1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "x86-isel"
17 #include "X86InstrBuilder.h"
18 #include "X86ISelLowering.h"
19 #include "X86TargetMachine.h"
20 #include "X86TargetObjectFile.h"
21 #include "llvm/CallingConv.h"
22 #include "llvm/Constants.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/GlobalAlias.h"
25 #include "llvm/GlobalVariable.h"
26 #include "llvm/Function.h"
27 #include "llvm/Instructions.h"
28 #include "llvm/Intrinsics.h"
29 #include "llvm/LLVMContext.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/PseudoSourceValue.h"
37 #include "llvm/MC/MCAsmInfo.h"
38 #include "llvm/MC/MCContext.h"
39 #include "llvm/MC/MCExpr.h"
40 #include "llvm/MC/MCSymbol.h"
41 #include "llvm/ADT/BitVector.h"
42 #include "llvm/ADT/SmallSet.h"
43 #include "llvm/ADT/Statistic.h"
44 #include "llvm/ADT/StringExtras.h"
45 #include "llvm/ADT/VectorExtras.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/ErrorHandling.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Support/raw_ostream.h"
53 STATISTIC(NumTailCalls, "Number of tail calls");
56 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
58 // Disable16Bit - 16-bit operations typically have a larger encoding than
59 // corresponding 32-bit instructions, and 16-bit code is slow on some
60 // processors. This is an experimental flag to disable 16-bit operations
61 // (which forces them to be Legalized to 32-bit operations).
63 Disable16Bit("disable-16bit", cl::Hidden,
64 cl::desc("Disable use of 16-bit instructions"));
66 // Forward declarations.
67 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
70 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
71 switch (TM.getSubtarget<X86Subtarget>().TargetType) {
72 default: llvm_unreachable("unknown subtarget type");
73 case X86Subtarget::isDarwin:
74 if (TM.getSubtarget<X86Subtarget>().is64Bit())
75 return new X8664_MachoTargetObjectFile();
76 return new X8632_MachoTargetObjectFile();
77 case X86Subtarget::isELF:
78 return new TargetLoweringObjectFileELF();
79 case X86Subtarget::isMingw:
80 case X86Subtarget::isCygwin:
81 case X86Subtarget::isWindows:
82 return new TargetLoweringObjectFileCOFF();
87 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
88 : TargetLowering(TM, createTLOF(TM)) {
89 Subtarget = &TM.getSubtarget<X86Subtarget>();
90 X86ScalarSSEf64 = Subtarget->hasSSE2();
91 X86ScalarSSEf32 = Subtarget->hasSSE1();
92 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
94 RegInfo = TM.getRegisterInfo();
97 // Set up the TargetLowering object.
99 // X86 is weird, it always uses i8 for shift amounts and setcc results.
100 setShiftAmountType(MVT::i8);
101 setBooleanContents(ZeroOrOneBooleanContent);
102 setSchedulingPreference(SchedulingForRegPressure);
103 setStackPointerRegisterToSaveRestore(X86StackPtr);
105 if (Subtarget->isTargetDarwin()) {
106 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
107 setUseUnderscoreSetJmp(false);
108 setUseUnderscoreLongJmp(false);
109 } else if (Subtarget->isTargetMingw()) {
110 // MS runtime is weird: it exports _setjmp, but longjmp!
111 setUseUnderscoreSetJmp(true);
112 setUseUnderscoreLongJmp(false);
114 setUseUnderscoreSetJmp(true);
115 setUseUnderscoreLongJmp(true);
118 // Set up the register classes.
119 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
121 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
122 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
123 if (Subtarget->is64Bit())
124 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
126 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
128 // We don't accept any truncstore of integer registers.
129 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
131 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
132 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
134 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
135 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
136 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
138 // SETOEQ and SETUNE require checking two conditions.
139 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
140 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
141 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
142 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
143 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
144 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
146 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
148 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
149 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
150 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
152 if (Subtarget->is64Bit()) {
153 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
154 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
155 } else if (!UseSoftFloat) {
156 if (X86ScalarSSEf64) {
157 // We have an impenetrably clever algorithm for ui64->double only.
158 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
160 // We have an algorithm for SSE2, and we turn this into a 64-bit
161 // FILD for other targets.
162 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
165 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
167 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
168 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
171 // SSE has no i16 to fp conversion, only i32
172 if (X86ScalarSSEf32) {
173 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
174 // f32 and f64 cases are Legal, f80 case is not
175 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
177 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
178 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
181 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
182 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
185 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
186 // are Legal, f80 is custom lowered.
187 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
188 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
190 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
192 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
193 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
195 if (X86ScalarSSEf32) {
196 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
197 // f32 and f64 cases are Legal, f80 case is not
198 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
200 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
201 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
204 // Handle FP_TO_UINT by promoting the destination to a larger signed
206 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
207 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
208 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
210 if (Subtarget->is64Bit()) {
211 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
212 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
213 } else if (!UseSoftFloat) {
214 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
215 // Expand FP_TO_UINT into a select.
216 // FIXME: We would like to use a Custom expander here eventually to do
217 // the optimal thing for SSE vs. the default expansion in the legalizer.
218 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
220 // With SSE3 we can use fisttpll to convert to a signed i64; without
221 // SSE, we're stuck with a fistpll.
222 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
225 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
226 if (!X86ScalarSSEf64) {
227 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
228 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
231 // Scalar integer divide and remainder are lowered to use operations that
232 // produce two results, to match the available instructions. This exposes
233 // the two-result form to trivial CSE, which is able to combine x/y and x%y
234 // into a single instruction.
236 // Scalar integer multiply-high is also lowered to use two-result
237 // operations, to match the available instructions. However, plain multiply
238 // (low) operations are left as Legal, as there are single-result
239 // instructions for this in x86. Using the two-result multiply instructions
240 // when both high and low results are needed must be arranged by dagcombine.
241 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
242 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
243 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
244 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
245 setOperationAction(ISD::SREM , MVT::i8 , Expand);
246 setOperationAction(ISD::UREM , MVT::i8 , Expand);
247 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
248 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
249 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
250 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
251 setOperationAction(ISD::SREM , MVT::i16 , Expand);
252 setOperationAction(ISD::UREM , MVT::i16 , Expand);
253 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
254 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
255 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
256 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
257 setOperationAction(ISD::SREM , MVT::i32 , Expand);
258 setOperationAction(ISD::UREM , MVT::i32 , Expand);
259 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
260 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
261 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
262 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
263 setOperationAction(ISD::SREM , MVT::i64 , Expand);
264 setOperationAction(ISD::UREM , MVT::i64 , Expand);
266 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
267 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
268 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
269 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
270 if (Subtarget->is64Bit())
271 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
272 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
273 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
274 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
275 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
276 setOperationAction(ISD::FREM , MVT::f32 , Expand);
277 setOperationAction(ISD::FREM , MVT::f64 , Expand);
278 setOperationAction(ISD::FREM , MVT::f80 , Expand);
279 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
281 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
282 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
283 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
284 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
286 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
287 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
289 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
290 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
292 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
293 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
294 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
295 if (Subtarget->is64Bit()) {
296 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
297 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
298 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
301 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
302 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
304 // These should be promoted to a larger select which is supported.
305 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
306 // X86 wants to expand cmov itself.
307 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
309 setOperationAction(ISD::SELECT , MVT::i16 , Expand);
311 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
312 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
313 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
314 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
315 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
316 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
318 setOperationAction(ISD::SETCC , MVT::i16 , Expand);
320 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
321 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
322 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
323 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
324 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
325 if (Subtarget->is64Bit()) {
326 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
327 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
329 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
332 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
333 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
334 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
335 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
336 if (Subtarget->is64Bit())
337 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
338 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
339 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
340 if (Subtarget->is64Bit()) {
341 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
342 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
343 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
344 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
345 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
347 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
348 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
349 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
350 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
351 if (Subtarget->is64Bit()) {
352 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
353 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
354 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
357 if (Subtarget->hasSSE1())
358 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
360 if (!Subtarget->hasSSE2())
361 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
363 // Expand certain atomics
364 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
365 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
366 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
367 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
369 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
370 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
371 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
372 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
374 if (!Subtarget->is64Bit()) {
375 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
376 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
377 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
378 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
379 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
380 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
381 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
384 // FIXME - use subtarget debug flags
385 if (!Subtarget->isTargetDarwin() &&
386 !Subtarget->isTargetELF() &&
387 !Subtarget->isTargetCygMing()) {
388 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
391 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
392 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
393 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
394 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
395 if (Subtarget->is64Bit()) {
396 setExceptionPointerRegister(X86::RAX);
397 setExceptionSelectorRegister(X86::RDX);
399 setExceptionPointerRegister(X86::EAX);
400 setExceptionSelectorRegister(X86::EDX);
402 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
403 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
405 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
407 setOperationAction(ISD::TRAP, MVT::Other, Legal);
409 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
410 setOperationAction(ISD::VASTART , MVT::Other, Custom);
411 setOperationAction(ISD::VAEND , MVT::Other, Expand);
412 if (Subtarget->is64Bit()) {
413 setOperationAction(ISD::VAARG , MVT::Other, Custom);
414 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
416 setOperationAction(ISD::VAARG , MVT::Other, Expand);
417 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
420 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
421 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
422 if (Subtarget->is64Bit())
423 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
424 if (Subtarget->isTargetCygMing())
425 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
427 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
429 if (!UseSoftFloat && X86ScalarSSEf64) {
430 // f32 and f64 use SSE.
431 // Set up the FP register classes.
432 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
433 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
435 // Use ANDPD to simulate FABS.
436 setOperationAction(ISD::FABS , MVT::f64, Custom);
437 setOperationAction(ISD::FABS , MVT::f32, Custom);
439 // Use XORP to simulate FNEG.
440 setOperationAction(ISD::FNEG , MVT::f64, Custom);
441 setOperationAction(ISD::FNEG , MVT::f32, Custom);
443 // Use ANDPD and ORPD to simulate FCOPYSIGN.
444 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
445 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
447 // We don't support sin/cos/fmod
448 setOperationAction(ISD::FSIN , MVT::f64, Expand);
449 setOperationAction(ISD::FCOS , MVT::f64, Expand);
450 setOperationAction(ISD::FSIN , MVT::f32, Expand);
451 setOperationAction(ISD::FCOS , MVT::f32, Expand);
453 // Expand FP immediates into loads from the stack, except for the special
455 addLegalFPImmediate(APFloat(+0.0)); // xorpd
456 addLegalFPImmediate(APFloat(+0.0f)); // xorps
457 } else if (!UseSoftFloat && X86ScalarSSEf32) {
458 // Use SSE for f32, x87 for f64.
459 // Set up the FP register classes.
460 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
461 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
463 // Use ANDPS to simulate FABS.
464 setOperationAction(ISD::FABS , MVT::f32, Custom);
466 // Use XORP to simulate FNEG.
467 setOperationAction(ISD::FNEG , MVT::f32, Custom);
469 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
471 // Use ANDPS and ORPS to simulate FCOPYSIGN.
472 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
473 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
475 // We don't support sin/cos/fmod
476 setOperationAction(ISD::FSIN , MVT::f32, Expand);
477 setOperationAction(ISD::FCOS , MVT::f32, Expand);
479 // Special cases we handle for FP constants.
480 addLegalFPImmediate(APFloat(+0.0f)); // xorps
481 addLegalFPImmediate(APFloat(+0.0)); // FLD0
482 addLegalFPImmediate(APFloat(+1.0)); // FLD1
483 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
484 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
487 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
488 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
490 } else if (!UseSoftFloat) {
491 // f32 and f64 in x87.
492 // Set up the FP register classes.
493 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
494 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
496 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
497 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
498 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
499 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
502 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
503 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
505 addLegalFPImmediate(APFloat(+0.0)); // FLD0
506 addLegalFPImmediate(APFloat(+1.0)); // FLD1
507 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
508 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
509 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
510 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
511 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
512 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
515 // Long double always uses X87.
517 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
518 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
519 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
522 APFloat TmpFlt(+0.0);
523 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
525 addLegalFPImmediate(TmpFlt); // FLD0
527 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
528 APFloat TmpFlt2(+1.0);
529 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
531 addLegalFPImmediate(TmpFlt2); // FLD1
532 TmpFlt2.changeSign();
533 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
537 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
538 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
542 // Always use a library call for pow.
543 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
544 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
545 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
547 setOperationAction(ISD::FLOG, MVT::f80, Expand);
548 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
549 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
550 setOperationAction(ISD::FEXP, MVT::f80, Expand);
551 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
553 // First set operation action for all vector types to either promote
554 // (for widening) or expand (for scalarization). Then we will selectively
555 // turn on ones that can be effectively codegen'd.
556 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
557 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
558 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
573 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
574 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
599 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
601 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
603 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
604 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
605 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
606 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
607 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
608 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
609 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
610 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
611 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
612 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
613 setTruncStoreAction((MVT::SimpleValueType)VT,
614 (MVT::SimpleValueType)InnerVT, Expand);
615 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
616 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
617 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
620 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
621 // with -msoft-float, disable use of MMX as well.
622 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
623 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
624 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
625 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
626 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
627 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
629 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
630 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
631 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
632 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
634 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
635 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
636 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
637 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
639 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
640 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
642 setOperationAction(ISD::AND, MVT::v8i8, Promote);
643 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
644 setOperationAction(ISD::AND, MVT::v4i16, Promote);
645 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
646 setOperationAction(ISD::AND, MVT::v2i32, Promote);
647 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
648 setOperationAction(ISD::AND, MVT::v1i64, Legal);
650 setOperationAction(ISD::OR, MVT::v8i8, Promote);
651 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
652 setOperationAction(ISD::OR, MVT::v4i16, Promote);
653 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
654 setOperationAction(ISD::OR, MVT::v2i32, Promote);
655 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
656 setOperationAction(ISD::OR, MVT::v1i64, Legal);
658 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
659 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
660 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
661 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
662 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
663 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
664 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
666 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
667 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
668 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
669 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
670 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
671 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
672 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
673 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
674 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
676 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
677 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
678 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
679 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
680 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
682 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
683 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
684 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
685 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
687 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
688 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
689 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
690 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
692 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
694 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
695 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
696 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
697 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
698 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
699 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
700 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
703 if (!UseSoftFloat && Subtarget->hasSSE1()) {
704 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
706 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
707 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
708 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
709 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
710 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
711 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
712 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
713 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
714 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
715 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
716 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
717 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
720 if (!UseSoftFloat && Subtarget->hasSSE2()) {
721 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
723 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
724 // registers cannot be used even for integer operations.
725 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
726 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
727 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
728 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
730 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
731 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
732 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
733 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
734 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
735 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
736 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
737 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
738 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
739 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
740 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
741 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
742 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
743 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
744 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
745 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
747 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
748 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
749 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
750 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
752 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
753 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
754 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
755 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
756 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
758 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
759 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
760 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
761 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
762 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
764 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
765 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
766 EVT VT = (MVT::SimpleValueType)i;
767 // Do not attempt to custom lower non-power-of-2 vectors
768 if (!isPowerOf2_32(VT.getVectorNumElements()))
770 // Do not attempt to custom lower non-128-bit vectors
771 if (!VT.is128BitVector())
773 setOperationAction(ISD::BUILD_VECTOR,
774 VT.getSimpleVT().SimpleTy, Custom);
775 setOperationAction(ISD::VECTOR_SHUFFLE,
776 VT.getSimpleVT().SimpleTy, Custom);
777 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
778 VT.getSimpleVT().SimpleTy, Custom);
781 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
782 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
783 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
784 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
785 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
786 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
788 if (Subtarget->is64Bit()) {
789 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
790 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
793 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
794 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
795 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
798 // Do not attempt to promote non-128-bit vectors
799 if (!VT.is128BitVector()) {
802 setOperationAction(ISD::AND, SVT, Promote);
803 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
804 setOperationAction(ISD::OR, SVT, Promote);
805 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
806 setOperationAction(ISD::XOR, SVT, Promote);
807 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
808 setOperationAction(ISD::LOAD, SVT, Promote);
809 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
810 setOperationAction(ISD::SELECT, SVT, Promote);
811 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
814 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
816 // Custom lower v2i64 and v2f64 selects.
817 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
818 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
819 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
820 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
822 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
823 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
824 if (!DisableMMX && Subtarget->hasMMX()) {
825 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
826 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
830 if (Subtarget->hasSSE41()) {
831 // FIXME: Do we need to handle scalar-to-vector here?
832 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
834 // i8 and i16 vectors are custom , because the source register and source
835 // source memory operand types are not the same width. f32 vectors are
836 // custom since the immediate controlling the insert encodes additional
838 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
839 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
840 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
841 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
843 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
844 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
845 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
846 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
848 if (Subtarget->is64Bit()) {
849 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
850 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
854 if (Subtarget->hasSSE42()) {
855 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
858 if (!UseSoftFloat && Subtarget->hasAVX()) {
859 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
860 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
861 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
862 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
864 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
865 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
866 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
867 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
868 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
869 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
870 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
871 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
872 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
873 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
874 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
875 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
876 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
877 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
878 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
880 // Operations to consider commented out -v16i16 v32i8
881 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
882 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
883 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
884 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
885 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
886 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
887 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
888 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
889 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
890 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
891 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
892 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
893 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
894 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
896 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
897 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
898 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
899 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
901 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
902 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
903 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
904 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
905 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
907 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
908 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
909 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
910 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
911 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
912 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
915 // Not sure we want to do this since there are no 256-bit integer
918 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
919 // This includes 256-bit vectors
920 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
921 EVT VT = (MVT::SimpleValueType)i;
923 // Do not attempt to custom lower non-power-of-2 vectors
924 if (!isPowerOf2_32(VT.getVectorNumElements()))
927 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
928 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
929 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
932 if (Subtarget->is64Bit()) {
933 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
934 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
939 // Not sure we want to do this since there are no 256-bit integer
942 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
943 // Including 256-bit vectors
944 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
945 EVT VT = (MVT::SimpleValueType)i;
947 if (!VT.is256BitVector()) {
950 setOperationAction(ISD::AND, VT, Promote);
951 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
952 setOperationAction(ISD::OR, VT, Promote);
953 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
954 setOperationAction(ISD::XOR, VT, Promote);
955 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
956 setOperationAction(ISD::LOAD, VT, Promote);
957 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
958 setOperationAction(ISD::SELECT, VT, Promote);
959 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
962 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
966 // We want to custom lower some of our intrinsics.
967 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
969 // Add/Sub/Mul with overflow operations are custom lowered.
970 setOperationAction(ISD::SADDO, MVT::i32, Custom);
971 setOperationAction(ISD::SADDO, MVT::i64, Custom);
972 setOperationAction(ISD::UADDO, MVT::i32, Custom);
973 setOperationAction(ISD::UADDO, MVT::i64, Custom);
974 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
975 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
976 setOperationAction(ISD::USUBO, MVT::i32, Custom);
977 setOperationAction(ISD::USUBO, MVT::i64, Custom);
978 setOperationAction(ISD::SMULO, MVT::i32, Custom);
979 setOperationAction(ISD::SMULO, MVT::i64, Custom);
981 if (!Subtarget->is64Bit()) {
982 // These libcalls are not available in 32-bit.
983 setLibcallName(RTLIB::SHL_I128, 0);
984 setLibcallName(RTLIB::SRL_I128, 0);
985 setLibcallName(RTLIB::SRA_I128, 0);
988 // We have target-specific dag combine patterns for the following nodes:
989 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
990 setTargetDAGCombine(ISD::BUILD_VECTOR);
991 setTargetDAGCombine(ISD::SELECT);
992 setTargetDAGCombine(ISD::SHL);
993 setTargetDAGCombine(ISD::SRA);
994 setTargetDAGCombine(ISD::SRL);
995 setTargetDAGCombine(ISD::OR);
996 setTargetDAGCombine(ISD::STORE);
997 setTargetDAGCombine(ISD::MEMBARRIER);
998 setTargetDAGCombine(ISD::ZERO_EXTEND);
999 if (Subtarget->is64Bit())
1000 setTargetDAGCombine(ISD::MUL);
1002 computeRegisterProperties();
1004 // Divide and reminder operations have no vector equivalent and can
1005 // trap. Do a custom widening for these operations in which we never
1006 // generate more divides/remainder than the original vector width.
1007 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1008 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
1009 if (!isTypeLegal((MVT::SimpleValueType)VT)) {
1010 setOperationAction(ISD::SDIV, (MVT::SimpleValueType) VT, Custom);
1011 setOperationAction(ISD::UDIV, (MVT::SimpleValueType) VT, Custom);
1012 setOperationAction(ISD::SREM, (MVT::SimpleValueType) VT, Custom);
1013 setOperationAction(ISD::UREM, (MVT::SimpleValueType) VT, Custom);
1017 // FIXME: These should be based on subtarget info. Plus, the values should
1018 // be smaller when we are in optimizing for size mode.
1019 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1020 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
1021 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1022 setPrefLoopAlignment(16);
1023 benefitFromCodePlacementOpt = true;
1027 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1032 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1033 /// the desired ByVal argument alignment.
1034 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1037 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1038 if (VTy->getBitWidth() == 128)
1040 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1041 unsigned EltAlign = 0;
1042 getMaxByValAlign(ATy->getElementType(), EltAlign);
1043 if (EltAlign > MaxAlign)
1044 MaxAlign = EltAlign;
1045 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1046 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1047 unsigned EltAlign = 0;
1048 getMaxByValAlign(STy->getElementType(i), EltAlign);
1049 if (EltAlign > MaxAlign)
1050 MaxAlign = EltAlign;
1058 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1059 /// function arguments in the caller parameter area. For X86, aggregates
1060 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1061 /// are at 4-byte boundaries.
1062 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1063 if (Subtarget->is64Bit()) {
1064 // Max of 8 and alignment of type.
1065 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1072 if (Subtarget->hasSSE1())
1073 getMaxByValAlign(Ty, Align);
1077 /// getOptimalMemOpType - Returns the target specific optimal type for load
1078 /// and store operations as a result of memset, memcpy, and memmove
1079 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
1082 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1083 bool isSrcConst, bool isSrcStr,
1084 SelectionDAG &DAG) const {
1085 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1086 // linux. This is because the stack realignment code can't handle certain
1087 // cases like PR2962. This should be removed when PR2962 is fixed.
1088 const Function *F = DAG.getMachineFunction().getFunction();
1089 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1090 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1091 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1093 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1096 if (Subtarget->is64Bit() && Size >= 8)
1101 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1102 /// current function. The returned value is a member of the
1103 /// MachineJumpTableInfo::JTEntryKind enum.
1104 unsigned X86TargetLowering::getJumpTableEncoding() const {
1105 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1107 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1108 Subtarget->isPICStyleGOT())
1109 return MachineJumpTableInfo::EK_Custom32;
1111 // Otherwise, use the normal jump table encoding heuristics.
1112 return TargetLowering::getJumpTableEncoding();
1115 /// getPICBaseSymbol - Return the X86-32 PIC base.
1117 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1118 MCContext &Ctx) const {
1119 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1120 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1121 Twine(MF->getFunctionNumber())+"$pb");
1126 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1127 const MachineBasicBlock *MBB,
1128 unsigned uid,MCContext &Ctx) const{
1129 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1130 Subtarget->isPICStyleGOT());
1131 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1134 // FIXME: @GOTOFF should be a property of MCSymbolRefExpr not in the MCSymbol.
1135 std::string Name = MBB->getSymbol(Ctx)->getName() + "@GOTOFF";
1136 return MCSymbolRefExpr::Create(Ctx.GetOrCreateSymbol(StringRef(Name)), Ctx);
1139 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1141 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1142 SelectionDAG &DAG) const {
1143 if (!Subtarget->is64Bit())
1144 // This doesn't have DebugLoc associated with it, but is not really the
1145 // same as a Register.
1146 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1151 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1152 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1154 const MCExpr *X86TargetLowering::
1155 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1156 MCContext &Ctx) const {
1157 // X86-64 uses RIP relative addressing based on the jump table label.
1158 if (Subtarget->isPICStyleRIPRel())
1159 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1161 // Otherwise, the reference is relative to the PIC base.
1162 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1165 /// getFunctionAlignment - Return the Log2 alignment of this function.
1166 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1167 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1170 //===----------------------------------------------------------------------===//
1171 // Return Value Calling Convention Implementation
1172 //===----------------------------------------------------------------------===//
1174 #include "X86GenCallingConv.inc"
1177 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1178 const SmallVectorImpl<EVT> &OutTys,
1179 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1180 SelectionDAG &DAG) {
1181 SmallVector<CCValAssign, 16> RVLocs;
1182 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1183 RVLocs, *DAG.getContext());
1184 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1188 X86TargetLowering::LowerReturn(SDValue Chain,
1189 CallingConv::ID CallConv, bool isVarArg,
1190 const SmallVectorImpl<ISD::OutputArg> &Outs,
1191 DebugLoc dl, SelectionDAG &DAG) {
1193 SmallVector<CCValAssign, 16> RVLocs;
1194 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1195 RVLocs, *DAG.getContext());
1196 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1198 // If this is the first return lowered for this function, add the regs to the
1199 // liveout set for the function.
1200 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1201 for (unsigned i = 0; i != RVLocs.size(); ++i)
1202 if (RVLocs[i].isRegLoc())
1203 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1208 SmallVector<SDValue, 6> RetOps;
1209 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1210 // Operand #1 = Bytes To Pop
1211 RetOps.push_back(DAG.getTargetConstant(getBytesToPopOnReturn(), MVT::i16));
1213 // Copy the result values into the output registers.
1214 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1215 CCValAssign &VA = RVLocs[i];
1216 assert(VA.isRegLoc() && "Can only return in registers!");
1217 SDValue ValToCopy = Outs[i].Val;
1219 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1220 // the RET instruction and handled by the FP Stackifier.
1221 if (VA.getLocReg() == X86::ST0 ||
1222 VA.getLocReg() == X86::ST1) {
1223 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1224 // change the value to the FP stack register class.
1225 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1226 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1227 RetOps.push_back(ValToCopy);
1228 // Don't emit a copytoreg.
1232 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1233 // which is returned in RAX / RDX.
1234 if (Subtarget->is64Bit()) {
1235 EVT ValVT = ValToCopy.getValueType();
1236 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1237 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1238 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1239 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1243 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1244 Flag = Chain.getValue(1);
1247 // The x86-64 ABI for returning structs by value requires that we copy
1248 // the sret argument into %rax for the return. We saved the argument into
1249 // a virtual register in the entry block, so now we copy the value out
1251 if (Subtarget->is64Bit() &&
1252 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1253 MachineFunction &MF = DAG.getMachineFunction();
1254 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1255 unsigned Reg = FuncInfo->getSRetReturnReg();
1257 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1258 FuncInfo->setSRetReturnReg(Reg);
1260 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1262 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1263 Flag = Chain.getValue(1);
1265 // RAX now acts like a return value.
1266 MF.getRegInfo().addLiveOut(X86::RAX);
1269 RetOps[0] = Chain; // Update chain.
1271 // Add the flag if we have it.
1273 RetOps.push_back(Flag);
1275 return DAG.getNode(X86ISD::RET_FLAG, dl,
1276 MVT::Other, &RetOps[0], RetOps.size());
1279 /// LowerCallResult - Lower the result values of a call into the
1280 /// appropriate copies out of appropriate physical registers.
1283 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1284 CallingConv::ID CallConv, bool isVarArg,
1285 const SmallVectorImpl<ISD::InputArg> &Ins,
1286 DebugLoc dl, SelectionDAG &DAG,
1287 SmallVectorImpl<SDValue> &InVals) {
1289 // Assign locations to each value returned by this call.
1290 SmallVector<CCValAssign, 16> RVLocs;
1291 bool Is64Bit = Subtarget->is64Bit();
1292 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1293 RVLocs, *DAG.getContext());
1294 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1296 // Copy all of the result registers out of their specified physreg.
1297 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1298 CCValAssign &VA = RVLocs[i];
1299 EVT CopyVT = VA.getValVT();
1301 // If this is x86-64, and we disabled SSE, we can't return FP values
1302 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1303 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1304 llvm_report_error("SSE register return with SSE disabled");
1307 // If this is a call to a function that returns an fp value on the floating
1308 // point stack, but where we prefer to use the value in xmm registers, copy
1309 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1310 if ((VA.getLocReg() == X86::ST0 ||
1311 VA.getLocReg() == X86::ST1) &&
1312 isScalarFPTypeInSSEReg(VA.getValVT())) {
1317 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1318 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1319 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1320 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1321 MVT::v2i64, InFlag).getValue(1);
1322 Val = Chain.getValue(0);
1323 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1324 Val, DAG.getConstant(0, MVT::i64));
1326 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1327 MVT::i64, InFlag).getValue(1);
1328 Val = Chain.getValue(0);
1330 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1332 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1333 CopyVT, InFlag).getValue(1);
1334 Val = Chain.getValue(0);
1336 InFlag = Chain.getValue(2);
1338 if (CopyVT != VA.getValVT()) {
1339 // Round the F80 the right size, which also moves to the appropriate xmm
1341 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1342 // This truncation won't change the value.
1343 DAG.getIntPtrConstant(1));
1346 InVals.push_back(Val);
1353 //===----------------------------------------------------------------------===//
1354 // C & StdCall & Fast Calling Convention implementation
1355 //===----------------------------------------------------------------------===//
1356 // StdCall calling convention seems to be standard for many Windows' API
1357 // routines and around. It differs from C calling convention just a little:
1358 // callee should clean up the stack, not caller. Symbols should be also
1359 // decorated in some fancy way :) It doesn't support any vector arguments.
1360 // For info on fast calling convention see Fast Calling Convention (tail call)
1361 // implementation LowerX86_32FastCCCallTo.
1363 /// CallIsStructReturn - Determines whether a call uses struct return
1365 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1369 return Outs[0].Flags.isSRet();
1372 /// ArgsAreStructReturn - Determines whether a function uses struct
1373 /// return semantics.
1375 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1379 return Ins[0].Flags.isSRet();
1382 /// IsCalleePop - Determines whether the callee is required to pop its
1383 /// own arguments. Callee pop is necessary to support tail calls.
1384 bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
1388 switch (CallingConv) {
1391 case CallingConv::X86_StdCall:
1392 return !Subtarget->is64Bit();
1393 case CallingConv::X86_FastCall:
1394 return !Subtarget->is64Bit();
1395 case CallingConv::Fast:
1396 return PerformTailCallOpt;
1400 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1401 /// given CallingConvention value.
1402 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1403 if (Subtarget->is64Bit()) {
1404 if (Subtarget->isTargetWin64())
1405 return CC_X86_Win64_C;
1410 if (CC == CallingConv::X86_FastCall)
1411 return CC_X86_32_FastCall;
1412 else if (CC == CallingConv::Fast)
1413 return CC_X86_32_FastCC;
1418 /// NameDecorationForCallConv - Selects the appropriate decoration to
1419 /// apply to a MachineFunction containing a given calling convention.
1421 X86TargetLowering::NameDecorationForCallConv(CallingConv::ID CallConv) {
1422 if (CallConv == CallingConv::X86_FastCall)
1424 else if (CallConv == CallingConv::X86_StdCall)
1430 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1431 /// by "Src" to address "Dst" with size and alignment information specified by
1432 /// the specific parameter attribute. The copy will be passed as a byval
1433 /// function parameter.
1435 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1436 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1438 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1439 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1440 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1443 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1444 /// a tailcall target by changing its ABI.
1445 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1446 return PerformTailCallOpt && CC == CallingConv::Fast;
1450 X86TargetLowering::LowerMemArgument(SDValue Chain,
1451 CallingConv::ID CallConv,
1452 const SmallVectorImpl<ISD::InputArg> &Ins,
1453 DebugLoc dl, SelectionDAG &DAG,
1454 const CCValAssign &VA,
1455 MachineFrameInfo *MFI,
1457 // Create the nodes corresponding to a load from this parameter slot.
1458 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1459 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1460 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1463 // If value is passed by pointer we have address passed instead of the value
1465 if (VA.getLocInfo() == CCValAssign::Indirect)
1466 ValVT = VA.getLocVT();
1468 ValVT = VA.getValVT();
1470 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1471 // changed with more analysis.
1472 // In case of tail call optimization mark all arguments mutable. Since they
1473 // could be overwritten by lowering of arguments in case of a tail call.
1474 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1475 VA.getLocMemOffset(), isImmutable, false);
1476 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1477 if (Flags.isByVal())
1479 return DAG.getLoad(ValVT, dl, Chain, FIN,
1480 PseudoSourceValue::getFixedStack(FI), 0);
1484 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1485 CallingConv::ID CallConv,
1487 const SmallVectorImpl<ISD::InputArg> &Ins,
1490 SmallVectorImpl<SDValue> &InVals) {
1492 MachineFunction &MF = DAG.getMachineFunction();
1493 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1495 const Function* Fn = MF.getFunction();
1496 if (Fn->hasExternalLinkage() &&
1497 Subtarget->isTargetCygMing() &&
1498 Fn->getName() == "main")
1499 FuncInfo->setForceFramePointer(true);
1501 // Decorate the function name.
1502 FuncInfo->setDecorationStyle(NameDecorationForCallConv(CallConv));
1504 MachineFrameInfo *MFI = MF.getFrameInfo();
1505 bool Is64Bit = Subtarget->is64Bit();
1506 bool IsWin64 = Subtarget->isTargetWin64();
1508 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1509 "Var args not supported with calling convention fastcc");
1511 // Assign locations to all of the incoming arguments.
1512 SmallVector<CCValAssign, 16> ArgLocs;
1513 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1514 ArgLocs, *DAG.getContext());
1515 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1517 unsigned LastVal = ~0U;
1519 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1520 CCValAssign &VA = ArgLocs[i];
1521 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1523 assert(VA.getValNo() != LastVal &&
1524 "Don't support value assigned to multiple locs yet");
1525 LastVal = VA.getValNo();
1527 if (VA.isRegLoc()) {
1528 EVT RegVT = VA.getLocVT();
1529 TargetRegisterClass *RC = NULL;
1530 if (RegVT == MVT::i32)
1531 RC = X86::GR32RegisterClass;
1532 else if (Is64Bit && RegVT == MVT::i64)
1533 RC = X86::GR64RegisterClass;
1534 else if (RegVT == MVT::f32)
1535 RC = X86::FR32RegisterClass;
1536 else if (RegVT == MVT::f64)
1537 RC = X86::FR64RegisterClass;
1538 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1539 RC = X86::VR128RegisterClass;
1540 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1541 RC = X86::VR64RegisterClass;
1543 llvm_unreachable("Unknown argument type!");
1545 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1546 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1548 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1549 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1551 if (VA.getLocInfo() == CCValAssign::SExt)
1552 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1553 DAG.getValueType(VA.getValVT()));
1554 else if (VA.getLocInfo() == CCValAssign::ZExt)
1555 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1556 DAG.getValueType(VA.getValVT()));
1557 else if (VA.getLocInfo() == CCValAssign::BCvt)
1558 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1560 if (VA.isExtInLoc()) {
1561 // Handle MMX values passed in XMM regs.
1562 if (RegVT.isVector()) {
1563 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1564 ArgValue, DAG.getConstant(0, MVT::i64));
1565 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1567 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1570 assert(VA.isMemLoc());
1571 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1574 // If value is passed via pointer - do a load.
1575 if (VA.getLocInfo() == CCValAssign::Indirect)
1576 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0);
1578 InVals.push_back(ArgValue);
1581 // The x86-64 ABI for returning structs by value requires that we copy
1582 // the sret argument into %rax for the return. Save the argument into
1583 // a virtual register so that we can access it from the return points.
1584 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1585 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1586 unsigned Reg = FuncInfo->getSRetReturnReg();
1588 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1589 FuncInfo->setSRetReturnReg(Reg);
1591 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1592 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1595 unsigned StackSize = CCInfo.getNextStackOffset();
1596 // Align stack specially for tail calls.
1597 if (FuncIsMadeTailCallSafe(CallConv))
1598 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1600 // If the function takes variable number of arguments, make a frame index for
1601 // the start of the first vararg value... for expansion of llvm.va_start.
1603 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1604 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize, true, false);
1607 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1609 // FIXME: We should really autogenerate these arrays
1610 static const unsigned GPR64ArgRegsWin64[] = {
1611 X86::RCX, X86::RDX, X86::R8, X86::R9
1613 static const unsigned XMMArgRegsWin64[] = {
1614 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1616 static const unsigned GPR64ArgRegs64Bit[] = {
1617 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1619 static const unsigned XMMArgRegs64Bit[] = {
1620 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1621 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1623 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1626 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1627 GPR64ArgRegs = GPR64ArgRegsWin64;
1628 XMMArgRegs = XMMArgRegsWin64;
1630 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1631 GPR64ArgRegs = GPR64ArgRegs64Bit;
1632 XMMArgRegs = XMMArgRegs64Bit;
1634 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1636 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1639 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1640 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1641 "SSE register cannot be used when SSE is disabled!");
1642 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1643 "SSE register cannot be used when SSE is disabled!");
1644 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1645 // Kernel mode asks for SSE to be disabled, so don't push them
1647 TotalNumXMMRegs = 0;
1649 // For X86-64, if there are vararg parameters that are passed via
1650 // registers, then we must store them to their spots on the stack so they
1651 // may be loaded by deferencing the result of va_next.
1652 VarArgsGPOffset = NumIntRegs * 8;
1653 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1654 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1655 TotalNumXMMRegs * 16, 16,
1658 // Store the integer parameter registers.
1659 SmallVector<SDValue, 8> MemOps;
1660 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1661 unsigned Offset = VarArgsGPOffset;
1662 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1663 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1664 DAG.getIntPtrConstant(Offset));
1665 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1666 X86::GR64RegisterClass);
1667 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1669 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1670 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1672 MemOps.push_back(Store);
1676 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1677 // Now store the XMM (fp + vector) parameter registers.
1678 SmallVector<SDValue, 11> SaveXMMOps;
1679 SaveXMMOps.push_back(Chain);
1681 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1682 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1683 SaveXMMOps.push_back(ALVal);
1685 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1686 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1688 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1689 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1690 X86::VR128RegisterClass);
1691 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1692 SaveXMMOps.push_back(Val);
1694 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1696 &SaveXMMOps[0], SaveXMMOps.size()));
1699 if (!MemOps.empty())
1700 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1701 &MemOps[0], MemOps.size());
1705 // Some CCs need callee pop.
1706 if (IsCalleePop(isVarArg, CallConv)) {
1707 BytesToPopOnReturn = StackSize; // Callee pops everything.
1709 BytesToPopOnReturn = 0; // Callee pops nothing.
1710 // If this is an sret function, the return should pop the hidden pointer.
1711 if (!Is64Bit && CallConv != CallingConv::Fast && ArgsAreStructReturn(Ins))
1712 BytesToPopOnReturn = 4;
1716 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1717 if (CallConv == CallingConv::X86_FastCall)
1718 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1721 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1727 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1728 SDValue StackPtr, SDValue Arg,
1729 DebugLoc dl, SelectionDAG &DAG,
1730 const CCValAssign &VA,
1731 ISD::ArgFlagsTy Flags) {
1732 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1733 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1734 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1735 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1736 if (Flags.isByVal()) {
1737 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1739 return DAG.getStore(Chain, dl, Arg, PtrOff,
1740 PseudoSourceValue::getStack(), LocMemOffset);
1743 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1744 /// optimization is performed and it is required.
1746 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1747 SDValue &OutRetAddr, SDValue Chain,
1748 bool IsTailCall, bool Is64Bit,
1749 int FPDiff, DebugLoc dl) {
1750 if (!IsTailCall || FPDiff==0) return Chain;
1752 // Adjust the Return address stack slot.
1753 EVT VT = getPointerTy();
1754 OutRetAddr = getReturnAddressFrameIndex(DAG);
1756 // Load the "old" Return address.
1757 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1758 return SDValue(OutRetAddr.getNode(), 1);
1761 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1762 /// optimization is performed and it is required (FPDiff!=0).
1764 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1765 SDValue Chain, SDValue RetAddrFrIdx,
1766 bool Is64Bit, int FPDiff, DebugLoc dl) {
1767 // Store the return address to the appropriate stack slot.
1768 if (!FPDiff) return Chain;
1769 // Calculate the new stack slot for the return address.
1770 int SlotSize = Is64Bit ? 8 : 4;
1771 int NewReturnAddrFI =
1772 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, true,false);
1773 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1774 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1775 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1776 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1781 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1782 CallingConv::ID CallConv, bool isVarArg,
1784 const SmallVectorImpl<ISD::OutputArg> &Outs,
1785 const SmallVectorImpl<ISD::InputArg> &Ins,
1786 DebugLoc dl, SelectionDAG &DAG,
1787 SmallVectorImpl<SDValue> &InVals) {
1788 MachineFunction &MF = DAG.getMachineFunction();
1789 bool Is64Bit = Subtarget->is64Bit();
1790 bool IsStructRet = CallIsStructReturn(Outs);
1793 // Check if it's really possible to do a tail call.
1794 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
1797 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1798 "Var args not supported with calling convention fastcc");
1800 // Analyze operands of the call, assigning locations to each operand.
1801 SmallVector<CCValAssign, 16> ArgLocs;
1802 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1803 ArgLocs, *DAG.getContext());
1804 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1806 // Get a count of how many bytes are to be pushed on the stack.
1807 unsigned NumBytes = CCInfo.getNextStackOffset();
1808 if (FuncIsMadeTailCallSafe(CallConv))
1809 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1815 // Lower arguments at fp - stackoffset + fpdiff.
1816 unsigned NumBytesCallerPushed =
1817 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1818 FPDiff = NumBytesCallerPushed - NumBytes;
1820 // Set the delta of movement of the returnaddr stackslot.
1821 // But only set if delta is greater than previous delta.
1822 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1823 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1826 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1828 SDValue RetAddrFrIdx;
1829 // Load return adress for tail calls.
1830 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall, Is64Bit,
1833 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1834 SmallVector<SDValue, 8> MemOpChains;
1837 // Walk the register/memloc assignments, inserting copies/loads. In the case
1838 // of tail call optimization arguments are handle later.
1839 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1840 CCValAssign &VA = ArgLocs[i];
1841 EVT RegVT = VA.getLocVT();
1842 SDValue Arg = Outs[i].Val;
1843 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1844 bool isByVal = Flags.isByVal();
1846 // Promote the value if needed.
1847 switch (VA.getLocInfo()) {
1848 default: llvm_unreachable("Unknown loc info!");
1849 case CCValAssign::Full: break;
1850 case CCValAssign::SExt:
1851 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1853 case CCValAssign::ZExt:
1854 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1856 case CCValAssign::AExt:
1857 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1858 // Special case: passing MMX values in XMM registers.
1859 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1860 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1861 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1863 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1865 case CCValAssign::BCvt:
1866 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1868 case CCValAssign::Indirect: {
1869 // Store the argument.
1870 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1871 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1872 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1873 PseudoSourceValue::getFixedStack(FI), 0);
1879 if (VA.isRegLoc()) {
1880 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1882 if (!isTailCall || (isTailCall && isByVal)) {
1883 assert(VA.isMemLoc());
1884 if (StackPtr.getNode() == 0)
1885 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1887 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1888 dl, DAG, VA, Flags));
1893 if (!MemOpChains.empty())
1894 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1895 &MemOpChains[0], MemOpChains.size());
1897 // Build a sequence of copy-to-reg nodes chained together with token chain
1898 // and flag operands which copy the outgoing args into registers.
1900 // Tail call byval lowering might overwrite argument registers so in case of
1901 // tail call optimization the copies to registers are lowered later.
1903 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1904 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1905 RegsToPass[i].second, InFlag);
1906 InFlag = Chain.getValue(1);
1910 if (Subtarget->isPICStyleGOT()) {
1911 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1914 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1915 DAG.getNode(X86ISD::GlobalBaseReg,
1916 DebugLoc::getUnknownLoc(),
1919 InFlag = Chain.getValue(1);
1921 // If we are tail calling and generating PIC/GOT style code load the
1922 // address of the callee into ECX. The value in ecx is used as target of
1923 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1924 // for tail calls on PIC/GOT architectures. Normally we would just put the
1925 // address of GOT into ebx and then call target@PLT. But for tail calls
1926 // ebx would be restored (since ebx is callee saved) before jumping to the
1929 // Note: The actual moving to ECX is done further down.
1930 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1931 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1932 !G->getGlobal()->hasProtectedVisibility())
1933 Callee = LowerGlobalAddress(Callee, DAG);
1934 else if (isa<ExternalSymbolSDNode>(Callee))
1935 Callee = LowerExternalSymbol(Callee, DAG);
1939 if (Is64Bit && isVarArg) {
1940 // From AMD64 ABI document:
1941 // For calls that may call functions that use varargs or stdargs
1942 // (prototype-less calls or calls to functions containing ellipsis (...) in
1943 // the declaration) %al is used as hidden argument to specify the number
1944 // of SSE registers used. The contents of %al do not need to match exactly
1945 // the number of registers, but must be an ubound on the number of SSE
1946 // registers used and is in the range 0 - 8 inclusive.
1948 // FIXME: Verify this on Win64
1949 // Count the number of XMM registers allocated.
1950 static const unsigned XMMArgRegs[] = {
1951 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1952 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1954 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1955 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1956 && "SSE registers cannot be used when SSE is disabled");
1958 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1959 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1960 InFlag = Chain.getValue(1);
1964 // For tail calls lower the arguments to the 'real' stack slot.
1966 // Force all the incoming stack arguments to be loaded from the stack
1967 // before any new outgoing arguments are stored to the stack, because the
1968 // outgoing stack slots may alias the incoming argument stack slots, and
1969 // the alias isn't otherwise explicit. This is slightly more conservative
1970 // than necessary, because it means that each store effectively depends
1971 // on every argument instead of just those arguments it would clobber.
1972 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1974 SmallVector<SDValue, 8> MemOpChains2;
1977 // Do not flag preceeding copytoreg stuff together with the following stuff.
1979 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1980 CCValAssign &VA = ArgLocs[i];
1981 if (!VA.isRegLoc()) {
1982 assert(VA.isMemLoc());
1983 SDValue Arg = Outs[i].Val;
1984 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1985 // Create frame index.
1986 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1987 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1988 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
1989 FIN = DAG.getFrameIndex(FI, getPointerTy());
1991 if (Flags.isByVal()) {
1992 // Copy relative to framepointer.
1993 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1994 if (StackPtr.getNode() == 0)
1995 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1997 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1999 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2003 // Store relative to framepointer.
2004 MemOpChains2.push_back(
2005 DAG.getStore(ArgChain, dl, Arg, FIN,
2006 PseudoSourceValue::getFixedStack(FI), 0));
2011 if (!MemOpChains2.empty())
2012 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2013 &MemOpChains2[0], MemOpChains2.size());
2015 // Copy arguments to their registers.
2016 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2017 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2018 RegsToPass[i].second, InFlag);
2019 InFlag = Chain.getValue(1);
2023 // Store the return address to the appropriate stack slot.
2024 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2028 bool WasGlobalOrExternal = false;
2029 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2030 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2031 // In the 64-bit large code model, we have to make all calls
2032 // through a register, since the call instruction's 32-bit
2033 // pc-relative offset may not be large enough to hold the whole
2035 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2036 WasGlobalOrExternal = true;
2037 // If the callee is a GlobalAddress node (quite common, every direct call
2038 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2041 // We should use extra load for direct calls to dllimported functions in
2043 GlobalValue *GV = G->getGlobal();
2044 if (!GV->hasDLLImportLinkage()) {
2045 unsigned char OpFlags = 0;
2047 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2048 // external symbols most go through the PLT in PIC mode. If the symbol
2049 // has hidden or protected visibility, or if it is static or local, then
2050 // we don't need to use the PLT - we can directly call it.
2051 if (Subtarget->isTargetELF() &&
2052 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2053 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2054 OpFlags = X86II::MO_PLT;
2055 } else if (Subtarget->isPICStyleStubAny() &&
2056 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2057 Subtarget->getDarwinVers() < 9) {
2058 // PC-relative references to external symbols should go through $stub,
2059 // unless we're building with the leopard linker or later, which
2060 // automatically synthesizes these stubs.
2061 OpFlags = X86II::MO_DARWIN_STUB;
2064 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2065 G->getOffset(), OpFlags);
2067 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2068 WasGlobalOrExternal = true;
2069 unsigned char OpFlags = 0;
2071 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2072 // symbols should go through the PLT.
2073 if (Subtarget->isTargetELF() &&
2074 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2075 OpFlags = X86II::MO_PLT;
2076 } else if (Subtarget->isPICStyleStubAny() &&
2077 Subtarget->getDarwinVers() < 9) {
2078 // PC-relative references to external symbols should go through $stub,
2079 // unless we're building with the leopard linker or later, which
2080 // automatically synthesizes these stubs.
2081 OpFlags = X86II::MO_DARWIN_STUB;
2084 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2088 if (isTailCall && !WasGlobalOrExternal) {
2089 unsigned Opc = Is64Bit ? X86::R11 : X86::EAX;
2091 Chain = DAG.getCopyToReg(Chain, dl,
2092 DAG.getRegister(Opc, getPointerTy()),
2094 Callee = DAG.getRegister(Opc, getPointerTy());
2095 // Add register as live out.
2096 MF.getRegInfo().addLiveOut(Opc);
2099 // Returns a chain & a flag for retval copy to use.
2100 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2101 SmallVector<SDValue, 8> Ops;
2104 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2105 DAG.getIntPtrConstant(0, true), InFlag);
2106 InFlag = Chain.getValue(1);
2109 Ops.push_back(Chain);
2110 Ops.push_back(Callee);
2113 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2115 // Add argument registers to the end of the list so that they are known live
2117 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2118 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2119 RegsToPass[i].second.getValueType()));
2121 // Add an implicit use GOT pointer in EBX.
2122 if (!isTailCall && Subtarget->isPICStyleGOT())
2123 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2125 // Add an implicit use of AL for x86 vararg functions.
2126 if (Is64Bit && isVarArg)
2127 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2129 if (InFlag.getNode())
2130 Ops.push_back(InFlag);
2133 // If this is the first return lowered for this function, add the regs
2134 // to the liveout set for the function.
2135 if (MF.getRegInfo().liveout_empty()) {
2136 SmallVector<CCValAssign, 16> RVLocs;
2137 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2139 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2140 for (unsigned i = 0; i != RVLocs.size(); ++i)
2141 if (RVLocs[i].isRegLoc())
2142 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2145 assert(((Callee.getOpcode() == ISD::Register &&
2146 (cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
2147 cast<RegisterSDNode>(Callee)->getReg() == X86::R11)) ||
2148 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2149 Callee.getOpcode() == ISD::TargetGlobalAddress) &&
2150 "Expecting a global address, external symbol, or scratch register");
2152 return DAG.getNode(X86ISD::TC_RETURN, dl,
2153 NodeTys, &Ops[0], Ops.size());
2156 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2157 InFlag = Chain.getValue(1);
2159 // Create the CALLSEQ_END node.
2160 unsigned NumBytesForCalleeToPush;
2161 if (IsCalleePop(isVarArg, CallConv))
2162 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2163 else if (!Is64Bit && CallConv != CallingConv::Fast && IsStructRet)
2164 // If this is is a call to a struct-return function, the callee
2165 // pops the hidden struct pointer, so we have to push it back.
2166 // This is common for Darwin/X86, Linux & Mingw32 targets.
2167 NumBytesForCalleeToPush = 4;
2169 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2171 // Returns a flag for retval copy to use.
2172 Chain = DAG.getCALLSEQ_END(Chain,
2173 DAG.getIntPtrConstant(NumBytes, true),
2174 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2177 InFlag = Chain.getValue(1);
2179 // Handle result values, copying them out of physregs into vregs that we
2181 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2182 Ins, dl, DAG, InVals);
2186 //===----------------------------------------------------------------------===//
2187 // Fast Calling Convention (tail call) implementation
2188 //===----------------------------------------------------------------------===//
2190 // Like std call, callee cleans arguments, convention except that ECX is
2191 // reserved for storing the tail called function address. Only 2 registers are
2192 // free for argument passing (inreg). Tail call optimization is performed
2194 // * tailcallopt is enabled
2195 // * caller/callee are fastcc
2196 // On X86_64 architecture with GOT-style position independent code only local
2197 // (within module) calls are supported at the moment.
2198 // To keep the stack aligned according to platform abi the function
2199 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2200 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2201 // If a tail called function callee has more arguments than the caller the
2202 // caller needs to make sure that there is room to move the RETADDR to. This is
2203 // achieved by reserving an area the size of the argument delta right after the
2204 // original REtADDR, but before the saved framepointer or the spilled registers
2205 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2217 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2218 /// for a 16 byte align requirement.
2219 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2220 SelectionDAG& DAG) {
2221 MachineFunction &MF = DAG.getMachineFunction();
2222 const TargetMachine &TM = MF.getTarget();
2223 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2224 unsigned StackAlignment = TFI.getStackAlignment();
2225 uint64_t AlignMask = StackAlignment - 1;
2226 int64_t Offset = StackSize;
2227 uint64_t SlotSize = TD->getPointerSize();
2228 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2229 // Number smaller than 12 so just add the difference.
2230 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2232 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2233 Offset = ((~AlignMask) & Offset) + StackAlignment +
2234 (StackAlignment-SlotSize);
2239 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2240 /// for tail call optimization. Targets which want to do tail call
2241 /// optimization should implement this function.
2243 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2244 CallingConv::ID CalleeCC,
2246 const SmallVectorImpl<ISD::OutputArg> &Outs,
2247 const SmallVectorImpl<ISD::InputArg> &Ins,
2248 SelectionDAG& DAG) const {
2249 // If -tailcallopt is specified, make fastcc functions tail-callable.
2250 const Function *F = DAG.getMachineFunction().getFunction();
2251 if (PerformTailCallOpt &&
2252 CalleeCC == CallingConv::Fast && F->getCallingConv() == CalleeCC)
2255 if (CalleeCC != CallingConv::Fast &&
2256 CalleeCC != CallingConv::C)
2259 // Look for obvious safe cases to perform tail call optimization.
2260 // For now, only consider callees which take no arguments and no return
2266 // If the caller does not return a value, then this is obviously safe.
2267 return F->getReturnType()->isVoidTy();
2273 X86TargetLowering::createFastISel(MachineFunction &mf, MachineModuleInfo *mmo,
2275 DenseMap<const Value *, unsigned> &vm,
2276 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2277 DenseMap<const AllocaInst *, int> &am
2279 , SmallSet<Instruction*, 8> &cil
2282 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2290 //===----------------------------------------------------------------------===//
2291 // Other Lowering Hooks
2292 //===----------------------------------------------------------------------===//
2295 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2296 MachineFunction &MF = DAG.getMachineFunction();
2297 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2298 int ReturnAddrIndex = FuncInfo->getRAIndex();
2300 if (ReturnAddrIndex == 0) {
2301 // Set up a frame object for the return address.
2302 uint64_t SlotSize = TD->getPointerSize();
2303 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2305 FuncInfo->setRAIndex(ReturnAddrIndex);
2308 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2312 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2313 bool hasSymbolicDisplacement) {
2314 // Offset should fit into 32 bit immediate field.
2315 if (!isInt32(Offset))
2318 // If we don't have a symbolic displacement - we don't have any extra
2320 if (!hasSymbolicDisplacement)
2323 // FIXME: Some tweaks might be needed for medium code model.
2324 if (M != CodeModel::Small && M != CodeModel::Kernel)
2327 // For small code model we assume that latest object is 16MB before end of 31
2328 // bits boundary. We may also accept pretty large negative constants knowing
2329 // that all objects are in the positive half of address space.
2330 if (M == CodeModel::Small && Offset < 16*1024*1024)
2333 // For kernel code model we know that all object resist in the negative half
2334 // of 32bits address space. We may not accept negative offsets, since they may
2335 // be just off and we may accept pretty large positive ones.
2336 if (M == CodeModel::Kernel && Offset > 0)
2342 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2343 /// specific condition code, returning the condition code and the LHS/RHS of the
2344 /// comparison to make.
2345 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2346 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2348 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2349 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2350 // X > -1 -> X == 0, jump !sign.
2351 RHS = DAG.getConstant(0, RHS.getValueType());
2352 return X86::COND_NS;
2353 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2354 // X < 0 -> X == 0, jump on sign.
2356 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2358 RHS = DAG.getConstant(0, RHS.getValueType());
2359 return X86::COND_LE;
2363 switch (SetCCOpcode) {
2364 default: llvm_unreachable("Invalid integer condition!");
2365 case ISD::SETEQ: return X86::COND_E;
2366 case ISD::SETGT: return X86::COND_G;
2367 case ISD::SETGE: return X86::COND_GE;
2368 case ISD::SETLT: return X86::COND_L;
2369 case ISD::SETLE: return X86::COND_LE;
2370 case ISD::SETNE: return X86::COND_NE;
2371 case ISD::SETULT: return X86::COND_B;
2372 case ISD::SETUGT: return X86::COND_A;
2373 case ISD::SETULE: return X86::COND_BE;
2374 case ISD::SETUGE: return X86::COND_AE;
2378 // First determine if it is required or is profitable to flip the operands.
2380 // If LHS is a foldable load, but RHS is not, flip the condition.
2381 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2382 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2383 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2384 std::swap(LHS, RHS);
2387 switch (SetCCOpcode) {
2393 std::swap(LHS, RHS);
2397 // On a floating point condition, the flags are set as follows:
2399 // 0 | 0 | 0 | X > Y
2400 // 0 | 0 | 1 | X < Y
2401 // 1 | 0 | 0 | X == Y
2402 // 1 | 1 | 1 | unordered
2403 switch (SetCCOpcode) {
2404 default: llvm_unreachable("Condcode should be pre-legalized away");
2406 case ISD::SETEQ: return X86::COND_E;
2407 case ISD::SETOLT: // flipped
2409 case ISD::SETGT: return X86::COND_A;
2410 case ISD::SETOLE: // flipped
2412 case ISD::SETGE: return X86::COND_AE;
2413 case ISD::SETUGT: // flipped
2415 case ISD::SETLT: return X86::COND_B;
2416 case ISD::SETUGE: // flipped
2418 case ISD::SETLE: return X86::COND_BE;
2420 case ISD::SETNE: return X86::COND_NE;
2421 case ISD::SETUO: return X86::COND_P;
2422 case ISD::SETO: return X86::COND_NP;
2424 case ISD::SETUNE: return X86::COND_INVALID;
2428 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2429 /// code. Current x86 isa includes the following FP cmov instructions:
2430 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2431 static bool hasFPCMov(unsigned X86CC) {
2447 /// isFPImmLegal - Returns true if the target can instruction select the
2448 /// specified FP immediate natively. If false, the legalizer will
2449 /// materialize the FP immediate as a load from a constant pool.
2450 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2451 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2452 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2458 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2459 /// the specified range (L, H].
2460 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2461 return (Val < 0) || (Val >= Low && Val < Hi);
2464 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2465 /// specified value.
2466 static bool isUndefOrEqual(int Val, int CmpVal) {
2467 if (Val < 0 || Val == CmpVal)
2472 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2473 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2474 /// the second operand.
2475 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2476 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2477 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2478 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2479 return (Mask[0] < 2 && Mask[1] < 2);
2483 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2484 SmallVector<int, 8> M;
2486 return ::isPSHUFDMask(M, N->getValueType(0));
2489 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2490 /// is suitable for input to PSHUFHW.
2491 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2492 if (VT != MVT::v8i16)
2495 // Lower quadword copied in order or undef.
2496 for (int i = 0; i != 4; ++i)
2497 if (Mask[i] >= 0 && Mask[i] != i)
2500 // Upper quadword shuffled.
2501 for (int i = 4; i != 8; ++i)
2502 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2508 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2509 SmallVector<int, 8> M;
2511 return ::isPSHUFHWMask(M, N->getValueType(0));
2514 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2515 /// is suitable for input to PSHUFLW.
2516 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2517 if (VT != MVT::v8i16)
2520 // Upper quadword copied in order.
2521 for (int i = 4; i != 8; ++i)
2522 if (Mask[i] >= 0 && Mask[i] != i)
2525 // Lower quadword shuffled.
2526 for (int i = 0; i != 4; ++i)
2533 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2534 SmallVector<int, 8> M;
2536 return ::isPSHUFLWMask(M, N->getValueType(0));
2539 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2540 /// is suitable for input to PALIGNR.
2541 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2543 int i, e = VT.getVectorNumElements();
2545 // Do not handle v2i64 / v2f64 shuffles with palignr.
2546 if (e < 4 || !hasSSSE3)
2549 for (i = 0; i != e; ++i)
2553 // All undef, not a palignr.
2557 // Determine if it's ok to perform a palignr with only the LHS, since we
2558 // don't have access to the actual shuffle elements to see if RHS is undef.
2559 bool Unary = Mask[i] < (int)e;
2560 bool NeedsUnary = false;
2562 int s = Mask[i] - i;
2564 // Check the rest of the elements to see if they are consecutive.
2565 for (++i; i != e; ++i) {
2570 Unary = Unary && (m < (int)e);
2571 NeedsUnary = NeedsUnary || (m < s);
2573 if (NeedsUnary && !Unary)
2575 if (Unary && m != ((s+i) & (e-1)))
2577 if (!Unary && m != (s+i))
2583 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2584 SmallVector<int, 8> M;
2586 return ::isPALIGNRMask(M, N->getValueType(0), true);
2589 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2590 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2591 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2592 int NumElems = VT.getVectorNumElements();
2593 if (NumElems != 2 && NumElems != 4)
2596 int Half = NumElems / 2;
2597 for (int i = 0; i < Half; ++i)
2598 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2600 for (int i = Half; i < NumElems; ++i)
2601 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2607 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2608 SmallVector<int, 8> M;
2610 return ::isSHUFPMask(M, N->getValueType(0));
2613 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2614 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2615 /// half elements to come from vector 1 (which would equal the dest.) and
2616 /// the upper half to come from vector 2.
2617 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2618 int NumElems = VT.getVectorNumElements();
2620 if (NumElems != 2 && NumElems != 4)
2623 int Half = NumElems / 2;
2624 for (int i = 0; i < Half; ++i)
2625 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2627 for (int i = Half; i < NumElems; ++i)
2628 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2633 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2634 SmallVector<int, 8> M;
2636 return isCommutedSHUFPMask(M, N->getValueType(0));
2639 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2640 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2641 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2642 if (N->getValueType(0).getVectorNumElements() != 4)
2645 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2646 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2647 isUndefOrEqual(N->getMaskElt(1), 7) &&
2648 isUndefOrEqual(N->getMaskElt(2), 2) &&
2649 isUndefOrEqual(N->getMaskElt(3), 3);
2652 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2653 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2655 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2656 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2661 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2662 isUndefOrEqual(N->getMaskElt(1), 3) &&
2663 isUndefOrEqual(N->getMaskElt(2), 2) &&
2664 isUndefOrEqual(N->getMaskElt(3), 3);
2667 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2668 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2669 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2670 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2672 if (NumElems != 2 && NumElems != 4)
2675 for (unsigned i = 0; i < NumElems/2; ++i)
2676 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2679 for (unsigned i = NumElems/2; i < NumElems; ++i)
2680 if (!isUndefOrEqual(N->getMaskElt(i), i))
2686 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2687 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2688 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2689 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2691 if (NumElems != 2 && NumElems != 4)
2694 for (unsigned i = 0; i < NumElems/2; ++i)
2695 if (!isUndefOrEqual(N->getMaskElt(i), i))
2698 for (unsigned i = 0; i < NumElems/2; ++i)
2699 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2705 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2706 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2707 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2708 bool V2IsSplat = false) {
2709 int NumElts = VT.getVectorNumElements();
2710 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2713 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2715 int BitI1 = Mask[i+1];
2716 if (!isUndefOrEqual(BitI, j))
2719 if (!isUndefOrEqual(BitI1, NumElts))
2722 if (!isUndefOrEqual(BitI1, j + NumElts))
2729 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2730 SmallVector<int, 8> M;
2732 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2735 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2736 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2737 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2738 bool V2IsSplat = false) {
2739 int NumElts = VT.getVectorNumElements();
2740 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2743 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2745 int BitI1 = Mask[i+1];
2746 if (!isUndefOrEqual(BitI, j + NumElts/2))
2749 if (isUndefOrEqual(BitI1, NumElts))
2752 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2759 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2760 SmallVector<int, 8> M;
2762 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2765 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2766 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2768 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2769 int NumElems = VT.getVectorNumElements();
2770 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2773 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2775 int BitI1 = Mask[i+1];
2776 if (!isUndefOrEqual(BitI, j))
2778 if (!isUndefOrEqual(BitI1, j))
2784 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2785 SmallVector<int, 8> M;
2787 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2790 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2791 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2793 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2794 int NumElems = VT.getVectorNumElements();
2795 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2798 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2800 int BitI1 = Mask[i+1];
2801 if (!isUndefOrEqual(BitI, j))
2803 if (!isUndefOrEqual(BitI1, j))
2809 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2810 SmallVector<int, 8> M;
2812 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2815 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2816 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2817 /// MOVSD, and MOVD, i.e. setting the lowest element.
2818 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2819 if (VT.getVectorElementType().getSizeInBits() < 32)
2822 int NumElts = VT.getVectorNumElements();
2824 if (!isUndefOrEqual(Mask[0], NumElts))
2827 for (int i = 1; i < NumElts; ++i)
2828 if (!isUndefOrEqual(Mask[i], i))
2834 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2835 SmallVector<int, 8> M;
2837 return ::isMOVLMask(M, N->getValueType(0));
2840 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2841 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2842 /// element of vector 2 and the other elements to come from vector 1 in order.
2843 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2844 bool V2IsSplat = false, bool V2IsUndef = false) {
2845 int NumOps = VT.getVectorNumElements();
2846 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2849 if (!isUndefOrEqual(Mask[0], 0))
2852 for (int i = 1; i < NumOps; ++i)
2853 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2854 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2855 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2861 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2862 bool V2IsUndef = false) {
2863 SmallVector<int, 8> M;
2865 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2868 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2869 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2870 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2871 if (N->getValueType(0).getVectorNumElements() != 4)
2874 // Expect 1, 1, 3, 3
2875 for (unsigned i = 0; i < 2; ++i) {
2876 int Elt = N->getMaskElt(i);
2877 if (Elt >= 0 && Elt != 1)
2882 for (unsigned i = 2; i < 4; ++i) {
2883 int Elt = N->getMaskElt(i);
2884 if (Elt >= 0 && Elt != 3)
2889 // Don't use movshdup if it can be done with a shufps.
2890 // FIXME: verify that matching u, u, 3, 3 is what we want.
2894 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2895 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2896 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2897 if (N->getValueType(0).getVectorNumElements() != 4)
2900 // Expect 0, 0, 2, 2
2901 for (unsigned i = 0; i < 2; ++i)
2902 if (N->getMaskElt(i) > 0)
2906 for (unsigned i = 2; i < 4; ++i) {
2907 int Elt = N->getMaskElt(i);
2908 if (Elt >= 0 && Elt != 2)
2913 // Don't use movsldup if it can be done with a shufps.
2917 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2918 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2919 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2920 int e = N->getValueType(0).getVectorNumElements() / 2;
2922 for (int i = 0; i < e; ++i)
2923 if (!isUndefOrEqual(N->getMaskElt(i), i))
2925 for (int i = 0; i < e; ++i)
2926 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2931 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2932 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
2933 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2934 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2935 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2937 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2939 for (int i = 0; i < NumOperands; ++i) {
2940 int Val = SVOp->getMaskElt(NumOperands-i-1);
2941 if (Val < 0) Val = 0;
2942 if (Val >= NumOperands) Val -= NumOperands;
2944 if (i != NumOperands - 1)
2950 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2951 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
2952 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2953 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2955 // 8 nodes, but we only care about the last 4.
2956 for (unsigned i = 7; i >= 4; --i) {
2957 int Val = SVOp->getMaskElt(i);
2966 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2967 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
2968 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2969 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2971 // 8 nodes, but we only care about the first 4.
2972 for (int i = 3; i >= 0; --i) {
2973 int Val = SVOp->getMaskElt(i);
2982 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
2983 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
2984 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
2985 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2986 EVT VVT = N->getValueType(0);
2987 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
2991 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
2992 Val = SVOp->getMaskElt(i);
2996 return (Val - i) * EltSize;
2999 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3001 bool X86::isZeroNode(SDValue Elt) {
3002 return ((isa<ConstantSDNode>(Elt) &&
3003 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
3004 (isa<ConstantFPSDNode>(Elt) &&
3005 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3008 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3009 /// their permute mask.
3010 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3011 SelectionDAG &DAG) {
3012 EVT VT = SVOp->getValueType(0);
3013 unsigned NumElems = VT.getVectorNumElements();
3014 SmallVector<int, 8> MaskVec;
3016 for (unsigned i = 0; i != NumElems; ++i) {
3017 int idx = SVOp->getMaskElt(i);
3019 MaskVec.push_back(idx);
3020 else if (idx < (int)NumElems)
3021 MaskVec.push_back(idx + NumElems);
3023 MaskVec.push_back(idx - NumElems);
3025 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3026 SVOp->getOperand(0), &MaskVec[0]);
3029 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3030 /// the two vector operands have swapped position.
3031 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3032 unsigned NumElems = VT.getVectorNumElements();
3033 for (unsigned i = 0; i != NumElems; ++i) {
3037 else if (idx < (int)NumElems)
3038 Mask[i] = idx + NumElems;
3040 Mask[i] = idx - NumElems;
3044 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3045 /// match movhlps. The lower half elements should come from upper half of
3046 /// V1 (and in order), and the upper half elements should come from the upper
3047 /// half of V2 (and in order).
3048 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3049 if (Op->getValueType(0).getVectorNumElements() != 4)
3051 for (unsigned i = 0, e = 2; i != e; ++i)
3052 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3054 for (unsigned i = 2; i != 4; ++i)
3055 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3060 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3061 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3063 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3064 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3066 N = N->getOperand(0).getNode();
3067 if (!ISD::isNON_EXTLoad(N))
3070 *LD = cast<LoadSDNode>(N);
3074 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3075 /// match movlp{s|d}. The lower half elements should come from lower half of
3076 /// V1 (and in order), and the upper half elements should come from the upper
3077 /// half of V2 (and in order). And since V1 will become the source of the
3078 /// MOVLP, it must be either a vector load or a scalar load to vector.
3079 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3080 ShuffleVectorSDNode *Op) {
3081 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3083 // Is V2 is a vector load, don't do this transformation. We will try to use
3084 // load folding shufps op.
3085 if (ISD::isNON_EXTLoad(V2))
3088 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3090 if (NumElems != 2 && NumElems != 4)
3092 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3093 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3095 for (unsigned i = NumElems/2; i != NumElems; ++i)
3096 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3101 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3103 static bool isSplatVector(SDNode *N) {
3104 if (N->getOpcode() != ISD::BUILD_VECTOR)
3107 SDValue SplatValue = N->getOperand(0);
3108 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3109 if (N->getOperand(i) != SplatValue)
3114 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3115 /// to an zero vector.
3116 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3117 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3118 SDValue V1 = N->getOperand(0);
3119 SDValue V2 = N->getOperand(1);
3120 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3121 for (unsigned i = 0; i != NumElems; ++i) {
3122 int Idx = N->getMaskElt(i);
3123 if (Idx >= (int)NumElems) {
3124 unsigned Opc = V2.getOpcode();
3125 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3127 if (Opc != ISD::BUILD_VECTOR ||
3128 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3130 } else if (Idx >= 0) {
3131 unsigned Opc = V1.getOpcode();
3132 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3134 if (Opc != ISD::BUILD_VECTOR ||
3135 !X86::isZeroNode(V1.getOperand(Idx)))
3142 /// getZeroVector - Returns a vector of specified type with all zero elements.
3144 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3146 assert(VT.isVector() && "Expected a vector type");
3148 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3149 // type. This ensures they get CSE'd.
3151 if (VT.getSizeInBits() == 64) { // MMX
3152 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3153 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3154 } else if (HasSSE2) { // SSE2
3155 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3156 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3158 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3159 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3161 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3164 /// getOnesVector - Returns a vector of specified type with all bits set.
3166 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3167 assert(VT.isVector() && "Expected a vector type");
3169 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3170 // type. This ensures they get CSE'd.
3171 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3173 if (VT.getSizeInBits() == 64) // MMX
3174 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3176 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3177 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3181 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3182 /// that point to V2 points to its first element.
3183 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3184 EVT VT = SVOp->getValueType(0);
3185 unsigned NumElems = VT.getVectorNumElements();
3187 bool Changed = false;
3188 SmallVector<int, 8> MaskVec;
3189 SVOp->getMask(MaskVec);
3191 for (unsigned i = 0; i != NumElems; ++i) {
3192 if (MaskVec[i] > (int)NumElems) {
3193 MaskVec[i] = NumElems;
3198 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3199 SVOp->getOperand(1), &MaskVec[0]);
3200 return SDValue(SVOp, 0);
3203 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3204 /// operation of specified width.
3205 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3207 unsigned NumElems = VT.getVectorNumElements();
3208 SmallVector<int, 8> Mask;
3209 Mask.push_back(NumElems);
3210 for (unsigned i = 1; i != NumElems; ++i)
3212 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3215 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3216 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3218 unsigned NumElems = VT.getVectorNumElements();
3219 SmallVector<int, 8> Mask;
3220 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3222 Mask.push_back(i + NumElems);
3224 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3227 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3228 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3230 unsigned NumElems = VT.getVectorNumElements();
3231 unsigned Half = NumElems/2;
3232 SmallVector<int, 8> Mask;
3233 for (unsigned i = 0; i != Half; ++i) {
3234 Mask.push_back(i + Half);
3235 Mask.push_back(i + NumElems + Half);
3237 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3240 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3241 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3243 if (SV->getValueType(0).getVectorNumElements() <= 4)
3244 return SDValue(SV, 0);
3246 EVT PVT = MVT::v4f32;
3247 EVT VT = SV->getValueType(0);
3248 DebugLoc dl = SV->getDebugLoc();
3249 SDValue V1 = SV->getOperand(0);
3250 int NumElems = VT.getVectorNumElements();
3251 int EltNo = SV->getSplatIndex();
3253 // unpack elements to the correct location
3254 while (NumElems > 4) {
3255 if (EltNo < NumElems/2) {
3256 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3258 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3259 EltNo -= NumElems/2;
3264 // Perform the splat.
3265 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3266 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3267 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3268 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3271 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3272 /// vector of zero or undef vector. This produces a shuffle where the low
3273 /// element of V2 is swizzled into the zero/undef vector, landing at element
3274 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3275 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3276 bool isZero, bool HasSSE2,
3277 SelectionDAG &DAG) {
3278 EVT VT = V2.getValueType();
3280 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3281 unsigned NumElems = VT.getVectorNumElements();
3282 SmallVector<int, 16> MaskVec;
3283 for (unsigned i = 0; i != NumElems; ++i)
3284 // If this is the insertion idx, put the low elt of V2 here.
3285 MaskVec.push_back(i == Idx ? NumElems : i);
3286 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3289 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3290 /// a shuffle that is zero.
3292 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3293 bool Low, SelectionDAG &DAG) {
3294 unsigned NumZeros = 0;
3295 for (int i = 0; i < NumElems; ++i) {
3296 unsigned Index = Low ? i : NumElems-i-1;
3297 int Idx = SVOp->getMaskElt(Index);
3302 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3303 if (Elt.getNode() && X86::isZeroNode(Elt))
3311 /// isVectorShift - Returns true if the shuffle can be implemented as a
3312 /// logical left or right shift of a vector.
3313 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3314 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3315 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3316 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3319 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3322 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3326 bool SeenV1 = false;
3327 bool SeenV2 = false;
3328 for (int i = NumZeros; i < NumElems; ++i) {
3329 int Val = isLeft ? (i - NumZeros) : i;
3330 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3342 if (SeenV1 && SeenV2)
3345 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3351 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3353 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3354 unsigned NumNonZero, unsigned NumZero,
3355 SelectionDAG &DAG, TargetLowering &TLI) {
3359 DebugLoc dl = Op.getDebugLoc();
3362 for (unsigned i = 0; i < 16; ++i) {
3363 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3364 if (ThisIsNonZero && First) {
3366 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3368 V = DAG.getUNDEF(MVT::v8i16);
3373 SDValue ThisElt(0, 0), LastElt(0, 0);
3374 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3375 if (LastIsNonZero) {
3376 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3377 MVT::i16, Op.getOperand(i-1));
3379 if (ThisIsNonZero) {
3380 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3381 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3382 ThisElt, DAG.getConstant(8, MVT::i8));
3384 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3388 if (ThisElt.getNode())
3389 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3390 DAG.getIntPtrConstant(i/2));
3394 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3397 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3399 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3400 unsigned NumNonZero, unsigned NumZero,
3401 SelectionDAG &DAG, TargetLowering &TLI) {
3405 DebugLoc dl = Op.getDebugLoc();
3408 for (unsigned i = 0; i < 8; ++i) {
3409 bool isNonZero = (NonZeros & (1 << i)) != 0;
3413 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3415 V = DAG.getUNDEF(MVT::v8i16);
3418 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3419 MVT::v8i16, V, Op.getOperand(i),
3420 DAG.getIntPtrConstant(i));
3427 /// getVShift - Return a vector logical shift node.
3429 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3430 unsigned NumBits, SelectionDAG &DAG,
3431 const TargetLowering &TLI, DebugLoc dl) {
3432 bool isMMX = VT.getSizeInBits() == 64;
3433 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3434 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3435 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3436 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3437 DAG.getNode(Opc, dl, ShVT, SrcOp,
3438 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3442 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3443 SelectionDAG &DAG) {
3445 // Check if the scalar load can be widened into a vector load. And if
3446 // the address is "base + cst" see if the cst can be "absorbed" into
3447 // the shuffle mask.
3448 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3449 SDValue Ptr = LD->getBasePtr();
3450 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3452 EVT PVT = LD->getValueType(0);
3453 if (PVT != MVT::i32 && PVT != MVT::f32)
3458 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3459 FI = FINode->getIndex();
3461 } else if (Ptr.getOpcode() == ISD::ADD &&
3462 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3463 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3464 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3465 Offset = Ptr.getConstantOperandVal(1);
3466 Ptr = Ptr.getOperand(0);
3471 SDValue Chain = LD->getChain();
3472 // Make sure the stack object alignment is at least 16.
3473 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3474 if (DAG.InferPtrAlignment(Ptr) < 16) {
3475 if (MFI->isFixedObjectIndex(FI)) {
3476 // Can't change the alignment. FIXME: It's possible to compute
3477 // the exact stack offset and reference FI + adjust offset instead.
3478 // If someone *really* cares about this. That's the way to implement it.
3481 MFI->setObjectAlignment(FI, 16);
3485 // (Offset % 16) must be multiple of 4. Then address is then
3486 // Ptr + (Offset & ~15).
3489 if ((Offset % 16) & 3)
3491 int64_t StartOffset = Offset & ~15;
3493 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3494 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3496 int EltNo = (Offset - StartOffset) >> 2;
3497 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3498 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3499 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0);
3500 // Canonicalize it to a v4i32 shuffle.
3501 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3502 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3503 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3504 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3511 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3512 DebugLoc dl = Op.getDebugLoc();
3513 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3514 if (ISD::isBuildVectorAllZeros(Op.getNode())
3515 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3516 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3517 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3518 // eliminated on x86-32 hosts.
3519 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3522 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3523 return getOnesVector(Op.getValueType(), DAG, dl);
3524 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3527 EVT VT = Op.getValueType();
3528 EVT ExtVT = VT.getVectorElementType();
3529 unsigned EVTBits = ExtVT.getSizeInBits();
3531 unsigned NumElems = Op.getNumOperands();
3532 unsigned NumZero = 0;
3533 unsigned NumNonZero = 0;
3534 unsigned NonZeros = 0;
3535 bool IsAllConstants = true;
3536 SmallSet<SDValue, 8> Values;
3537 for (unsigned i = 0; i < NumElems; ++i) {
3538 SDValue Elt = Op.getOperand(i);
3539 if (Elt.getOpcode() == ISD::UNDEF)
3542 if (Elt.getOpcode() != ISD::Constant &&
3543 Elt.getOpcode() != ISD::ConstantFP)
3544 IsAllConstants = false;
3545 if (X86::isZeroNode(Elt))
3548 NonZeros |= (1 << i);
3553 if (NumNonZero == 0) {
3554 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3555 return DAG.getUNDEF(VT);
3558 // Special case for single non-zero, non-undef, element.
3559 if (NumNonZero == 1) {
3560 unsigned Idx = CountTrailingZeros_32(NonZeros);
3561 SDValue Item = Op.getOperand(Idx);
3563 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3564 // the value are obviously zero, truncate the value to i32 and do the
3565 // insertion that way. Only do this if the value is non-constant or if the
3566 // value is a constant being inserted into element 0. It is cheaper to do
3567 // a constant pool load than it is to do a movd + shuffle.
3568 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3569 (!IsAllConstants || Idx == 0)) {
3570 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3571 // Handle MMX and SSE both.
3572 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3573 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3575 // Truncate the value (which may itself be a constant) to i32, and
3576 // convert it to a vector with movd (S2V+shuffle to zero extend).
3577 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3578 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3579 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3580 Subtarget->hasSSE2(), DAG);
3582 // Now we have our 32-bit value zero extended in the low element of
3583 // a vector. If Idx != 0, swizzle it into place.
3585 SmallVector<int, 4> Mask;
3586 Mask.push_back(Idx);
3587 for (unsigned i = 1; i != VecElts; ++i)
3589 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3590 DAG.getUNDEF(Item.getValueType()),
3593 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3597 // If we have a constant or non-constant insertion into the low element of
3598 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3599 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3600 // depending on what the source datatype is.
3603 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3604 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3605 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3606 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3607 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3608 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3610 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3611 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3612 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3613 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3614 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3615 Subtarget->hasSSE2(), DAG);
3616 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3620 // Is it a vector logical left shift?
3621 if (NumElems == 2 && Idx == 1 &&
3622 X86::isZeroNode(Op.getOperand(0)) &&
3623 !X86::isZeroNode(Op.getOperand(1))) {
3624 unsigned NumBits = VT.getSizeInBits();
3625 return getVShift(true, VT,
3626 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3627 VT, Op.getOperand(1)),
3628 NumBits/2, DAG, *this, dl);
3631 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3634 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3635 // is a non-constant being inserted into an element other than the low one,
3636 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3637 // movd/movss) to move this into the low element, then shuffle it into
3639 if (EVTBits == 32) {
3640 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3642 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3643 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3644 Subtarget->hasSSE2(), DAG);
3645 SmallVector<int, 8> MaskVec;
3646 for (unsigned i = 0; i < NumElems; i++)
3647 MaskVec.push_back(i == Idx ? 0 : 1);
3648 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3652 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3653 if (Values.size() == 1) {
3654 if (EVTBits == 32) {
3655 // Instead of a shuffle like this:
3656 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3657 // Check if it's possible to issue this instead.
3658 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3659 unsigned Idx = CountTrailingZeros_32(NonZeros);
3660 SDValue Item = Op.getOperand(Idx);
3661 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3662 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3667 // A vector full of immediates; various special cases are already
3668 // handled, so this is best done with a single constant-pool load.
3672 // Let legalizer expand 2-wide build_vectors.
3673 if (EVTBits == 64) {
3674 if (NumNonZero == 1) {
3675 // One half is zero or undef.
3676 unsigned Idx = CountTrailingZeros_32(NonZeros);
3677 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3678 Op.getOperand(Idx));
3679 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3680 Subtarget->hasSSE2(), DAG);
3685 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3686 if (EVTBits == 8 && NumElems == 16) {
3687 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3689 if (V.getNode()) return V;
3692 if (EVTBits == 16 && NumElems == 8) {
3693 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3695 if (V.getNode()) return V;
3698 // If element VT is == 32 bits, turn it into a number of shuffles.
3699 SmallVector<SDValue, 8> V;
3701 if (NumElems == 4 && NumZero > 0) {
3702 for (unsigned i = 0; i < 4; ++i) {
3703 bool isZero = !(NonZeros & (1 << i));
3705 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3707 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3710 for (unsigned i = 0; i < 2; ++i) {
3711 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3714 V[i] = V[i*2]; // Must be a zero vector.
3717 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3720 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3723 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3728 SmallVector<int, 8> MaskVec;
3729 bool Reverse = (NonZeros & 0x3) == 2;
3730 for (unsigned i = 0; i < 2; ++i)
3731 MaskVec.push_back(Reverse ? 1-i : i);
3732 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3733 for (unsigned i = 0; i < 2; ++i)
3734 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3735 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3738 if (Values.size() > 2) {
3739 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3740 // values to be inserted is equal to the number of elements, in which case
3741 // use the unpack code below in the hopes of matching the consecutive elts
3742 // load merge pattern for shuffles.
3743 // FIXME: We could probably just check that here directly.
3744 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3745 getSubtarget()->hasSSE41()) {
3746 V[0] = DAG.getUNDEF(VT);
3747 for (unsigned i = 0; i < NumElems; ++i)
3748 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3749 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3750 Op.getOperand(i), DAG.getIntPtrConstant(i));
3753 // Expand into a number of unpckl*.
3755 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3756 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3757 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3758 for (unsigned i = 0; i < NumElems; ++i)
3759 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3761 while (NumElems != 0) {
3762 for (unsigned i = 0; i < NumElems; ++i)
3763 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3773 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3774 // We support concatenate two MMX registers and place them in a MMX
3775 // register. This is better than doing a stack convert.
3776 DebugLoc dl = Op.getDebugLoc();
3777 EVT ResVT = Op.getValueType();
3778 assert(Op.getNumOperands() == 2);
3779 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
3780 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
3782 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
3783 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3784 InVec = Op.getOperand(1);
3785 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3786 unsigned NumElts = ResVT.getVectorNumElements();
3787 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3788 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
3789 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
3791 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
3792 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3793 Mask[0] = 0; Mask[1] = 2;
3794 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
3796 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3799 // v8i16 shuffles - Prefer shuffles in the following order:
3800 // 1. [all] pshuflw, pshufhw, optional move
3801 // 2. [ssse3] 1 x pshufb
3802 // 3. [ssse3] 2 x pshufb + 1 x por
3803 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3805 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3806 SelectionDAG &DAG, X86TargetLowering &TLI) {
3807 SDValue V1 = SVOp->getOperand(0);
3808 SDValue V2 = SVOp->getOperand(1);
3809 DebugLoc dl = SVOp->getDebugLoc();
3810 SmallVector<int, 8> MaskVals;
3812 // Determine if more than 1 of the words in each of the low and high quadwords
3813 // of the result come from the same quadword of one of the two inputs. Undef
3814 // mask values count as coming from any quadword, for better codegen.
3815 SmallVector<unsigned, 4> LoQuad(4);
3816 SmallVector<unsigned, 4> HiQuad(4);
3817 BitVector InputQuads(4);
3818 for (unsigned i = 0; i < 8; ++i) {
3819 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3820 int EltIdx = SVOp->getMaskElt(i);
3821 MaskVals.push_back(EltIdx);
3830 InputQuads.set(EltIdx / 4);
3833 int BestLoQuad = -1;
3834 unsigned MaxQuad = 1;
3835 for (unsigned i = 0; i < 4; ++i) {
3836 if (LoQuad[i] > MaxQuad) {
3838 MaxQuad = LoQuad[i];
3842 int BestHiQuad = -1;
3844 for (unsigned i = 0; i < 4; ++i) {
3845 if (HiQuad[i] > MaxQuad) {
3847 MaxQuad = HiQuad[i];
3851 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3852 // of the two input vectors, shuffle them into one input vector so only a
3853 // single pshufb instruction is necessary. If There are more than 2 input
3854 // quads, disable the next transformation since it does not help SSSE3.
3855 bool V1Used = InputQuads[0] || InputQuads[1];
3856 bool V2Used = InputQuads[2] || InputQuads[3];
3857 if (TLI.getSubtarget()->hasSSSE3()) {
3858 if (InputQuads.count() == 2 && V1Used && V2Used) {
3859 BestLoQuad = InputQuads.find_first();
3860 BestHiQuad = InputQuads.find_next(BestLoQuad);
3862 if (InputQuads.count() > 2) {
3868 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3869 // the shuffle mask. If a quad is scored as -1, that means that it contains
3870 // words from all 4 input quadwords.
3872 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3873 SmallVector<int, 8> MaskV;
3874 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3875 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3876 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3877 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3878 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3879 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3881 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3882 // source words for the shuffle, to aid later transformations.
3883 bool AllWordsInNewV = true;
3884 bool InOrder[2] = { true, true };
3885 for (unsigned i = 0; i != 8; ++i) {
3886 int idx = MaskVals[i];
3888 InOrder[i/4] = false;
3889 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3891 AllWordsInNewV = false;
3895 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3896 if (AllWordsInNewV) {
3897 for (int i = 0; i != 8; ++i) {
3898 int idx = MaskVals[i];
3901 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3902 if ((idx != i) && idx < 4)
3904 if ((idx != i) && idx > 3)
3913 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3914 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3915 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3916 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3917 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3921 // If we have SSSE3, and all words of the result are from 1 input vector,
3922 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3923 // is present, fall back to case 4.
3924 if (TLI.getSubtarget()->hasSSSE3()) {
3925 SmallVector<SDValue,16> pshufbMask;
3927 // If we have elements from both input vectors, set the high bit of the
3928 // shuffle mask element to zero out elements that come from V2 in the V1
3929 // mask, and elements that come from V1 in the V2 mask, so that the two
3930 // results can be OR'd together.
3931 bool TwoInputs = V1Used && V2Used;
3932 for (unsigned i = 0; i != 8; ++i) {
3933 int EltIdx = MaskVals[i] * 2;
3934 if (TwoInputs && (EltIdx >= 16)) {
3935 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3936 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3939 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3940 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3942 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3943 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3944 DAG.getNode(ISD::BUILD_VECTOR, dl,
3945 MVT::v16i8, &pshufbMask[0], 16));
3947 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3949 // Calculate the shuffle mask for the second input, shuffle it, and
3950 // OR it with the first shuffled input.
3952 for (unsigned i = 0; i != 8; ++i) {
3953 int EltIdx = MaskVals[i] * 2;
3955 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3956 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3959 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3960 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3962 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3963 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3964 DAG.getNode(ISD::BUILD_VECTOR, dl,
3965 MVT::v16i8, &pshufbMask[0], 16));
3966 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3967 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3970 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3971 // and update MaskVals with new element order.
3972 BitVector InOrder(8);
3973 if (BestLoQuad >= 0) {
3974 SmallVector<int, 8> MaskV;
3975 for (int i = 0; i != 4; ++i) {
3976 int idx = MaskVals[i];
3978 MaskV.push_back(-1);
3980 } else if ((idx / 4) == BestLoQuad) {
3981 MaskV.push_back(idx & 3);
3984 MaskV.push_back(-1);
3987 for (unsigned i = 4; i != 8; ++i)
3989 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3993 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
3994 // and update MaskVals with the new element order.
3995 if (BestHiQuad >= 0) {
3996 SmallVector<int, 8> MaskV;
3997 for (unsigned i = 0; i != 4; ++i)
3999 for (unsigned i = 4; i != 8; ++i) {
4000 int idx = MaskVals[i];
4002 MaskV.push_back(-1);
4004 } else if ((idx / 4) == BestHiQuad) {
4005 MaskV.push_back((idx & 3) + 4);
4008 MaskV.push_back(-1);
4011 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4015 // In case BestHi & BestLo were both -1, which means each quadword has a word
4016 // from each of the four input quadwords, calculate the InOrder bitvector now
4017 // before falling through to the insert/extract cleanup.
4018 if (BestLoQuad == -1 && BestHiQuad == -1) {
4020 for (int i = 0; i != 8; ++i)
4021 if (MaskVals[i] < 0 || MaskVals[i] == i)
4025 // The other elements are put in the right place using pextrw and pinsrw.
4026 for (unsigned i = 0; i != 8; ++i) {
4029 int EltIdx = MaskVals[i];
4032 SDValue ExtOp = (EltIdx < 8)
4033 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4034 DAG.getIntPtrConstant(EltIdx))
4035 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4036 DAG.getIntPtrConstant(EltIdx - 8));
4037 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4038 DAG.getIntPtrConstant(i));
4043 // v16i8 shuffles - Prefer shuffles in the following order:
4044 // 1. [ssse3] 1 x pshufb
4045 // 2. [ssse3] 2 x pshufb + 1 x por
4046 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4048 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4049 SelectionDAG &DAG, X86TargetLowering &TLI) {
4050 SDValue V1 = SVOp->getOperand(0);
4051 SDValue V2 = SVOp->getOperand(1);
4052 DebugLoc dl = SVOp->getDebugLoc();
4053 SmallVector<int, 16> MaskVals;
4054 SVOp->getMask(MaskVals);
4056 // If we have SSSE3, case 1 is generated when all result bytes come from
4057 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4058 // present, fall back to case 3.
4059 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4062 for (unsigned i = 0; i < 16; ++i) {
4063 int EltIdx = MaskVals[i];
4072 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4073 if (TLI.getSubtarget()->hasSSSE3()) {
4074 SmallVector<SDValue,16> pshufbMask;
4076 // If all result elements are from one input vector, then only translate
4077 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4079 // Otherwise, we have elements from both input vectors, and must zero out
4080 // elements that come from V2 in the first mask, and V1 in the second mask
4081 // so that we can OR them together.
4082 bool TwoInputs = !(V1Only || V2Only);
4083 for (unsigned i = 0; i != 16; ++i) {
4084 int EltIdx = MaskVals[i];
4085 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4086 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4089 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4091 // If all the elements are from V2, assign it to V1 and return after
4092 // building the first pshufb.
4095 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4096 DAG.getNode(ISD::BUILD_VECTOR, dl,
4097 MVT::v16i8, &pshufbMask[0], 16));
4101 // Calculate the shuffle mask for the second input, shuffle it, and
4102 // OR it with the first shuffled input.
4104 for (unsigned i = 0; i != 16; ++i) {
4105 int EltIdx = MaskVals[i];
4107 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4110 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4112 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4113 DAG.getNode(ISD::BUILD_VECTOR, dl,
4114 MVT::v16i8, &pshufbMask[0], 16));
4115 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4118 // No SSSE3 - Calculate in place words and then fix all out of place words
4119 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4120 // the 16 different words that comprise the two doublequadword input vectors.
4121 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4122 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4123 SDValue NewV = V2Only ? V2 : V1;
4124 for (int i = 0; i != 8; ++i) {
4125 int Elt0 = MaskVals[i*2];
4126 int Elt1 = MaskVals[i*2+1];
4128 // This word of the result is all undef, skip it.
4129 if (Elt0 < 0 && Elt1 < 0)
4132 // This word of the result is already in the correct place, skip it.
4133 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4135 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4138 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4139 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4142 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4143 // using a single extract together, load it and store it.
4144 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4145 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4146 DAG.getIntPtrConstant(Elt1 / 2));
4147 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4148 DAG.getIntPtrConstant(i));
4152 // If Elt1 is defined, extract it from the appropriate source. If the
4153 // source byte is not also odd, shift the extracted word left 8 bits
4154 // otherwise clear the bottom 8 bits if we need to do an or.
4156 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4157 DAG.getIntPtrConstant(Elt1 / 2));
4158 if ((Elt1 & 1) == 0)
4159 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4160 DAG.getConstant(8, TLI.getShiftAmountTy()));
4162 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4163 DAG.getConstant(0xFF00, MVT::i16));
4165 // If Elt0 is defined, extract it from the appropriate source. If the
4166 // source byte is not also even, shift the extracted word right 8 bits. If
4167 // Elt1 was also defined, OR the extracted values together before
4168 // inserting them in the result.
4170 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4171 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4172 if ((Elt0 & 1) != 0)
4173 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4174 DAG.getConstant(8, TLI.getShiftAmountTy()));
4176 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4177 DAG.getConstant(0x00FF, MVT::i16));
4178 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4181 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4182 DAG.getIntPtrConstant(i));
4184 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4187 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4188 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4189 /// done when every pair / quad of shuffle mask elements point to elements in
4190 /// the right sequence. e.g.
4191 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4193 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4195 TargetLowering &TLI, DebugLoc dl) {
4196 EVT VT = SVOp->getValueType(0);
4197 SDValue V1 = SVOp->getOperand(0);
4198 SDValue V2 = SVOp->getOperand(1);
4199 unsigned NumElems = VT.getVectorNumElements();
4200 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4201 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4202 EVT MaskEltVT = MaskVT.getVectorElementType();
4204 switch (VT.getSimpleVT().SimpleTy) {
4205 default: assert(false && "Unexpected!");
4206 case MVT::v4f32: NewVT = MVT::v2f64; break;
4207 case MVT::v4i32: NewVT = MVT::v2i64; break;
4208 case MVT::v8i16: NewVT = MVT::v4i32; break;
4209 case MVT::v16i8: NewVT = MVT::v4i32; break;
4212 if (NewWidth == 2) {
4218 int Scale = NumElems / NewWidth;
4219 SmallVector<int, 8> MaskVec;
4220 for (unsigned i = 0; i < NumElems; i += Scale) {
4222 for (int j = 0; j < Scale; ++j) {
4223 int EltIdx = SVOp->getMaskElt(i+j);
4227 StartIdx = EltIdx - (EltIdx % Scale);
4228 if (EltIdx != StartIdx + j)
4232 MaskVec.push_back(-1);
4234 MaskVec.push_back(StartIdx / Scale);
4237 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4238 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4239 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4242 /// getVZextMovL - Return a zero-extending vector move low node.
4244 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4245 SDValue SrcOp, SelectionDAG &DAG,
4246 const X86Subtarget *Subtarget, DebugLoc dl) {
4247 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4248 LoadSDNode *LD = NULL;
4249 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4250 LD = dyn_cast<LoadSDNode>(SrcOp);
4252 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4254 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4255 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4256 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4257 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4258 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4260 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4261 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4262 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4263 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4271 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4272 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4273 DAG.getNode(ISD::BIT_CONVERT, dl,
4277 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4280 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4281 SDValue V1 = SVOp->getOperand(0);
4282 SDValue V2 = SVOp->getOperand(1);
4283 DebugLoc dl = SVOp->getDebugLoc();
4284 EVT VT = SVOp->getValueType(0);
4286 SmallVector<std::pair<int, int>, 8> Locs;
4288 SmallVector<int, 8> Mask1(4U, -1);
4289 SmallVector<int, 8> PermMask;
4290 SVOp->getMask(PermMask);
4294 for (unsigned i = 0; i != 4; ++i) {
4295 int Idx = PermMask[i];
4297 Locs[i] = std::make_pair(-1, -1);
4299 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4301 Locs[i] = std::make_pair(0, NumLo);
4305 Locs[i] = std::make_pair(1, NumHi);
4307 Mask1[2+NumHi] = Idx;
4313 if (NumLo <= 2 && NumHi <= 2) {
4314 // If no more than two elements come from either vector. This can be
4315 // implemented with two shuffles. First shuffle gather the elements.
4316 // The second shuffle, which takes the first shuffle as both of its
4317 // vector operands, put the elements into the right order.
4318 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4320 SmallVector<int, 8> Mask2(4U, -1);
4322 for (unsigned i = 0; i != 4; ++i) {
4323 if (Locs[i].first == -1)
4326 unsigned Idx = (i < 2) ? 0 : 4;
4327 Idx += Locs[i].first * 2 + Locs[i].second;
4332 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4333 } else if (NumLo == 3 || NumHi == 3) {
4334 // Otherwise, we must have three elements from one vector, call it X, and
4335 // one element from the other, call it Y. First, use a shufps to build an
4336 // intermediate vector with the one element from Y and the element from X
4337 // that will be in the same half in the final destination (the indexes don't
4338 // matter). Then, use a shufps to build the final vector, taking the half
4339 // containing the element from Y from the intermediate, and the other half
4342 // Normalize it so the 3 elements come from V1.
4343 CommuteVectorShuffleMask(PermMask, VT);
4347 // Find the element from V2.
4349 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4350 int Val = PermMask[HiIndex];
4357 Mask1[0] = PermMask[HiIndex];
4359 Mask1[2] = PermMask[HiIndex^1];
4361 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4364 Mask1[0] = PermMask[0];
4365 Mask1[1] = PermMask[1];
4366 Mask1[2] = HiIndex & 1 ? 6 : 4;
4367 Mask1[3] = HiIndex & 1 ? 4 : 6;
4368 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4370 Mask1[0] = HiIndex & 1 ? 2 : 0;
4371 Mask1[1] = HiIndex & 1 ? 0 : 2;
4372 Mask1[2] = PermMask[2];
4373 Mask1[3] = PermMask[3];
4378 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4382 // Break it into (shuffle shuffle_hi, shuffle_lo).
4384 SmallVector<int,8> LoMask(4U, -1);
4385 SmallVector<int,8> HiMask(4U, -1);
4387 SmallVector<int,8> *MaskPtr = &LoMask;
4388 unsigned MaskIdx = 0;
4391 for (unsigned i = 0; i != 4; ++i) {
4398 int Idx = PermMask[i];
4400 Locs[i] = std::make_pair(-1, -1);
4401 } else if (Idx < 4) {
4402 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4403 (*MaskPtr)[LoIdx] = Idx;
4406 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4407 (*MaskPtr)[HiIdx] = Idx;
4412 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4413 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4414 SmallVector<int, 8> MaskOps;
4415 for (unsigned i = 0; i != 4; ++i) {
4416 if (Locs[i].first == -1) {
4417 MaskOps.push_back(-1);
4419 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4420 MaskOps.push_back(Idx);
4423 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4427 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4428 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4429 SDValue V1 = Op.getOperand(0);
4430 SDValue V2 = Op.getOperand(1);
4431 EVT VT = Op.getValueType();
4432 DebugLoc dl = Op.getDebugLoc();
4433 unsigned NumElems = VT.getVectorNumElements();
4434 bool isMMX = VT.getSizeInBits() == 64;
4435 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4436 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4437 bool V1IsSplat = false;
4438 bool V2IsSplat = false;
4440 if (isZeroShuffle(SVOp))
4441 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4443 // Promote splats to v4f32.
4444 if (SVOp->isSplat()) {
4445 if (isMMX || NumElems < 4)
4447 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4450 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4452 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4453 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4454 if (NewOp.getNode())
4455 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4456 LowerVECTOR_SHUFFLE(NewOp, DAG));
4457 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4458 // FIXME: Figure out a cleaner way to do this.
4459 // Try to make use of movq to zero out the top part.
4460 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4461 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4462 if (NewOp.getNode()) {
4463 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4464 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4465 DAG, Subtarget, dl);
4467 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4468 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4469 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4470 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4471 DAG, Subtarget, dl);
4475 if (X86::isPSHUFDMask(SVOp))
4478 // Check if this can be converted into a logical shift.
4479 bool isLeft = false;
4482 bool isShift = getSubtarget()->hasSSE2() &&
4483 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4484 if (isShift && ShVal.hasOneUse()) {
4485 // If the shifted value has multiple uses, it may be cheaper to use
4486 // v_set0 + movlhps or movhlps, etc.
4487 EVT EltVT = VT.getVectorElementType();
4488 ShAmt *= EltVT.getSizeInBits();
4489 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4492 if (X86::isMOVLMask(SVOp)) {
4495 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4496 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4501 // FIXME: fold these into legal mask.
4502 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4503 X86::isMOVSLDUPMask(SVOp) ||
4504 X86::isMOVHLPSMask(SVOp) ||
4505 X86::isMOVLHPSMask(SVOp) ||
4506 X86::isMOVLPMask(SVOp)))
4509 if (ShouldXformToMOVHLPS(SVOp) ||
4510 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4511 return CommuteVectorShuffle(SVOp, DAG);
4514 // No better options. Use a vshl / vsrl.
4515 EVT EltVT = VT.getVectorElementType();
4516 ShAmt *= EltVT.getSizeInBits();
4517 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4520 bool Commuted = false;
4521 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4522 // 1,1,1,1 -> v8i16 though.
4523 V1IsSplat = isSplatVector(V1.getNode());
4524 V2IsSplat = isSplatVector(V2.getNode());
4526 // Canonicalize the splat or undef, if present, to be on the RHS.
4527 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4528 Op = CommuteVectorShuffle(SVOp, DAG);
4529 SVOp = cast<ShuffleVectorSDNode>(Op);
4530 V1 = SVOp->getOperand(0);
4531 V2 = SVOp->getOperand(1);
4532 std::swap(V1IsSplat, V2IsSplat);
4533 std::swap(V1IsUndef, V2IsUndef);
4537 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4538 // Shuffling low element of v1 into undef, just return v1.
4541 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4542 // the instruction selector will not match, so get a canonical MOVL with
4543 // swapped operands to undo the commute.
4544 return getMOVL(DAG, dl, VT, V2, V1);
4547 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4548 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4549 X86::isUNPCKLMask(SVOp) ||
4550 X86::isUNPCKHMask(SVOp))
4554 // Normalize mask so all entries that point to V2 points to its first
4555 // element then try to match unpck{h|l} again. If match, return a
4556 // new vector_shuffle with the corrected mask.
4557 SDValue NewMask = NormalizeMask(SVOp, DAG);
4558 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4559 if (NSVOp != SVOp) {
4560 if (X86::isUNPCKLMask(NSVOp, true)) {
4562 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4569 // Commute is back and try unpck* again.
4570 // FIXME: this seems wrong.
4571 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4572 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4573 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4574 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4575 X86::isUNPCKLMask(NewSVOp) ||
4576 X86::isUNPCKHMask(NewSVOp))
4580 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4582 // Normalize the node to match x86 shuffle ops if needed
4583 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4584 return CommuteVectorShuffle(SVOp, DAG);
4586 // Check for legal shuffle and return?
4587 SmallVector<int, 16> PermMask;
4588 SVOp->getMask(PermMask);
4589 if (isShuffleMaskLegal(PermMask, VT))
4592 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4593 if (VT == MVT::v8i16) {
4594 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4595 if (NewOp.getNode())
4599 if (VT == MVT::v16i8) {
4600 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4601 if (NewOp.getNode())
4605 // Handle all 4 wide cases with a number of shuffles except for MMX.
4606 if (NumElems == 4 && !isMMX)
4607 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4613 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4614 SelectionDAG &DAG) {
4615 EVT VT = Op.getValueType();
4616 DebugLoc dl = Op.getDebugLoc();
4617 if (VT.getSizeInBits() == 8) {
4618 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4619 Op.getOperand(0), Op.getOperand(1));
4620 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4621 DAG.getValueType(VT));
4622 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4623 } else if (VT.getSizeInBits() == 16) {
4624 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4625 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4627 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4628 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4629 DAG.getNode(ISD::BIT_CONVERT, dl,
4633 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4634 Op.getOperand(0), Op.getOperand(1));
4635 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4636 DAG.getValueType(VT));
4637 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4638 } else if (VT == MVT::f32) {
4639 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4640 // the result back to FR32 register. It's only worth matching if the
4641 // result has a single use which is a store or a bitcast to i32. And in
4642 // the case of a store, it's not worth it if the index is a constant 0,
4643 // because a MOVSSmr can be used instead, which is smaller and faster.
4644 if (!Op.hasOneUse())
4646 SDNode *User = *Op.getNode()->use_begin();
4647 if ((User->getOpcode() != ISD::STORE ||
4648 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4649 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4650 (User->getOpcode() != ISD::BIT_CONVERT ||
4651 User->getValueType(0) != MVT::i32))
4653 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4654 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4657 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4658 } else if (VT == MVT::i32) {
4659 // ExtractPS works with constant index.
4660 if (isa<ConstantSDNode>(Op.getOperand(1)))
4668 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4669 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4672 if (Subtarget->hasSSE41()) {
4673 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4678 EVT VT = Op.getValueType();
4679 DebugLoc dl = Op.getDebugLoc();
4680 // TODO: handle v16i8.
4681 if (VT.getSizeInBits() == 16) {
4682 SDValue Vec = Op.getOperand(0);
4683 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4685 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4686 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4687 DAG.getNode(ISD::BIT_CONVERT, dl,
4690 // Transform it so it match pextrw which produces a 32-bit result.
4691 EVT EltVT = MVT::i32;
4692 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4693 Op.getOperand(0), Op.getOperand(1));
4694 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4695 DAG.getValueType(VT));
4696 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4697 } else if (VT.getSizeInBits() == 32) {
4698 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4702 // SHUFPS the element to the lowest double word, then movss.
4703 int Mask[4] = { Idx, -1, -1, -1 };
4704 EVT VVT = Op.getOperand(0).getValueType();
4705 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4706 DAG.getUNDEF(VVT), Mask);
4707 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4708 DAG.getIntPtrConstant(0));
4709 } else if (VT.getSizeInBits() == 64) {
4710 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4711 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4712 // to match extract_elt for f64.
4713 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4717 // UNPCKHPD the element to the lowest double word, then movsd.
4718 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4719 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4720 int Mask[2] = { 1, -1 };
4721 EVT VVT = Op.getOperand(0).getValueType();
4722 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4723 DAG.getUNDEF(VVT), Mask);
4724 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4725 DAG.getIntPtrConstant(0));
4732 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4733 EVT VT = Op.getValueType();
4734 EVT EltVT = VT.getVectorElementType();
4735 DebugLoc dl = Op.getDebugLoc();
4737 SDValue N0 = Op.getOperand(0);
4738 SDValue N1 = Op.getOperand(1);
4739 SDValue N2 = Op.getOperand(2);
4741 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4742 isa<ConstantSDNode>(N2)) {
4743 unsigned Opc = (EltVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4745 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4747 if (N1.getValueType() != MVT::i32)
4748 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4749 if (N2.getValueType() != MVT::i32)
4750 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4751 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4752 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4753 // Bits [7:6] of the constant are the source select. This will always be
4754 // zero here. The DAG Combiner may combine an extract_elt index into these
4755 // bits. For example (insert (extract, 3), 2) could be matched by putting
4756 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4757 // Bits [5:4] of the constant are the destination select. This is the
4758 // value of the incoming immediate.
4759 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4760 // combine either bitwise AND or insert of float 0.0 to set these bits.
4761 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4762 // Create this as a scalar to vector..
4763 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4764 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4765 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4766 // PINSR* works with constant index.
4773 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4774 EVT VT = Op.getValueType();
4775 EVT EltVT = VT.getVectorElementType();
4777 if (Subtarget->hasSSE41())
4778 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4780 if (EltVT == MVT::i8)
4783 DebugLoc dl = Op.getDebugLoc();
4784 SDValue N0 = Op.getOperand(0);
4785 SDValue N1 = Op.getOperand(1);
4786 SDValue N2 = Op.getOperand(2);
4788 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4789 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4790 // as its second argument.
4791 if (N1.getValueType() != MVT::i32)
4792 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4793 if (N2.getValueType() != MVT::i32)
4794 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4795 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4801 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4802 DebugLoc dl = Op.getDebugLoc();
4803 if (Op.getValueType() == MVT::v2f32)
4804 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4805 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4806 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4807 Op.getOperand(0))));
4809 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4810 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4812 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4813 EVT VT = MVT::v2i32;
4814 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4821 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4822 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4825 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4826 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4827 // one of the above mentioned nodes. It has to be wrapped because otherwise
4828 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4829 // be used to form addressing mode. These wrapped nodes will be selected
4832 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4833 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4835 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4837 unsigned char OpFlag = 0;
4838 unsigned WrapperKind = X86ISD::Wrapper;
4839 CodeModel::Model M = getTargetMachine().getCodeModel();
4841 if (Subtarget->isPICStyleRIPRel() &&
4842 (M == CodeModel::Small || M == CodeModel::Kernel))
4843 WrapperKind = X86ISD::WrapperRIP;
4844 else if (Subtarget->isPICStyleGOT())
4845 OpFlag = X86II::MO_GOTOFF;
4846 else if (Subtarget->isPICStyleStubPIC())
4847 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4849 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4851 CP->getOffset(), OpFlag);
4852 DebugLoc DL = CP->getDebugLoc();
4853 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4854 // With PIC, the address is actually $g + Offset.
4856 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4857 DAG.getNode(X86ISD::GlobalBaseReg,
4858 DebugLoc::getUnknownLoc(), getPointerTy()),
4865 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4866 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4868 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4870 unsigned char OpFlag = 0;
4871 unsigned WrapperKind = X86ISD::Wrapper;
4872 CodeModel::Model M = getTargetMachine().getCodeModel();
4874 if (Subtarget->isPICStyleRIPRel() &&
4875 (M == CodeModel::Small || M == CodeModel::Kernel))
4876 WrapperKind = X86ISD::WrapperRIP;
4877 else if (Subtarget->isPICStyleGOT())
4878 OpFlag = X86II::MO_GOTOFF;
4879 else if (Subtarget->isPICStyleStubPIC())
4880 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4882 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4884 DebugLoc DL = JT->getDebugLoc();
4885 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4887 // With PIC, the address is actually $g + Offset.
4889 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4890 DAG.getNode(X86ISD::GlobalBaseReg,
4891 DebugLoc::getUnknownLoc(), getPointerTy()),
4899 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4900 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4902 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4904 unsigned char OpFlag = 0;
4905 unsigned WrapperKind = X86ISD::Wrapper;
4906 CodeModel::Model M = getTargetMachine().getCodeModel();
4908 if (Subtarget->isPICStyleRIPRel() &&
4909 (M == CodeModel::Small || M == CodeModel::Kernel))
4910 WrapperKind = X86ISD::WrapperRIP;
4911 else if (Subtarget->isPICStyleGOT())
4912 OpFlag = X86II::MO_GOTOFF;
4913 else if (Subtarget->isPICStyleStubPIC())
4914 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4916 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4918 DebugLoc DL = Op.getDebugLoc();
4919 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4922 // With PIC, the address is actually $g + Offset.
4923 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4924 !Subtarget->is64Bit()) {
4925 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4926 DAG.getNode(X86ISD::GlobalBaseReg,
4927 DebugLoc::getUnknownLoc(),
4936 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
4937 // Create the TargetBlockAddressAddress node.
4938 unsigned char OpFlags =
4939 Subtarget->ClassifyBlockAddressReference();
4940 CodeModel::Model M = getTargetMachine().getCodeModel();
4941 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
4942 DebugLoc dl = Op.getDebugLoc();
4943 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
4944 /*isTarget=*/true, OpFlags);
4946 if (Subtarget->isPICStyleRIPRel() &&
4947 (M == CodeModel::Small || M == CodeModel::Kernel))
4948 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4950 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4952 // With PIC, the address is actually $g + Offset.
4953 if (isGlobalRelativeToPICBase(OpFlags)) {
4954 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4955 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4963 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4965 SelectionDAG &DAG) const {
4966 // Create the TargetGlobalAddress node, folding in the constant
4967 // offset if it is legal.
4968 unsigned char OpFlags =
4969 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
4970 CodeModel::Model M = getTargetMachine().getCodeModel();
4972 if (OpFlags == X86II::MO_NO_FLAG &&
4973 X86::isOffsetSuitableForCodeModel(Offset, M)) {
4974 // A direct static reference to a global.
4975 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4978 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
4981 if (Subtarget->isPICStyleRIPRel() &&
4982 (M == CodeModel::Small || M == CodeModel::Kernel))
4983 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4985 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4987 // With PIC, the address is actually $g + Offset.
4988 if (isGlobalRelativeToPICBase(OpFlags)) {
4989 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4990 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4994 // For globals that require a load from a stub to get the address, emit the
4996 if (isGlobalStubReference(OpFlags))
4997 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4998 PseudoSourceValue::getGOT(), 0);
5000 // If there was a non-zero offset that we didn't fold, create an explicit
5003 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5004 DAG.getConstant(Offset, getPointerTy()));
5010 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
5011 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5012 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5013 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5017 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5018 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5019 unsigned char OperandFlags) {
5020 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5021 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5022 DebugLoc dl = GA->getDebugLoc();
5023 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5024 GA->getValueType(0),
5028 SDValue Ops[] = { Chain, TGA, *InFlag };
5029 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5031 SDValue Ops[] = { Chain, TGA };
5032 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5035 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5036 MFI->setHasCalls(true);
5038 SDValue Flag = Chain.getValue(1);
5039 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5042 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5044 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5047 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5048 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5049 DAG.getNode(X86ISD::GlobalBaseReg,
5050 DebugLoc::getUnknownLoc(),
5052 InFlag = Chain.getValue(1);
5054 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5057 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5059 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5061 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5062 X86::RAX, X86II::MO_TLSGD);
5065 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5066 // "local exec" model.
5067 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5068 const EVT PtrVT, TLSModel::Model model,
5070 DebugLoc dl = GA->getDebugLoc();
5071 // Get the Thread Pointer
5072 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5073 DebugLoc::getUnknownLoc(), PtrVT,
5074 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5077 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5080 unsigned char OperandFlags = 0;
5081 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5083 unsigned WrapperKind = X86ISD::Wrapper;
5084 if (model == TLSModel::LocalExec) {
5085 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5086 } else if (is64Bit) {
5087 assert(model == TLSModel::InitialExec);
5088 OperandFlags = X86II::MO_GOTTPOFF;
5089 WrapperKind = X86ISD::WrapperRIP;
5091 assert(model == TLSModel::InitialExec);
5092 OperandFlags = X86II::MO_INDNTPOFF;
5095 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5097 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5098 GA->getOffset(), OperandFlags);
5099 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5101 if (model == TLSModel::InitialExec)
5102 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5103 PseudoSourceValue::getGOT(), 0);
5105 // The address of the thread local variable is the add of the thread
5106 // pointer with the offset of the variable.
5107 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5111 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
5112 // TODO: implement the "local dynamic" model
5113 // TODO: implement the "initial exec"model for pic executables
5114 assert(Subtarget->isTargetELF() &&
5115 "TLS not implemented for non-ELF targets");
5116 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5117 const GlobalValue *GV = GA->getGlobal();
5119 // If GV is an alias then use the aliasee for determining
5120 // thread-localness.
5121 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5122 GV = GA->resolveAliasedGlobal(false);
5124 TLSModel::Model model = getTLSModel(GV,
5125 getTargetMachine().getRelocationModel());
5128 case TLSModel::GeneralDynamic:
5129 case TLSModel::LocalDynamic: // not implemented
5130 if (Subtarget->is64Bit())
5131 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5132 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5134 case TLSModel::InitialExec:
5135 case TLSModel::LocalExec:
5136 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5137 Subtarget->is64Bit());
5140 llvm_unreachable("Unreachable");
5145 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5146 /// take a 2 x i32 value to shift plus a shift amount.
5147 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
5148 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5149 EVT VT = Op.getValueType();
5150 unsigned VTBits = VT.getSizeInBits();
5151 DebugLoc dl = Op.getDebugLoc();
5152 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5153 SDValue ShOpLo = Op.getOperand(0);
5154 SDValue ShOpHi = Op.getOperand(1);
5155 SDValue ShAmt = Op.getOperand(2);
5156 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5157 DAG.getConstant(VTBits - 1, MVT::i8))
5158 : DAG.getConstant(0, VT);
5161 if (Op.getOpcode() == ISD::SHL_PARTS) {
5162 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5163 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5165 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5166 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5169 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5170 DAG.getConstant(VTBits, MVT::i8));
5171 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
5172 AndNode, DAG.getConstant(0, MVT::i8));
5175 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5176 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5177 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5179 if (Op.getOpcode() == ISD::SHL_PARTS) {
5180 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5181 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5183 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5184 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5187 SDValue Ops[2] = { Lo, Hi };
5188 return DAG.getMergeValues(Ops, 2, dl);
5191 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5192 EVT SrcVT = Op.getOperand(0).getValueType();
5194 if (SrcVT.isVector()) {
5195 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5201 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5202 "Unknown SINT_TO_FP to lower!");
5204 // These are really Legal; return the operand so the caller accepts it as
5206 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5208 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5209 Subtarget->is64Bit()) {
5213 DebugLoc dl = Op.getDebugLoc();
5214 unsigned Size = SrcVT.getSizeInBits()/8;
5215 MachineFunction &MF = DAG.getMachineFunction();
5216 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5217 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5218 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5220 PseudoSourceValue::getFixedStack(SSFI), 0);
5221 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5224 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5226 SelectionDAG &DAG) {
5228 DebugLoc dl = Op.getDebugLoc();
5230 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5232 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5234 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5235 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5236 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5237 Tys, Ops, array_lengthof(Ops));
5240 Chain = Result.getValue(1);
5241 SDValue InFlag = Result.getValue(2);
5243 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5244 // shouldn't be necessary except that RFP cannot be live across
5245 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5246 MachineFunction &MF = DAG.getMachineFunction();
5247 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5248 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5249 Tys = DAG.getVTList(MVT::Other);
5251 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5253 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5254 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5255 PseudoSourceValue::getFixedStack(SSFI), 0);
5261 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5262 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5263 // This algorithm is not obvious. Here it is in C code, more or less:
5265 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5266 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5267 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5269 // Copy ints to xmm registers.
5270 __m128i xh = _mm_cvtsi32_si128( hi );
5271 __m128i xl = _mm_cvtsi32_si128( lo );
5273 // Combine into low half of a single xmm register.
5274 __m128i x = _mm_unpacklo_epi32( xh, xl );
5278 // Merge in appropriate exponents to give the integer bits the right
5280 x = _mm_unpacklo_epi32( x, exp );
5282 // Subtract away the biases to deal with the IEEE-754 double precision
5284 d = _mm_sub_pd( (__m128d) x, bias );
5286 // All conversions up to here are exact. The correctly rounded result is
5287 // calculated using the current rounding mode using the following
5289 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5290 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5291 // store doesn't really need to be here (except
5292 // maybe to zero the other double)
5297 DebugLoc dl = Op.getDebugLoc();
5298 LLVMContext *Context = DAG.getContext();
5300 // Build some magic constants.
5301 std::vector<Constant*> CV0;
5302 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5303 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5304 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5305 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5306 Constant *C0 = ConstantVector::get(CV0);
5307 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5309 std::vector<Constant*> CV1;
5311 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5313 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5314 Constant *C1 = ConstantVector::get(CV1);
5315 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5317 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5318 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5320 DAG.getIntPtrConstant(1)));
5321 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5322 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5324 DAG.getIntPtrConstant(0)));
5325 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5326 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5327 PseudoSourceValue::getConstantPool(), 0,
5329 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5330 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5331 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5332 PseudoSourceValue::getConstantPool(), 0,
5334 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5336 // Add the halves; easiest way is to swap them into another reg first.
5337 int ShufMask[2] = { 1, -1 };
5338 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5339 DAG.getUNDEF(MVT::v2f64), ShufMask);
5340 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5341 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5342 DAG.getIntPtrConstant(0));
5345 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5346 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5347 DebugLoc dl = Op.getDebugLoc();
5348 // FP constant to bias correct the final result.
5349 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5352 // Load the 32-bit value into an XMM register.
5353 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5354 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5356 DAG.getIntPtrConstant(0)));
5358 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5359 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5360 DAG.getIntPtrConstant(0));
5362 // Or the load with the bias.
5363 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5364 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5365 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5367 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5368 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5369 MVT::v2f64, Bias)));
5370 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5371 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5372 DAG.getIntPtrConstant(0));
5374 // Subtract the bias.
5375 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5377 // Handle final rounding.
5378 EVT DestVT = Op.getValueType();
5380 if (DestVT.bitsLT(MVT::f64)) {
5381 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5382 DAG.getIntPtrConstant(0));
5383 } else if (DestVT.bitsGT(MVT::f64)) {
5384 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5387 // Handle final rounding.
5391 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5392 SDValue N0 = Op.getOperand(0);
5393 DebugLoc dl = Op.getDebugLoc();
5395 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5396 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5397 // the optimization here.
5398 if (DAG.SignBitIsZero(N0))
5399 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5401 EVT SrcVT = N0.getValueType();
5402 if (SrcVT == MVT::i64) {
5403 // We only handle SSE2 f64 target here; caller can expand the rest.
5404 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5407 return LowerUINT_TO_FP_i64(Op, DAG);
5408 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5409 return LowerUINT_TO_FP_i32(Op, DAG);
5412 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5414 // Make a 64-bit buffer, and use it to build an FILD.
5415 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5416 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5417 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5418 getPointerTy(), StackSlot, WordOff);
5419 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5420 StackSlot, NULL, 0);
5421 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5422 OffsetSlot, NULL, 0);
5423 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5426 std::pair<SDValue,SDValue> X86TargetLowering::
5427 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5428 DebugLoc dl = Op.getDebugLoc();
5430 EVT DstTy = Op.getValueType();
5433 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5437 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5438 DstTy.getSimpleVT() >= MVT::i16 &&
5439 "Unknown FP_TO_SINT to lower!");
5441 // These are really Legal.
5442 if (DstTy == MVT::i32 &&
5443 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5444 return std::make_pair(SDValue(), SDValue());
5445 if (Subtarget->is64Bit() &&
5446 DstTy == MVT::i64 &&
5447 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5448 return std::make_pair(SDValue(), SDValue());
5450 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5452 MachineFunction &MF = DAG.getMachineFunction();
5453 unsigned MemSize = DstTy.getSizeInBits()/8;
5454 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5455 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5458 switch (DstTy.getSimpleVT().SimpleTy) {
5459 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5460 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5461 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5462 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5465 SDValue Chain = DAG.getEntryNode();
5466 SDValue Value = Op.getOperand(0);
5467 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5468 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5469 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5470 PseudoSourceValue::getFixedStack(SSFI), 0);
5471 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5473 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5475 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5476 Chain = Value.getValue(1);
5477 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5478 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5481 // Build the FP_TO_INT*_IN_MEM
5482 SDValue Ops[] = { Chain, Value, StackSlot };
5483 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5485 return std::make_pair(FIST, StackSlot);
5488 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5489 if (Op.getValueType().isVector()) {
5490 if (Op.getValueType() == MVT::v2i32 &&
5491 Op.getOperand(0).getValueType() == MVT::v2f64) {
5497 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5498 SDValue FIST = Vals.first, StackSlot = Vals.second;
5499 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5500 if (FIST.getNode() == 0) return Op;
5503 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5504 FIST, StackSlot, NULL, 0);
5507 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5508 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5509 SDValue FIST = Vals.first, StackSlot = Vals.second;
5510 assert(FIST.getNode() && "Unexpected failure");
5513 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5514 FIST, StackSlot, NULL, 0);
5517 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5518 LLVMContext *Context = DAG.getContext();
5519 DebugLoc dl = Op.getDebugLoc();
5520 EVT VT = Op.getValueType();
5523 EltVT = VT.getVectorElementType();
5524 std::vector<Constant*> CV;
5525 if (EltVT == MVT::f64) {
5526 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5530 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5536 Constant *C = ConstantVector::get(CV);
5537 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5538 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5539 PseudoSourceValue::getConstantPool(), 0,
5541 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5544 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5545 LLVMContext *Context = DAG.getContext();
5546 DebugLoc dl = Op.getDebugLoc();
5547 EVT VT = Op.getValueType();
5550 EltVT = VT.getVectorElementType();
5551 std::vector<Constant*> CV;
5552 if (EltVT == MVT::f64) {
5553 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5557 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5563 Constant *C = ConstantVector::get(CV);
5564 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5565 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5566 PseudoSourceValue::getConstantPool(), 0,
5568 if (VT.isVector()) {
5569 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5570 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5571 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5573 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5575 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5579 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5580 LLVMContext *Context = DAG.getContext();
5581 SDValue Op0 = Op.getOperand(0);
5582 SDValue Op1 = Op.getOperand(1);
5583 DebugLoc dl = Op.getDebugLoc();
5584 EVT VT = Op.getValueType();
5585 EVT SrcVT = Op1.getValueType();
5587 // If second operand is smaller, extend it first.
5588 if (SrcVT.bitsLT(VT)) {
5589 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5592 // And if it is bigger, shrink it first.
5593 if (SrcVT.bitsGT(VT)) {
5594 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5598 // At this point the operands and the result should have the same
5599 // type, and that won't be f80 since that is not custom lowered.
5601 // First get the sign bit of second operand.
5602 std::vector<Constant*> CV;
5603 if (SrcVT == MVT::f64) {
5604 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5605 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5607 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5608 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5609 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5610 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5612 Constant *C = ConstantVector::get(CV);
5613 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5614 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5615 PseudoSourceValue::getConstantPool(), 0,
5617 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5619 // Shift sign bit right or left if the two operands have different types.
5620 if (SrcVT.bitsGT(VT)) {
5621 // Op0 is MVT::f32, Op1 is MVT::f64.
5622 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5623 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5624 DAG.getConstant(32, MVT::i32));
5625 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5626 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5627 DAG.getIntPtrConstant(0));
5630 // Clear first operand sign bit.
5632 if (VT == MVT::f64) {
5633 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5634 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5636 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5637 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5638 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5639 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5641 C = ConstantVector::get(CV);
5642 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5643 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5644 PseudoSourceValue::getConstantPool(), 0,
5646 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5648 // Or the value with the sign bit.
5649 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5652 /// Emit nodes that will be selected as "test Op0,Op0", or something
5654 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5655 SelectionDAG &DAG) {
5656 DebugLoc dl = Op.getDebugLoc();
5658 // CF and OF aren't always set the way we want. Determine which
5659 // of these we need.
5660 bool NeedCF = false;
5661 bool NeedOF = false;
5663 case X86::COND_A: case X86::COND_AE:
5664 case X86::COND_B: case X86::COND_BE:
5667 case X86::COND_G: case X86::COND_GE:
5668 case X86::COND_L: case X86::COND_LE:
5669 case X86::COND_O: case X86::COND_NO:
5675 // See if we can use the EFLAGS value from the operand instead of
5676 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5677 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5678 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5679 unsigned Opcode = 0;
5680 unsigned NumOperands = 0;
5681 switch (Op.getNode()->getOpcode()) {
5683 // Due to an isel shortcoming, be conservative if this add is likely to
5684 // be selected as part of a load-modify-store instruction. When the root
5685 // node in a match is a store, isel doesn't know how to remap non-chain
5686 // non-flag uses of other nodes in the match, such as the ADD in this
5687 // case. This leads to the ADD being left around and reselected, with
5688 // the result being two adds in the output.
5689 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5690 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5691 if (UI->getOpcode() == ISD::STORE)
5693 if (ConstantSDNode *C =
5694 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5695 // An add of one will be selected as an INC.
5696 if (C->getAPIntValue() == 1) {
5697 Opcode = X86ISD::INC;
5701 // An add of negative one (subtract of one) will be selected as a DEC.
5702 if (C->getAPIntValue().isAllOnesValue()) {
5703 Opcode = X86ISD::DEC;
5708 // Otherwise use a regular EFLAGS-setting add.
5709 Opcode = X86ISD::ADD;
5713 // If the primary and result isn't used, don't bother using X86ISD::AND,
5714 // because a TEST instruction will be better.
5715 bool NonFlagUse = false;
5716 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5717 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
5719 unsigned UOpNo = UI.getOperandNo();
5720 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
5721 // Look pass truncate.
5722 UOpNo = User->use_begin().getOperandNo();
5723 User = *User->use_begin();
5725 if (User->getOpcode() != ISD::BRCOND &&
5726 User->getOpcode() != ISD::SETCC &&
5727 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
5739 // Due to the ISEL shortcoming noted above, be conservative if this op is
5740 // likely to be selected as part of a load-modify-store instruction.
5741 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5742 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5743 if (UI->getOpcode() == ISD::STORE)
5745 // Otherwise use a regular EFLAGS-setting instruction.
5746 switch (Op.getNode()->getOpcode()) {
5747 case ISD::SUB: Opcode = X86ISD::SUB; break;
5748 case ISD::OR: Opcode = X86ISD::OR; break;
5749 case ISD::XOR: Opcode = X86ISD::XOR; break;
5750 case ISD::AND: Opcode = X86ISD::AND; break;
5751 default: llvm_unreachable("unexpected operator!");
5762 return SDValue(Op.getNode(), 1);
5768 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5769 SmallVector<SDValue, 4> Ops;
5770 for (unsigned i = 0; i != NumOperands; ++i)
5771 Ops.push_back(Op.getOperand(i));
5772 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5773 DAG.ReplaceAllUsesWith(Op, New);
5774 return SDValue(New.getNode(), 1);
5778 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5779 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5780 DAG.getConstant(0, Op.getValueType()));
5783 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5785 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5786 SelectionDAG &DAG) {
5787 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5788 if (C->getAPIntValue() == 0)
5789 return EmitTest(Op0, X86CC, DAG);
5791 DebugLoc dl = Op0.getDebugLoc();
5792 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5795 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
5796 /// if it's possible.
5797 static SDValue LowerToBT(SDValue Op0, ISD::CondCode CC,
5798 DebugLoc dl, SelectionDAG &DAG) {
5800 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5801 if (ConstantSDNode *Op010C =
5802 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5803 if (Op010C->getZExtValue() == 1) {
5804 LHS = Op0.getOperand(0);
5805 RHS = Op0.getOperand(1).getOperand(1);
5807 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5808 if (ConstantSDNode *Op000C =
5809 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5810 if (Op000C->getZExtValue() == 1) {
5811 LHS = Op0.getOperand(1);
5812 RHS = Op0.getOperand(0).getOperand(1);
5814 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5815 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5816 SDValue AndLHS = Op0.getOperand(0);
5817 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5818 LHS = AndLHS.getOperand(0);
5819 RHS = AndLHS.getOperand(1);
5823 if (LHS.getNode()) {
5824 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5825 // instruction. Since the shift amount is in-range-or-undefined, we know
5826 // that doing a bittest on the i16 value is ok. We extend to i32 because
5827 // the encoding for the i16 version is larger than the i32 version.
5828 if (LHS.getValueType() == MVT::i8)
5829 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5831 // If the operand types disagree, extend the shift amount to match. Since
5832 // BT ignores high bits (like shifts) we can use anyextend.
5833 if (LHS.getValueType() != RHS.getValueType())
5834 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5836 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5837 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5838 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5839 DAG.getConstant(Cond, MVT::i8), BT);
5845 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5846 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5847 SDValue Op0 = Op.getOperand(0);
5848 SDValue Op1 = Op.getOperand(1);
5849 DebugLoc dl = Op.getDebugLoc();
5850 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5852 // Optimize to BT if possible.
5853 // Lower (X & (1 << N)) == 0 to BT(X, N).
5854 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5855 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5856 if (Op0.getOpcode() == ISD::AND &&
5858 Op1.getOpcode() == ISD::Constant &&
5859 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5860 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5861 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
5862 if (NewSetCC.getNode())
5866 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5867 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5868 if (X86CC == X86::COND_INVALID)
5871 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5873 // Use sbb x, x to materialize carry bit into a GPR.
5874 if (X86CC == X86::COND_B)
5875 return DAG.getNode(ISD::AND, dl, MVT::i8,
5876 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
5877 DAG.getConstant(X86CC, MVT::i8), Cond),
5878 DAG.getConstant(1, MVT::i8));
5880 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5881 DAG.getConstant(X86CC, MVT::i8), Cond);
5884 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5886 SDValue Op0 = Op.getOperand(0);
5887 SDValue Op1 = Op.getOperand(1);
5888 SDValue CC = Op.getOperand(2);
5889 EVT VT = Op.getValueType();
5890 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5891 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5892 DebugLoc dl = Op.getDebugLoc();
5896 EVT VT0 = Op0.getValueType();
5897 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5898 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5901 switch (SetCCOpcode) {
5904 case ISD::SETEQ: SSECC = 0; break;
5906 case ISD::SETGT: Swap = true; // Fallthrough
5908 case ISD::SETOLT: SSECC = 1; break;
5910 case ISD::SETGE: Swap = true; // Fallthrough
5912 case ISD::SETOLE: SSECC = 2; break;
5913 case ISD::SETUO: SSECC = 3; break;
5915 case ISD::SETNE: SSECC = 4; break;
5916 case ISD::SETULE: Swap = true;
5917 case ISD::SETUGE: SSECC = 5; break;
5918 case ISD::SETULT: Swap = true;
5919 case ISD::SETUGT: SSECC = 6; break;
5920 case ISD::SETO: SSECC = 7; break;
5923 std::swap(Op0, Op1);
5925 // In the two special cases we can't handle, emit two comparisons.
5927 if (SetCCOpcode == ISD::SETUEQ) {
5929 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5930 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5931 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5933 else if (SetCCOpcode == ISD::SETONE) {
5935 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5936 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5937 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5939 llvm_unreachable("Illegal FP comparison");
5941 // Handle all other FP comparisons here.
5942 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5945 // We are handling one of the integer comparisons here. Since SSE only has
5946 // GT and EQ comparisons for integer, swapping operands and multiple
5947 // operations may be required for some comparisons.
5948 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5949 bool Swap = false, Invert = false, FlipSigns = false;
5951 switch (VT.getSimpleVT().SimpleTy) {
5954 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5956 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5958 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5959 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5962 switch (SetCCOpcode) {
5964 case ISD::SETNE: Invert = true;
5965 case ISD::SETEQ: Opc = EQOpc; break;
5966 case ISD::SETLT: Swap = true;
5967 case ISD::SETGT: Opc = GTOpc; break;
5968 case ISD::SETGE: Swap = true;
5969 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5970 case ISD::SETULT: Swap = true;
5971 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5972 case ISD::SETUGE: Swap = true;
5973 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5976 std::swap(Op0, Op1);
5978 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5979 // bits of the inputs before performing those operations.
5981 EVT EltVT = VT.getVectorElementType();
5982 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5984 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5985 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5987 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5988 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5991 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5993 // If the logical-not of the result is required, perform that now.
5995 Result = DAG.getNOT(dl, Result, VT);
6000 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6001 static bool isX86LogicalCmp(SDValue Op) {
6002 unsigned Opc = Op.getNode()->getOpcode();
6003 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6005 if (Op.getResNo() == 1 &&
6006 (Opc == X86ISD::ADD ||
6007 Opc == X86ISD::SUB ||
6008 Opc == X86ISD::SMUL ||
6009 Opc == X86ISD::UMUL ||
6010 Opc == X86ISD::INC ||
6011 Opc == X86ISD::DEC ||
6012 Opc == X86ISD::OR ||
6013 Opc == X86ISD::XOR ||
6014 Opc == X86ISD::AND))
6020 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
6021 bool addTest = true;
6022 SDValue Cond = Op.getOperand(0);
6023 DebugLoc dl = Op.getDebugLoc();
6026 if (Cond.getOpcode() == ISD::SETCC) {
6027 SDValue NewCond = LowerSETCC(Cond, DAG);
6028 if (NewCond.getNode())
6032 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6033 SDValue Op1 = Op.getOperand(1);
6034 SDValue Op2 = Op.getOperand(2);
6035 if (Cond.getOpcode() == X86ISD::SETCC &&
6036 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6037 SDValue Cmp = Cond.getOperand(1);
6038 if (Cmp.getOpcode() == X86ISD::CMP) {
6039 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6040 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6041 ConstantSDNode *RHSC =
6042 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6043 if (N1C && N1C->isAllOnesValue() &&
6044 N2C && N2C->isNullValue() &&
6045 RHSC && RHSC->isNullValue()) {
6046 SDValue CmpOp0 = Cmp.getOperand(0);
6047 Cmp = DAG.getNode(X86ISD::CMP, dl, CmpOp0.getValueType(),
6048 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6049 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6050 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6055 // Look pass (and (setcc_carry (cmp ...)), 1).
6056 if (Cond.getOpcode() == ISD::AND &&
6057 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6058 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6059 if (C && C->getAPIntValue() == 1)
6060 Cond = Cond.getOperand(0);
6063 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6064 // setting operand in place of the X86ISD::SETCC.
6065 if (Cond.getOpcode() == X86ISD::SETCC ||
6066 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6067 CC = Cond.getOperand(0);
6069 SDValue Cmp = Cond.getOperand(1);
6070 unsigned Opc = Cmp.getOpcode();
6071 EVT VT = Op.getValueType();
6073 bool IllegalFPCMov = false;
6074 if (VT.isFloatingPoint() && !VT.isVector() &&
6075 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6076 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6078 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6079 Opc == X86ISD::BT) { // FIXME
6086 // Look pass the truncate.
6087 if (Cond.getOpcode() == ISD::TRUNCATE)
6088 Cond = Cond.getOperand(0);
6090 // We know the result of AND is compared against zero. Try to match
6092 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6093 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6094 if (NewSetCC.getNode()) {
6095 CC = NewSetCC.getOperand(0);
6096 Cond = NewSetCC.getOperand(1);
6103 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6104 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6107 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6108 // condition is true.
6109 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6110 SDValue Ops[] = { Op2, Op1, CC, Cond };
6111 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6114 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6115 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6116 // from the AND / OR.
6117 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6118 Opc = Op.getOpcode();
6119 if (Opc != ISD::OR && Opc != ISD::AND)
6121 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6122 Op.getOperand(0).hasOneUse() &&
6123 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6124 Op.getOperand(1).hasOneUse());
6127 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6128 // 1 and that the SETCC node has a single use.
6129 static bool isXor1OfSetCC(SDValue Op) {
6130 if (Op.getOpcode() != ISD::XOR)
6132 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6133 if (N1C && N1C->getAPIntValue() == 1) {
6134 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6135 Op.getOperand(0).hasOneUse();
6140 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
6141 bool addTest = true;
6142 SDValue Chain = Op.getOperand(0);
6143 SDValue Cond = Op.getOperand(1);
6144 SDValue Dest = Op.getOperand(2);
6145 DebugLoc dl = Op.getDebugLoc();
6148 if (Cond.getOpcode() == ISD::SETCC) {
6149 SDValue NewCond = LowerSETCC(Cond, DAG);
6150 if (NewCond.getNode())
6154 // FIXME: LowerXALUO doesn't handle these!!
6155 else if (Cond.getOpcode() == X86ISD::ADD ||
6156 Cond.getOpcode() == X86ISD::SUB ||
6157 Cond.getOpcode() == X86ISD::SMUL ||
6158 Cond.getOpcode() == X86ISD::UMUL)
6159 Cond = LowerXALUO(Cond, DAG);
6162 // Look pass (and (setcc_carry (cmp ...)), 1).
6163 if (Cond.getOpcode() == ISD::AND &&
6164 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6165 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6166 if (C && C->getAPIntValue() == 1)
6167 Cond = Cond.getOperand(0);
6170 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6171 // setting operand in place of the X86ISD::SETCC.
6172 if (Cond.getOpcode() == X86ISD::SETCC ||
6173 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6174 CC = Cond.getOperand(0);
6176 SDValue Cmp = Cond.getOperand(1);
6177 unsigned Opc = Cmp.getOpcode();
6178 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6179 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6183 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6187 // These can only come from an arithmetic instruction with overflow,
6188 // e.g. SADDO, UADDO.
6189 Cond = Cond.getNode()->getOperand(1);
6196 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6197 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6198 if (CondOpc == ISD::OR) {
6199 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6200 // two branches instead of an explicit OR instruction with a
6202 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6203 isX86LogicalCmp(Cmp)) {
6204 CC = Cond.getOperand(0).getOperand(0);
6205 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6206 Chain, Dest, CC, Cmp);
6207 CC = Cond.getOperand(1).getOperand(0);
6211 } else { // ISD::AND
6212 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6213 // two branches instead of an explicit AND instruction with a
6214 // separate test. However, we only do this if this block doesn't
6215 // have a fall-through edge, because this requires an explicit
6216 // jmp when the condition is false.
6217 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6218 isX86LogicalCmp(Cmp) &&
6219 Op.getNode()->hasOneUse()) {
6220 X86::CondCode CCode =
6221 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6222 CCode = X86::GetOppositeBranchCondition(CCode);
6223 CC = DAG.getConstant(CCode, MVT::i8);
6224 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6225 // Look for an unconditional branch following this conditional branch.
6226 // We need this because we need to reverse the successors in order
6227 // to implement FCMP_OEQ.
6228 if (User.getOpcode() == ISD::BR) {
6229 SDValue FalseBB = User.getOperand(1);
6231 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6232 assert(NewBR == User);
6235 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6236 Chain, Dest, CC, Cmp);
6237 X86::CondCode CCode =
6238 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6239 CCode = X86::GetOppositeBranchCondition(CCode);
6240 CC = DAG.getConstant(CCode, MVT::i8);
6246 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6247 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6248 // It should be transformed during dag combiner except when the condition
6249 // is set by a arithmetics with overflow node.
6250 X86::CondCode CCode =
6251 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6252 CCode = X86::GetOppositeBranchCondition(CCode);
6253 CC = DAG.getConstant(CCode, MVT::i8);
6254 Cond = Cond.getOperand(0).getOperand(1);
6260 // Look pass the truncate.
6261 if (Cond.getOpcode() == ISD::TRUNCATE)
6262 Cond = Cond.getOperand(0);
6264 // We know the result of AND is compared against zero. Try to match
6266 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6267 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6268 if (NewSetCC.getNode()) {
6269 CC = NewSetCC.getOperand(0);
6270 Cond = NewSetCC.getOperand(1);
6277 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6278 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6280 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6281 Chain, Dest, CC, Cond);
6285 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6286 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6287 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6288 // that the guard pages used by the OS virtual memory manager are allocated in
6289 // correct sequence.
6291 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6292 SelectionDAG &DAG) {
6293 assert(Subtarget->isTargetCygMing() &&
6294 "This should be used only on Cygwin/Mingw targets");
6295 DebugLoc dl = Op.getDebugLoc();
6298 SDValue Chain = Op.getOperand(0);
6299 SDValue Size = Op.getOperand(1);
6300 // FIXME: Ensure alignment here
6304 EVT IntPtr = getPointerTy();
6305 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6307 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
6309 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6310 Flag = Chain.getValue(1);
6312 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6313 SDValue Ops[] = { Chain,
6314 DAG.getTargetExternalSymbol("_alloca", IntPtr),
6315 DAG.getRegister(X86::EAX, IntPtr),
6316 DAG.getRegister(X86StackPtr, SPTy),
6318 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
6319 Flag = Chain.getValue(1);
6321 Chain = DAG.getCALLSEQ_END(Chain,
6322 DAG.getIntPtrConstant(0, true),
6323 DAG.getIntPtrConstant(0, true),
6326 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6328 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6329 return DAG.getMergeValues(Ops1, 2, dl);
6333 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6335 SDValue Dst, SDValue Src,
6336 SDValue Size, unsigned Align,
6338 uint64_t DstSVOff) {
6339 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6341 // If not DWORD aligned or size is more than the threshold, call the library.
6342 // The libc version is likely to be faster for these cases. It can use the
6343 // address value and run time information about the CPU.
6344 if ((Align & 3) != 0 ||
6346 ConstantSize->getZExtValue() >
6347 getSubtarget()->getMaxInlineSizeThreshold()) {
6348 SDValue InFlag(0, 0);
6350 // Check to see if there is a specialized entry-point for memory zeroing.
6351 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6353 if (const char *bzeroEntry = V &&
6354 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6355 EVT IntPtr = getPointerTy();
6356 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6357 TargetLowering::ArgListTy Args;
6358 TargetLowering::ArgListEntry Entry;
6360 Entry.Ty = IntPtrTy;
6361 Args.push_back(Entry);
6363 Args.push_back(Entry);
6364 std::pair<SDValue,SDValue> CallResult =
6365 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6366 false, false, false, false,
6367 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6368 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl,
6369 DAG.GetOrdering(Chain.getNode()));
6370 return CallResult.second;
6373 // Otherwise have the target-independent code call memset.
6377 uint64_t SizeVal = ConstantSize->getZExtValue();
6378 SDValue InFlag(0, 0);
6381 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6382 unsigned BytesLeft = 0;
6383 bool TwoRepStos = false;
6386 uint64_t Val = ValC->getZExtValue() & 255;
6388 // If the value is a constant, then we can potentially use larger sets.
6389 switch (Align & 3) {
6390 case 2: // WORD aligned
6393 Val = (Val << 8) | Val;
6395 case 0: // DWORD aligned
6398 Val = (Val << 8) | Val;
6399 Val = (Val << 16) | Val;
6400 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6403 Val = (Val << 32) | Val;
6406 default: // Byte aligned
6409 Count = DAG.getIntPtrConstant(SizeVal);
6413 if (AVT.bitsGT(MVT::i8)) {
6414 unsigned UBytes = AVT.getSizeInBits() / 8;
6415 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6416 BytesLeft = SizeVal % UBytes;
6419 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6421 InFlag = Chain.getValue(1);
6424 Count = DAG.getIntPtrConstant(SizeVal);
6425 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6426 InFlag = Chain.getValue(1);
6429 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6432 InFlag = Chain.getValue(1);
6433 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6436 InFlag = Chain.getValue(1);
6438 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6439 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6440 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6443 InFlag = Chain.getValue(1);
6445 EVT CVT = Count.getValueType();
6446 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6447 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6448 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6451 InFlag = Chain.getValue(1);
6452 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6453 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
6454 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6455 } else if (BytesLeft) {
6456 // Handle the last 1 - 7 bytes.
6457 unsigned Offset = SizeVal - BytesLeft;
6458 EVT AddrVT = Dst.getValueType();
6459 EVT SizeVT = Size.getValueType();
6461 Chain = DAG.getMemset(Chain, dl,
6462 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6463 DAG.getConstant(Offset, AddrVT)),
6465 DAG.getConstant(BytesLeft, SizeVT),
6466 Align, DstSV, DstSVOff + Offset);
6469 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6474 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6475 SDValue Chain, SDValue Dst, SDValue Src,
6476 SDValue Size, unsigned Align,
6478 const Value *DstSV, uint64_t DstSVOff,
6479 const Value *SrcSV, uint64_t SrcSVOff) {
6480 // This requires the copy size to be a constant, preferrably
6481 // within a subtarget-specific limit.
6482 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6485 uint64_t SizeVal = ConstantSize->getZExtValue();
6486 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6489 /// If not DWORD aligned, call the library.
6490 if ((Align & 3) != 0)
6495 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6498 unsigned UBytes = AVT.getSizeInBits() / 8;
6499 unsigned CountVal = SizeVal / UBytes;
6500 SDValue Count = DAG.getIntPtrConstant(CountVal);
6501 unsigned BytesLeft = SizeVal % UBytes;
6503 SDValue InFlag(0, 0);
6504 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6507 InFlag = Chain.getValue(1);
6508 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6511 InFlag = Chain.getValue(1);
6512 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6515 InFlag = Chain.getValue(1);
6517 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6518 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6519 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
6520 array_lengthof(Ops));
6522 SmallVector<SDValue, 4> Results;
6523 Results.push_back(RepMovs);
6525 // Handle the last 1 - 7 bytes.
6526 unsigned Offset = SizeVal - BytesLeft;
6527 EVT DstVT = Dst.getValueType();
6528 EVT SrcVT = Src.getValueType();
6529 EVT SizeVT = Size.getValueType();
6530 Results.push_back(DAG.getMemcpy(Chain, dl,
6531 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6532 DAG.getConstant(Offset, DstVT)),
6533 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6534 DAG.getConstant(Offset, SrcVT)),
6535 DAG.getConstant(BytesLeft, SizeVT),
6536 Align, AlwaysInline,
6537 DstSV, DstSVOff + Offset,
6538 SrcSV, SrcSVOff + Offset));
6541 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6542 &Results[0], Results.size());
6545 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6546 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6547 DebugLoc dl = Op.getDebugLoc();
6549 if (!Subtarget->is64Bit()) {
6550 // vastart just stores the address of the VarArgsFrameIndex slot into the
6551 // memory location argument.
6552 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6553 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6557 // gp_offset (0 - 6 * 8)
6558 // fp_offset (48 - 48 + 8 * 16)
6559 // overflow_arg_area (point to parameters coming in memory).
6561 SmallVector<SDValue, 8> MemOps;
6562 SDValue FIN = Op.getOperand(1);
6564 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6565 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6567 MemOps.push_back(Store);
6570 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6571 FIN, DAG.getIntPtrConstant(4));
6572 Store = DAG.getStore(Op.getOperand(0), dl,
6573 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6575 MemOps.push_back(Store);
6577 // Store ptr to overflow_arg_area
6578 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6579 FIN, DAG.getIntPtrConstant(4));
6580 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6581 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6582 MemOps.push_back(Store);
6584 // Store ptr to reg_save_area.
6585 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6586 FIN, DAG.getIntPtrConstant(8));
6587 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6588 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6589 MemOps.push_back(Store);
6590 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6591 &MemOps[0], MemOps.size());
6594 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6595 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6596 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6597 SDValue Chain = Op.getOperand(0);
6598 SDValue SrcPtr = Op.getOperand(1);
6599 SDValue SrcSV = Op.getOperand(2);
6601 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6605 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6606 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6607 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6608 SDValue Chain = Op.getOperand(0);
6609 SDValue DstPtr = Op.getOperand(1);
6610 SDValue SrcPtr = Op.getOperand(2);
6611 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6612 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6613 DebugLoc dl = Op.getDebugLoc();
6615 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6616 DAG.getIntPtrConstant(24), 8, false,
6617 DstSV, 0, SrcSV, 0);
6621 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6622 DebugLoc dl = Op.getDebugLoc();
6623 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6625 default: return SDValue(); // Don't custom lower most intrinsics.
6626 // Comparison intrinsics.
6627 case Intrinsic::x86_sse_comieq_ss:
6628 case Intrinsic::x86_sse_comilt_ss:
6629 case Intrinsic::x86_sse_comile_ss:
6630 case Intrinsic::x86_sse_comigt_ss:
6631 case Intrinsic::x86_sse_comige_ss:
6632 case Intrinsic::x86_sse_comineq_ss:
6633 case Intrinsic::x86_sse_ucomieq_ss:
6634 case Intrinsic::x86_sse_ucomilt_ss:
6635 case Intrinsic::x86_sse_ucomile_ss:
6636 case Intrinsic::x86_sse_ucomigt_ss:
6637 case Intrinsic::x86_sse_ucomige_ss:
6638 case Intrinsic::x86_sse_ucomineq_ss:
6639 case Intrinsic::x86_sse2_comieq_sd:
6640 case Intrinsic::x86_sse2_comilt_sd:
6641 case Intrinsic::x86_sse2_comile_sd:
6642 case Intrinsic::x86_sse2_comigt_sd:
6643 case Intrinsic::x86_sse2_comige_sd:
6644 case Intrinsic::x86_sse2_comineq_sd:
6645 case Intrinsic::x86_sse2_ucomieq_sd:
6646 case Intrinsic::x86_sse2_ucomilt_sd:
6647 case Intrinsic::x86_sse2_ucomile_sd:
6648 case Intrinsic::x86_sse2_ucomigt_sd:
6649 case Intrinsic::x86_sse2_ucomige_sd:
6650 case Intrinsic::x86_sse2_ucomineq_sd: {
6652 ISD::CondCode CC = ISD::SETCC_INVALID;
6655 case Intrinsic::x86_sse_comieq_ss:
6656 case Intrinsic::x86_sse2_comieq_sd:
6660 case Intrinsic::x86_sse_comilt_ss:
6661 case Intrinsic::x86_sse2_comilt_sd:
6665 case Intrinsic::x86_sse_comile_ss:
6666 case Intrinsic::x86_sse2_comile_sd:
6670 case Intrinsic::x86_sse_comigt_ss:
6671 case Intrinsic::x86_sse2_comigt_sd:
6675 case Intrinsic::x86_sse_comige_ss:
6676 case Intrinsic::x86_sse2_comige_sd:
6680 case Intrinsic::x86_sse_comineq_ss:
6681 case Intrinsic::x86_sse2_comineq_sd:
6685 case Intrinsic::x86_sse_ucomieq_ss:
6686 case Intrinsic::x86_sse2_ucomieq_sd:
6687 Opc = X86ISD::UCOMI;
6690 case Intrinsic::x86_sse_ucomilt_ss:
6691 case Intrinsic::x86_sse2_ucomilt_sd:
6692 Opc = X86ISD::UCOMI;
6695 case Intrinsic::x86_sse_ucomile_ss:
6696 case Intrinsic::x86_sse2_ucomile_sd:
6697 Opc = X86ISD::UCOMI;
6700 case Intrinsic::x86_sse_ucomigt_ss:
6701 case Intrinsic::x86_sse2_ucomigt_sd:
6702 Opc = X86ISD::UCOMI;
6705 case Intrinsic::x86_sse_ucomige_ss:
6706 case Intrinsic::x86_sse2_ucomige_sd:
6707 Opc = X86ISD::UCOMI;
6710 case Intrinsic::x86_sse_ucomineq_ss:
6711 case Intrinsic::x86_sse2_ucomineq_sd:
6712 Opc = X86ISD::UCOMI;
6717 SDValue LHS = Op.getOperand(1);
6718 SDValue RHS = Op.getOperand(2);
6719 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6720 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6721 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6722 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6723 DAG.getConstant(X86CC, MVT::i8), Cond);
6724 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6726 // ptest intrinsics. The intrinsic these come from are designed to return
6727 // an integer value, not just an instruction so lower it to the ptest
6728 // pattern and a setcc for the result.
6729 case Intrinsic::x86_sse41_ptestz:
6730 case Intrinsic::x86_sse41_ptestc:
6731 case Intrinsic::x86_sse41_ptestnzc:{
6734 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6735 case Intrinsic::x86_sse41_ptestz:
6737 X86CC = X86::COND_E;
6739 case Intrinsic::x86_sse41_ptestc:
6741 X86CC = X86::COND_B;
6743 case Intrinsic::x86_sse41_ptestnzc:
6745 X86CC = X86::COND_A;
6749 SDValue LHS = Op.getOperand(1);
6750 SDValue RHS = Op.getOperand(2);
6751 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6752 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6753 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6754 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6757 // Fix vector shift instructions where the last operand is a non-immediate
6759 case Intrinsic::x86_sse2_pslli_w:
6760 case Intrinsic::x86_sse2_pslli_d:
6761 case Intrinsic::x86_sse2_pslli_q:
6762 case Intrinsic::x86_sse2_psrli_w:
6763 case Intrinsic::x86_sse2_psrli_d:
6764 case Intrinsic::x86_sse2_psrli_q:
6765 case Intrinsic::x86_sse2_psrai_w:
6766 case Intrinsic::x86_sse2_psrai_d:
6767 case Intrinsic::x86_mmx_pslli_w:
6768 case Intrinsic::x86_mmx_pslli_d:
6769 case Intrinsic::x86_mmx_pslli_q:
6770 case Intrinsic::x86_mmx_psrli_w:
6771 case Intrinsic::x86_mmx_psrli_d:
6772 case Intrinsic::x86_mmx_psrli_q:
6773 case Intrinsic::x86_mmx_psrai_w:
6774 case Intrinsic::x86_mmx_psrai_d: {
6775 SDValue ShAmt = Op.getOperand(2);
6776 if (isa<ConstantSDNode>(ShAmt))
6779 unsigned NewIntNo = 0;
6780 EVT ShAmtVT = MVT::v4i32;
6782 case Intrinsic::x86_sse2_pslli_w:
6783 NewIntNo = Intrinsic::x86_sse2_psll_w;
6785 case Intrinsic::x86_sse2_pslli_d:
6786 NewIntNo = Intrinsic::x86_sse2_psll_d;
6788 case Intrinsic::x86_sse2_pslli_q:
6789 NewIntNo = Intrinsic::x86_sse2_psll_q;
6791 case Intrinsic::x86_sse2_psrli_w:
6792 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6794 case Intrinsic::x86_sse2_psrli_d:
6795 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6797 case Intrinsic::x86_sse2_psrli_q:
6798 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6800 case Intrinsic::x86_sse2_psrai_w:
6801 NewIntNo = Intrinsic::x86_sse2_psra_w;
6803 case Intrinsic::x86_sse2_psrai_d:
6804 NewIntNo = Intrinsic::x86_sse2_psra_d;
6807 ShAmtVT = MVT::v2i32;
6809 case Intrinsic::x86_mmx_pslli_w:
6810 NewIntNo = Intrinsic::x86_mmx_psll_w;
6812 case Intrinsic::x86_mmx_pslli_d:
6813 NewIntNo = Intrinsic::x86_mmx_psll_d;
6815 case Intrinsic::x86_mmx_pslli_q:
6816 NewIntNo = Intrinsic::x86_mmx_psll_q;
6818 case Intrinsic::x86_mmx_psrli_w:
6819 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6821 case Intrinsic::x86_mmx_psrli_d:
6822 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6824 case Intrinsic::x86_mmx_psrli_q:
6825 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6827 case Intrinsic::x86_mmx_psrai_w:
6828 NewIntNo = Intrinsic::x86_mmx_psra_w;
6830 case Intrinsic::x86_mmx_psrai_d:
6831 NewIntNo = Intrinsic::x86_mmx_psra_d;
6833 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6839 // The vector shift intrinsics with scalars uses 32b shift amounts but
6840 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
6844 ShOps[1] = DAG.getConstant(0, MVT::i32);
6845 if (ShAmtVT == MVT::v4i32) {
6846 ShOps[2] = DAG.getUNDEF(MVT::i32);
6847 ShOps[3] = DAG.getUNDEF(MVT::i32);
6848 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
6850 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
6853 EVT VT = Op.getValueType();
6854 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
6855 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6856 DAG.getConstant(NewIntNo, MVT::i32),
6857 Op.getOperand(1), ShAmt);
6862 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6863 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6864 DebugLoc dl = Op.getDebugLoc();
6867 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6869 DAG.getConstant(TD->getPointerSize(),
6870 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6871 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6872 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6877 // Just load the return address.
6878 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6879 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6880 RetAddrFI, NULL, 0);
6883 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6884 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6885 MFI->setFrameAddressIsTaken(true);
6886 EVT VT = Op.getValueType();
6887 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6888 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6889 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6890 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6892 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6896 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6897 SelectionDAG &DAG) {
6898 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6901 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6903 MachineFunction &MF = DAG.getMachineFunction();
6904 SDValue Chain = Op.getOperand(0);
6905 SDValue Offset = Op.getOperand(1);
6906 SDValue Handler = Op.getOperand(2);
6907 DebugLoc dl = Op.getDebugLoc();
6909 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6911 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6913 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6914 DAG.getIntPtrConstant(-TD->getPointerSize()));
6915 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6916 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6917 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6918 MF.getRegInfo().addLiveOut(StoreAddrReg);
6920 return DAG.getNode(X86ISD::EH_RETURN, dl,
6922 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6925 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6926 SelectionDAG &DAG) {
6927 SDValue Root = Op.getOperand(0);
6928 SDValue Trmp = Op.getOperand(1); // trampoline
6929 SDValue FPtr = Op.getOperand(2); // nested function
6930 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6931 DebugLoc dl = Op.getDebugLoc();
6933 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6935 const X86InstrInfo *TII =
6936 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6938 if (Subtarget->is64Bit()) {
6939 SDValue OutChains[6];
6941 // Large code-model.
6943 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6944 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6946 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6947 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6949 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6951 // Load the pointer to the nested function into R11.
6952 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6953 SDValue Addr = Trmp;
6954 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6957 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6958 DAG.getConstant(2, MVT::i64));
6959 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6961 // Load the 'nest' parameter value into R10.
6962 // R10 is specified in X86CallingConv.td
6963 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6964 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6965 DAG.getConstant(10, MVT::i64));
6966 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6967 Addr, TrmpAddr, 10);
6969 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6970 DAG.getConstant(12, MVT::i64));
6971 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6973 // Jump to the nested function.
6974 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6975 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6976 DAG.getConstant(20, MVT::i64));
6977 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6978 Addr, TrmpAddr, 20);
6980 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6981 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6982 DAG.getConstant(22, MVT::i64));
6983 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6987 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6988 return DAG.getMergeValues(Ops, 2, dl);
6990 const Function *Func =
6991 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6992 CallingConv::ID CC = Func->getCallingConv();
6997 llvm_unreachable("Unsupported calling convention");
6998 case CallingConv::C:
6999 case CallingConv::X86_StdCall: {
7000 // Pass 'nest' parameter in ECX.
7001 // Must be kept in sync with X86CallingConv.td
7004 // Check that ECX wasn't needed by an 'inreg' parameter.
7005 const FunctionType *FTy = Func->getFunctionType();
7006 const AttrListPtr &Attrs = Func->getAttributes();
7008 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7009 unsigned InRegCount = 0;
7012 for (FunctionType::param_iterator I = FTy->param_begin(),
7013 E = FTy->param_end(); I != E; ++I, ++Idx)
7014 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7015 // FIXME: should only count parameters that are lowered to integers.
7016 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7018 if (InRegCount > 2) {
7019 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
7024 case CallingConv::X86_FastCall:
7025 case CallingConv::Fast:
7026 // Pass 'nest' parameter in EAX.
7027 // Must be kept in sync with X86CallingConv.td
7032 SDValue OutChains[4];
7035 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7036 DAG.getConstant(10, MVT::i32));
7037 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7039 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
7040 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7041 OutChains[0] = DAG.getStore(Root, dl,
7042 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7045 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7046 DAG.getConstant(1, MVT::i32));
7047 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
7049 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
7050 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7051 DAG.getConstant(5, MVT::i32));
7052 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7053 TrmpAddr, 5, false, 1);
7055 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7056 DAG.getConstant(6, MVT::i32));
7057 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
7060 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7061 return DAG.getMergeValues(Ops, 2, dl);
7065 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
7067 The rounding mode is in bits 11:10 of FPSR, and has the following
7074 FLT_ROUNDS, on the other hand, expects the following:
7081 To perform the conversion, we do:
7082 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7085 MachineFunction &MF = DAG.getMachineFunction();
7086 const TargetMachine &TM = MF.getTarget();
7087 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7088 unsigned StackAlignment = TFI.getStackAlignment();
7089 EVT VT = Op.getValueType();
7090 DebugLoc dl = Op.getDebugLoc();
7092 // Save FP Control Word to stack slot
7093 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7094 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7096 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7097 DAG.getEntryNode(), StackSlot);
7099 // Load FP Control Word from stack slot
7100 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
7102 // Transform as necessary
7104 DAG.getNode(ISD::SRL, dl, MVT::i16,
7105 DAG.getNode(ISD::AND, dl, MVT::i16,
7106 CWD, DAG.getConstant(0x800, MVT::i16)),
7107 DAG.getConstant(11, MVT::i8));
7109 DAG.getNode(ISD::SRL, dl, MVT::i16,
7110 DAG.getNode(ISD::AND, dl, MVT::i16,
7111 CWD, DAG.getConstant(0x400, MVT::i16)),
7112 DAG.getConstant(9, MVT::i8));
7115 DAG.getNode(ISD::AND, dl, MVT::i16,
7116 DAG.getNode(ISD::ADD, dl, MVT::i16,
7117 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7118 DAG.getConstant(1, MVT::i16)),
7119 DAG.getConstant(3, MVT::i16));
7122 return DAG.getNode((VT.getSizeInBits() < 16 ?
7123 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7126 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
7127 EVT VT = Op.getValueType();
7129 unsigned NumBits = VT.getSizeInBits();
7130 DebugLoc dl = Op.getDebugLoc();
7132 Op = Op.getOperand(0);
7133 if (VT == MVT::i8) {
7134 // Zero extend to i32 since there is not an i8 bsr.
7136 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7139 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7140 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7141 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7143 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7146 DAG.getConstant(NumBits+NumBits-1, OpVT),
7147 DAG.getConstant(X86::COND_E, MVT::i8),
7150 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7152 // Finally xor with NumBits-1.
7153 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7156 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7160 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
7161 EVT VT = Op.getValueType();
7163 unsigned NumBits = VT.getSizeInBits();
7164 DebugLoc dl = Op.getDebugLoc();
7166 Op = Op.getOperand(0);
7167 if (VT == MVT::i8) {
7169 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7172 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7173 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7174 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7176 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7179 DAG.getConstant(NumBits, OpVT),
7180 DAG.getConstant(X86::COND_E, MVT::i8),
7183 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7186 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7190 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
7191 EVT VT = Op.getValueType();
7192 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7193 DebugLoc dl = Op.getDebugLoc();
7195 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7196 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7197 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7198 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7199 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7201 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7202 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7203 // return AloBlo + AloBhi + AhiBlo;
7205 SDValue A = Op.getOperand(0);
7206 SDValue B = Op.getOperand(1);
7208 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7209 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7210 A, DAG.getConstant(32, MVT::i32));
7211 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7212 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7213 B, DAG.getConstant(32, MVT::i32));
7214 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7215 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7217 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7218 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7220 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7221 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7223 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7224 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7225 AloBhi, DAG.getConstant(32, MVT::i32));
7226 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7227 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7228 AhiBlo, DAG.getConstant(32, MVT::i32));
7229 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7230 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7235 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
7236 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7237 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7238 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7239 // has only one use.
7240 SDNode *N = Op.getNode();
7241 SDValue LHS = N->getOperand(0);
7242 SDValue RHS = N->getOperand(1);
7243 unsigned BaseOp = 0;
7245 DebugLoc dl = Op.getDebugLoc();
7247 switch (Op.getOpcode()) {
7248 default: llvm_unreachable("Unknown ovf instruction!");
7250 // A subtract of one will be selected as a INC. Note that INC doesn't
7251 // set CF, so we can't do this for UADDO.
7252 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7253 if (C->getAPIntValue() == 1) {
7254 BaseOp = X86ISD::INC;
7258 BaseOp = X86ISD::ADD;
7262 BaseOp = X86ISD::ADD;
7266 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7267 // set CF, so we can't do this for USUBO.
7268 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7269 if (C->getAPIntValue() == 1) {
7270 BaseOp = X86ISD::DEC;
7274 BaseOp = X86ISD::SUB;
7278 BaseOp = X86ISD::SUB;
7282 BaseOp = X86ISD::SMUL;
7286 BaseOp = X86ISD::UMUL;
7291 // Also sets EFLAGS.
7292 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7293 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7296 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7297 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7299 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7303 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
7304 EVT T = Op.getValueType();
7305 DebugLoc dl = Op.getDebugLoc();
7308 switch(T.getSimpleVT().SimpleTy) {
7310 assert(false && "Invalid value type!");
7311 case MVT::i8: Reg = X86::AL; size = 1; break;
7312 case MVT::i16: Reg = X86::AX; size = 2; break;
7313 case MVT::i32: Reg = X86::EAX; size = 4; break;
7315 assert(Subtarget->is64Bit() && "Node not type legal!");
7316 Reg = X86::RAX; size = 8;
7319 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7320 Op.getOperand(2), SDValue());
7321 SDValue Ops[] = { cpIn.getValue(0),
7324 DAG.getTargetConstant(size, MVT::i8),
7326 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7327 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7329 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7333 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7334 SelectionDAG &DAG) {
7335 assert(Subtarget->is64Bit() && "Result not type legalized?");
7336 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7337 SDValue TheChain = Op.getOperand(0);
7338 DebugLoc dl = Op.getDebugLoc();
7339 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7340 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7341 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7343 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7344 DAG.getConstant(32, MVT::i8));
7346 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7349 return DAG.getMergeValues(Ops, 2, dl);
7352 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7353 SDNode *Node = Op.getNode();
7354 DebugLoc dl = Node->getDebugLoc();
7355 EVT T = Node->getValueType(0);
7356 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7357 DAG.getConstant(0, T), Node->getOperand(2));
7358 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7359 cast<AtomicSDNode>(Node)->getMemoryVT(),
7360 Node->getOperand(0),
7361 Node->getOperand(1), negOp,
7362 cast<AtomicSDNode>(Node)->getSrcValue(),
7363 cast<AtomicSDNode>(Node)->getAlignment());
7366 /// LowerOperation - Provide custom lowering hooks for some operations.
7368 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7369 switch (Op.getOpcode()) {
7370 default: llvm_unreachable("Should not custom lower this!");
7371 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7372 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7373 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7374 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7375 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7376 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7377 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7378 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7379 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7380 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7381 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7382 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7383 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7384 case ISD::SHL_PARTS:
7385 case ISD::SRA_PARTS:
7386 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7387 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7388 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7389 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7390 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7391 case ISD::FABS: return LowerFABS(Op, DAG);
7392 case ISD::FNEG: return LowerFNEG(Op, DAG);
7393 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7394 case ISD::SETCC: return LowerSETCC(Op, DAG);
7395 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7396 case ISD::SELECT: return LowerSELECT(Op, DAG);
7397 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7398 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7399 case ISD::VASTART: return LowerVASTART(Op, DAG);
7400 case ISD::VAARG: return LowerVAARG(Op, DAG);
7401 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7402 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7403 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7404 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7405 case ISD::FRAME_TO_ARGS_OFFSET:
7406 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7407 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7408 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7409 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7410 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7411 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7412 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7413 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7419 case ISD::UMULO: return LowerXALUO(Op, DAG);
7420 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7424 void X86TargetLowering::
7425 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7426 SelectionDAG &DAG, unsigned NewOp) {
7427 EVT T = Node->getValueType(0);
7428 DebugLoc dl = Node->getDebugLoc();
7429 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7431 SDValue Chain = Node->getOperand(0);
7432 SDValue In1 = Node->getOperand(1);
7433 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7434 Node->getOperand(2), DAG.getIntPtrConstant(0));
7435 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7436 Node->getOperand(2), DAG.getIntPtrConstant(1));
7437 SDValue Ops[] = { Chain, In1, In2L, In2H };
7438 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7440 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7441 cast<MemSDNode>(Node)->getMemOperand());
7442 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7443 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7444 Results.push_back(Result.getValue(2));
7447 /// ReplaceNodeResults - Replace a node with an illegal result type
7448 /// with a new node built out of custom code.
7449 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7450 SmallVectorImpl<SDValue>&Results,
7451 SelectionDAG &DAG) {
7452 DebugLoc dl = N->getDebugLoc();
7453 switch (N->getOpcode()) {
7455 assert(false && "Do not know how to custom type legalize this operation!");
7457 case ISD::FP_TO_SINT: {
7458 std::pair<SDValue,SDValue> Vals =
7459 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7460 SDValue FIST = Vals.first, StackSlot = Vals.second;
7461 if (FIST.getNode() != 0) {
7462 EVT VT = N->getValueType(0);
7463 // Return a load from the stack slot.
7464 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
7468 case ISD::READCYCLECOUNTER: {
7469 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7470 SDValue TheChain = N->getOperand(0);
7471 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7472 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7474 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7476 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7477 SDValue Ops[] = { eax, edx };
7478 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7479 Results.push_back(edx.getValue(1));
7486 EVT WidenVT = getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
7487 Results.push_back(DAG.UnrollVectorOp(N, WidenVT.getVectorNumElements()));
7490 case ISD::ATOMIC_CMP_SWAP: {
7491 EVT T = N->getValueType(0);
7492 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7493 SDValue cpInL, cpInH;
7494 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7495 DAG.getConstant(0, MVT::i32));
7496 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7497 DAG.getConstant(1, MVT::i32));
7498 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7499 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7501 SDValue swapInL, swapInH;
7502 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7503 DAG.getConstant(0, MVT::i32));
7504 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7505 DAG.getConstant(1, MVT::i32));
7506 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7508 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7509 swapInL.getValue(1));
7510 SDValue Ops[] = { swapInH.getValue(0),
7512 swapInH.getValue(1) };
7513 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7514 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7515 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7516 MVT::i32, Result.getValue(1));
7517 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7518 MVT::i32, cpOutL.getValue(2));
7519 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7520 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7521 Results.push_back(cpOutH.getValue(1));
7524 case ISD::ATOMIC_LOAD_ADD:
7525 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7527 case ISD::ATOMIC_LOAD_AND:
7528 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7530 case ISD::ATOMIC_LOAD_NAND:
7531 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7533 case ISD::ATOMIC_LOAD_OR:
7534 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7536 case ISD::ATOMIC_LOAD_SUB:
7537 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7539 case ISD::ATOMIC_LOAD_XOR:
7540 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7542 case ISD::ATOMIC_SWAP:
7543 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7548 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7550 default: return NULL;
7551 case X86ISD::BSF: return "X86ISD::BSF";
7552 case X86ISD::BSR: return "X86ISD::BSR";
7553 case X86ISD::SHLD: return "X86ISD::SHLD";
7554 case X86ISD::SHRD: return "X86ISD::SHRD";
7555 case X86ISD::FAND: return "X86ISD::FAND";
7556 case X86ISD::FOR: return "X86ISD::FOR";
7557 case X86ISD::FXOR: return "X86ISD::FXOR";
7558 case X86ISD::FSRL: return "X86ISD::FSRL";
7559 case X86ISD::FILD: return "X86ISD::FILD";
7560 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7561 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7562 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7563 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7564 case X86ISD::FLD: return "X86ISD::FLD";
7565 case X86ISD::FST: return "X86ISD::FST";
7566 case X86ISD::CALL: return "X86ISD::CALL";
7567 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7568 case X86ISD::BT: return "X86ISD::BT";
7569 case X86ISD::CMP: return "X86ISD::CMP";
7570 case X86ISD::COMI: return "X86ISD::COMI";
7571 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7572 case X86ISD::SETCC: return "X86ISD::SETCC";
7573 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7574 case X86ISD::CMOV: return "X86ISD::CMOV";
7575 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7576 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7577 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7578 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7579 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7580 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7581 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7582 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7583 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7584 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7585 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7586 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7587 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7588 case X86ISD::FMAX: return "X86ISD::FMAX";
7589 case X86ISD::FMIN: return "X86ISD::FMIN";
7590 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7591 case X86ISD::FRCP: return "X86ISD::FRCP";
7592 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7593 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7594 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7595 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7596 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7597 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7598 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7599 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7600 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7601 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7602 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7603 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7604 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7605 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7606 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7607 case X86ISD::VSHL: return "X86ISD::VSHL";
7608 case X86ISD::VSRL: return "X86ISD::VSRL";
7609 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7610 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7611 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7612 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7613 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7614 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7615 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7616 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7617 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7618 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7619 case X86ISD::ADD: return "X86ISD::ADD";
7620 case X86ISD::SUB: return "X86ISD::SUB";
7621 case X86ISD::SMUL: return "X86ISD::SMUL";
7622 case X86ISD::UMUL: return "X86ISD::UMUL";
7623 case X86ISD::INC: return "X86ISD::INC";
7624 case X86ISD::DEC: return "X86ISD::DEC";
7625 case X86ISD::OR: return "X86ISD::OR";
7626 case X86ISD::XOR: return "X86ISD::XOR";
7627 case X86ISD::AND: return "X86ISD::AND";
7628 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7629 case X86ISD::PTEST: return "X86ISD::PTEST";
7630 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7634 // isLegalAddressingMode - Return true if the addressing mode represented
7635 // by AM is legal for this target, for a load/store of the specified type.
7636 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7637 const Type *Ty) const {
7638 // X86 supports extremely general addressing modes.
7639 CodeModel::Model M = getTargetMachine().getCodeModel();
7641 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7642 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7647 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7649 // If a reference to this global requires an extra load, we can't fold it.
7650 if (isGlobalStubReference(GVFlags))
7653 // If BaseGV requires a register for the PIC base, we cannot also have a
7654 // BaseReg specified.
7655 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7658 // If lower 4G is not available, then we must use rip-relative addressing.
7659 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7669 // These scales always work.
7674 // These scales are formed with basereg+scalereg. Only accept if there is
7679 default: // Other stuff never works.
7687 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7688 if (!Ty1->isInteger() || !Ty2->isInteger())
7690 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7691 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7692 if (NumBits1 <= NumBits2)
7694 return Subtarget->is64Bit() || NumBits1 < 64;
7697 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7698 if (!VT1.isInteger() || !VT2.isInteger())
7700 unsigned NumBits1 = VT1.getSizeInBits();
7701 unsigned NumBits2 = VT2.getSizeInBits();
7702 if (NumBits1 <= NumBits2)
7704 return Subtarget->is64Bit() || NumBits1 < 64;
7707 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7708 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7709 return Ty1->isInteger(32) && Ty2->isInteger(64) && Subtarget->is64Bit();
7712 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7713 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7714 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7717 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7718 // i16 instructions are longer (0x66 prefix) and potentially slower.
7719 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7722 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7723 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7724 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7725 /// are assumed to be legal.
7727 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7729 // Only do shuffles on 128-bit vector types for now.
7730 if (VT.getSizeInBits() == 64)
7733 // FIXME: pshufb, blends, shifts.
7734 return (VT.getVectorNumElements() == 2 ||
7735 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7736 isMOVLMask(M, VT) ||
7737 isSHUFPMask(M, VT) ||
7738 isPSHUFDMask(M, VT) ||
7739 isPSHUFHWMask(M, VT) ||
7740 isPSHUFLWMask(M, VT) ||
7741 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7742 isUNPCKLMask(M, VT) ||
7743 isUNPCKHMask(M, VT) ||
7744 isUNPCKL_v_undef_Mask(M, VT) ||
7745 isUNPCKH_v_undef_Mask(M, VT));
7749 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7751 unsigned NumElts = VT.getVectorNumElements();
7752 // FIXME: This collection of masks seems suspect.
7755 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7756 return (isMOVLMask(Mask, VT) ||
7757 isCommutedMOVLMask(Mask, VT, true) ||
7758 isSHUFPMask(Mask, VT) ||
7759 isCommutedSHUFPMask(Mask, VT));
7764 //===----------------------------------------------------------------------===//
7765 // X86 Scheduler Hooks
7766 //===----------------------------------------------------------------------===//
7768 // private utility function
7770 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7771 MachineBasicBlock *MBB,
7779 TargetRegisterClass *RC,
7780 bool invSrc) const {
7781 // For the atomic bitwise operator, we generate
7784 // ld t1 = [bitinstr.addr]
7785 // op t2 = t1, [bitinstr.val]
7787 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7789 // fallthrough -->nextMBB
7790 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7791 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7792 MachineFunction::iterator MBBIter = MBB;
7795 /// First build the CFG
7796 MachineFunction *F = MBB->getParent();
7797 MachineBasicBlock *thisMBB = MBB;
7798 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7799 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7800 F->insert(MBBIter, newMBB);
7801 F->insert(MBBIter, nextMBB);
7803 // Move all successors to thisMBB to nextMBB
7804 nextMBB->transferSuccessors(thisMBB);
7806 // Update thisMBB to fall through to newMBB
7807 thisMBB->addSuccessor(newMBB);
7809 // newMBB jumps to itself and fall through to nextMBB
7810 newMBB->addSuccessor(nextMBB);
7811 newMBB->addSuccessor(newMBB);
7813 // Insert instructions into newMBB based on incoming instruction
7814 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7815 "unexpected number of operands");
7816 DebugLoc dl = bInstr->getDebugLoc();
7817 MachineOperand& destOper = bInstr->getOperand(0);
7818 MachineOperand* argOpers[2 + X86AddrNumOperands];
7819 int numArgs = bInstr->getNumOperands() - 1;
7820 for (int i=0; i < numArgs; ++i)
7821 argOpers[i] = &bInstr->getOperand(i+1);
7823 // x86 address has 4 operands: base, index, scale, and displacement
7824 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7825 int valArgIndx = lastAddrIndx + 1;
7827 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7828 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7829 for (int i=0; i <= lastAddrIndx; ++i)
7830 (*MIB).addOperand(*argOpers[i]);
7832 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7834 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7839 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7840 assert((argOpers[valArgIndx]->isReg() ||
7841 argOpers[valArgIndx]->isImm()) &&
7843 if (argOpers[valArgIndx]->isReg())
7844 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7846 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7848 (*MIB).addOperand(*argOpers[valArgIndx]);
7850 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7853 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7854 for (int i=0; i <= lastAddrIndx; ++i)
7855 (*MIB).addOperand(*argOpers[i]);
7857 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7858 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7859 bInstr->memoperands_end());
7861 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7865 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7867 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7871 // private utility function: 64 bit atomics on 32 bit host.
7873 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7874 MachineBasicBlock *MBB,
7879 bool invSrc) const {
7880 // For the atomic bitwise operator, we generate
7881 // thisMBB (instructions are in pairs, except cmpxchg8b)
7882 // ld t1,t2 = [bitinstr.addr]
7884 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7885 // op t5, t6 <- out1, out2, [bitinstr.val]
7886 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7887 // mov ECX, EBX <- t5, t6
7888 // mov EAX, EDX <- t1, t2
7889 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7890 // mov t3, t4 <- EAX, EDX
7892 // result in out1, out2
7893 // fallthrough -->nextMBB
7895 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7896 const unsigned LoadOpc = X86::MOV32rm;
7897 const unsigned copyOpc = X86::MOV32rr;
7898 const unsigned NotOpc = X86::NOT32r;
7899 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7900 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7901 MachineFunction::iterator MBBIter = MBB;
7904 /// First build the CFG
7905 MachineFunction *F = MBB->getParent();
7906 MachineBasicBlock *thisMBB = MBB;
7907 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7908 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7909 F->insert(MBBIter, newMBB);
7910 F->insert(MBBIter, nextMBB);
7912 // Move all successors to thisMBB to nextMBB
7913 nextMBB->transferSuccessors(thisMBB);
7915 // Update thisMBB to fall through to newMBB
7916 thisMBB->addSuccessor(newMBB);
7918 // newMBB jumps to itself and fall through to nextMBB
7919 newMBB->addSuccessor(nextMBB);
7920 newMBB->addSuccessor(newMBB);
7922 DebugLoc dl = bInstr->getDebugLoc();
7923 // Insert instructions into newMBB based on incoming instruction
7924 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7925 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7926 "unexpected number of operands");
7927 MachineOperand& dest1Oper = bInstr->getOperand(0);
7928 MachineOperand& dest2Oper = bInstr->getOperand(1);
7929 MachineOperand* argOpers[2 + X86AddrNumOperands];
7930 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7931 argOpers[i] = &bInstr->getOperand(i+2);
7933 // x86 address has 5 operands: base, index, scale, displacement, and segment.
7934 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7936 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7937 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7938 for (int i=0; i <= lastAddrIndx; ++i)
7939 (*MIB).addOperand(*argOpers[i]);
7940 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7941 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7942 // add 4 to displacement.
7943 for (int i=0; i <= lastAddrIndx-2; ++i)
7944 (*MIB).addOperand(*argOpers[i]);
7945 MachineOperand newOp3 = *(argOpers[3]);
7947 newOp3.setImm(newOp3.getImm()+4);
7949 newOp3.setOffset(newOp3.getOffset()+4);
7950 (*MIB).addOperand(newOp3);
7951 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7953 // t3/4 are defined later, at the bottom of the loop
7954 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7955 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7956 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7957 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7958 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7959 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7961 // The subsequent operations should be using the destination registers of
7962 //the PHI instructions.
7964 t1 = F->getRegInfo().createVirtualRegister(RC);
7965 t2 = F->getRegInfo().createVirtualRegister(RC);
7966 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
7967 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
7969 t1 = dest1Oper.getReg();
7970 t2 = dest2Oper.getReg();
7973 int valArgIndx = lastAddrIndx + 1;
7974 assert((argOpers[valArgIndx]->isReg() ||
7975 argOpers[valArgIndx]->isImm()) &&
7977 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7978 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7979 if (argOpers[valArgIndx]->isReg())
7980 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7982 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7983 if (regOpcL != X86::MOV32rr)
7985 (*MIB).addOperand(*argOpers[valArgIndx]);
7986 assert(argOpers[valArgIndx + 1]->isReg() ==
7987 argOpers[valArgIndx]->isReg());
7988 assert(argOpers[valArgIndx + 1]->isImm() ==
7989 argOpers[valArgIndx]->isImm());
7990 if (argOpers[valArgIndx + 1]->isReg())
7991 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7993 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7994 if (regOpcH != X86::MOV32rr)
7996 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
7998 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8000 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8003 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8005 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8008 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8009 for (int i=0; i <= lastAddrIndx; ++i)
8010 (*MIB).addOperand(*argOpers[i]);
8012 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8013 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8014 bInstr->memoperands_end());
8016 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8017 MIB.addReg(X86::EAX);
8018 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8019 MIB.addReg(X86::EDX);
8022 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
8024 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8028 // private utility function
8030 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8031 MachineBasicBlock *MBB,
8032 unsigned cmovOpc) const {
8033 // For the atomic min/max operator, we generate
8036 // ld t1 = [min/max.addr]
8037 // mov t2 = [min/max.val]
8039 // cmov[cond] t2 = t1
8041 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8043 // fallthrough -->nextMBB
8045 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8046 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8047 MachineFunction::iterator MBBIter = MBB;
8050 /// First build the CFG
8051 MachineFunction *F = MBB->getParent();
8052 MachineBasicBlock *thisMBB = MBB;
8053 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8054 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8055 F->insert(MBBIter, newMBB);
8056 F->insert(MBBIter, nextMBB);
8058 // Move all successors of thisMBB to nextMBB
8059 nextMBB->transferSuccessors(thisMBB);
8061 // Update thisMBB to fall through to newMBB
8062 thisMBB->addSuccessor(newMBB);
8064 // newMBB jumps to newMBB and fall through to nextMBB
8065 newMBB->addSuccessor(nextMBB);
8066 newMBB->addSuccessor(newMBB);
8068 DebugLoc dl = mInstr->getDebugLoc();
8069 // Insert instructions into newMBB based on incoming instruction
8070 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8071 "unexpected number of operands");
8072 MachineOperand& destOper = mInstr->getOperand(0);
8073 MachineOperand* argOpers[2 + X86AddrNumOperands];
8074 int numArgs = mInstr->getNumOperands() - 1;
8075 for (int i=0; i < numArgs; ++i)
8076 argOpers[i] = &mInstr->getOperand(i+1);
8078 // x86 address has 4 operands: base, index, scale, and displacement
8079 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8080 int valArgIndx = lastAddrIndx + 1;
8082 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8083 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8084 for (int i=0; i <= lastAddrIndx; ++i)
8085 (*MIB).addOperand(*argOpers[i]);
8087 // We only support register and immediate values
8088 assert((argOpers[valArgIndx]->isReg() ||
8089 argOpers[valArgIndx]->isImm()) &&
8092 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8093 if (argOpers[valArgIndx]->isReg())
8094 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8096 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8097 (*MIB).addOperand(*argOpers[valArgIndx]);
8099 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8102 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8107 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8108 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8112 // Cmp and exchange if none has modified the memory location
8113 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8114 for (int i=0; i <= lastAddrIndx; ++i)
8115 (*MIB).addOperand(*argOpers[i]);
8117 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8118 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8119 mInstr->memoperands_end());
8121 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8122 MIB.addReg(X86::EAX);
8125 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
8127 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8131 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8132 // all of this code can be replaced with that in the .td file.
8134 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8135 unsigned numArgs, bool memArg) const {
8137 MachineFunction *F = BB->getParent();
8138 DebugLoc dl = MI->getDebugLoc();
8139 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8143 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8145 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8147 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8149 for (unsigned i = 0; i < numArgs; ++i) {
8150 MachineOperand &Op = MI->getOperand(i+1);
8152 if (!(Op.isReg() && Op.isImplicit()))
8156 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8159 F->DeleteMachineInstr(MI);
8165 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8167 MachineBasicBlock *MBB) const {
8168 // Emit code to save XMM registers to the stack. The ABI says that the
8169 // number of registers to save is given in %al, so it's theoretically
8170 // possible to do an indirect jump trick to avoid saving all of them,
8171 // however this code takes a simpler approach and just executes all
8172 // of the stores if %al is non-zero. It's less code, and it's probably
8173 // easier on the hardware branch predictor, and stores aren't all that
8174 // expensive anyway.
8176 // Create the new basic blocks. One block contains all the XMM stores,
8177 // and one block is the final destination regardless of whether any
8178 // stores were performed.
8179 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8180 MachineFunction *F = MBB->getParent();
8181 MachineFunction::iterator MBBIter = MBB;
8183 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8184 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8185 F->insert(MBBIter, XMMSaveMBB);
8186 F->insert(MBBIter, EndMBB);
8189 // Move any original successors of MBB to the end block.
8190 EndMBB->transferSuccessors(MBB);
8191 // The original block will now fall through to the XMM save block.
8192 MBB->addSuccessor(XMMSaveMBB);
8193 // The XMMSaveMBB will fall through to the end block.
8194 XMMSaveMBB->addSuccessor(EndMBB);
8196 // Now add the instructions.
8197 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8198 DebugLoc DL = MI->getDebugLoc();
8200 unsigned CountReg = MI->getOperand(0).getReg();
8201 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8202 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8204 if (!Subtarget->isTargetWin64()) {
8205 // If %al is 0, branch around the XMM save block.
8206 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8207 BuildMI(MBB, DL, TII->get(X86::JE)).addMBB(EndMBB);
8208 MBB->addSuccessor(EndMBB);
8211 // In the XMM save block, save all the XMM argument registers.
8212 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8213 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8214 MachineMemOperand *MMO =
8215 F->getMachineMemOperand(
8216 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8217 MachineMemOperand::MOStore, Offset,
8218 /*Size=*/16, /*Align=*/16);
8219 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8220 .addFrameIndex(RegSaveFrameIndex)
8221 .addImm(/*Scale=*/1)
8222 .addReg(/*IndexReg=*/0)
8223 .addImm(/*Disp=*/Offset)
8224 .addReg(/*Segment=*/0)
8225 .addReg(MI->getOperand(i).getReg())
8226 .addMemOperand(MMO);
8229 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8235 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8236 MachineBasicBlock *BB,
8237 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8238 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8239 DebugLoc DL = MI->getDebugLoc();
8241 // To "insert" a SELECT_CC instruction, we actually have to insert the
8242 // diamond control-flow pattern. The incoming instruction knows the
8243 // destination vreg to set, the condition code register to branch on, the
8244 // true/false values to select between, and a branch opcode to use.
8245 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8246 MachineFunction::iterator It = BB;
8252 // cmpTY ccX, r1, r2
8254 // fallthrough --> copy0MBB
8255 MachineBasicBlock *thisMBB = BB;
8256 MachineFunction *F = BB->getParent();
8257 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8258 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8260 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8261 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8262 F->insert(It, copy0MBB);
8263 F->insert(It, sinkMBB);
8264 // Update machine-CFG edges by first adding all successors of the current
8265 // block to the new block which will contain the Phi node for the select.
8266 // Also inform sdisel of the edge changes.
8267 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8268 E = BB->succ_end(); I != E; ++I) {
8269 EM->insert(std::make_pair(*I, sinkMBB));
8270 sinkMBB->addSuccessor(*I);
8272 // Next, remove all successors of the current block, and add the true
8273 // and fallthrough blocks as its successors.
8274 while (!BB->succ_empty())
8275 BB->removeSuccessor(BB->succ_begin());
8276 // Add the true and fallthrough blocks as its successors.
8277 BB->addSuccessor(copy0MBB);
8278 BB->addSuccessor(sinkMBB);
8281 // %FalseValue = ...
8282 // # fallthrough to sinkMBB
8285 // Update machine-CFG edges
8286 BB->addSuccessor(sinkMBB);
8289 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8292 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8293 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8294 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8296 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8302 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8303 MachineBasicBlock *BB,
8304 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8305 switch (MI->getOpcode()) {
8306 default: assert(false && "Unexpected instr type to insert");
8308 case X86::CMOV_V1I64:
8309 case X86::CMOV_FR32:
8310 case X86::CMOV_FR64:
8311 case X86::CMOV_V4F32:
8312 case X86::CMOV_V2F64:
8313 case X86::CMOV_V2I64:
8314 return EmitLoweredSelect(MI, BB, EM);
8316 case X86::FP32_TO_INT16_IN_MEM:
8317 case X86::FP32_TO_INT32_IN_MEM:
8318 case X86::FP32_TO_INT64_IN_MEM:
8319 case X86::FP64_TO_INT16_IN_MEM:
8320 case X86::FP64_TO_INT32_IN_MEM:
8321 case X86::FP64_TO_INT64_IN_MEM:
8322 case X86::FP80_TO_INT16_IN_MEM:
8323 case X86::FP80_TO_INT32_IN_MEM:
8324 case X86::FP80_TO_INT64_IN_MEM: {
8325 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8326 DebugLoc DL = MI->getDebugLoc();
8328 // Change the floating point control register to use "round towards zero"
8329 // mode when truncating to an integer value.
8330 MachineFunction *F = BB->getParent();
8331 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8332 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8334 // Load the old value of the high byte of the control word...
8336 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8337 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8340 // Set the high part to be round to zero...
8341 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8344 // Reload the modified control word now...
8345 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8347 // Restore the memory image of control word to original value
8348 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8351 // Get the X86 opcode to use.
8353 switch (MI->getOpcode()) {
8354 default: llvm_unreachable("illegal opcode!");
8355 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8356 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8357 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8358 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8359 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8360 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8361 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8362 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8363 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8367 MachineOperand &Op = MI->getOperand(0);
8369 AM.BaseType = X86AddressMode::RegBase;
8370 AM.Base.Reg = Op.getReg();
8372 AM.BaseType = X86AddressMode::FrameIndexBase;
8373 AM.Base.FrameIndex = Op.getIndex();
8375 Op = MI->getOperand(1);
8377 AM.Scale = Op.getImm();
8378 Op = MI->getOperand(2);
8380 AM.IndexReg = Op.getImm();
8381 Op = MI->getOperand(3);
8382 if (Op.isGlobal()) {
8383 AM.GV = Op.getGlobal();
8385 AM.Disp = Op.getImm();
8387 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8388 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8390 // Reload the original control word now.
8391 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8393 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8396 // String/text processing lowering.
8397 case X86::PCMPISTRM128REG:
8398 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8399 case X86::PCMPISTRM128MEM:
8400 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8401 case X86::PCMPESTRM128REG:
8402 return EmitPCMP(MI, BB, 5, false /* in mem */);
8403 case X86::PCMPESTRM128MEM:
8404 return EmitPCMP(MI, BB, 5, true /* in mem */);
8407 case X86::ATOMAND32:
8408 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8409 X86::AND32ri, X86::MOV32rm,
8410 X86::LCMPXCHG32, X86::MOV32rr,
8411 X86::NOT32r, X86::EAX,
8412 X86::GR32RegisterClass);
8414 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8415 X86::OR32ri, X86::MOV32rm,
8416 X86::LCMPXCHG32, X86::MOV32rr,
8417 X86::NOT32r, X86::EAX,
8418 X86::GR32RegisterClass);
8419 case X86::ATOMXOR32:
8420 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8421 X86::XOR32ri, X86::MOV32rm,
8422 X86::LCMPXCHG32, X86::MOV32rr,
8423 X86::NOT32r, X86::EAX,
8424 X86::GR32RegisterClass);
8425 case X86::ATOMNAND32:
8426 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8427 X86::AND32ri, X86::MOV32rm,
8428 X86::LCMPXCHG32, X86::MOV32rr,
8429 X86::NOT32r, X86::EAX,
8430 X86::GR32RegisterClass, true);
8431 case X86::ATOMMIN32:
8432 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8433 case X86::ATOMMAX32:
8434 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8435 case X86::ATOMUMIN32:
8436 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8437 case X86::ATOMUMAX32:
8438 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8440 case X86::ATOMAND16:
8441 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8442 X86::AND16ri, X86::MOV16rm,
8443 X86::LCMPXCHG16, X86::MOV16rr,
8444 X86::NOT16r, X86::AX,
8445 X86::GR16RegisterClass);
8447 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8448 X86::OR16ri, X86::MOV16rm,
8449 X86::LCMPXCHG16, X86::MOV16rr,
8450 X86::NOT16r, X86::AX,
8451 X86::GR16RegisterClass);
8452 case X86::ATOMXOR16:
8453 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8454 X86::XOR16ri, X86::MOV16rm,
8455 X86::LCMPXCHG16, X86::MOV16rr,
8456 X86::NOT16r, X86::AX,
8457 X86::GR16RegisterClass);
8458 case X86::ATOMNAND16:
8459 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8460 X86::AND16ri, X86::MOV16rm,
8461 X86::LCMPXCHG16, X86::MOV16rr,
8462 X86::NOT16r, X86::AX,
8463 X86::GR16RegisterClass, true);
8464 case X86::ATOMMIN16:
8465 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8466 case X86::ATOMMAX16:
8467 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8468 case X86::ATOMUMIN16:
8469 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8470 case X86::ATOMUMAX16:
8471 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8474 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8475 X86::AND8ri, X86::MOV8rm,
8476 X86::LCMPXCHG8, X86::MOV8rr,
8477 X86::NOT8r, X86::AL,
8478 X86::GR8RegisterClass);
8480 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8481 X86::OR8ri, X86::MOV8rm,
8482 X86::LCMPXCHG8, X86::MOV8rr,
8483 X86::NOT8r, X86::AL,
8484 X86::GR8RegisterClass);
8486 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8487 X86::XOR8ri, X86::MOV8rm,
8488 X86::LCMPXCHG8, X86::MOV8rr,
8489 X86::NOT8r, X86::AL,
8490 X86::GR8RegisterClass);
8491 case X86::ATOMNAND8:
8492 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8493 X86::AND8ri, X86::MOV8rm,
8494 X86::LCMPXCHG8, X86::MOV8rr,
8495 X86::NOT8r, X86::AL,
8496 X86::GR8RegisterClass, true);
8497 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8498 // This group is for 64-bit host.
8499 case X86::ATOMAND64:
8500 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8501 X86::AND64ri32, X86::MOV64rm,
8502 X86::LCMPXCHG64, X86::MOV64rr,
8503 X86::NOT64r, X86::RAX,
8504 X86::GR64RegisterClass);
8506 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8507 X86::OR64ri32, X86::MOV64rm,
8508 X86::LCMPXCHG64, X86::MOV64rr,
8509 X86::NOT64r, X86::RAX,
8510 X86::GR64RegisterClass);
8511 case X86::ATOMXOR64:
8512 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8513 X86::XOR64ri32, X86::MOV64rm,
8514 X86::LCMPXCHG64, X86::MOV64rr,
8515 X86::NOT64r, X86::RAX,
8516 X86::GR64RegisterClass);
8517 case X86::ATOMNAND64:
8518 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8519 X86::AND64ri32, X86::MOV64rm,
8520 X86::LCMPXCHG64, X86::MOV64rr,
8521 X86::NOT64r, X86::RAX,
8522 X86::GR64RegisterClass, true);
8523 case X86::ATOMMIN64:
8524 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8525 case X86::ATOMMAX64:
8526 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8527 case X86::ATOMUMIN64:
8528 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8529 case X86::ATOMUMAX64:
8530 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8532 // This group does 64-bit operations on a 32-bit host.
8533 case X86::ATOMAND6432:
8534 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8535 X86::AND32rr, X86::AND32rr,
8536 X86::AND32ri, X86::AND32ri,
8538 case X86::ATOMOR6432:
8539 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8540 X86::OR32rr, X86::OR32rr,
8541 X86::OR32ri, X86::OR32ri,
8543 case X86::ATOMXOR6432:
8544 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8545 X86::XOR32rr, X86::XOR32rr,
8546 X86::XOR32ri, X86::XOR32ri,
8548 case X86::ATOMNAND6432:
8549 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8550 X86::AND32rr, X86::AND32rr,
8551 X86::AND32ri, X86::AND32ri,
8553 case X86::ATOMADD6432:
8554 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8555 X86::ADD32rr, X86::ADC32rr,
8556 X86::ADD32ri, X86::ADC32ri,
8558 case X86::ATOMSUB6432:
8559 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8560 X86::SUB32rr, X86::SBB32rr,
8561 X86::SUB32ri, X86::SBB32ri,
8563 case X86::ATOMSWAP6432:
8564 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8565 X86::MOV32rr, X86::MOV32rr,
8566 X86::MOV32ri, X86::MOV32ri,
8568 case X86::VASTART_SAVE_XMM_REGS:
8569 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8573 //===----------------------------------------------------------------------===//
8574 // X86 Optimization Hooks
8575 //===----------------------------------------------------------------------===//
8577 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8581 const SelectionDAG &DAG,
8582 unsigned Depth) const {
8583 unsigned Opc = Op.getOpcode();
8584 assert((Opc >= ISD::BUILTIN_OP_END ||
8585 Opc == ISD::INTRINSIC_WO_CHAIN ||
8586 Opc == ISD::INTRINSIC_W_CHAIN ||
8587 Opc == ISD::INTRINSIC_VOID) &&
8588 "Should use MaskedValueIsZero if you don't know whether Op"
8589 " is a target node!");
8591 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8603 // These nodes' second result is a boolean.
8604 if (Op.getResNo() == 0)
8608 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8609 Mask.getBitWidth() - 1);
8614 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8615 /// node is a GlobalAddress + offset.
8616 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8617 GlobalValue* &GA, int64_t &Offset) const{
8618 if (N->getOpcode() == X86ISD::Wrapper) {
8619 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8620 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8621 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8625 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8628 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8629 EVT EltVT, LoadSDNode *&LDBase,
8630 unsigned &LastLoadedElt,
8631 SelectionDAG &DAG, MachineFrameInfo *MFI,
8632 const TargetLowering &TLI) {
8634 LastLoadedElt = -1U;
8635 for (unsigned i = 0; i < NumElems; ++i) {
8636 if (N->getMaskElt(i) < 0) {
8642 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8643 if (!Elt.getNode() ||
8644 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8647 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8649 LDBase = cast<LoadSDNode>(Elt.getNode());
8653 if (Elt.getOpcode() == ISD::UNDEF)
8656 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8657 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
8664 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8665 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8666 /// if the load addresses are consecutive, non-overlapping, and in the right
8667 /// order. In the case of v2i64, it will see if it can rewrite the
8668 /// shuffle to be an appropriate build vector so it can take advantage of
8669 // performBuildVectorCombine.
8670 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8671 const TargetLowering &TLI) {
8672 DebugLoc dl = N->getDebugLoc();
8673 EVT VT = N->getValueType(0);
8674 EVT EltVT = VT.getVectorElementType();
8675 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8676 unsigned NumElems = VT.getVectorNumElements();
8678 if (VT.getSizeInBits() != 128)
8681 // Try to combine a vector_shuffle into a 128-bit load.
8682 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8683 LoadSDNode *LD = NULL;
8684 unsigned LastLoadedElt;
8685 if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
8689 if (LastLoadedElt == NumElems - 1) {
8690 if (DAG.InferPtrAlignment(LD->getBasePtr()) >= 16)
8691 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8692 LD->getSrcValue(), LD->getSrcValueOffset(),
8694 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8695 LD->getSrcValue(), LD->getSrcValueOffset(),
8696 LD->isVolatile(), LD->getAlignment());
8697 } else if (NumElems == 4 && LastLoadedElt == 1) {
8698 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8699 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8700 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8701 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8706 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8707 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8708 const X86Subtarget *Subtarget) {
8709 DebugLoc DL = N->getDebugLoc();
8710 SDValue Cond = N->getOperand(0);
8711 // Get the LHS/RHS of the select.
8712 SDValue LHS = N->getOperand(1);
8713 SDValue RHS = N->getOperand(2);
8715 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8716 // instructions have the peculiarity that if either operand is a NaN,
8717 // they chose what we call the RHS operand (and as such are not symmetric).
8718 // It happens that this matches the semantics of the common C idiom
8719 // x<y?x:y and related forms, so we can recognize these cases.
8720 if (Subtarget->hasSSE2() &&
8721 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8722 Cond.getOpcode() == ISD::SETCC) {
8723 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8725 unsigned Opcode = 0;
8726 // Check for x CC y ? x : y.
8727 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8731 // This can be a min if we can prove that at least one of the operands
8733 if (!FiniteOnlyFPMath()) {
8734 if (DAG.isKnownNeverNaN(RHS)) {
8735 // Put the potential NaN in the RHS so that SSE will preserve it.
8736 std::swap(LHS, RHS);
8737 } else if (!DAG.isKnownNeverNaN(LHS))
8740 Opcode = X86ISD::FMIN;
8743 // This can be a min if we can prove that at least one of the operands
8745 if (!FiniteOnlyFPMath()) {
8746 if (DAG.isKnownNeverNaN(LHS)) {
8747 // Put the potential NaN in the RHS so that SSE will preserve it.
8748 std::swap(LHS, RHS);
8749 } else if (!DAG.isKnownNeverNaN(RHS))
8752 Opcode = X86ISD::FMIN;
8755 // This can be a min, but if either operand is a NaN we need it to
8756 // preserve the original LHS.
8757 std::swap(LHS, RHS);
8761 Opcode = X86ISD::FMIN;
8765 // This can be a max if we can prove that at least one of the operands
8767 if (!FiniteOnlyFPMath()) {
8768 if (DAG.isKnownNeverNaN(LHS)) {
8769 // Put the potential NaN in the RHS so that SSE will preserve it.
8770 std::swap(LHS, RHS);
8771 } else if (!DAG.isKnownNeverNaN(RHS))
8774 Opcode = X86ISD::FMAX;
8777 // This can be a max if we can prove that at least one of the operands
8779 if (!FiniteOnlyFPMath()) {
8780 if (DAG.isKnownNeverNaN(RHS)) {
8781 // Put the potential NaN in the RHS so that SSE will preserve it.
8782 std::swap(LHS, RHS);
8783 } else if (!DAG.isKnownNeverNaN(LHS))
8786 Opcode = X86ISD::FMAX;
8789 // This can be a max, but if either operand is a NaN we need it to
8790 // preserve the original LHS.
8791 std::swap(LHS, RHS);
8795 Opcode = X86ISD::FMAX;
8798 // Check for x CC y ? y : x -- a min/max with reversed arms.
8799 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8803 // This can be a min if we can prove that at least one of the operands
8805 if (!FiniteOnlyFPMath()) {
8806 if (DAG.isKnownNeverNaN(RHS)) {
8807 // Put the potential NaN in the RHS so that SSE will preserve it.
8808 std::swap(LHS, RHS);
8809 } else if (!DAG.isKnownNeverNaN(LHS))
8812 Opcode = X86ISD::FMIN;
8815 // This can be a min if we can prove that at least one of the operands
8817 if (!FiniteOnlyFPMath()) {
8818 if (DAG.isKnownNeverNaN(LHS)) {
8819 // Put the potential NaN in the RHS so that SSE will preserve it.
8820 std::swap(LHS, RHS);
8821 } else if (!DAG.isKnownNeverNaN(RHS))
8824 Opcode = X86ISD::FMIN;
8827 // This can be a min, but if either operand is a NaN we need it to
8828 // preserve the original LHS.
8829 std::swap(LHS, RHS);
8833 Opcode = X86ISD::FMIN;
8837 // This can be a max if we can prove that at least one of the operands
8839 if (!FiniteOnlyFPMath()) {
8840 if (DAG.isKnownNeverNaN(LHS)) {
8841 // Put the potential NaN in the RHS so that SSE will preserve it.
8842 std::swap(LHS, RHS);
8843 } else if (!DAG.isKnownNeverNaN(RHS))
8846 Opcode = X86ISD::FMAX;
8849 // This can be a max if we can prove that at least one of the operands
8851 if (!FiniteOnlyFPMath()) {
8852 if (DAG.isKnownNeverNaN(RHS)) {
8853 // Put the potential NaN in the RHS so that SSE will preserve it.
8854 std::swap(LHS, RHS);
8855 } else if (!DAG.isKnownNeverNaN(LHS))
8858 Opcode = X86ISD::FMAX;
8861 // This can be a max, but if either operand is a NaN we need it to
8862 // preserve the original LHS.
8863 std::swap(LHS, RHS);
8867 Opcode = X86ISD::FMAX;
8873 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8876 // If this is a select between two integer constants, try to do some
8878 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8879 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8880 // Don't do this for crazy integer types.
8881 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8882 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8883 // so that TrueC (the true value) is larger than FalseC.
8884 bool NeedsCondInvert = false;
8886 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8887 // Efficiently invertible.
8888 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8889 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8890 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8891 NeedsCondInvert = true;
8892 std::swap(TrueC, FalseC);
8895 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8896 if (FalseC->getAPIntValue() == 0 &&
8897 TrueC->getAPIntValue().isPowerOf2()) {
8898 if (NeedsCondInvert) // Invert the condition if needed.
8899 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8900 DAG.getConstant(1, Cond.getValueType()));
8902 // Zero extend the condition if needed.
8903 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8905 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8906 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8907 DAG.getConstant(ShAmt, MVT::i8));
8910 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8911 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8912 if (NeedsCondInvert) // Invert the condition if needed.
8913 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8914 DAG.getConstant(1, Cond.getValueType()));
8916 // Zero extend the condition if needed.
8917 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8918 FalseC->getValueType(0), Cond);
8919 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8920 SDValue(FalseC, 0));
8923 // Optimize cases that will turn into an LEA instruction. This requires
8924 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8925 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8926 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8927 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8929 bool isFastMultiplier = false;
8931 switch ((unsigned char)Diff) {
8933 case 1: // result = add base, cond
8934 case 2: // result = lea base( , cond*2)
8935 case 3: // result = lea base(cond, cond*2)
8936 case 4: // result = lea base( , cond*4)
8937 case 5: // result = lea base(cond, cond*4)
8938 case 8: // result = lea base( , cond*8)
8939 case 9: // result = lea base(cond, cond*8)
8940 isFastMultiplier = true;
8945 if (isFastMultiplier) {
8946 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8947 if (NeedsCondInvert) // Invert the condition if needed.
8948 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8949 DAG.getConstant(1, Cond.getValueType()));
8951 // Zero extend the condition if needed.
8952 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8954 // Scale the condition by the difference.
8956 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8957 DAG.getConstant(Diff, Cond.getValueType()));
8959 // Add the base if non-zero.
8960 if (FalseC->getAPIntValue() != 0)
8961 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8962 SDValue(FalseC, 0));
8972 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
8973 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8974 TargetLowering::DAGCombinerInfo &DCI) {
8975 DebugLoc DL = N->getDebugLoc();
8977 // If the flag operand isn't dead, don't touch this CMOV.
8978 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8981 // If this is a select between two integer constants, try to do some
8982 // optimizations. Note that the operands are ordered the opposite of SELECT
8984 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
8985 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8986 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
8987 // larger than FalseC (the false value).
8988 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
8990 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
8991 CC = X86::GetOppositeBranchCondition(CC);
8992 std::swap(TrueC, FalseC);
8995 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
8996 // This is efficient for any integer data type (including i8/i16) and
8998 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
8999 SDValue Cond = N->getOperand(3);
9000 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9001 DAG.getConstant(CC, MVT::i8), Cond);
9003 // Zero extend the condition if needed.
9004 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9006 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9007 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9008 DAG.getConstant(ShAmt, MVT::i8));
9009 if (N->getNumValues() == 2) // Dead flag value?
9010 return DCI.CombineTo(N, Cond, SDValue());
9014 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9015 // for any integer data type, including i8/i16.
9016 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9017 SDValue Cond = N->getOperand(3);
9018 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9019 DAG.getConstant(CC, MVT::i8), Cond);
9021 // Zero extend the condition if needed.
9022 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9023 FalseC->getValueType(0), Cond);
9024 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9025 SDValue(FalseC, 0));
9027 if (N->getNumValues() == 2) // Dead flag value?
9028 return DCI.CombineTo(N, Cond, SDValue());
9032 // Optimize cases that will turn into an LEA instruction. This requires
9033 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9034 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9035 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9036 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9038 bool isFastMultiplier = false;
9040 switch ((unsigned char)Diff) {
9042 case 1: // result = add base, cond
9043 case 2: // result = lea base( , cond*2)
9044 case 3: // result = lea base(cond, cond*2)
9045 case 4: // result = lea base( , cond*4)
9046 case 5: // result = lea base(cond, cond*4)
9047 case 8: // result = lea base( , cond*8)
9048 case 9: // result = lea base(cond, cond*8)
9049 isFastMultiplier = true;
9054 if (isFastMultiplier) {
9055 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9056 SDValue Cond = N->getOperand(3);
9057 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9058 DAG.getConstant(CC, MVT::i8), Cond);
9059 // Zero extend the condition if needed.
9060 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9062 // Scale the condition by the difference.
9064 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9065 DAG.getConstant(Diff, Cond.getValueType()));
9067 // Add the base if non-zero.
9068 if (FalseC->getAPIntValue() != 0)
9069 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9070 SDValue(FalseC, 0));
9071 if (N->getNumValues() == 2) // Dead flag value?
9072 return DCI.CombineTo(N, Cond, SDValue());
9082 /// PerformMulCombine - Optimize a single multiply with constant into two
9083 /// in order to implement it with two cheaper instructions, e.g.
9084 /// LEA + SHL, LEA + LEA.
9085 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9086 TargetLowering::DAGCombinerInfo &DCI) {
9087 if (DAG.getMachineFunction().
9088 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
9091 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9094 EVT VT = N->getValueType(0);
9098 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9101 uint64_t MulAmt = C->getZExtValue();
9102 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9105 uint64_t MulAmt1 = 0;
9106 uint64_t MulAmt2 = 0;
9107 if ((MulAmt % 9) == 0) {
9109 MulAmt2 = MulAmt / 9;
9110 } else if ((MulAmt % 5) == 0) {
9112 MulAmt2 = MulAmt / 5;
9113 } else if ((MulAmt % 3) == 0) {
9115 MulAmt2 = MulAmt / 3;
9118 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9119 DebugLoc DL = N->getDebugLoc();
9121 if (isPowerOf2_64(MulAmt2) &&
9122 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9123 // If second multiplifer is pow2, issue it first. We want the multiply by
9124 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9126 std::swap(MulAmt1, MulAmt2);
9129 if (isPowerOf2_64(MulAmt1))
9130 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9131 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9133 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9134 DAG.getConstant(MulAmt1, VT));
9136 if (isPowerOf2_64(MulAmt2))
9137 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9138 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9140 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9141 DAG.getConstant(MulAmt2, VT));
9143 // Do not add new nodes to DAG combiner worklist.
9144 DCI.CombineTo(N, NewMul, false);
9149 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9150 SDValue N0 = N->getOperand(0);
9151 SDValue N1 = N->getOperand(1);
9152 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9153 EVT VT = N0.getValueType();
9155 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9156 // since the result of setcc_c is all zero's or all ones.
9157 if (N1C && N0.getOpcode() == ISD::AND &&
9158 N0.getOperand(1).getOpcode() == ISD::Constant) {
9159 SDValue N00 = N0.getOperand(0);
9160 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9161 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9162 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9163 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9164 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9165 APInt ShAmt = N1C->getAPIntValue();
9166 Mask = Mask.shl(ShAmt);
9168 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9169 N00, DAG.getConstant(Mask, VT));
9176 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9178 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9179 const X86Subtarget *Subtarget) {
9180 EVT VT = N->getValueType(0);
9181 if (!VT.isVector() && VT.isInteger() &&
9182 N->getOpcode() == ISD::SHL)
9183 return PerformSHLCombine(N, DAG);
9185 // On X86 with SSE2 support, we can transform this to a vector shift if
9186 // all elements are shifted by the same amount. We can't do this in legalize
9187 // because the a constant vector is typically transformed to a constant pool
9188 // so we have no knowledge of the shift amount.
9189 if (!Subtarget->hasSSE2())
9192 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9195 SDValue ShAmtOp = N->getOperand(1);
9196 EVT EltVT = VT.getVectorElementType();
9197 DebugLoc DL = N->getDebugLoc();
9198 SDValue BaseShAmt = SDValue();
9199 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9200 unsigned NumElts = VT.getVectorNumElements();
9202 for (; i != NumElts; ++i) {
9203 SDValue Arg = ShAmtOp.getOperand(i);
9204 if (Arg.getOpcode() == ISD::UNDEF) continue;
9208 for (; i != NumElts; ++i) {
9209 SDValue Arg = ShAmtOp.getOperand(i);
9210 if (Arg.getOpcode() == ISD::UNDEF) continue;
9211 if (Arg != BaseShAmt) {
9215 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9216 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9217 SDValue InVec = ShAmtOp.getOperand(0);
9218 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9219 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9221 for (; i != NumElts; ++i) {
9222 SDValue Arg = InVec.getOperand(i);
9223 if (Arg.getOpcode() == ISD::UNDEF) continue;
9227 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9228 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9229 unsigned SplatIdx = cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9230 if (C->getZExtValue() == SplatIdx)
9231 BaseShAmt = InVec.getOperand(1);
9234 if (BaseShAmt.getNode() == 0)
9235 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9236 DAG.getIntPtrConstant(0));
9240 // The shift amount is an i32.
9241 if (EltVT.bitsGT(MVT::i32))
9242 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9243 else if (EltVT.bitsLT(MVT::i32))
9244 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9246 // The shift amount is identical so we can do a vector shift.
9247 SDValue ValOp = N->getOperand(0);
9248 switch (N->getOpcode()) {
9250 llvm_unreachable("Unknown shift opcode!");
9253 if (VT == MVT::v2i64)
9254 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9255 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9257 if (VT == MVT::v4i32)
9258 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9259 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9261 if (VT == MVT::v8i16)
9262 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9263 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9267 if (VT == MVT::v4i32)
9268 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9269 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9271 if (VT == MVT::v8i16)
9272 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9273 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9277 if (VT == MVT::v2i64)
9278 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9279 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9281 if (VT == MVT::v4i32)
9282 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9283 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9285 if (VT == MVT::v8i16)
9286 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9287 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9294 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9295 const X86Subtarget *Subtarget) {
9296 EVT VT = N->getValueType(0);
9297 if (VT != MVT::i64 || !Subtarget->is64Bit())
9300 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9301 SDValue N0 = N->getOperand(0);
9302 SDValue N1 = N->getOperand(1);
9303 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9305 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9308 SDValue ShAmt0 = N0.getOperand(1);
9309 if (ShAmt0.getValueType() != MVT::i8)
9311 SDValue ShAmt1 = N1.getOperand(1);
9312 if (ShAmt1.getValueType() != MVT::i8)
9314 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9315 ShAmt0 = ShAmt0.getOperand(0);
9316 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9317 ShAmt1 = ShAmt1.getOperand(0);
9319 DebugLoc DL = N->getDebugLoc();
9320 unsigned Opc = X86ISD::SHLD;
9321 SDValue Op0 = N0.getOperand(0);
9322 SDValue Op1 = N1.getOperand(0);
9323 if (ShAmt0.getOpcode() == ISD::SUB) {
9325 std::swap(Op0, Op1);
9326 std::swap(ShAmt0, ShAmt1);
9329 if (ShAmt1.getOpcode() == ISD::SUB) {
9330 SDValue Sum = ShAmt1.getOperand(0);
9331 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9332 if (SumC->getSExtValue() == 64 &&
9333 ShAmt1.getOperand(1) == ShAmt0)
9334 return DAG.getNode(Opc, DL, VT,
9336 DAG.getNode(ISD::TRUNCATE, DL,
9339 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9340 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9342 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
9343 return DAG.getNode(Opc, DL, VT,
9344 N0.getOperand(0), N1.getOperand(0),
9345 DAG.getNode(ISD::TRUNCATE, DL,
9352 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9353 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9354 const X86Subtarget *Subtarget) {
9355 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9356 // the FP state in cases where an emms may be missing.
9357 // A preferable solution to the general problem is to figure out the right
9358 // places to insert EMMS. This qualifies as a quick hack.
9360 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9361 StoreSDNode *St = cast<StoreSDNode>(N);
9362 EVT VT = St->getValue().getValueType();
9363 if (VT.getSizeInBits() != 64)
9366 const Function *F = DAG.getMachineFunction().getFunction();
9367 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9368 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9369 && Subtarget->hasSSE2();
9370 if ((VT.isVector() ||
9371 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9372 isa<LoadSDNode>(St->getValue()) &&
9373 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9374 St->getChain().hasOneUse() && !St->isVolatile()) {
9375 SDNode* LdVal = St->getValue().getNode();
9377 int TokenFactorIndex = -1;
9378 SmallVector<SDValue, 8> Ops;
9379 SDNode* ChainVal = St->getChain().getNode();
9380 // Must be a store of a load. We currently handle two cases: the load
9381 // is a direct child, and it's under an intervening TokenFactor. It is
9382 // possible to dig deeper under nested TokenFactors.
9383 if (ChainVal == LdVal)
9384 Ld = cast<LoadSDNode>(St->getChain());
9385 else if (St->getValue().hasOneUse() &&
9386 ChainVal->getOpcode() == ISD::TokenFactor) {
9387 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9388 if (ChainVal->getOperand(i).getNode() == LdVal) {
9389 TokenFactorIndex = i;
9390 Ld = cast<LoadSDNode>(St->getValue());
9392 Ops.push_back(ChainVal->getOperand(i));
9396 if (!Ld || !ISD::isNormalLoad(Ld))
9399 // If this is not the MMX case, i.e. we are just turning i64 load/store
9400 // into f64 load/store, avoid the transformation if there are multiple
9401 // uses of the loaded value.
9402 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9405 DebugLoc LdDL = Ld->getDebugLoc();
9406 DebugLoc StDL = N->getDebugLoc();
9407 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9408 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9410 if (Subtarget->is64Bit() || F64IsLegal) {
9411 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9412 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9413 Ld->getBasePtr(), Ld->getSrcValue(),
9414 Ld->getSrcValueOffset(), Ld->isVolatile(),
9415 Ld->getAlignment());
9416 SDValue NewChain = NewLd.getValue(1);
9417 if (TokenFactorIndex != -1) {
9418 Ops.push_back(NewChain);
9419 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9422 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9423 St->getSrcValue(), St->getSrcValueOffset(),
9424 St->isVolatile(), St->getAlignment());
9427 // Otherwise, lower to two pairs of 32-bit loads / stores.
9428 SDValue LoAddr = Ld->getBasePtr();
9429 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9430 DAG.getConstant(4, MVT::i32));
9432 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9433 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9434 Ld->isVolatile(), Ld->getAlignment());
9435 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9436 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9438 MinAlign(Ld->getAlignment(), 4));
9440 SDValue NewChain = LoLd.getValue(1);
9441 if (TokenFactorIndex != -1) {
9442 Ops.push_back(LoLd);
9443 Ops.push_back(HiLd);
9444 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9448 LoAddr = St->getBasePtr();
9449 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9450 DAG.getConstant(4, MVT::i32));
9452 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9453 St->getSrcValue(), St->getSrcValueOffset(),
9454 St->isVolatile(), St->getAlignment());
9455 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9457 St->getSrcValueOffset() + 4,
9459 MinAlign(St->getAlignment(), 4));
9460 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9465 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9466 /// X86ISD::FXOR nodes.
9467 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9468 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9469 // F[X]OR(0.0, x) -> x
9470 // F[X]OR(x, 0.0) -> x
9471 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9472 if (C->getValueAPF().isPosZero())
9473 return N->getOperand(1);
9474 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9475 if (C->getValueAPF().isPosZero())
9476 return N->getOperand(0);
9480 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9481 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9482 // FAND(0.0, x) -> 0.0
9483 // FAND(x, 0.0) -> 0.0
9484 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9485 if (C->getValueAPF().isPosZero())
9486 return N->getOperand(0);
9487 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9488 if (C->getValueAPF().isPosZero())
9489 return N->getOperand(1);
9493 static SDValue PerformBTCombine(SDNode *N,
9495 TargetLowering::DAGCombinerInfo &DCI) {
9496 // BT ignores high bits in the bit index operand.
9497 SDValue Op1 = N->getOperand(1);
9498 if (Op1.hasOneUse()) {
9499 unsigned BitWidth = Op1.getValueSizeInBits();
9500 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9501 APInt KnownZero, KnownOne;
9502 TargetLowering::TargetLoweringOpt TLO(DAG);
9503 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9504 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9505 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9506 DCI.CommitTargetLoweringOpt(TLO);
9511 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9512 SDValue Op = N->getOperand(0);
9513 if (Op.getOpcode() == ISD::BIT_CONVERT)
9514 Op = Op.getOperand(0);
9515 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9516 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9517 VT.getVectorElementType().getSizeInBits() ==
9518 OpVT.getVectorElementType().getSizeInBits()) {
9519 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9524 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9525 // Locked instructions, in turn, have implicit fence semantics (all memory
9526 // operations are flushed before issuing the locked instruction, and the
9527 // are not buffered), so we can fold away the common pattern of
9528 // fence-atomic-fence.
9529 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9530 SDValue atomic = N->getOperand(0);
9531 switch (atomic.getOpcode()) {
9532 case ISD::ATOMIC_CMP_SWAP:
9533 case ISD::ATOMIC_SWAP:
9534 case ISD::ATOMIC_LOAD_ADD:
9535 case ISD::ATOMIC_LOAD_SUB:
9536 case ISD::ATOMIC_LOAD_AND:
9537 case ISD::ATOMIC_LOAD_OR:
9538 case ISD::ATOMIC_LOAD_XOR:
9539 case ISD::ATOMIC_LOAD_NAND:
9540 case ISD::ATOMIC_LOAD_MIN:
9541 case ISD::ATOMIC_LOAD_MAX:
9542 case ISD::ATOMIC_LOAD_UMIN:
9543 case ISD::ATOMIC_LOAD_UMAX:
9549 SDValue fence = atomic.getOperand(0);
9550 if (fence.getOpcode() != ISD::MEMBARRIER)
9553 switch (atomic.getOpcode()) {
9554 case ISD::ATOMIC_CMP_SWAP:
9555 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9556 atomic.getOperand(1), atomic.getOperand(2),
9557 atomic.getOperand(3));
9558 case ISD::ATOMIC_SWAP:
9559 case ISD::ATOMIC_LOAD_ADD:
9560 case ISD::ATOMIC_LOAD_SUB:
9561 case ISD::ATOMIC_LOAD_AND:
9562 case ISD::ATOMIC_LOAD_OR:
9563 case ISD::ATOMIC_LOAD_XOR:
9564 case ISD::ATOMIC_LOAD_NAND:
9565 case ISD::ATOMIC_LOAD_MIN:
9566 case ISD::ATOMIC_LOAD_MAX:
9567 case ISD::ATOMIC_LOAD_UMIN:
9568 case ISD::ATOMIC_LOAD_UMAX:
9569 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9570 atomic.getOperand(1), atomic.getOperand(2));
9576 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9577 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9578 // (and (i32 x86isd::setcc_carry), 1)
9579 // This eliminates the zext. This transformation is necessary because
9580 // ISD::SETCC is always legalized to i8.
9581 DebugLoc dl = N->getDebugLoc();
9582 SDValue N0 = N->getOperand(0);
9583 EVT VT = N->getValueType(0);
9584 if (N0.getOpcode() == ISD::AND &&
9586 N0.getOperand(0).hasOneUse()) {
9587 SDValue N00 = N0.getOperand(0);
9588 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9590 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9591 if (!C || C->getZExtValue() != 1)
9593 return DAG.getNode(ISD::AND, dl, VT,
9594 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9595 N00.getOperand(0), N00.getOperand(1)),
9596 DAG.getConstant(1, VT));
9602 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9603 DAGCombinerInfo &DCI) const {
9604 SelectionDAG &DAG = DCI.DAG;
9605 switch (N->getOpcode()) {
9607 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9608 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9609 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9610 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9613 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9614 case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
9615 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9617 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9618 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9619 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9620 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9621 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9622 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9628 //===----------------------------------------------------------------------===//
9629 // X86 Inline Assembly Support
9630 //===----------------------------------------------------------------------===//
9632 static bool LowerToBSwap(CallInst *CI) {
9633 // FIXME: this should verify that we are targetting a 486 or better. If not,
9634 // we will turn this bswap into something that will be lowered to logical ops
9635 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9636 // so don't worry about this.
9638 // Verify this is a simple bswap.
9639 if (CI->getNumOperands() != 2 ||
9640 CI->getType() != CI->getOperand(1)->getType() ||
9641 !CI->getType()->isInteger())
9644 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9645 if (!Ty || Ty->getBitWidth() % 16 != 0)
9648 // Okay, we can do this xform, do so now.
9649 const Type *Tys[] = { Ty };
9650 Module *M = CI->getParent()->getParent()->getParent();
9651 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9653 Value *Op = CI->getOperand(1);
9654 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9656 CI->replaceAllUsesWith(Op);
9657 CI->eraseFromParent();
9661 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9662 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9663 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9665 std::string AsmStr = IA->getAsmString();
9667 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9668 SmallVector<StringRef, 4> AsmPieces;
9669 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9671 switch (AsmPieces.size()) {
9672 default: return false;
9674 AsmStr = AsmPieces[0];
9676 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9679 if (AsmPieces.size() == 2 &&
9680 (AsmPieces[0] == "bswap" ||
9681 AsmPieces[0] == "bswapq" ||
9682 AsmPieces[0] == "bswapl") &&
9683 (AsmPieces[1] == "$0" ||
9684 AsmPieces[1] == "${0:q}")) {
9685 // No need to check constraints, nothing other than the equivalent of
9686 // "=r,0" would be valid here.
9687 return LowerToBSwap(CI);
9689 // rorw $$8, ${0:w} --> llvm.bswap.i16
9690 if (CI->getType()->isInteger(16) &&
9691 AsmPieces.size() == 3 &&
9692 AsmPieces[0] == "rorw" &&
9693 AsmPieces[1] == "$$8," &&
9694 AsmPieces[2] == "${0:w}" &&
9695 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
9696 return LowerToBSwap(CI);
9700 if (CI->getType()->isInteger(64) &&
9701 Constraints.size() >= 2 &&
9702 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9703 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9704 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9705 SmallVector<StringRef, 4> Words;
9706 SplitString(AsmPieces[0], Words, " \t");
9707 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9709 SplitString(AsmPieces[1], Words, " \t");
9710 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9712 SplitString(AsmPieces[2], Words, " \t,");
9713 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9714 Words[2] == "%edx") {
9715 return LowerToBSwap(CI);
9727 /// getConstraintType - Given a constraint letter, return the type of
9728 /// constraint it is for this target.
9729 X86TargetLowering::ConstraintType
9730 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
9731 if (Constraint.size() == 1) {
9732 switch (Constraint[0]) {
9744 return C_RegisterClass;
9752 return TargetLowering::getConstraintType(Constraint);
9755 /// LowerXConstraint - try to replace an X constraint, which matches anything,
9756 /// with another that has more specific requirements based on the type of the
9757 /// corresponding operand.
9758 const char *X86TargetLowering::
9759 LowerXConstraint(EVT ConstraintVT) const {
9760 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
9761 // 'f' like normal targets.
9762 if (ConstraintVT.isFloatingPoint()) {
9763 if (Subtarget->hasSSE2())
9765 if (Subtarget->hasSSE1())
9769 return TargetLowering::LowerXConstraint(ConstraintVT);
9772 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9773 /// vector. If it is invalid, don't add anything to Ops.
9774 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9777 std::vector<SDValue>&Ops,
9778 SelectionDAG &DAG) const {
9779 SDValue Result(0, 0);
9781 switch (Constraint) {
9784 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9785 if (C->getZExtValue() <= 31) {
9786 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9792 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9793 if (C->getZExtValue() <= 63) {
9794 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9800 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9801 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9802 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9808 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9809 if (C->getZExtValue() <= 255) {
9810 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9816 // 32-bit signed value
9817 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9818 const ConstantInt *CI = C->getConstantIntValue();
9819 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9820 C->getSExtValue())) {
9821 // Widen to 64 bits here to get it sign extended.
9822 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9825 // FIXME gcc accepts some relocatable values here too, but only in certain
9826 // memory models; it's complicated.
9831 // 32-bit unsigned value
9832 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9833 const ConstantInt *CI = C->getConstantIntValue();
9834 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9835 C->getZExtValue())) {
9836 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9840 // FIXME gcc accepts some relocatable values here too, but only in certain
9841 // memory models; it's complicated.
9845 // Literal immediates are always ok.
9846 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
9847 // Widen to 64 bits here to get it sign extended.
9848 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
9852 // If we are in non-pic codegen mode, we allow the address of a global (with
9853 // an optional displacement) to be used with 'i'.
9854 GlobalAddressSDNode *GA = 0;
9857 // Match either (GA), (GA+C), (GA+C1+C2), etc.
9859 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
9860 Offset += GA->getOffset();
9862 } else if (Op.getOpcode() == ISD::ADD) {
9863 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9864 Offset += C->getZExtValue();
9865 Op = Op.getOperand(0);
9868 } else if (Op.getOpcode() == ISD::SUB) {
9869 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9870 Offset += -C->getZExtValue();
9871 Op = Op.getOperand(0);
9876 // Otherwise, this isn't something we can handle, reject it.
9880 GlobalValue *GV = GA->getGlobal();
9881 // If we require an extra load to get this address, as in PIC mode, we
9883 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
9884 getTargetMachine())))
9888 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
9890 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
9896 if (Result.getNode()) {
9897 Ops.push_back(Result);
9900 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
9904 std::vector<unsigned> X86TargetLowering::
9905 getRegClassForInlineAsmConstraint(const std::string &Constraint,
9907 if (Constraint.size() == 1) {
9908 // FIXME: not handling fp-stack yet!
9909 switch (Constraint[0]) { // GCC X86 Constraint Letters
9910 default: break; // Unknown constraint letter
9911 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
9912 if (Subtarget->is64Bit()) {
9914 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
9915 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
9916 X86::R10D,X86::R11D,X86::R12D,
9917 X86::R13D,X86::R14D,X86::R15D,
9918 X86::EBP, X86::ESP, 0);
9919 else if (VT == MVT::i16)
9920 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
9921 X86::SI, X86::DI, X86::R8W,X86::R9W,
9922 X86::R10W,X86::R11W,X86::R12W,
9923 X86::R13W,X86::R14W,X86::R15W,
9924 X86::BP, X86::SP, 0);
9925 else if (VT == MVT::i8)
9926 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
9927 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
9928 X86::R10B,X86::R11B,X86::R12B,
9929 X86::R13B,X86::R14B,X86::R15B,
9930 X86::BPL, X86::SPL, 0);
9932 else if (VT == MVT::i64)
9933 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
9934 X86::RSI, X86::RDI, X86::R8, X86::R9,
9935 X86::R10, X86::R11, X86::R12,
9936 X86::R13, X86::R14, X86::R15,
9937 X86::RBP, X86::RSP, 0);
9941 // 32-bit fallthrough
9944 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
9945 else if (VT == MVT::i16)
9946 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
9947 else if (VT == MVT::i8)
9948 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
9949 else if (VT == MVT::i64)
9950 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
9955 return std::vector<unsigned>();
9958 std::pair<unsigned, const TargetRegisterClass*>
9959 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9961 // First, see if this is a constraint that directly corresponds to an LLVM
9963 if (Constraint.size() == 1) {
9964 // GCC Constraint Letters
9965 switch (Constraint[0]) {
9967 case 'r': // GENERAL_REGS
9968 case 'l': // INDEX_REGS
9970 return std::make_pair(0U, X86::GR8RegisterClass);
9972 return std::make_pair(0U, X86::GR16RegisterClass);
9973 if (VT == MVT::i32 || !Subtarget->is64Bit())
9974 return std::make_pair(0U, X86::GR32RegisterClass);
9975 return std::make_pair(0U, X86::GR64RegisterClass);
9976 case 'R': // LEGACY_REGS
9978 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
9980 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
9981 if (VT == MVT::i32 || !Subtarget->is64Bit())
9982 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
9983 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
9984 case 'f': // FP Stack registers.
9985 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
9986 // value to the correct fpstack register class.
9987 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
9988 return std::make_pair(0U, X86::RFP32RegisterClass);
9989 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
9990 return std::make_pair(0U, X86::RFP64RegisterClass);
9991 return std::make_pair(0U, X86::RFP80RegisterClass);
9992 case 'y': // MMX_REGS if MMX allowed.
9993 if (!Subtarget->hasMMX()) break;
9994 return std::make_pair(0U, X86::VR64RegisterClass);
9995 case 'Y': // SSE_REGS if SSE2 allowed
9996 if (!Subtarget->hasSSE2()) break;
9998 case 'x': // SSE_REGS if SSE1 allowed
9999 if (!Subtarget->hasSSE1()) break;
10001 switch (VT.getSimpleVT().SimpleTy) {
10003 // Scalar SSE types.
10006 return std::make_pair(0U, X86::FR32RegisterClass);
10009 return std::make_pair(0U, X86::FR64RegisterClass);
10017 return std::make_pair(0U, X86::VR128RegisterClass);
10023 // Use the default implementation in TargetLowering to convert the register
10024 // constraint into a member of a register class.
10025 std::pair<unsigned, const TargetRegisterClass*> Res;
10026 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10028 // Not found as a standard register?
10029 if (Res.second == 0) {
10030 // Map st(0) -> st(7) -> ST0
10031 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10032 tolower(Constraint[1]) == 's' &&
10033 tolower(Constraint[2]) == 't' &&
10034 Constraint[3] == '(' &&
10035 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10036 Constraint[5] == ')' &&
10037 Constraint[6] == '}') {
10039 Res.first = X86::ST0+Constraint[4]-'0';
10040 Res.second = X86::RFP80RegisterClass;
10044 // GCC allows "st(0)" to be called just plain "st".
10045 if (StringRef("{st}").equals_lower(Constraint)) {
10046 Res.first = X86::ST0;
10047 Res.second = X86::RFP80RegisterClass;
10052 if (StringRef("{flags}").equals_lower(Constraint)) {
10053 Res.first = X86::EFLAGS;
10054 Res.second = X86::CCRRegisterClass;
10058 // 'A' means EAX + EDX.
10059 if (Constraint == "A") {
10060 Res.first = X86::EAX;
10061 Res.second = X86::GR32_ADRegisterClass;
10067 // Otherwise, check to see if this is a register class of the wrong value
10068 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10069 // turn into {ax},{dx}.
10070 if (Res.second->hasType(VT))
10071 return Res; // Correct type already, nothing to do.
10073 // All of the single-register GCC register classes map their values onto
10074 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10075 // really want an 8-bit or 32-bit register, map to the appropriate register
10076 // class and return the appropriate register.
10077 if (Res.second == X86::GR16RegisterClass) {
10078 if (VT == MVT::i8) {
10079 unsigned DestReg = 0;
10080 switch (Res.first) {
10082 case X86::AX: DestReg = X86::AL; break;
10083 case X86::DX: DestReg = X86::DL; break;
10084 case X86::CX: DestReg = X86::CL; break;
10085 case X86::BX: DestReg = X86::BL; break;
10088 Res.first = DestReg;
10089 Res.second = X86::GR8RegisterClass;
10091 } else if (VT == MVT::i32) {
10092 unsigned DestReg = 0;
10093 switch (Res.first) {
10095 case X86::AX: DestReg = X86::EAX; break;
10096 case X86::DX: DestReg = X86::EDX; break;
10097 case X86::CX: DestReg = X86::ECX; break;
10098 case X86::BX: DestReg = X86::EBX; break;
10099 case X86::SI: DestReg = X86::ESI; break;
10100 case X86::DI: DestReg = X86::EDI; break;
10101 case X86::BP: DestReg = X86::EBP; break;
10102 case X86::SP: DestReg = X86::ESP; break;
10105 Res.first = DestReg;
10106 Res.second = X86::GR32RegisterClass;
10108 } else if (VT == MVT::i64) {
10109 unsigned DestReg = 0;
10110 switch (Res.first) {
10112 case X86::AX: DestReg = X86::RAX; break;
10113 case X86::DX: DestReg = X86::RDX; break;
10114 case X86::CX: DestReg = X86::RCX; break;
10115 case X86::BX: DestReg = X86::RBX; break;
10116 case X86::SI: DestReg = X86::RSI; break;
10117 case X86::DI: DestReg = X86::RDI; break;
10118 case X86::BP: DestReg = X86::RBP; break;
10119 case X86::SP: DestReg = X86::RSP; break;
10122 Res.first = DestReg;
10123 Res.second = X86::GR64RegisterClass;
10126 } else if (Res.second == X86::FR32RegisterClass ||
10127 Res.second == X86::FR64RegisterClass ||
10128 Res.second == X86::VR128RegisterClass) {
10129 // Handle references to XMM physical registers that got mapped into the
10130 // wrong class. This can happen with constraints like {xmm0} where the
10131 // target independent register mapper will just pick the first match it can
10132 // find, ignoring the required type.
10133 if (VT == MVT::f32)
10134 Res.second = X86::FR32RegisterClass;
10135 else if (VT == MVT::f64)
10136 Res.second = X86::FR64RegisterClass;
10137 else if (X86::VR128RegisterClass->hasType(VT))
10138 Res.second = X86::VR128RegisterClass;
10144 //===----------------------------------------------------------------------===//
10145 // X86 Widen vector type
10146 //===----------------------------------------------------------------------===//
10148 /// getWidenVectorType: given a vector type, returns the type to widen
10149 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
10150 /// If there is no vector type that we want to widen to, returns MVT::Other
10151 /// When and where to widen is target dependent based on the cost of
10152 /// scalarizing vs using the wider vector type.
10154 EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
10155 assert(VT.isVector());
10156 if (isTypeLegal(VT))
10159 // TODO: In computeRegisterProperty, we can compute the list of legal vector
10160 // type based on element type. This would speed up our search (though
10161 // it may not be worth it since the size of the list is relatively
10163 EVT EltVT = VT.getVectorElementType();
10164 unsigned NElts = VT.getVectorNumElements();
10166 // On X86, it make sense to widen any vector wider than 1
10170 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
10171 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
10172 EVT SVT = (MVT::SimpleValueType)nVT;
10174 if (isTypeLegal(SVT) &&
10175 SVT.getVectorElementType() == EltVT &&
10176 SVT.getVectorNumElements() > NElts)