break;
9832 case X86::CX: DestReg = X86::CL; break;
9833 case X86::BX: DestReg = X86::BL; break;
9836 Res.first = DestReg;
9837 Res.second = X86::GR8RegisterClass;
9839 } else if (VT == MVT::i32) {
9840 unsigned DestReg = 0;
9841 switch (Res.first) {
9843 case X86::AX: DestReg = X86::EAX; break;
9844 case X86::DX: DestReg = X86::EDX; break;
9845 case X86::CX: DestReg = X86::ECX; break;
9846 case X86::BX: DestReg = X86::EBX; break;
9847 case X86::SI: DestReg = X86::ESI; break;
9848 case X86::DI: DestReg = X86::EDI; break;
9849 case X86::BP: DestReg = X86::EBP; break;
9850 case X86::SP: DestReg = X86::ESP; break;
9853 Res.first = DestReg;
9854 Res.second = X86::GR32RegisterClass;
9856 } else if (VT == MVT::i64) {
9857 unsigned DestReg = 0;
9858 switch (Res.first) {
9860 case X86::AX: DestReg = X86::RAX; break;
9861 case X86::DX: DestReg = X86::RDX; break;
9862 case X86::CX: DestReg = X86::RCX; break;
9863 case X86::BX: DestReg = X86::RBX; break;
9864 case X86::SI: DestReg = X86::RSI; break;
9865 case X86::DI: DestReg = X86::RDI; break;
9866 case X86::BP: DestReg = X86::RBP; break;
9867 case X86::SP: DestReg = X86::RSP; break;
9870 Res.first = DestReg;
9871 Res.second = X86::GR64RegisterClass;
9874 } else if (Res.second == X86::FR32RegisterClass ||
9875 Res.second == X86::FR64RegisterClass ||
9876 Res.second == X86::VR128RegisterClass) {
9877 // Handle references to XMM physical registers that got mapped into the
9878 // wrong class. This can happen with constraints like {xmm0} where the
9879 // target independent register mapper will just pick the first match it can
9880 // find, ignoring the required type.
9882 Res.second = X86::FR32RegisterClass;
9883 else if (VT == MVT::f64)
9884 Res.second = X86::FR64RegisterClass;
9885 else if (X86::VR128RegisterClass->hasType(VT))
9886 Res.second = X86::VR128RegisterClass;
9892 //===----------------------------------------------------------------------===//
9893 // X86 Widen vector type
9894 //===----------------------------------------------------------------------===//
9896 /// getWidenVectorType: given a vector type, returns the type to widen
9897 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
9898 /// If there is no vector type that we want to widen to, returns MVT::Other
9899 /// When and where to widen is target dependent based on the cost of
9900 /// scalarizing vs using the wider vector type.
9902 EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
9903 assert(VT.isVector());
9904 if (isTypeLegal(VT))
9907 // TODO: In computeRegisterProperty, we can compute the list of legal vector
9908 // type based on element type. This would speed up our search (though
9909 // it may not be worth it since the size of the list is relatively
9911 EVT EltVT = VT.getVectorElementType();
9912 unsigned NElts = VT.getVectorNumElements();
9914 // On X86, it make sense to widen any vector wider than 1
9918 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
9919 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
9920 EVT SVT = (MVT::SimpleValueType)nVT;
9922 if (isTypeLegal(SVT) &&
9923 SVT.getVectorElementType() == EltVT &&
9924 SVT.getVectorNumElements() > NElts)