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 (CalleeCC != CallingConv::Fast &&
2250 CalleeCC != CallingConv::C)
2253 // If -tailcallopt is specified, make fastcc functions tail-callable.
2254 const Function *CallerF = DAG.getMachineFunction().getFunction();
2255 if (PerformTailCallOpt &&
2256 CalleeCC == CallingConv::Fast &&
2257 CallerF->getCallingConv() == CalleeCC)
2260 // Look for obvious safe cases to perform tail call optimization.
2261 // If the callee takes no arguments then go on to check the results of the
2263 if (!Outs.empty()) {
2264 // Check if stack adjustment is needed. For now, do not do this if any
2265 // argument is passed on the stack.
2266 SmallVector<CCValAssign, 16> ArgLocs;
2267 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2268 ArgLocs, *DAG.getContext());
2269 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2270 if (CCInfo.getNextStackOffset())
2274 // If the caller does not return a value, then this is obviously safe.
2275 // This is one case where it's safe to perform this optimization even
2276 // if the return types do not match.
2277 const Type *CallerRetTy = CallerF->getReturnType();
2278 if (CallerRetTy->isVoidTy())
2281 // If the return types match, then it's safe.
2282 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2283 if (!G) return false; // FIXME: common external symbols?
2284 Function *CalleeF = cast<Function>(G->getGlobal());
2285 const Type *CalleeRetTy = CalleeF->getReturnType();
2286 return CallerRetTy == CalleeRetTy;
2290 X86TargetLowering::createFastISel(MachineFunction &mf, MachineModuleInfo *mmo,
2292 DenseMap<const Value *, unsigned> &vm,
2293 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2294 DenseMap<const AllocaInst *, int> &am
2296 , SmallSet<Instruction*, 8> &cil
2299 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2307 //===----------------------------------------------------------------------===//
2308 // Other Lowering Hooks
2309 //===----------------------------------------------------------------------===//
2312 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2313 MachineFunction &MF = DAG.getMachineFunction();
2314 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2315 int ReturnAddrIndex = FuncInfo->getRAIndex();
2317 if (ReturnAddrIndex == 0) {
2318 // Set up a frame object for the return address.
2319 uint64_t SlotSize = TD->getPointerSize();
2320 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2322 FuncInfo->setRAIndex(ReturnAddrIndex);
2325 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2329 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2330 bool hasSymbolicDisplacement) {
2331 // Offset should fit into 32 bit immediate field.
2332 if (!isInt32(Offset))
2335 // If we don't have a symbolic displacement - we don't have any extra
2337 if (!hasSymbolicDisplacement)
2340 // FIXME: Some tweaks might be needed for medium code model.
2341 if (M != CodeModel::Small && M != CodeModel::Kernel)
2344 // For small code model we assume that latest object is 16MB before end of 31
2345 // bits boundary. We may also accept pretty large negative constants knowing
2346 // that all objects are in the positive half of address space.
2347 if (M == CodeModel::Small && Offset < 16*1024*1024)
2350 // For kernel code model we know that all object resist in the negative half
2351 // of 32bits address space. We may not accept negative offsets, since they may
2352 // be just off and we may accept pretty large positive ones.
2353 if (M == CodeModel::Kernel && Offset > 0)
2359 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2360 /// specific condition code, returning the condition code and the LHS/RHS of the
2361 /// comparison to make.
2362 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2363 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2365 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2366 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2367 // X > -1 -> X == 0, jump !sign.
2368 RHS = DAG.getConstant(0, RHS.getValueType());
2369 return X86::COND_NS;
2370 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2371 // X < 0 -> X == 0, jump on sign.
2373 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2375 RHS = DAG.getConstant(0, RHS.getValueType());
2376 return X86::COND_LE;
2380 switch (SetCCOpcode) {
2381 default: llvm_unreachable("Invalid integer condition!");
2382 case ISD::SETEQ: return X86::COND_E;
2383 case ISD::SETGT: return X86::COND_G;
2384 case ISD::SETGE: return X86::COND_GE;
2385 case ISD::SETLT: return X86::COND_L;
2386 case ISD::SETLE: return X86::COND_LE;
2387 case ISD::SETNE: return X86::COND_NE;
2388 case ISD::SETULT: return X86::COND_B;
2389 case ISD::SETUGT: return X86::COND_A;
2390 case ISD::SETULE: return X86::COND_BE;
2391 case ISD::SETUGE: return X86::COND_AE;
2395 // First determine if it is required or is profitable to flip the operands.
2397 // If LHS is a foldable load, but RHS is not, flip the condition.
2398 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2399 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2400 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2401 std::swap(LHS, RHS);
2404 switch (SetCCOpcode) {
2410 std::swap(LHS, RHS);
2414 // On a floating point condition, the flags are set as follows:
2416 // 0 | 0 | 0 | X > Y
2417 // 0 | 0 | 1 | X < Y
2418 // 1 | 0 | 0 | X == Y
2419 // 1 | 1 | 1 | unordered
2420 switch (SetCCOpcode) {
2421 default: llvm_unreachable("Condcode should be pre-legalized away");
2423 case ISD::SETEQ: return X86::COND_E;
2424 case ISD::SETOLT: // flipped
2426 case ISD::SETGT: return X86::COND_A;
2427 case ISD::SETOLE: // flipped
2429 case ISD::SETGE: return X86::COND_AE;
2430 case ISD::SETUGT: // flipped
2432 case ISD::SETLT: return X86::COND_B;
2433 case ISD::SETUGE: // flipped
2435 case ISD::SETLE: return X86::COND_BE;
2437 case ISD::SETNE: return X86::COND_NE;
2438 case ISD::SETUO: return X86::COND_P;
2439 case ISD::SETO: return X86::COND_NP;
2441 case ISD::SETUNE: return X86::COND_INVALID;
2445 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2446 /// code. Current x86 isa includes the following FP cmov instructions:
2447 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2448 static bool hasFPCMov(unsigned X86CC) {
2464 /// isFPImmLegal - Returns true if the target can instruction select the
2465 /// specified FP immediate natively. If false, the legalizer will
2466 /// materialize the FP immediate as a load from a constant pool.
2467 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2468 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2469 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2475 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2476 /// the specified range (L, H].
2477 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2478 return (Val < 0) || (Val >= Low && Val < Hi);
2481 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2482 /// specified value.
2483 static bool isUndefOrEqual(int Val, int CmpVal) {
2484 if (Val < 0 || Val == CmpVal)
2489 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2490 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2491 /// the second operand.
2492 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2493 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2494 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2495 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2496 return (Mask[0] < 2 && Mask[1] < 2);
2500 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2501 SmallVector<int, 8> M;
2503 return ::isPSHUFDMask(M, N->getValueType(0));
2506 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2507 /// is suitable for input to PSHUFHW.
2508 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2509 if (VT != MVT::v8i16)
2512 // Lower quadword copied in order or undef.
2513 for (int i = 0; i != 4; ++i)
2514 if (Mask[i] >= 0 && Mask[i] != i)
2517 // Upper quadword shuffled.
2518 for (int i = 4; i != 8; ++i)
2519 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2525 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2526 SmallVector<int, 8> M;
2528 return ::isPSHUFHWMask(M, N->getValueType(0));
2531 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2532 /// is suitable for input to PSHUFLW.
2533 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2534 if (VT != MVT::v8i16)
2537 // Upper quadword copied in order.
2538 for (int i = 4; i != 8; ++i)
2539 if (Mask[i] >= 0 && Mask[i] != i)
2542 // Lower quadword shuffled.
2543 for (int i = 0; i != 4; ++i)
2550 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2551 SmallVector<int, 8> M;
2553 return ::isPSHUFLWMask(M, N->getValueType(0));
2556 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2557 /// is suitable for input to PALIGNR.
2558 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2560 int i, e = VT.getVectorNumElements();
2562 // Do not handle v2i64 / v2f64 shuffles with palignr.
2563 if (e < 4 || !hasSSSE3)
2566 for (i = 0; i != e; ++i)
2570 // All undef, not a palignr.
2574 // Determine if it's ok to perform a palignr with only the LHS, since we
2575 // don't have access to the actual shuffle elements to see if RHS is undef.
2576 bool Unary = Mask[i] < (int)e;
2577 bool NeedsUnary = false;
2579 int s = Mask[i] - i;
2581 // Check the rest of the elements to see if they are consecutive.
2582 for (++i; i != e; ++i) {
2587 Unary = Unary && (m < (int)e);
2588 NeedsUnary = NeedsUnary || (m < s);
2590 if (NeedsUnary && !Unary)
2592 if (Unary && m != ((s+i) & (e-1)))
2594 if (!Unary && m != (s+i))
2600 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2601 SmallVector<int, 8> M;
2603 return ::isPALIGNRMask(M, N->getValueType(0), true);
2606 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2607 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2608 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2609 int NumElems = VT.getVectorNumElements();
2610 if (NumElems != 2 && NumElems != 4)
2613 int Half = NumElems / 2;
2614 for (int i = 0; i < Half; ++i)
2615 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2617 for (int i = Half; i < NumElems; ++i)
2618 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2624 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2625 SmallVector<int, 8> M;
2627 return ::isSHUFPMask(M, N->getValueType(0));
2630 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2631 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2632 /// half elements to come from vector 1 (which would equal the dest.) and
2633 /// the upper half to come from vector 2.
2634 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2635 int NumElems = VT.getVectorNumElements();
2637 if (NumElems != 2 && NumElems != 4)
2640 int Half = NumElems / 2;
2641 for (int i = 0; i < Half; ++i)
2642 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2644 for (int i = Half; i < NumElems; ++i)
2645 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2650 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2651 SmallVector<int, 8> M;
2653 return isCommutedSHUFPMask(M, N->getValueType(0));
2656 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2657 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2658 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2659 if (N->getValueType(0).getVectorNumElements() != 4)
2662 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2663 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2664 isUndefOrEqual(N->getMaskElt(1), 7) &&
2665 isUndefOrEqual(N->getMaskElt(2), 2) &&
2666 isUndefOrEqual(N->getMaskElt(3), 3);
2669 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2670 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2672 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2673 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2678 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2679 isUndefOrEqual(N->getMaskElt(1), 3) &&
2680 isUndefOrEqual(N->getMaskElt(2), 2) &&
2681 isUndefOrEqual(N->getMaskElt(3), 3);
2684 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2685 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2686 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2687 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2689 if (NumElems != 2 && NumElems != 4)
2692 for (unsigned i = 0; i < NumElems/2; ++i)
2693 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2696 for (unsigned i = NumElems/2; i < NumElems; ++i)
2697 if (!isUndefOrEqual(N->getMaskElt(i), i))
2703 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2704 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2705 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2706 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2708 if (NumElems != 2 && NumElems != 4)
2711 for (unsigned i = 0; i < NumElems/2; ++i)
2712 if (!isUndefOrEqual(N->getMaskElt(i), i))
2715 for (unsigned i = 0; i < NumElems/2; ++i)
2716 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2722 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2723 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2724 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2725 bool V2IsSplat = false) {
2726 int NumElts = VT.getVectorNumElements();
2727 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2730 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2732 int BitI1 = Mask[i+1];
2733 if (!isUndefOrEqual(BitI, j))
2736 if (!isUndefOrEqual(BitI1, NumElts))
2739 if (!isUndefOrEqual(BitI1, j + NumElts))
2746 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2747 SmallVector<int, 8> M;
2749 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2752 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2753 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2754 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2755 bool V2IsSplat = false) {
2756 int NumElts = VT.getVectorNumElements();
2757 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2760 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2762 int BitI1 = Mask[i+1];
2763 if (!isUndefOrEqual(BitI, j + NumElts/2))
2766 if (isUndefOrEqual(BitI1, NumElts))
2769 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2776 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2777 SmallVector<int, 8> M;
2779 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2782 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2783 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2785 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2786 int NumElems = VT.getVectorNumElements();
2787 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2790 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2792 int BitI1 = Mask[i+1];
2793 if (!isUndefOrEqual(BitI, j))
2795 if (!isUndefOrEqual(BitI1, j))
2801 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2802 SmallVector<int, 8> M;
2804 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2807 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2808 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2810 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2811 int NumElems = VT.getVectorNumElements();
2812 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2815 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2817 int BitI1 = Mask[i+1];
2818 if (!isUndefOrEqual(BitI, j))
2820 if (!isUndefOrEqual(BitI1, j))
2826 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2827 SmallVector<int, 8> M;
2829 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2832 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2833 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2834 /// MOVSD, and MOVD, i.e. setting the lowest element.
2835 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2836 if (VT.getVectorElementType().getSizeInBits() < 32)
2839 int NumElts = VT.getVectorNumElements();
2841 if (!isUndefOrEqual(Mask[0], NumElts))
2844 for (int i = 1; i < NumElts; ++i)
2845 if (!isUndefOrEqual(Mask[i], i))
2851 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2852 SmallVector<int, 8> M;
2854 return ::isMOVLMask(M, N->getValueType(0));
2857 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2858 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2859 /// element of vector 2 and the other elements to come from vector 1 in order.
2860 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2861 bool V2IsSplat = false, bool V2IsUndef = false) {
2862 int NumOps = VT.getVectorNumElements();
2863 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2866 if (!isUndefOrEqual(Mask[0], 0))
2869 for (int i = 1; i < NumOps; ++i)
2870 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2871 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2872 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2878 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2879 bool V2IsUndef = false) {
2880 SmallVector<int, 8> M;
2882 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2885 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2886 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2887 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2888 if (N->getValueType(0).getVectorNumElements() != 4)
2891 // Expect 1, 1, 3, 3
2892 for (unsigned i = 0; i < 2; ++i) {
2893 int Elt = N->getMaskElt(i);
2894 if (Elt >= 0 && Elt != 1)
2899 for (unsigned i = 2; i < 4; ++i) {
2900 int Elt = N->getMaskElt(i);
2901 if (Elt >= 0 && Elt != 3)
2906 // Don't use movshdup if it can be done with a shufps.
2907 // FIXME: verify that matching u, u, 3, 3 is what we want.
2911 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2912 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2913 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2914 if (N->getValueType(0).getVectorNumElements() != 4)
2917 // Expect 0, 0, 2, 2
2918 for (unsigned i = 0; i < 2; ++i)
2919 if (N->getMaskElt(i) > 0)
2923 for (unsigned i = 2; i < 4; ++i) {
2924 int Elt = N->getMaskElt(i);
2925 if (Elt >= 0 && Elt != 2)
2930 // Don't use movsldup if it can be done with a shufps.
2934 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2935 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2936 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2937 int e = N->getValueType(0).getVectorNumElements() / 2;
2939 for (int i = 0; i < e; ++i)
2940 if (!isUndefOrEqual(N->getMaskElt(i), i))
2942 for (int i = 0; i < e; ++i)
2943 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2948 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2949 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
2950 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2951 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2952 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2954 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2956 for (int i = 0; i < NumOperands; ++i) {
2957 int Val = SVOp->getMaskElt(NumOperands-i-1);
2958 if (Val < 0) Val = 0;
2959 if (Val >= NumOperands) Val -= NumOperands;
2961 if (i != NumOperands - 1)
2967 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2968 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
2969 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2970 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2972 // 8 nodes, but we only care about the last 4.
2973 for (unsigned i = 7; i >= 4; --i) {
2974 int Val = SVOp->getMaskElt(i);
2983 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2984 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
2985 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2986 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2988 // 8 nodes, but we only care about the first 4.
2989 for (int i = 3; i >= 0; --i) {
2990 int Val = SVOp->getMaskElt(i);
2999 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3000 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3001 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3002 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3003 EVT VVT = N->getValueType(0);
3004 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3008 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3009 Val = SVOp->getMaskElt(i);
3013 return (Val - i) * EltSize;
3016 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3018 bool X86::isZeroNode(SDValue Elt) {
3019 return ((isa<ConstantSDNode>(Elt) &&
3020 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
3021 (isa<ConstantFPSDNode>(Elt) &&
3022 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3025 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3026 /// their permute mask.
3027 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3028 SelectionDAG &DAG) {
3029 EVT VT = SVOp->getValueType(0);
3030 unsigned NumElems = VT.getVectorNumElements();
3031 SmallVector<int, 8> MaskVec;
3033 for (unsigned i = 0; i != NumElems; ++i) {
3034 int idx = SVOp->getMaskElt(i);
3036 MaskVec.push_back(idx);
3037 else if (idx < (int)NumElems)
3038 MaskVec.push_back(idx + NumElems);
3040 MaskVec.push_back(idx - NumElems);
3042 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3043 SVOp->getOperand(0), &MaskVec[0]);
3046 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3047 /// the two vector operands have swapped position.
3048 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3049 unsigned NumElems = VT.getVectorNumElements();
3050 for (unsigned i = 0; i != NumElems; ++i) {
3054 else if (idx < (int)NumElems)
3055 Mask[i] = idx + NumElems;
3057 Mask[i] = idx - NumElems;
3061 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3062 /// match movhlps. The lower half elements should come from upper half of
3063 /// V1 (and in order), and the upper half elements should come from the upper
3064 /// half of V2 (and in order).
3065 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3066 if (Op->getValueType(0).getVectorNumElements() != 4)
3068 for (unsigned i = 0, e = 2; i != e; ++i)
3069 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3071 for (unsigned i = 2; i != 4; ++i)
3072 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3077 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3078 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3080 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3081 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3083 N = N->getOperand(0).getNode();
3084 if (!ISD::isNON_EXTLoad(N))
3087 *LD = cast<LoadSDNode>(N);
3091 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3092 /// match movlp{s|d}. The lower half elements should come from lower half of
3093 /// V1 (and in order), and the upper half elements should come from the upper
3094 /// half of V2 (and in order). And since V1 will become the source of the
3095 /// MOVLP, it must be either a vector load or a scalar load to vector.
3096 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3097 ShuffleVectorSDNode *Op) {
3098 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3100 // Is V2 is a vector load, don't do this transformation. We will try to use
3101 // load folding shufps op.
3102 if (ISD::isNON_EXTLoad(V2))
3105 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3107 if (NumElems != 2 && NumElems != 4)
3109 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3110 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3112 for (unsigned i = NumElems/2; i != NumElems; ++i)
3113 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3118 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3120 static bool isSplatVector(SDNode *N) {
3121 if (N->getOpcode() != ISD::BUILD_VECTOR)
3124 SDValue SplatValue = N->getOperand(0);
3125 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3126 if (N->getOperand(i) != SplatValue)
3131 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3132 /// to an zero vector.
3133 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3134 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3135 SDValue V1 = N->getOperand(0);
3136 SDValue V2 = N->getOperand(1);
3137 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3138 for (unsigned i = 0; i != NumElems; ++i) {
3139 int Idx = N->getMaskElt(i);
3140 if (Idx >= (int)NumElems) {
3141 unsigned Opc = V2.getOpcode();
3142 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3144 if (Opc != ISD::BUILD_VECTOR ||
3145 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3147 } else if (Idx >= 0) {
3148 unsigned Opc = V1.getOpcode();
3149 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3151 if (Opc != ISD::BUILD_VECTOR ||
3152 !X86::isZeroNode(V1.getOperand(Idx)))
3159 /// getZeroVector - Returns a vector of specified type with all zero elements.
3161 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3163 assert(VT.isVector() && "Expected a vector type");
3165 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3166 // type. This ensures they get CSE'd.
3168 if (VT.getSizeInBits() == 64) { // MMX
3169 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3170 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3171 } else if (HasSSE2) { // SSE2
3172 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3173 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3175 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3176 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3178 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3181 /// getOnesVector - Returns a vector of specified type with all bits set.
3183 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3184 assert(VT.isVector() && "Expected a vector type");
3186 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3187 // type. This ensures they get CSE'd.
3188 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3190 if (VT.getSizeInBits() == 64) // MMX
3191 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3193 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3194 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3198 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3199 /// that point to V2 points to its first element.
3200 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3201 EVT VT = SVOp->getValueType(0);
3202 unsigned NumElems = VT.getVectorNumElements();
3204 bool Changed = false;
3205 SmallVector<int, 8> MaskVec;
3206 SVOp->getMask(MaskVec);
3208 for (unsigned i = 0; i != NumElems; ++i) {
3209 if (MaskVec[i] > (int)NumElems) {
3210 MaskVec[i] = NumElems;
3215 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3216 SVOp->getOperand(1), &MaskVec[0]);
3217 return SDValue(SVOp, 0);
3220 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3221 /// operation of specified width.
3222 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3224 unsigned NumElems = VT.getVectorNumElements();
3225 SmallVector<int, 8> Mask;
3226 Mask.push_back(NumElems);
3227 for (unsigned i = 1; i != NumElems; ++i)
3229 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3232 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3233 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3235 unsigned NumElems = VT.getVectorNumElements();
3236 SmallVector<int, 8> Mask;
3237 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3239 Mask.push_back(i + NumElems);
3241 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3244 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3245 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3247 unsigned NumElems = VT.getVectorNumElements();
3248 unsigned Half = NumElems/2;
3249 SmallVector<int, 8> Mask;
3250 for (unsigned i = 0; i != Half; ++i) {
3251 Mask.push_back(i + Half);
3252 Mask.push_back(i + NumElems + Half);
3254 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3257 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3258 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3260 if (SV->getValueType(0).getVectorNumElements() <= 4)
3261 return SDValue(SV, 0);
3263 EVT PVT = MVT::v4f32;
3264 EVT VT = SV->getValueType(0);
3265 DebugLoc dl = SV->getDebugLoc();
3266 SDValue V1 = SV->getOperand(0);
3267 int NumElems = VT.getVectorNumElements();
3268 int EltNo = SV->getSplatIndex();
3270 // unpack elements to the correct location
3271 while (NumElems > 4) {
3272 if (EltNo < NumElems/2) {
3273 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3275 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3276 EltNo -= NumElems/2;
3281 // Perform the splat.
3282 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3283 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3284 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3285 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3288 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3289 /// vector of zero or undef vector. This produces a shuffle where the low
3290 /// element of V2 is swizzled into the zero/undef vector, landing at element
3291 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3292 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3293 bool isZero, bool HasSSE2,
3294 SelectionDAG &DAG) {
3295 EVT VT = V2.getValueType();
3297 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3298 unsigned NumElems = VT.getVectorNumElements();
3299 SmallVector<int, 16> MaskVec;
3300 for (unsigned i = 0; i != NumElems; ++i)
3301 // If this is the insertion idx, put the low elt of V2 here.
3302 MaskVec.push_back(i == Idx ? NumElems : i);
3303 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3306 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3307 /// a shuffle that is zero.
3309 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3310 bool Low, SelectionDAG &DAG) {
3311 unsigned NumZeros = 0;
3312 for (int i = 0; i < NumElems; ++i) {
3313 unsigned Index = Low ? i : NumElems-i-1;
3314 int Idx = SVOp->getMaskElt(Index);
3319 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3320 if (Elt.getNode() && X86::isZeroNode(Elt))
3328 /// isVectorShift - Returns true if the shuffle can be implemented as a
3329 /// logical left or right shift of a vector.
3330 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3331 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3332 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3333 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3336 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3339 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3343 bool SeenV1 = false;
3344 bool SeenV2 = false;
3345 for (int i = NumZeros; i < NumElems; ++i) {
3346 int Val = isLeft ? (i - NumZeros) : i;
3347 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3359 if (SeenV1 && SeenV2)
3362 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3368 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3370 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3371 unsigned NumNonZero, unsigned NumZero,
3372 SelectionDAG &DAG, TargetLowering &TLI) {
3376 DebugLoc dl = Op.getDebugLoc();
3379 for (unsigned i = 0; i < 16; ++i) {
3380 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3381 if (ThisIsNonZero && First) {
3383 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3385 V = DAG.getUNDEF(MVT::v8i16);
3390 SDValue ThisElt(0, 0), LastElt(0, 0);
3391 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3392 if (LastIsNonZero) {
3393 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3394 MVT::i16, Op.getOperand(i-1));
3396 if (ThisIsNonZero) {
3397 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3398 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3399 ThisElt, DAG.getConstant(8, MVT::i8));
3401 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3405 if (ThisElt.getNode())
3406 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3407 DAG.getIntPtrConstant(i/2));
3411 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3414 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3416 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3417 unsigned NumNonZero, unsigned NumZero,
3418 SelectionDAG &DAG, TargetLowering &TLI) {
3422 DebugLoc dl = Op.getDebugLoc();
3425 for (unsigned i = 0; i < 8; ++i) {
3426 bool isNonZero = (NonZeros & (1 << i)) != 0;
3430 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3432 V = DAG.getUNDEF(MVT::v8i16);
3435 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3436 MVT::v8i16, V, Op.getOperand(i),
3437 DAG.getIntPtrConstant(i));
3444 /// getVShift - Return a vector logical shift node.
3446 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3447 unsigned NumBits, SelectionDAG &DAG,
3448 const TargetLowering &TLI, DebugLoc dl) {
3449 bool isMMX = VT.getSizeInBits() == 64;
3450 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3451 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3452 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3453 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3454 DAG.getNode(Opc, dl, ShVT, SrcOp,
3455 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3459 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3460 SelectionDAG &DAG) {
3462 // Check if the scalar load can be widened into a vector load. And if
3463 // the address is "base + cst" see if the cst can be "absorbed" into
3464 // the shuffle mask.
3465 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3466 SDValue Ptr = LD->getBasePtr();
3467 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3469 EVT PVT = LD->getValueType(0);
3470 if (PVT != MVT::i32 && PVT != MVT::f32)
3475 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3476 FI = FINode->getIndex();
3478 } else if (Ptr.getOpcode() == ISD::ADD &&
3479 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3480 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3481 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3482 Offset = Ptr.getConstantOperandVal(1);
3483 Ptr = Ptr.getOperand(0);
3488 SDValue Chain = LD->getChain();
3489 // Make sure the stack object alignment is at least 16.
3490 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3491 if (DAG.InferPtrAlignment(Ptr) < 16) {
3492 if (MFI->isFixedObjectIndex(FI)) {
3493 // Can't change the alignment. FIXME: It's possible to compute
3494 // the exact stack offset and reference FI + adjust offset instead.
3495 // If someone *really* cares about this. That's the way to implement it.
3498 MFI->setObjectAlignment(FI, 16);
3502 // (Offset % 16) must be multiple of 4. Then address is then
3503 // Ptr + (Offset & ~15).
3506 if ((Offset % 16) & 3)
3508 int64_t StartOffset = Offset & ~15;
3510 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3511 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3513 int EltNo = (Offset - StartOffset) >> 2;
3514 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3515 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3516 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0);
3517 // Canonicalize it to a v4i32 shuffle.
3518 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3519 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3520 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3521 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3528 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3529 DebugLoc dl = Op.getDebugLoc();
3530 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3531 if (ISD::isBuildVectorAllZeros(Op.getNode())
3532 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3533 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3534 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3535 // eliminated on x86-32 hosts.
3536 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3539 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3540 return getOnesVector(Op.getValueType(), DAG, dl);
3541 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3544 EVT VT = Op.getValueType();
3545 EVT ExtVT = VT.getVectorElementType();
3546 unsigned EVTBits = ExtVT.getSizeInBits();
3548 unsigned NumElems = Op.getNumOperands();
3549 unsigned NumZero = 0;
3550 unsigned NumNonZero = 0;
3551 unsigned NonZeros = 0;
3552 bool IsAllConstants = true;
3553 SmallSet<SDValue, 8> Values;
3554 for (unsigned i = 0; i < NumElems; ++i) {
3555 SDValue Elt = Op.getOperand(i);
3556 if (Elt.getOpcode() == ISD::UNDEF)
3559 if (Elt.getOpcode() != ISD::Constant &&
3560 Elt.getOpcode() != ISD::ConstantFP)
3561 IsAllConstants = false;
3562 if (X86::isZeroNode(Elt))
3565 NonZeros |= (1 << i);
3570 if (NumNonZero == 0) {
3571 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3572 return DAG.getUNDEF(VT);
3575 // Special case for single non-zero, non-undef, element.
3576 if (NumNonZero == 1) {
3577 unsigned Idx = CountTrailingZeros_32(NonZeros);
3578 SDValue Item = Op.getOperand(Idx);
3580 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3581 // the value are obviously zero, truncate the value to i32 and do the
3582 // insertion that way. Only do this if the value is non-constant or if the
3583 // value is a constant being inserted into element 0. It is cheaper to do
3584 // a constant pool load than it is to do a movd + shuffle.
3585 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3586 (!IsAllConstants || Idx == 0)) {
3587 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3588 // Handle MMX and SSE both.
3589 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3590 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3592 // Truncate the value (which may itself be a constant) to i32, and
3593 // convert it to a vector with movd (S2V+shuffle to zero extend).
3594 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3595 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3596 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3597 Subtarget->hasSSE2(), DAG);
3599 // Now we have our 32-bit value zero extended in the low element of
3600 // a vector. If Idx != 0, swizzle it into place.
3602 SmallVector<int, 4> Mask;
3603 Mask.push_back(Idx);
3604 for (unsigned i = 1; i != VecElts; ++i)
3606 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3607 DAG.getUNDEF(Item.getValueType()),
3610 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3614 // If we have a constant or non-constant insertion into the low element of
3615 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3616 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3617 // depending on what the source datatype is.
3620 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3621 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3622 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3623 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3624 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3625 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3627 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3628 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3629 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3630 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3631 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3632 Subtarget->hasSSE2(), DAG);
3633 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3637 // Is it a vector logical left shift?
3638 if (NumElems == 2 && Idx == 1 &&
3639 X86::isZeroNode(Op.getOperand(0)) &&
3640 !X86::isZeroNode(Op.getOperand(1))) {
3641 unsigned NumBits = VT.getSizeInBits();
3642 return getVShift(true, VT,
3643 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3644 VT, Op.getOperand(1)),
3645 NumBits/2, DAG, *this, dl);
3648 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3651 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3652 // is a non-constant being inserted into an element other than the low one,
3653 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3654 // movd/movss) to move this into the low element, then shuffle it into
3656 if (EVTBits == 32) {
3657 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3659 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3660 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3661 Subtarget->hasSSE2(), DAG);
3662 SmallVector<int, 8> MaskVec;
3663 for (unsigned i = 0; i < NumElems; i++)
3664 MaskVec.push_back(i == Idx ? 0 : 1);
3665 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3669 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3670 if (Values.size() == 1) {
3671 if (EVTBits == 32) {
3672 // Instead of a shuffle like this:
3673 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3674 // Check if it's possible to issue this instead.
3675 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3676 unsigned Idx = CountTrailingZeros_32(NonZeros);
3677 SDValue Item = Op.getOperand(Idx);
3678 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3679 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3684 // A vector full of immediates; various special cases are already
3685 // handled, so this is best done with a single constant-pool load.
3689 // Let legalizer expand 2-wide build_vectors.
3690 if (EVTBits == 64) {
3691 if (NumNonZero == 1) {
3692 // One half is zero or undef.
3693 unsigned Idx = CountTrailingZeros_32(NonZeros);
3694 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3695 Op.getOperand(Idx));
3696 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3697 Subtarget->hasSSE2(), DAG);
3702 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3703 if (EVTBits == 8 && NumElems == 16) {
3704 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3706 if (V.getNode()) return V;
3709 if (EVTBits == 16 && NumElems == 8) {
3710 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3712 if (V.getNode()) return V;
3715 // If element VT is == 32 bits, turn it into a number of shuffles.
3716 SmallVector<SDValue, 8> V;
3718 if (NumElems == 4 && NumZero > 0) {
3719 for (unsigned i = 0; i < 4; ++i) {
3720 bool isZero = !(NonZeros & (1 << i));
3722 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3724 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3727 for (unsigned i = 0; i < 2; ++i) {
3728 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3731 V[i] = V[i*2]; // Must be a zero vector.
3734 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3737 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3740 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3745 SmallVector<int, 8> MaskVec;
3746 bool Reverse = (NonZeros & 0x3) == 2;
3747 for (unsigned i = 0; i < 2; ++i)
3748 MaskVec.push_back(Reverse ? 1-i : i);
3749 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3750 for (unsigned i = 0; i < 2; ++i)
3751 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3752 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3755 if (Values.size() > 2) {
3756 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3757 // values to be inserted is equal to the number of elements, in which case
3758 // use the unpack code below in the hopes of matching the consecutive elts
3759 // load merge pattern for shuffles.
3760 // FIXME: We could probably just check that here directly.
3761 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3762 getSubtarget()->hasSSE41()) {
3763 V[0] = DAG.getUNDEF(VT);
3764 for (unsigned i = 0; i < NumElems; ++i)
3765 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3766 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3767 Op.getOperand(i), DAG.getIntPtrConstant(i));
3770 // Expand into a number of unpckl*.
3772 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3773 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3774 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3775 for (unsigned i = 0; i < NumElems; ++i)
3776 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3778 while (NumElems != 0) {
3779 for (unsigned i = 0; i < NumElems; ++i)
3780 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3790 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3791 // We support concatenate two MMX registers and place them in a MMX
3792 // register. This is better than doing a stack convert.
3793 DebugLoc dl = Op.getDebugLoc();
3794 EVT ResVT = Op.getValueType();
3795 assert(Op.getNumOperands() == 2);
3796 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
3797 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
3799 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
3800 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3801 InVec = Op.getOperand(1);
3802 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3803 unsigned NumElts = ResVT.getVectorNumElements();
3804 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3805 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
3806 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
3808 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
3809 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3810 Mask[0] = 0; Mask[1] = 2;
3811 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
3813 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3816 // v8i16 shuffles - Prefer shuffles in the following order:
3817 // 1. [all] pshuflw, pshufhw, optional move
3818 // 2. [ssse3] 1 x pshufb
3819 // 3. [ssse3] 2 x pshufb + 1 x por
3820 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3822 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3823 SelectionDAG &DAG, X86TargetLowering &TLI) {
3824 SDValue V1 = SVOp->getOperand(0);
3825 SDValue V2 = SVOp->getOperand(1);
3826 DebugLoc dl = SVOp->getDebugLoc();
3827 SmallVector<int, 8> MaskVals;
3829 // Determine if more than 1 of the words in each of the low and high quadwords
3830 // of the result come from the same quadword of one of the two inputs. Undef
3831 // mask values count as coming from any quadword, for better codegen.
3832 SmallVector<unsigned, 4> LoQuad(4);
3833 SmallVector<unsigned, 4> HiQuad(4);
3834 BitVector InputQuads(4);
3835 for (unsigned i = 0; i < 8; ++i) {
3836 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3837 int EltIdx = SVOp->getMaskElt(i);
3838 MaskVals.push_back(EltIdx);
3847 InputQuads.set(EltIdx / 4);
3850 int BestLoQuad = -1;
3851 unsigned MaxQuad = 1;
3852 for (unsigned i = 0; i < 4; ++i) {
3853 if (LoQuad[i] > MaxQuad) {
3855 MaxQuad = LoQuad[i];
3859 int BestHiQuad = -1;
3861 for (unsigned i = 0; i < 4; ++i) {
3862 if (HiQuad[i] > MaxQuad) {
3864 MaxQuad = HiQuad[i];
3868 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3869 // of the two input vectors, shuffle them into one input vector so only a
3870 // single pshufb instruction is necessary. If There are more than 2 input
3871 // quads, disable the next transformation since it does not help SSSE3.
3872 bool V1Used = InputQuads[0] || InputQuads[1];
3873 bool V2Used = InputQuads[2] || InputQuads[3];
3874 if (TLI.getSubtarget()->hasSSSE3()) {
3875 if (InputQuads.count() == 2 && V1Used && V2Used) {
3876 BestLoQuad = InputQuads.find_first();
3877 BestHiQuad = InputQuads.find_next(BestLoQuad);
3879 if (InputQuads.count() > 2) {
3885 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3886 // the shuffle mask. If a quad is scored as -1, that means that it contains
3887 // words from all 4 input quadwords.
3889 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3890 SmallVector<int, 8> MaskV;
3891 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3892 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3893 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3894 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3895 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3896 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3898 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3899 // source words for the shuffle, to aid later transformations.
3900 bool AllWordsInNewV = true;
3901 bool InOrder[2] = { true, true };
3902 for (unsigned i = 0; i != 8; ++i) {
3903 int idx = MaskVals[i];
3905 InOrder[i/4] = false;
3906 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3908 AllWordsInNewV = false;
3912 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3913 if (AllWordsInNewV) {
3914 for (int i = 0; i != 8; ++i) {
3915 int idx = MaskVals[i];
3918 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3919 if ((idx != i) && idx < 4)
3921 if ((idx != i) && idx > 3)
3930 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3931 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3932 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3933 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3934 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3938 // If we have SSSE3, and all words of the result are from 1 input vector,
3939 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3940 // is present, fall back to case 4.
3941 if (TLI.getSubtarget()->hasSSSE3()) {
3942 SmallVector<SDValue,16> pshufbMask;
3944 // If we have elements from both input vectors, set the high bit of the
3945 // shuffle mask element to zero out elements that come from V2 in the V1
3946 // mask, and elements that come from V1 in the V2 mask, so that the two
3947 // results can be OR'd together.
3948 bool TwoInputs = V1Used && V2Used;
3949 for (unsigned i = 0; i != 8; ++i) {
3950 int EltIdx = MaskVals[i] * 2;
3951 if (TwoInputs && (EltIdx >= 16)) {
3952 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3953 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3956 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3957 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3959 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3960 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3961 DAG.getNode(ISD::BUILD_VECTOR, dl,
3962 MVT::v16i8, &pshufbMask[0], 16));
3964 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3966 // Calculate the shuffle mask for the second input, shuffle it, and
3967 // OR it with the first shuffled input.
3969 for (unsigned i = 0; i != 8; ++i) {
3970 int EltIdx = MaskVals[i] * 2;
3972 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3973 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3976 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3977 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3979 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3980 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3981 DAG.getNode(ISD::BUILD_VECTOR, dl,
3982 MVT::v16i8, &pshufbMask[0], 16));
3983 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3984 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3987 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3988 // and update MaskVals with new element order.
3989 BitVector InOrder(8);
3990 if (BestLoQuad >= 0) {
3991 SmallVector<int, 8> MaskV;
3992 for (int i = 0; i != 4; ++i) {
3993 int idx = MaskVals[i];
3995 MaskV.push_back(-1);
3997 } else if ((idx / 4) == BestLoQuad) {
3998 MaskV.push_back(idx & 3);
4001 MaskV.push_back(-1);
4004 for (unsigned i = 4; i != 8; ++i)
4006 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4010 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4011 // and update MaskVals with the new element order.
4012 if (BestHiQuad >= 0) {
4013 SmallVector<int, 8> MaskV;
4014 for (unsigned i = 0; i != 4; ++i)
4016 for (unsigned i = 4; i != 8; ++i) {
4017 int idx = MaskVals[i];
4019 MaskV.push_back(-1);
4021 } else if ((idx / 4) == BestHiQuad) {
4022 MaskV.push_back((idx & 3) + 4);
4025 MaskV.push_back(-1);
4028 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4032 // In case BestHi & BestLo were both -1, which means each quadword has a word
4033 // from each of the four input quadwords, calculate the InOrder bitvector now
4034 // before falling through to the insert/extract cleanup.
4035 if (BestLoQuad == -1 && BestHiQuad == -1) {
4037 for (int i = 0; i != 8; ++i)
4038 if (MaskVals[i] < 0 || MaskVals[i] == i)
4042 // The other elements are put in the right place using pextrw and pinsrw.
4043 for (unsigned i = 0; i != 8; ++i) {
4046 int EltIdx = MaskVals[i];
4049 SDValue ExtOp = (EltIdx < 8)
4050 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4051 DAG.getIntPtrConstant(EltIdx))
4052 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4053 DAG.getIntPtrConstant(EltIdx - 8));
4054 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4055 DAG.getIntPtrConstant(i));
4060 // v16i8 shuffles - Prefer shuffles in the following order:
4061 // 1. [ssse3] 1 x pshufb
4062 // 2. [ssse3] 2 x pshufb + 1 x por
4063 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4065 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4066 SelectionDAG &DAG, X86TargetLowering &TLI) {
4067 SDValue V1 = SVOp->getOperand(0);
4068 SDValue V2 = SVOp->getOperand(1);
4069 DebugLoc dl = SVOp->getDebugLoc();
4070 SmallVector<int, 16> MaskVals;
4071 SVOp->getMask(MaskVals);
4073 // If we have SSSE3, case 1 is generated when all result bytes come from
4074 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4075 // present, fall back to case 3.
4076 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4079 for (unsigned i = 0; i < 16; ++i) {
4080 int EltIdx = MaskVals[i];
4089 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4090 if (TLI.getSubtarget()->hasSSSE3()) {
4091 SmallVector<SDValue,16> pshufbMask;
4093 // If all result elements are from one input vector, then only translate
4094 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4096 // Otherwise, we have elements from both input vectors, and must zero out
4097 // elements that come from V2 in the first mask, and V1 in the second mask
4098 // so that we can OR them together.
4099 bool TwoInputs = !(V1Only || V2Only);
4100 for (unsigned i = 0; i != 16; ++i) {
4101 int EltIdx = MaskVals[i];
4102 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4103 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4106 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4108 // If all the elements are from V2, assign it to V1 and return after
4109 // building the first pshufb.
4112 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4113 DAG.getNode(ISD::BUILD_VECTOR, dl,
4114 MVT::v16i8, &pshufbMask[0], 16));
4118 // Calculate the shuffle mask for the second input, shuffle it, and
4119 // OR it with the first shuffled input.
4121 for (unsigned i = 0; i != 16; ++i) {
4122 int EltIdx = MaskVals[i];
4124 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4127 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4129 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4130 DAG.getNode(ISD::BUILD_VECTOR, dl,
4131 MVT::v16i8, &pshufbMask[0], 16));
4132 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4135 // No SSSE3 - Calculate in place words and then fix all out of place words
4136 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4137 // the 16 different words that comprise the two doublequadword input vectors.
4138 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4139 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4140 SDValue NewV = V2Only ? V2 : V1;
4141 for (int i = 0; i != 8; ++i) {
4142 int Elt0 = MaskVals[i*2];
4143 int Elt1 = MaskVals[i*2+1];
4145 // This word of the result is all undef, skip it.
4146 if (Elt0 < 0 && Elt1 < 0)
4149 // This word of the result is already in the correct place, skip it.
4150 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4152 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4155 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4156 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4159 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4160 // using a single extract together, load it and store it.
4161 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4162 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4163 DAG.getIntPtrConstant(Elt1 / 2));
4164 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4165 DAG.getIntPtrConstant(i));
4169 // If Elt1 is defined, extract it from the appropriate source. If the
4170 // source byte is not also odd, shift the extracted word left 8 bits
4171 // otherwise clear the bottom 8 bits if we need to do an or.
4173 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4174 DAG.getIntPtrConstant(Elt1 / 2));
4175 if ((Elt1 & 1) == 0)
4176 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4177 DAG.getConstant(8, TLI.getShiftAmountTy()));
4179 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4180 DAG.getConstant(0xFF00, MVT::i16));
4182 // If Elt0 is defined, extract it from the appropriate source. If the
4183 // source byte is not also even, shift the extracted word right 8 bits. If
4184 // Elt1 was also defined, OR the extracted values together before
4185 // inserting them in the result.
4187 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4188 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4189 if ((Elt0 & 1) != 0)
4190 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4191 DAG.getConstant(8, TLI.getShiftAmountTy()));
4193 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4194 DAG.getConstant(0x00FF, MVT::i16));
4195 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4198 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4199 DAG.getIntPtrConstant(i));
4201 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4204 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4205 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4206 /// done when every pair / quad of shuffle mask elements point to elements in
4207 /// the right sequence. e.g.
4208 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4210 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4212 TargetLowering &TLI, DebugLoc dl) {
4213 EVT VT = SVOp->getValueType(0);
4214 SDValue V1 = SVOp->getOperand(0);
4215 SDValue V2 = SVOp->getOperand(1);
4216 unsigned NumElems = VT.getVectorNumElements();
4217 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4218 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4219 EVT MaskEltVT = MaskVT.getVectorElementType();
4221 switch (VT.getSimpleVT().SimpleTy) {
4222 default: assert(false && "Unexpected!");
4223 case MVT::v4f32: NewVT = MVT::v2f64; break;
4224 case MVT::v4i32: NewVT = MVT::v2i64; break;
4225 case MVT::v8i16: NewVT = MVT::v4i32; break;
4226 case MVT::v16i8: NewVT = MVT::v4i32; break;
4229 if (NewWidth == 2) {
4235 int Scale = NumElems / NewWidth;
4236 SmallVector<int, 8> MaskVec;
4237 for (unsigned i = 0; i < NumElems; i += Scale) {
4239 for (int j = 0; j < Scale; ++j) {
4240 int EltIdx = SVOp->getMaskElt(i+j);
4244 StartIdx = EltIdx - (EltIdx % Scale);
4245 if (EltIdx != StartIdx + j)
4249 MaskVec.push_back(-1);
4251 MaskVec.push_back(StartIdx / Scale);
4254 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4255 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4256 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4259 /// getVZextMovL - Return a zero-extending vector move low node.
4261 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4262 SDValue SrcOp, SelectionDAG &DAG,
4263 const X86Subtarget *Subtarget, DebugLoc dl) {
4264 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4265 LoadSDNode *LD = NULL;
4266 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4267 LD = dyn_cast<LoadSDNode>(SrcOp);
4269 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4271 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4272 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4273 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4274 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4275 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4277 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4278 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4279 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4280 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4288 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4289 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4290 DAG.getNode(ISD::BIT_CONVERT, dl,
4294 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4297 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4298 SDValue V1 = SVOp->getOperand(0);
4299 SDValue V2 = SVOp->getOperand(1);
4300 DebugLoc dl = SVOp->getDebugLoc();
4301 EVT VT = SVOp->getValueType(0);
4303 SmallVector<std::pair<int, int>, 8> Locs;
4305 SmallVector<int, 8> Mask1(4U, -1);
4306 SmallVector<int, 8> PermMask;
4307 SVOp->getMask(PermMask);
4311 for (unsigned i = 0; i != 4; ++i) {
4312 int Idx = PermMask[i];
4314 Locs[i] = std::make_pair(-1, -1);
4316 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4318 Locs[i] = std::make_pair(0, NumLo);
4322 Locs[i] = std::make_pair(1, NumHi);
4324 Mask1[2+NumHi] = Idx;
4330 if (NumLo <= 2 && NumHi <= 2) {
4331 // If no more than two elements come from either vector. This can be
4332 // implemented with two shuffles. First shuffle gather the elements.
4333 // The second shuffle, which takes the first shuffle as both of its
4334 // vector operands, put the elements into the right order.
4335 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4337 SmallVector<int, 8> Mask2(4U, -1);
4339 for (unsigned i = 0; i != 4; ++i) {
4340 if (Locs[i].first == -1)
4343 unsigned Idx = (i < 2) ? 0 : 4;
4344 Idx += Locs[i].first * 2 + Locs[i].second;
4349 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4350 } else if (NumLo == 3 || NumHi == 3) {
4351 // Otherwise, we must have three elements from one vector, call it X, and
4352 // one element from the other, call it Y. First, use a shufps to build an
4353 // intermediate vector with the one element from Y and the element from X
4354 // that will be in the same half in the final destination (the indexes don't
4355 // matter). Then, use a shufps to build the final vector, taking the half
4356 // containing the element from Y from the intermediate, and the other half
4359 // Normalize it so the 3 elements come from V1.
4360 CommuteVectorShuffleMask(PermMask, VT);
4364 // Find the element from V2.
4366 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4367 int Val = PermMask[HiIndex];
4374 Mask1[0] = PermMask[HiIndex];
4376 Mask1[2] = PermMask[HiIndex^1];
4378 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4381 Mask1[0] = PermMask[0];
4382 Mask1[1] = PermMask[1];
4383 Mask1[2] = HiIndex & 1 ? 6 : 4;
4384 Mask1[3] = HiIndex & 1 ? 4 : 6;
4385 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4387 Mask1[0] = HiIndex & 1 ? 2 : 0;
4388 Mask1[1] = HiIndex & 1 ? 0 : 2;
4389 Mask1[2] = PermMask[2];
4390 Mask1[3] = PermMask[3];
4395 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4399 // Break it into (shuffle shuffle_hi, shuffle_lo).
4401 SmallVector<int,8> LoMask(4U, -1);
4402 SmallVector<int,8> HiMask(4U, -1);
4404 SmallVector<int,8> *MaskPtr = &LoMask;
4405 unsigned MaskIdx = 0;
4408 for (unsigned i = 0; i != 4; ++i) {
4415 int Idx = PermMask[i];
4417 Locs[i] = std::make_pair(-1, -1);
4418 } else if (Idx < 4) {
4419 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4420 (*MaskPtr)[LoIdx] = Idx;
4423 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4424 (*MaskPtr)[HiIdx] = Idx;
4429 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4430 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4431 SmallVector<int, 8> MaskOps;
4432 for (unsigned i = 0; i != 4; ++i) {
4433 if (Locs[i].first == -1) {
4434 MaskOps.push_back(-1);
4436 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4437 MaskOps.push_back(Idx);
4440 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4444 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4445 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4446 SDValue V1 = Op.getOperand(0);
4447 SDValue V2 = Op.getOperand(1);
4448 EVT VT = Op.getValueType();
4449 DebugLoc dl = Op.getDebugLoc();
4450 unsigned NumElems = VT.getVectorNumElements();
4451 bool isMMX = VT.getSizeInBits() == 64;
4452 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4453 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4454 bool V1IsSplat = false;
4455 bool V2IsSplat = false;
4457 if (isZeroShuffle(SVOp))
4458 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4460 // Promote splats to v4f32.
4461 if (SVOp->isSplat()) {
4462 if (isMMX || NumElems < 4)
4464 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4467 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4469 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4470 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4471 if (NewOp.getNode())
4472 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4473 LowerVECTOR_SHUFFLE(NewOp, DAG));
4474 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4475 // FIXME: Figure out a cleaner way to do this.
4476 // Try to make use of movq to zero out the top part.
4477 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4478 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4479 if (NewOp.getNode()) {
4480 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4481 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4482 DAG, Subtarget, dl);
4484 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4485 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4486 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4487 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4488 DAG, Subtarget, dl);
4492 if (X86::isPSHUFDMask(SVOp))
4495 // Check if this can be converted into a logical shift.
4496 bool isLeft = false;
4499 bool isShift = getSubtarget()->hasSSE2() &&
4500 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4501 if (isShift && ShVal.hasOneUse()) {
4502 // If the shifted value has multiple uses, it may be cheaper to use
4503 // v_set0 + movlhps or movhlps, etc.
4504 EVT EltVT = VT.getVectorElementType();
4505 ShAmt *= EltVT.getSizeInBits();
4506 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4509 if (X86::isMOVLMask(SVOp)) {
4512 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4513 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4518 // FIXME: fold these into legal mask.
4519 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4520 X86::isMOVSLDUPMask(SVOp) ||
4521 X86::isMOVHLPSMask(SVOp) ||
4522 X86::isMOVLHPSMask(SVOp) ||
4523 X86::isMOVLPMask(SVOp)))
4526 if (ShouldXformToMOVHLPS(SVOp) ||
4527 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4528 return CommuteVectorShuffle(SVOp, DAG);
4531 // No better options. Use a vshl / vsrl.
4532 EVT EltVT = VT.getVectorElementType();
4533 ShAmt *= EltVT.getSizeInBits();
4534 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4537 bool Commuted = false;
4538 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4539 // 1,1,1,1 -> v8i16 though.
4540 V1IsSplat = isSplatVector(V1.getNode());
4541 V2IsSplat = isSplatVector(V2.getNode());
4543 // Canonicalize the splat or undef, if present, to be on the RHS.
4544 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4545 Op = CommuteVectorShuffle(SVOp, DAG);
4546 SVOp = cast<ShuffleVectorSDNode>(Op);
4547 V1 = SVOp->getOperand(0);
4548 V2 = SVOp->getOperand(1);
4549 std::swap(V1IsSplat, V2IsSplat);
4550 std::swap(V1IsUndef, V2IsUndef);
4554 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4555 // Shuffling low element of v1 into undef, just return v1.
4558 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4559 // the instruction selector will not match, so get a canonical MOVL with
4560 // swapped operands to undo the commute.
4561 return getMOVL(DAG, dl, VT, V2, V1);
4564 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4565 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4566 X86::isUNPCKLMask(SVOp) ||
4567 X86::isUNPCKHMask(SVOp))
4571 // Normalize mask so all entries that point to V2 points to its first
4572 // element then try to match unpck{h|l} again. If match, return a
4573 // new vector_shuffle with the corrected mask.
4574 SDValue NewMask = NormalizeMask(SVOp, DAG);
4575 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4576 if (NSVOp != SVOp) {
4577 if (X86::isUNPCKLMask(NSVOp, true)) {
4579 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4586 // Commute is back and try unpck* again.
4587 // FIXME: this seems wrong.
4588 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4589 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4590 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4591 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4592 X86::isUNPCKLMask(NewSVOp) ||
4593 X86::isUNPCKHMask(NewSVOp))
4597 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4599 // Normalize the node to match x86 shuffle ops if needed
4600 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4601 return CommuteVectorShuffle(SVOp, DAG);
4603 // Check for legal shuffle and return?
4604 SmallVector<int, 16> PermMask;
4605 SVOp->getMask(PermMask);
4606 if (isShuffleMaskLegal(PermMask, VT))
4609 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4610 if (VT == MVT::v8i16) {
4611 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4612 if (NewOp.getNode())
4616 if (VT == MVT::v16i8) {
4617 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4618 if (NewOp.getNode())
4622 // Handle all 4 wide cases with a number of shuffles except for MMX.
4623 if (NumElems == 4 && !isMMX)
4624 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4630 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4631 SelectionDAG &DAG) {
4632 EVT VT = Op.getValueType();
4633 DebugLoc dl = Op.getDebugLoc();
4634 if (VT.getSizeInBits() == 8) {
4635 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4636 Op.getOperand(0), Op.getOperand(1));
4637 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4638 DAG.getValueType(VT));
4639 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4640 } else if (VT.getSizeInBits() == 16) {
4641 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4642 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4644 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4645 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4646 DAG.getNode(ISD::BIT_CONVERT, dl,
4650 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4651 Op.getOperand(0), Op.getOperand(1));
4652 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4653 DAG.getValueType(VT));
4654 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4655 } else if (VT == MVT::f32) {
4656 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4657 // the result back to FR32 register. It's only worth matching if the
4658 // result has a single use which is a store or a bitcast to i32. And in
4659 // the case of a store, it's not worth it if the index is a constant 0,
4660 // because a MOVSSmr can be used instead, which is smaller and faster.
4661 if (!Op.hasOneUse())
4663 SDNode *User = *Op.getNode()->use_begin();
4664 if ((User->getOpcode() != ISD::STORE ||
4665 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4666 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4667 (User->getOpcode() != ISD::BIT_CONVERT ||
4668 User->getValueType(0) != MVT::i32))
4670 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4671 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4674 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4675 } else if (VT == MVT::i32) {
4676 // ExtractPS works with constant index.
4677 if (isa<ConstantSDNode>(Op.getOperand(1)))
4685 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4686 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4689 if (Subtarget->hasSSE41()) {
4690 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4695 EVT VT = Op.getValueType();
4696 DebugLoc dl = Op.getDebugLoc();
4697 // TODO: handle v16i8.
4698 if (VT.getSizeInBits() == 16) {
4699 SDValue Vec = Op.getOperand(0);
4700 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4702 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4703 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4704 DAG.getNode(ISD::BIT_CONVERT, dl,
4707 // Transform it so it match pextrw which produces a 32-bit result.
4708 EVT EltVT = MVT::i32;
4709 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4710 Op.getOperand(0), Op.getOperand(1));
4711 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4712 DAG.getValueType(VT));
4713 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4714 } else if (VT.getSizeInBits() == 32) {
4715 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4719 // SHUFPS the element to the lowest double word, then movss.
4720 int Mask[4] = { Idx, -1, -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));
4726 } else if (VT.getSizeInBits() == 64) {
4727 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4728 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4729 // to match extract_elt for f64.
4730 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4734 // UNPCKHPD the element to the lowest double word, then movsd.
4735 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4736 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4737 int Mask[2] = { 1, -1 };
4738 EVT VVT = Op.getOperand(0).getValueType();
4739 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4740 DAG.getUNDEF(VVT), Mask);
4741 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4742 DAG.getIntPtrConstant(0));
4749 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4750 EVT VT = Op.getValueType();
4751 EVT EltVT = VT.getVectorElementType();
4752 DebugLoc dl = Op.getDebugLoc();
4754 SDValue N0 = Op.getOperand(0);
4755 SDValue N1 = Op.getOperand(1);
4756 SDValue N2 = Op.getOperand(2);
4758 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4759 isa<ConstantSDNode>(N2)) {
4760 unsigned Opc = (EltVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4762 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4764 if (N1.getValueType() != MVT::i32)
4765 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4766 if (N2.getValueType() != MVT::i32)
4767 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4768 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4769 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4770 // Bits [7:6] of the constant are the source select. This will always be
4771 // zero here. The DAG Combiner may combine an extract_elt index into these
4772 // bits. For example (insert (extract, 3), 2) could be matched by putting
4773 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4774 // Bits [5:4] of the constant are the destination select. This is the
4775 // value of the incoming immediate.
4776 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4777 // combine either bitwise AND or insert of float 0.0 to set these bits.
4778 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4779 // Create this as a scalar to vector..
4780 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4781 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4782 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4783 // PINSR* works with constant index.
4790 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4791 EVT VT = Op.getValueType();
4792 EVT EltVT = VT.getVectorElementType();
4794 if (Subtarget->hasSSE41())
4795 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4797 if (EltVT == MVT::i8)
4800 DebugLoc dl = Op.getDebugLoc();
4801 SDValue N0 = Op.getOperand(0);
4802 SDValue N1 = Op.getOperand(1);
4803 SDValue N2 = Op.getOperand(2);
4805 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4806 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4807 // as its second argument.
4808 if (N1.getValueType() != MVT::i32)
4809 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4810 if (N2.getValueType() != MVT::i32)
4811 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4812 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4818 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4819 DebugLoc dl = Op.getDebugLoc();
4820 if (Op.getValueType() == MVT::v2f32)
4821 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4822 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4823 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4824 Op.getOperand(0))));
4826 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4827 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4829 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4830 EVT VT = MVT::v2i32;
4831 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4838 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4839 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4842 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4843 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4844 // one of the above mentioned nodes. It has to be wrapped because otherwise
4845 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4846 // be used to form addressing mode. These wrapped nodes will be selected
4849 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4850 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4852 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4854 unsigned char OpFlag = 0;
4855 unsigned WrapperKind = X86ISD::Wrapper;
4856 CodeModel::Model M = getTargetMachine().getCodeModel();
4858 if (Subtarget->isPICStyleRIPRel() &&
4859 (M == CodeModel::Small || M == CodeModel::Kernel))
4860 WrapperKind = X86ISD::WrapperRIP;
4861 else if (Subtarget->isPICStyleGOT())
4862 OpFlag = X86II::MO_GOTOFF;
4863 else if (Subtarget->isPICStyleStubPIC())
4864 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4866 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4868 CP->getOffset(), OpFlag);
4869 DebugLoc DL = CP->getDebugLoc();
4870 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4871 // With PIC, the address is actually $g + Offset.
4873 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4874 DAG.getNode(X86ISD::GlobalBaseReg,
4875 DebugLoc::getUnknownLoc(), getPointerTy()),
4882 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4883 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4885 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4887 unsigned char OpFlag = 0;
4888 unsigned WrapperKind = X86ISD::Wrapper;
4889 CodeModel::Model M = getTargetMachine().getCodeModel();
4891 if (Subtarget->isPICStyleRIPRel() &&
4892 (M == CodeModel::Small || M == CodeModel::Kernel))
4893 WrapperKind = X86ISD::WrapperRIP;
4894 else if (Subtarget->isPICStyleGOT())
4895 OpFlag = X86II::MO_GOTOFF;
4896 else if (Subtarget->isPICStyleStubPIC())
4897 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4899 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4901 DebugLoc DL = JT->getDebugLoc();
4902 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4904 // With PIC, the address is actually $g + Offset.
4906 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4907 DAG.getNode(X86ISD::GlobalBaseReg,
4908 DebugLoc::getUnknownLoc(), getPointerTy()),
4916 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4917 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4919 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4921 unsigned char OpFlag = 0;
4922 unsigned WrapperKind = X86ISD::Wrapper;
4923 CodeModel::Model M = getTargetMachine().getCodeModel();
4925 if (Subtarget->isPICStyleRIPRel() &&
4926 (M == CodeModel::Small || M == CodeModel::Kernel))
4927 WrapperKind = X86ISD::WrapperRIP;
4928 else if (Subtarget->isPICStyleGOT())
4929 OpFlag = X86II::MO_GOTOFF;
4930 else if (Subtarget->isPICStyleStubPIC())
4931 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4933 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4935 DebugLoc DL = Op.getDebugLoc();
4936 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4939 // With PIC, the address is actually $g + Offset.
4940 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4941 !Subtarget->is64Bit()) {
4942 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4943 DAG.getNode(X86ISD::GlobalBaseReg,
4944 DebugLoc::getUnknownLoc(),
4953 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
4954 // Create the TargetBlockAddressAddress node.
4955 unsigned char OpFlags =
4956 Subtarget->ClassifyBlockAddressReference();
4957 CodeModel::Model M = getTargetMachine().getCodeModel();
4958 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
4959 DebugLoc dl = Op.getDebugLoc();
4960 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
4961 /*isTarget=*/true, OpFlags);
4963 if (Subtarget->isPICStyleRIPRel() &&
4964 (M == CodeModel::Small || M == CodeModel::Kernel))
4965 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4967 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4969 // With PIC, the address is actually $g + Offset.
4970 if (isGlobalRelativeToPICBase(OpFlags)) {
4971 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4972 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4980 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4982 SelectionDAG &DAG) const {
4983 // Create the TargetGlobalAddress node, folding in the constant
4984 // offset if it is legal.
4985 unsigned char OpFlags =
4986 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
4987 CodeModel::Model M = getTargetMachine().getCodeModel();
4989 if (OpFlags == X86II::MO_NO_FLAG &&
4990 X86::isOffsetSuitableForCodeModel(Offset, M)) {
4991 // A direct static reference to a global.
4992 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4995 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
4998 if (Subtarget->isPICStyleRIPRel() &&
4999 (M == CodeModel::Small || M == CodeModel::Kernel))
5000 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5002 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5004 // With PIC, the address is actually $g + Offset.
5005 if (isGlobalRelativeToPICBase(OpFlags)) {
5006 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5007 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5011 // For globals that require a load from a stub to get the address, emit the
5013 if (isGlobalStubReference(OpFlags))
5014 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5015 PseudoSourceValue::getGOT(), 0);
5017 // If there was a non-zero offset that we didn't fold, create an explicit
5020 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5021 DAG.getConstant(Offset, getPointerTy()));
5027 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
5028 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5029 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5030 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5034 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5035 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5036 unsigned char OperandFlags) {
5037 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5038 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5039 DebugLoc dl = GA->getDebugLoc();
5040 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5041 GA->getValueType(0),
5045 SDValue Ops[] = { Chain, TGA, *InFlag };
5046 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5048 SDValue Ops[] = { Chain, TGA };
5049 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5052 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5053 MFI->setHasCalls(true);
5055 SDValue Flag = Chain.getValue(1);
5056 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5059 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5061 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5064 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5065 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5066 DAG.getNode(X86ISD::GlobalBaseReg,
5067 DebugLoc::getUnknownLoc(),
5069 InFlag = Chain.getValue(1);
5071 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5074 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5076 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5078 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5079 X86::RAX, X86II::MO_TLSGD);
5082 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5083 // "local exec" model.
5084 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5085 const EVT PtrVT, TLSModel::Model model,
5087 DebugLoc dl = GA->getDebugLoc();
5088 // Get the Thread Pointer
5089 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5090 DebugLoc::getUnknownLoc(), PtrVT,
5091 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5094 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5097 unsigned char OperandFlags = 0;
5098 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5100 unsigned WrapperKind = X86ISD::Wrapper;
5101 if (model == TLSModel::LocalExec) {
5102 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5103 } else if (is64Bit) {
5104 assert(model == TLSModel::InitialExec);
5105 OperandFlags = X86II::MO_GOTTPOFF;
5106 WrapperKind = X86ISD::WrapperRIP;
5108 assert(model == TLSModel::InitialExec);
5109 OperandFlags = X86II::MO_INDNTPOFF;
5112 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5114 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5115 GA->getOffset(), OperandFlags);
5116 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5118 if (model == TLSModel::InitialExec)
5119 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5120 PseudoSourceValue::getGOT(), 0);
5122 // The address of the thread local variable is the add of the thread
5123 // pointer with the offset of the variable.
5124 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5128 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
5129 // TODO: implement the "local dynamic" model
5130 // TODO: implement the "initial exec"model for pic executables
5131 assert(Subtarget->isTargetELF() &&
5132 "TLS not implemented for non-ELF targets");
5133 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5134 const GlobalValue *GV = GA->getGlobal();
5136 // If GV is an alias then use the aliasee for determining
5137 // thread-localness.
5138 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5139 GV = GA->resolveAliasedGlobal(false);
5141 TLSModel::Model model = getTLSModel(GV,
5142 getTargetMachine().getRelocationModel());
5145 case TLSModel::GeneralDynamic:
5146 case TLSModel::LocalDynamic: // not implemented
5147 if (Subtarget->is64Bit())
5148 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5149 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5151 case TLSModel::InitialExec:
5152 case TLSModel::LocalExec:
5153 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5154 Subtarget->is64Bit());
5157 llvm_unreachable("Unreachable");
5162 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5163 /// take a 2 x i32 value to shift plus a shift amount.
5164 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
5165 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5166 EVT VT = Op.getValueType();
5167 unsigned VTBits = VT.getSizeInBits();
5168 DebugLoc dl = Op.getDebugLoc();
5169 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5170 SDValue ShOpLo = Op.getOperand(0);
5171 SDValue ShOpHi = Op.getOperand(1);
5172 SDValue ShAmt = Op.getOperand(2);
5173 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5174 DAG.getConstant(VTBits - 1, MVT::i8))
5175 : DAG.getConstant(0, VT);
5178 if (Op.getOpcode() == ISD::SHL_PARTS) {
5179 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5180 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5182 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5183 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5186 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5187 DAG.getConstant(VTBits, MVT::i8));
5188 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
5189 AndNode, DAG.getConstant(0, MVT::i8));
5192 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5193 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5194 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5196 if (Op.getOpcode() == ISD::SHL_PARTS) {
5197 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5198 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5200 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5201 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5204 SDValue Ops[2] = { Lo, Hi };
5205 return DAG.getMergeValues(Ops, 2, dl);
5208 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5209 EVT SrcVT = Op.getOperand(0).getValueType();
5211 if (SrcVT.isVector()) {
5212 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5218 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5219 "Unknown SINT_TO_FP to lower!");
5221 // These are really Legal; return the operand so the caller accepts it as
5223 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5225 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5226 Subtarget->is64Bit()) {
5230 DebugLoc dl = Op.getDebugLoc();
5231 unsigned Size = SrcVT.getSizeInBits()/8;
5232 MachineFunction &MF = DAG.getMachineFunction();
5233 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5234 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5235 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5237 PseudoSourceValue::getFixedStack(SSFI), 0);
5238 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5241 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5243 SelectionDAG &DAG) {
5245 DebugLoc dl = Op.getDebugLoc();
5247 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5249 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5251 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5252 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5253 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5254 Tys, Ops, array_lengthof(Ops));
5257 Chain = Result.getValue(1);
5258 SDValue InFlag = Result.getValue(2);
5260 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5261 // shouldn't be necessary except that RFP cannot be live across
5262 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5263 MachineFunction &MF = DAG.getMachineFunction();
5264 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5265 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5266 Tys = DAG.getVTList(MVT::Other);
5268 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5270 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5271 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5272 PseudoSourceValue::getFixedStack(SSFI), 0);
5278 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5279 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5280 // This algorithm is not obvious. Here it is in C code, more or less:
5282 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5283 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5284 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5286 // Copy ints to xmm registers.
5287 __m128i xh = _mm_cvtsi32_si128( hi );
5288 __m128i xl = _mm_cvtsi32_si128( lo );
5290 // Combine into low half of a single xmm register.
5291 __m128i x = _mm_unpacklo_epi32( xh, xl );
5295 // Merge in appropriate exponents to give the integer bits the right
5297 x = _mm_unpacklo_epi32( x, exp );
5299 // Subtract away the biases to deal with the IEEE-754 double precision
5301 d = _mm_sub_pd( (__m128d) x, bias );
5303 // All conversions up to here are exact. The correctly rounded result is
5304 // calculated using the current rounding mode using the following
5306 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5307 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5308 // store doesn't really need to be here (except
5309 // maybe to zero the other double)
5314 DebugLoc dl = Op.getDebugLoc();
5315 LLVMContext *Context = DAG.getContext();
5317 // Build some magic constants.
5318 std::vector<Constant*> CV0;
5319 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5320 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5321 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5322 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5323 Constant *C0 = ConstantVector::get(CV0);
5324 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5326 std::vector<Constant*> CV1;
5328 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5330 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5331 Constant *C1 = ConstantVector::get(CV1);
5332 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5334 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5335 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5337 DAG.getIntPtrConstant(1)));
5338 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5339 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5341 DAG.getIntPtrConstant(0)));
5342 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5343 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5344 PseudoSourceValue::getConstantPool(), 0,
5346 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5347 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5348 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5349 PseudoSourceValue::getConstantPool(), 0,
5351 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5353 // Add the halves; easiest way is to swap them into another reg first.
5354 int ShufMask[2] = { 1, -1 };
5355 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5356 DAG.getUNDEF(MVT::v2f64), ShufMask);
5357 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5358 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5359 DAG.getIntPtrConstant(0));
5362 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5363 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5364 DebugLoc dl = Op.getDebugLoc();
5365 // FP constant to bias correct the final result.
5366 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5369 // Load the 32-bit value into an XMM register.
5370 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5371 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5373 DAG.getIntPtrConstant(0)));
5375 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5376 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5377 DAG.getIntPtrConstant(0));
5379 // Or the load with the bias.
5380 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5381 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5382 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5384 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5385 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5386 MVT::v2f64, Bias)));
5387 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5388 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5389 DAG.getIntPtrConstant(0));
5391 // Subtract the bias.
5392 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5394 // Handle final rounding.
5395 EVT DestVT = Op.getValueType();
5397 if (DestVT.bitsLT(MVT::f64)) {
5398 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5399 DAG.getIntPtrConstant(0));
5400 } else if (DestVT.bitsGT(MVT::f64)) {
5401 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5404 // Handle final rounding.
5408 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5409 SDValue N0 = Op.getOperand(0);
5410 DebugLoc dl = Op.getDebugLoc();
5412 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5413 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5414 // the optimization here.
5415 if (DAG.SignBitIsZero(N0))
5416 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5418 EVT SrcVT = N0.getValueType();
5419 if (SrcVT == MVT::i64) {
5420 // We only handle SSE2 f64 target here; caller can expand the rest.
5421 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5424 return LowerUINT_TO_FP_i64(Op, DAG);
5425 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5426 return LowerUINT_TO_FP_i32(Op, DAG);
5429 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5431 // Make a 64-bit buffer, and use it to build an FILD.
5432 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5433 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5434 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5435 getPointerTy(), StackSlot, WordOff);
5436 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5437 StackSlot, NULL, 0);
5438 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5439 OffsetSlot, NULL, 0);
5440 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5443 std::pair<SDValue,SDValue> X86TargetLowering::
5444 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5445 DebugLoc dl = Op.getDebugLoc();
5447 EVT DstTy = Op.getValueType();
5450 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5454 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5455 DstTy.getSimpleVT() >= MVT::i16 &&
5456 "Unknown FP_TO_SINT to lower!");
5458 // These are really Legal.
5459 if (DstTy == MVT::i32 &&
5460 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5461 return std::make_pair(SDValue(), SDValue());
5462 if (Subtarget->is64Bit() &&
5463 DstTy == MVT::i64 &&
5464 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5465 return std::make_pair(SDValue(), SDValue());
5467 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5469 MachineFunction &MF = DAG.getMachineFunction();
5470 unsigned MemSize = DstTy.getSizeInBits()/8;
5471 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5472 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5475 switch (DstTy.getSimpleVT().SimpleTy) {
5476 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5477 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5478 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5479 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5482 SDValue Chain = DAG.getEntryNode();
5483 SDValue Value = Op.getOperand(0);
5484 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5485 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5486 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5487 PseudoSourceValue::getFixedStack(SSFI), 0);
5488 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5490 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5492 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5493 Chain = Value.getValue(1);
5494 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5495 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5498 // Build the FP_TO_INT*_IN_MEM
5499 SDValue Ops[] = { Chain, Value, StackSlot };
5500 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5502 return std::make_pair(FIST, StackSlot);
5505 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5506 if (Op.getValueType().isVector()) {
5507 if (Op.getValueType() == MVT::v2i32 &&
5508 Op.getOperand(0).getValueType() == MVT::v2f64) {
5514 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5515 SDValue FIST = Vals.first, StackSlot = Vals.second;
5516 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5517 if (FIST.getNode() == 0) return Op;
5520 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5521 FIST, StackSlot, NULL, 0);
5524 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5525 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5526 SDValue FIST = Vals.first, StackSlot = Vals.second;
5527 assert(FIST.getNode() && "Unexpected failure");
5530 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5531 FIST, StackSlot, NULL, 0);
5534 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5535 LLVMContext *Context = DAG.getContext();
5536 DebugLoc dl = Op.getDebugLoc();
5537 EVT VT = Op.getValueType();
5540 EltVT = VT.getVectorElementType();
5541 std::vector<Constant*> CV;
5542 if (EltVT == MVT::f64) {
5543 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5547 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5553 Constant *C = ConstantVector::get(CV);
5554 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5555 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5556 PseudoSourceValue::getConstantPool(), 0,
5558 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5561 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5562 LLVMContext *Context = DAG.getContext();
5563 DebugLoc dl = Op.getDebugLoc();
5564 EVT VT = Op.getValueType();
5567 EltVT = VT.getVectorElementType();
5568 std::vector<Constant*> CV;
5569 if (EltVT == MVT::f64) {
5570 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5574 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5580 Constant *C = ConstantVector::get(CV);
5581 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5582 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5583 PseudoSourceValue::getConstantPool(), 0,
5585 if (VT.isVector()) {
5586 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5587 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5588 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5590 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5592 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5596 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5597 LLVMContext *Context = DAG.getContext();
5598 SDValue Op0 = Op.getOperand(0);
5599 SDValue Op1 = Op.getOperand(1);
5600 DebugLoc dl = Op.getDebugLoc();
5601 EVT VT = Op.getValueType();
5602 EVT SrcVT = Op1.getValueType();
5604 // If second operand is smaller, extend it first.
5605 if (SrcVT.bitsLT(VT)) {
5606 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5609 // And if it is bigger, shrink it first.
5610 if (SrcVT.bitsGT(VT)) {
5611 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5615 // At this point the operands and the result should have the same
5616 // type, and that won't be f80 since that is not custom lowered.
5618 // First get the sign bit of second operand.
5619 std::vector<Constant*> CV;
5620 if (SrcVT == MVT::f64) {
5621 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5622 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5624 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5625 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5626 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5627 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5629 Constant *C = ConstantVector::get(CV);
5630 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5631 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5632 PseudoSourceValue::getConstantPool(), 0,
5634 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5636 // Shift sign bit right or left if the two operands have different types.
5637 if (SrcVT.bitsGT(VT)) {
5638 // Op0 is MVT::f32, Op1 is MVT::f64.
5639 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5640 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5641 DAG.getConstant(32, MVT::i32));
5642 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5643 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5644 DAG.getIntPtrConstant(0));
5647 // Clear first operand sign bit.
5649 if (VT == MVT::f64) {
5650 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5651 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5653 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5654 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5655 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5656 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5658 C = ConstantVector::get(CV);
5659 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5660 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5661 PseudoSourceValue::getConstantPool(), 0,
5663 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5665 // Or the value with the sign bit.
5666 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5669 /// Emit nodes that will be selected as "test Op0,Op0", or something
5671 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5672 SelectionDAG &DAG) {
5673 DebugLoc dl = Op.getDebugLoc();
5675 // CF and OF aren't always set the way we want. Determine which
5676 // of these we need.
5677 bool NeedCF = false;
5678 bool NeedOF = false;
5680 case X86::COND_A: case X86::COND_AE:
5681 case X86::COND_B: case X86::COND_BE:
5684 case X86::COND_G: case X86::COND_GE:
5685 case X86::COND_L: case X86::COND_LE:
5686 case X86::COND_O: case X86::COND_NO:
5692 // See if we can use the EFLAGS value from the operand instead of
5693 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5694 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5695 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5696 unsigned Opcode = 0;
5697 unsigned NumOperands = 0;
5698 switch (Op.getNode()->getOpcode()) {
5700 // Due to an isel shortcoming, be conservative if this add is likely to
5701 // be selected as part of a load-modify-store instruction. When the root
5702 // node in a match is a store, isel doesn't know how to remap non-chain
5703 // non-flag uses of other nodes in the match, such as the ADD in this
5704 // case. This leads to the ADD being left around and reselected, with
5705 // the result being two adds in the output.
5706 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5707 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5708 if (UI->getOpcode() == ISD::STORE)
5710 if (ConstantSDNode *C =
5711 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5712 // An add of one will be selected as an INC.
5713 if (C->getAPIntValue() == 1) {
5714 Opcode = X86ISD::INC;
5718 // An add of negative one (subtract of one) will be selected as a DEC.
5719 if (C->getAPIntValue().isAllOnesValue()) {
5720 Opcode = X86ISD::DEC;
5725 // Otherwise use a regular EFLAGS-setting add.
5726 Opcode = X86ISD::ADD;
5730 // If the primary and result isn't used, don't bother using X86ISD::AND,
5731 // because a TEST instruction will be better.
5732 bool NonFlagUse = false;
5733 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5734 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
5736 unsigned UOpNo = UI.getOperandNo();
5737 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
5738 // Look pass truncate.
5739 UOpNo = User->use_begin().getOperandNo();
5740 User = *User->use_begin();
5742 if (User->getOpcode() != ISD::BRCOND &&
5743 User->getOpcode() != ISD::SETCC &&
5744 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
5756 // Due to the ISEL shortcoming noted above, be conservative if this op is
5757 // likely to be selected as part of a load-modify-store instruction.
5758 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5759 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5760 if (UI->getOpcode() == ISD::STORE)
5762 // Otherwise use a regular EFLAGS-setting instruction.
5763 switch (Op.getNode()->getOpcode()) {
5764 case ISD::SUB: Opcode = X86ISD::SUB; break;
5765 case ISD::OR: Opcode = X86ISD::OR; break;
5766 case ISD::XOR: Opcode = X86ISD::XOR; break;
5767 case ISD::AND: Opcode = X86ISD::AND; break;
5768 default: llvm_unreachable("unexpected operator!");
5779 return SDValue(Op.getNode(), 1);
5785 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5786 SmallVector<SDValue, 4> Ops;
5787 for (unsigned i = 0; i != NumOperands; ++i)
5788 Ops.push_back(Op.getOperand(i));
5789 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5790 DAG.ReplaceAllUsesWith(Op, New);
5791 return SDValue(New.getNode(), 1);
5795 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5796 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5797 DAG.getConstant(0, Op.getValueType()));
5800 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5802 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5803 SelectionDAG &DAG) {
5804 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5805 if (C->getAPIntValue() == 0)
5806 return EmitTest(Op0, X86CC, DAG);
5808 DebugLoc dl = Op0.getDebugLoc();
5809 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5812 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
5813 /// if it's possible.
5814 static SDValue LowerToBT(SDValue Op0, ISD::CondCode CC,
5815 DebugLoc dl, SelectionDAG &DAG) {
5817 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5818 if (ConstantSDNode *Op010C =
5819 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5820 if (Op010C->getZExtValue() == 1) {
5821 LHS = Op0.getOperand(0);
5822 RHS = Op0.getOperand(1).getOperand(1);
5824 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5825 if (ConstantSDNode *Op000C =
5826 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5827 if (Op000C->getZExtValue() == 1) {
5828 LHS = Op0.getOperand(1);
5829 RHS = Op0.getOperand(0).getOperand(1);
5831 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5832 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5833 SDValue AndLHS = Op0.getOperand(0);
5834 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5835 LHS = AndLHS.getOperand(0);
5836 RHS = AndLHS.getOperand(1);
5840 if (LHS.getNode()) {
5841 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5842 // instruction. Since the shift amount is in-range-or-undefined, we know
5843 // that doing a bittest on the i16 value is ok. We extend to i32 because
5844 // the encoding for the i16 version is larger than the i32 version.
5845 if (LHS.getValueType() == MVT::i8)
5846 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5848 // If the operand types disagree, extend the shift amount to match. Since
5849 // BT ignores high bits (like shifts) we can use anyextend.
5850 if (LHS.getValueType() != RHS.getValueType())
5851 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5853 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5854 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5855 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5856 DAG.getConstant(Cond, MVT::i8), BT);
5862 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5863 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5864 SDValue Op0 = Op.getOperand(0);
5865 SDValue Op1 = Op.getOperand(1);
5866 DebugLoc dl = Op.getDebugLoc();
5867 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5869 // Optimize to BT if possible.
5870 // Lower (X & (1 << N)) == 0 to BT(X, N).
5871 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5872 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5873 if (Op0.getOpcode() == ISD::AND &&
5875 Op1.getOpcode() == ISD::Constant &&
5876 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5877 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5878 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
5879 if (NewSetCC.getNode())
5883 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5884 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5885 if (X86CC == X86::COND_INVALID)
5888 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5890 // Use sbb x, x to materialize carry bit into a GPR.
5891 if (X86CC == X86::COND_B)
5892 return DAG.getNode(ISD::AND, dl, MVT::i8,
5893 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
5894 DAG.getConstant(X86CC, MVT::i8), Cond),
5895 DAG.getConstant(1, MVT::i8));
5897 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5898 DAG.getConstant(X86CC, MVT::i8), Cond);
5901 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5903 SDValue Op0 = Op.getOperand(0);
5904 SDValue Op1 = Op.getOperand(1);
5905 SDValue CC = Op.getOperand(2);
5906 EVT VT = Op.getValueType();
5907 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5908 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5909 DebugLoc dl = Op.getDebugLoc();
5913 EVT VT0 = Op0.getValueType();
5914 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5915 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5918 switch (SetCCOpcode) {
5921 case ISD::SETEQ: SSECC = 0; break;
5923 case ISD::SETGT: Swap = true; // Fallthrough
5925 case ISD::SETOLT: SSECC = 1; break;
5927 case ISD::SETGE: Swap = true; // Fallthrough
5929 case ISD::SETOLE: SSECC = 2; break;
5930 case ISD::SETUO: SSECC = 3; break;
5932 case ISD::SETNE: SSECC = 4; break;
5933 case ISD::SETULE: Swap = true;
5934 case ISD::SETUGE: SSECC = 5; break;
5935 case ISD::SETULT: Swap = true;
5936 case ISD::SETUGT: SSECC = 6; break;
5937 case ISD::SETO: SSECC = 7; break;
5940 std::swap(Op0, Op1);
5942 // In the two special cases we can't handle, emit two comparisons.
5944 if (SetCCOpcode == ISD::SETUEQ) {
5946 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5947 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5948 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5950 else if (SetCCOpcode == ISD::SETONE) {
5952 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5953 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5954 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5956 llvm_unreachable("Illegal FP comparison");
5958 // Handle all other FP comparisons here.
5959 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5962 // We are handling one of the integer comparisons here. Since SSE only has
5963 // GT and EQ comparisons for integer, swapping operands and multiple
5964 // operations may be required for some comparisons.
5965 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5966 bool Swap = false, Invert = false, FlipSigns = false;
5968 switch (VT.getSimpleVT().SimpleTy) {
5971 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5973 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5975 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5976 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5979 switch (SetCCOpcode) {
5981 case ISD::SETNE: Invert = true;
5982 case ISD::SETEQ: Opc = EQOpc; break;
5983 case ISD::SETLT: Swap = true;
5984 case ISD::SETGT: Opc = GTOpc; break;
5985 case ISD::SETGE: Swap = true;
5986 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5987 case ISD::SETULT: Swap = true;
5988 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5989 case ISD::SETUGE: Swap = true;
5990 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5993 std::swap(Op0, Op1);
5995 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5996 // bits of the inputs before performing those operations.
5998 EVT EltVT = VT.getVectorElementType();
5999 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6001 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6002 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6004 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6005 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6008 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6010 // If the logical-not of the result is required, perform that now.
6012 Result = DAG.getNOT(dl, Result, VT);
6017 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6018 static bool isX86LogicalCmp(SDValue Op) {
6019 unsigned Opc = Op.getNode()->getOpcode();
6020 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6022 if (Op.getResNo() == 1 &&
6023 (Opc == X86ISD::ADD ||
6024 Opc == X86ISD::SUB ||
6025 Opc == X86ISD::SMUL ||
6026 Opc == X86ISD::UMUL ||
6027 Opc == X86ISD::INC ||
6028 Opc == X86ISD::DEC ||
6029 Opc == X86ISD::OR ||
6030 Opc == X86ISD::XOR ||
6031 Opc == X86ISD::AND))
6037 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
6038 bool addTest = true;
6039 SDValue Cond = Op.getOperand(0);
6040 DebugLoc dl = Op.getDebugLoc();
6043 if (Cond.getOpcode() == ISD::SETCC) {
6044 SDValue NewCond = LowerSETCC(Cond, DAG);
6045 if (NewCond.getNode())
6049 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6050 SDValue Op1 = Op.getOperand(1);
6051 SDValue Op2 = Op.getOperand(2);
6052 if (Cond.getOpcode() == X86ISD::SETCC &&
6053 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6054 SDValue Cmp = Cond.getOperand(1);
6055 if (Cmp.getOpcode() == X86ISD::CMP) {
6056 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6057 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6058 ConstantSDNode *RHSC =
6059 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6060 if (N1C && N1C->isAllOnesValue() &&
6061 N2C && N2C->isNullValue() &&
6062 RHSC && RHSC->isNullValue()) {
6063 SDValue CmpOp0 = Cmp.getOperand(0);
6064 Cmp = DAG.getNode(X86ISD::CMP, dl, CmpOp0.getValueType(),
6065 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6066 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6067 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6072 // Look pass (and (setcc_carry (cmp ...)), 1).
6073 if (Cond.getOpcode() == ISD::AND &&
6074 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6075 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6076 if (C && C->getAPIntValue() == 1)
6077 Cond = Cond.getOperand(0);
6080 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6081 // setting operand in place of the X86ISD::SETCC.
6082 if (Cond.getOpcode() == X86ISD::SETCC ||
6083 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6084 CC = Cond.getOperand(0);
6086 SDValue Cmp = Cond.getOperand(1);
6087 unsigned Opc = Cmp.getOpcode();
6088 EVT VT = Op.getValueType();
6090 bool IllegalFPCMov = false;
6091 if (VT.isFloatingPoint() && !VT.isVector() &&
6092 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6093 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6095 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6096 Opc == X86ISD::BT) { // FIXME
6103 // Look pass the truncate.
6104 if (Cond.getOpcode() == ISD::TRUNCATE)
6105 Cond = Cond.getOperand(0);
6107 // We know the result of AND is compared against zero. Try to match
6109 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6110 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6111 if (NewSetCC.getNode()) {
6112 CC = NewSetCC.getOperand(0);
6113 Cond = NewSetCC.getOperand(1);
6120 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6121 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6124 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6125 // condition is true.
6126 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6127 SDValue Ops[] = { Op2, Op1, CC, Cond };
6128 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6131 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6132 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6133 // from the AND / OR.
6134 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6135 Opc = Op.getOpcode();
6136 if (Opc != ISD::OR && Opc != ISD::AND)
6138 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6139 Op.getOperand(0).hasOneUse() &&
6140 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6141 Op.getOperand(1).hasOneUse());
6144 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6145 // 1 and that the SETCC node has a single use.
6146 static bool isXor1OfSetCC(SDValue Op) {
6147 if (Op.getOpcode() != ISD::XOR)
6149 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6150 if (N1C && N1C->getAPIntValue() == 1) {
6151 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6152 Op.getOperand(0).hasOneUse();
6157 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
6158 bool addTest = true;
6159 SDValue Chain = Op.getOperand(0);
6160 SDValue Cond = Op.getOperand(1);
6161 SDValue Dest = Op.getOperand(2);
6162 DebugLoc dl = Op.getDebugLoc();
6165 if (Cond.getOpcode() == ISD::SETCC) {
6166 SDValue NewCond = LowerSETCC(Cond, DAG);
6167 if (NewCond.getNode())
6171 // FIXME: LowerXALUO doesn't handle these!!
6172 else if (Cond.getOpcode() == X86ISD::ADD ||
6173 Cond.getOpcode() == X86ISD::SUB ||
6174 Cond.getOpcode() == X86ISD::SMUL ||
6175 Cond.getOpcode() == X86ISD::UMUL)
6176 Cond = LowerXALUO(Cond, DAG);
6179 // Look pass (and (setcc_carry (cmp ...)), 1).
6180 if (Cond.getOpcode() == ISD::AND &&
6181 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6182 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6183 if (C && C->getAPIntValue() == 1)
6184 Cond = Cond.getOperand(0);
6187 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6188 // setting operand in place of the X86ISD::SETCC.
6189 if (Cond.getOpcode() == X86ISD::SETCC ||
6190 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6191 CC = Cond.getOperand(0);
6193 SDValue Cmp = Cond.getOperand(1);
6194 unsigned Opc = Cmp.getOpcode();
6195 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6196 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6200 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6204 // These can only come from an arithmetic instruction with overflow,
6205 // e.g. SADDO, UADDO.
6206 Cond = Cond.getNode()->getOperand(1);
6213 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6214 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6215 if (CondOpc == ISD::OR) {
6216 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6217 // two branches instead of an explicit OR instruction with a
6219 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6220 isX86LogicalCmp(Cmp)) {
6221 CC = Cond.getOperand(0).getOperand(0);
6222 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6223 Chain, Dest, CC, Cmp);
6224 CC = Cond.getOperand(1).getOperand(0);
6228 } else { // ISD::AND
6229 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6230 // two branches instead of an explicit AND instruction with a
6231 // separate test. However, we only do this if this block doesn't
6232 // have a fall-through edge, because this requires an explicit
6233 // jmp when the condition is false.
6234 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6235 isX86LogicalCmp(Cmp) &&
6236 Op.getNode()->hasOneUse()) {
6237 X86::CondCode CCode =
6238 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6239 CCode = X86::GetOppositeBranchCondition(CCode);
6240 CC = DAG.getConstant(CCode, MVT::i8);
6241 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6242 // Look for an unconditional branch following this conditional branch.
6243 // We need this because we need to reverse the successors in order
6244 // to implement FCMP_OEQ.
6245 if (User.getOpcode() == ISD::BR) {
6246 SDValue FalseBB = User.getOperand(1);
6248 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6249 assert(NewBR == User);
6252 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6253 Chain, Dest, CC, Cmp);
6254 X86::CondCode CCode =
6255 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6256 CCode = X86::GetOppositeBranchCondition(CCode);
6257 CC = DAG.getConstant(CCode, MVT::i8);
6263 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6264 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6265 // It should be transformed during dag combiner except when the condition
6266 // is set by a arithmetics with overflow node.
6267 X86::CondCode CCode =
6268 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6269 CCode = X86::GetOppositeBranchCondition(CCode);
6270 CC = DAG.getConstant(CCode, MVT::i8);
6271 Cond = Cond.getOperand(0).getOperand(1);
6277 // Look pass the truncate.
6278 if (Cond.getOpcode() == ISD::TRUNCATE)
6279 Cond = Cond.getOperand(0);
6281 // We know the result of AND is compared against zero. Try to match
6283 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6284 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6285 if (NewSetCC.getNode()) {
6286 CC = NewSetCC.getOperand(0);
6287 Cond = NewSetCC.getOperand(1);
6294 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6295 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6297 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6298 Chain, Dest, CC, Cond);
6302 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6303 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6304 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6305 // that the guard pages used by the OS virtual memory manager are allocated in
6306 // correct sequence.
6308 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6309 SelectionDAG &DAG) {
6310 assert(Subtarget->isTargetCygMing() &&
6311 "This should be used only on Cygwin/Mingw targets");
6312 DebugLoc dl = Op.getDebugLoc();
6315 SDValue Chain = Op.getOperand(0);
6316 SDValue Size = Op.getOperand(1);
6317 // FIXME: Ensure alignment here
6321 EVT IntPtr = getPointerTy();
6322 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6324 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
6326 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6327 Flag = Chain.getValue(1);
6329 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6330 SDValue Ops[] = { Chain,
6331 DAG.getTargetExternalSymbol("_alloca", IntPtr),
6332 DAG.getRegister(X86::EAX, IntPtr),
6333 DAG.getRegister(X86StackPtr, SPTy),
6335 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
6336 Flag = Chain.getValue(1);
6338 Chain = DAG.getCALLSEQ_END(Chain,
6339 DAG.getIntPtrConstant(0, true),
6340 DAG.getIntPtrConstant(0, true),
6343 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6345 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6346 return DAG.getMergeValues(Ops1, 2, dl);
6350 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6352 SDValue Dst, SDValue Src,
6353 SDValue Size, unsigned Align,
6355 uint64_t DstSVOff) {
6356 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6358 // If not DWORD aligned or size is more than the threshold, call the library.
6359 // The libc version is likely to be faster for these cases. It can use the
6360 // address value and run time information about the CPU.
6361 if ((Align & 3) != 0 ||
6363 ConstantSize->getZExtValue() >
6364 getSubtarget()->getMaxInlineSizeThreshold()) {
6365 SDValue InFlag(0, 0);
6367 // Check to see if there is a specialized entry-point for memory zeroing.
6368 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6370 if (const char *bzeroEntry = V &&
6371 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6372 EVT IntPtr = getPointerTy();
6373 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6374 TargetLowering::ArgListTy Args;
6375 TargetLowering::ArgListEntry Entry;
6377 Entry.Ty = IntPtrTy;
6378 Args.push_back(Entry);
6380 Args.push_back(Entry);
6381 std::pair<SDValue,SDValue> CallResult =
6382 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6383 false, false, false, false,
6384 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6385 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl,
6386 DAG.GetOrdering(Chain.getNode()));
6387 return CallResult.second;
6390 // Otherwise have the target-independent code call memset.
6394 uint64_t SizeVal = ConstantSize->getZExtValue();
6395 SDValue InFlag(0, 0);
6398 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6399 unsigned BytesLeft = 0;
6400 bool TwoRepStos = false;
6403 uint64_t Val = ValC->getZExtValue() & 255;
6405 // If the value is a constant, then we can potentially use larger sets.
6406 switch (Align & 3) {
6407 case 2: // WORD aligned
6410 Val = (Val << 8) | Val;
6412 case 0: // DWORD aligned
6415 Val = (Val << 8) | Val;
6416 Val = (Val << 16) | Val;
6417 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6420 Val = (Val << 32) | Val;
6423 default: // Byte aligned
6426 Count = DAG.getIntPtrConstant(SizeVal);
6430 if (AVT.bitsGT(MVT::i8)) {
6431 unsigned UBytes = AVT.getSizeInBits() / 8;
6432 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6433 BytesLeft = SizeVal % UBytes;
6436 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6438 InFlag = Chain.getValue(1);
6441 Count = DAG.getIntPtrConstant(SizeVal);
6442 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6443 InFlag = Chain.getValue(1);
6446 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6449 InFlag = Chain.getValue(1);
6450 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6453 InFlag = Chain.getValue(1);
6455 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6456 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6457 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6460 InFlag = Chain.getValue(1);
6462 EVT CVT = Count.getValueType();
6463 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6464 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6465 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6468 InFlag = Chain.getValue(1);
6469 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6470 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
6471 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6472 } else if (BytesLeft) {
6473 // Handle the last 1 - 7 bytes.
6474 unsigned Offset = SizeVal - BytesLeft;
6475 EVT AddrVT = Dst.getValueType();
6476 EVT SizeVT = Size.getValueType();
6478 Chain = DAG.getMemset(Chain, dl,
6479 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6480 DAG.getConstant(Offset, AddrVT)),
6482 DAG.getConstant(BytesLeft, SizeVT),
6483 Align, DstSV, DstSVOff + Offset);
6486 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6491 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6492 SDValue Chain, SDValue Dst, SDValue Src,
6493 SDValue Size, unsigned Align,
6495 const Value *DstSV, uint64_t DstSVOff,
6496 const Value *SrcSV, uint64_t SrcSVOff) {
6497 // This requires the copy size to be a constant, preferrably
6498 // within a subtarget-specific limit.
6499 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6502 uint64_t SizeVal = ConstantSize->getZExtValue();
6503 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6506 /// If not DWORD aligned, call the library.
6507 if ((Align & 3) != 0)
6512 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6515 unsigned UBytes = AVT.getSizeInBits() / 8;
6516 unsigned CountVal = SizeVal / UBytes;
6517 SDValue Count = DAG.getIntPtrConstant(CountVal);
6518 unsigned BytesLeft = SizeVal % UBytes;
6520 SDValue InFlag(0, 0);
6521 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6524 InFlag = Chain.getValue(1);
6525 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6528 InFlag = Chain.getValue(1);
6529 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6532 InFlag = Chain.getValue(1);
6534 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6535 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6536 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
6537 array_lengthof(Ops));
6539 SmallVector<SDValue, 4> Results;
6540 Results.push_back(RepMovs);
6542 // Handle the last 1 - 7 bytes.
6543 unsigned Offset = SizeVal - BytesLeft;
6544 EVT DstVT = Dst.getValueType();
6545 EVT SrcVT = Src.getValueType();
6546 EVT SizeVT = Size.getValueType();
6547 Results.push_back(DAG.getMemcpy(Chain, dl,
6548 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6549 DAG.getConstant(Offset, DstVT)),
6550 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6551 DAG.getConstant(Offset, SrcVT)),
6552 DAG.getConstant(BytesLeft, SizeVT),
6553 Align, AlwaysInline,
6554 DstSV, DstSVOff + Offset,
6555 SrcSV, SrcSVOff + Offset));
6558 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6559 &Results[0], Results.size());
6562 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6563 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6564 DebugLoc dl = Op.getDebugLoc();
6566 if (!Subtarget->is64Bit()) {
6567 // vastart just stores the address of the VarArgsFrameIndex slot into the
6568 // memory location argument.
6569 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6570 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6574 // gp_offset (0 - 6 * 8)
6575 // fp_offset (48 - 48 + 8 * 16)
6576 // overflow_arg_area (point to parameters coming in memory).
6578 SmallVector<SDValue, 8> MemOps;
6579 SDValue FIN = Op.getOperand(1);
6581 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6582 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6584 MemOps.push_back(Store);
6587 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6588 FIN, DAG.getIntPtrConstant(4));
6589 Store = DAG.getStore(Op.getOperand(0), dl,
6590 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6592 MemOps.push_back(Store);
6594 // Store ptr to overflow_arg_area
6595 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6596 FIN, DAG.getIntPtrConstant(4));
6597 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6598 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6599 MemOps.push_back(Store);
6601 // Store ptr to reg_save_area.
6602 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6603 FIN, DAG.getIntPtrConstant(8));
6604 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6605 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6606 MemOps.push_back(Store);
6607 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6608 &MemOps[0], MemOps.size());
6611 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6612 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6613 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6614 SDValue Chain = Op.getOperand(0);
6615 SDValue SrcPtr = Op.getOperand(1);
6616 SDValue SrcSV = Op.getOperand(2);
6618 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6622 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6623 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6624 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6625 SDValue Chain = Op.getOperand(0);
6626 SDValue DstPtr = Op.getOperand(1);
6627 SDValue SrcPtr = Op.getOperand(2);
6628 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6629 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6630 DebugLoc dl = Op.getDebugLoc();
6632 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6633 DAG.getIntPtrConstant(24), 8, false,
6634 DstSV, 0, SrcSV, 0);
6638 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6639 DebugLoc dl = Op.getDebugLoc();
6640 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6642 default: return SDValue(); // Don't custom lower most intrinsics.
6643 // Comparison intrinsics.
6644 case Intrinsic::x86_sse_comieq_ss:
6645 case Intrinsic::x86_sse_comilt_ss:
6646 case Intrinsic::x86_sse_comile_ss:
6647 case Intrinsic::x86_sse_comigt_ss:
6648 case Intrinsic::x86_sse_comige_ss:
6649 case Intrinsic::x86_sse_comineq_ss:
6650 case Intrinsic::x86_sse_ucomieq_ss:
6651 case Intrinsic::x86_sse_ucomilt_ss:
6652 case Intrinsic::x86_sse_ucomile_ss:
6653 case Intrinsic::x86_sse_ucomigt_ss:
6654 case Intrinsic::x86_sse_ucomige_ss:
6655 case Intrinsic::x86_sse_ucomineq_ss:
6656 case Intrinsic::x86_sse2_comieq_sd:
6657 case Intrinsic::x86_sse2_comilt_sd:
6658 case Intrinsic::x86_sse2_comile_sd:
6659 case Intrinsic::x86_sse2_comigt_sd:
6660 case Intrinsic::x86_sse2_comige_sd:
6661 case Intrinsic::x86_sse2_comineq_sd:
6662 case Intrinsic::x86_sse2_ucomieq_sd:
6663 case Intrinsic::x86_sse2_ucomilt_sd:
6664 case Intrinsic::x86_sse2_ucomile_sd:
6665 case Intrinsic::x86_sse2_ucomigt_sd:
6666 case Intrinsic::x86_sse2_ucomige_sd:
6667 case Intrinsic::x86_sse2_ucomineq_sd: {
6669 ISD::CondCode CC = ISD::SETCC_INVALID;
6672 case Intrinsic::x86_sse_comieq_ss:
6673 case Intrinsic::x86_sse2_comieq_sd:
6677 case Intrinsic::x86_sse_comilt_ss:
6678 case Intrinsic::x86_sse2_comilt_sd:
6682 case Intrinsic::x86_sse_comile_ss:
6683 case Intrinsic::x86_sse2_comile_sd:
6687 case Intrinsic::x86_sse_comigt_ss:
6688 case Intrinsic::x86_sse2_comigt_sd:
6692 case Intrinsic::x86_sse_comige_ss:
6693 case Intrinsic::x86_sse2_comige_sd:
6697 case Intrinsic::x86_sse_comineq_ss:
6698 case Intrinsic::x86_sse2_comineq_sd:
6702 case Intrinsic::x86_sse_ucomieq_ss:
6703 case Intrinsic::x86_sse2_ucomieq_sd:
6704 Opc = X86ISD::UCOMI;
6707 case Intrinsic::x86_sse_ucomilt_ss:
6708 case Intrinsic::x86_sse2_ucomilt_sd:
6709 Opc = X86ISD::UCOMI;
6712 case Intrinsic::x86_sse_ucomile_ss:
6713 case Intrinsic::x86_sse2_ucomile_sd:
6714 Opc = X86ISD::UCOMI;
6717 case Intrinsic::x86_sse_ucomigt_ss:
6718 case Intrinsic::x86_sse2_ucomigt_sd:
6719 Opc = X86ISD::UCOMI;
6722 case Intrinsic::x86_sse_ucomige_ss:
6723 case Intrinsic::x86_sse2_ucomige_sd:
6724 Opc = X86ISD::UCOMI;
6727 case Intrinsic::x86_sse_ucomineq_ss:
6728 case Intrinsic::x86_sse2_ucomineq_sd:
6729 Opc = X86ISD::UCOMI;
6734 SDValue LHS = Op.getOperand(1);
6735 SDValue RHS = Op.getOperand(2);
6736 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6737 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6738 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6739 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6740 DAG.getConstant(X86CC, MVT::i8), Cond);
6741 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6743 // ptest intrinsics. The intrinsic these come from are designed to return
6744 // an integer value, not just an instruction so lower it to the ptest
6745 // pattern and a setcc for the result.
6746 case Intrinsic::x86_sse41_ptestz:
6747 case Intrinsic::x86_sse41_ptestc:
6748 case Intrinsic::x86_sse41_ptestnzc:{
6751 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6752 case Intrinsic::x86_sse41_ptestz:
6754 X86CC = X86::COND_E;
6756 case Intrinsic::x86_sse41_ptestc:
6758 X86CC = X86::COND_B;
6760 case Intrinsic::x86_sse41_ptestnzc:
6762 X86CC = X86::COND_A;
6766 SDValue LHS = Op.getOperand(1);
6767 SDValue RHS = Op.getOperand(2);
6768 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6769 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6770 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6771 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6774 // Fix vector shift instructions where the last operand is a non-immediate
6776 case Intrinsic::x86_sse2_pslli_w:
6777 case Intrinsic::x86_sse2_pslli_d:
6778 case Intrinsic::x86_sse2_pslli_q:
6779 case Intrinsic::x86_sse2_psrli_w:
6780 case Intrinsic::x86_sse2_psrli_d:
6781 case Intrinsic::x86_sse2_psrli_q:
6782 case Intrinsic::x86_sse2_psrai_w:
6783 case Intrinsic::x86_sse2_psrai_d:
6784 case Intrinsic::x86_mmx_pslli_w:
6785 case Intrinsic::x86_mmx_pslli_d:
6786 case Intrinsic::x86_mmx_pslli_q:
6787 case Intrinsic::x86_mmx_psrli_w:
6788 case Intrinsic::x86_mmx_psrli_d:
6789 case Intrinsic::x86_mmx_psrli_q:
6790 case Intrinsic::x86_mmx_psrai_w:
6791 case Intrinsic::x86_mmx_psrai_d: {
6792 SDValue ShAmt = Op.getOperand(2);
6793 if (isa<ConstantSDNode>(ShAmt))
6796 unsigned NewIntNo = 0;
6797 EVT ShAmtVT = MVT::v4i32;
6799 case Intrinsic::x86_sse2_pslli_w:
6800 NewIntNo = Intrinsic::x86_sse2_psll_w;
6802 case Intrinsic::x86_sse2_pslli_d:
6803 NewIntNo = Intrinsic::x86_sse2_psll_d;
6805 case Intrinsic::x86_sse2_pslli_q:
6806 NewIntNo = Intrinsic::x86_sse2_psll_q;
6808 case Intrinsic::x86_sse2_psrli_w:
6809 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6811 case Intrinsic::x86_sse2_psrli_d:
6812 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6814 case Intrinsic::x86_sse2_psrli_q:
6815 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6817 case Intrinsic::x86_sse2_psrai_w:
6818 NewIntNo = Intrinsic::x86_sse2_psra_w;
6820 case Intrinsic::x86_sse2_psrai_d:
6821 NewIntNo = Intrinsic::x86_sse2_psra_d;
6824 ShAmtVT = MVT::v2i32;
6826 case Intrinsic::x86_mmx_pslli_w:
6827 NewIntNo = Intrinsic::x86_mmx_psll_w;
6829 case Intrinsic::x86_mmx_pslli_d:
6830 NewIntNo = Intrinsic::x86_mmx_psll_d;
6832 case Intrinsic::x86_mmx_pslli_q:
6833 NewIntNo = Intrinsic::x86_mmx_psll_q;
6835 case Intrinsic::x86_mmx_psrli_w:
6836 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6838 case Intrinsic::x86_mmx_psrli_d:
6839 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6841 case Intrinsic::x86_mmx_psrli_q:
6842 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6844 case Intrinsic::x86_mmx_psrai_w:
6845 NewIntNo = Intrinsic::x86_mmx_psra_w;
6847 case Intrinsic::x86_mmx_psrai_d:
6848 NewIntNo = Intrinsic::x86_mmx_psra_d;
6850 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6856 // The vector shift intrinsics with scalars uses 32b shift amounts but
6857 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
6861 ShOps[1] = DAG.getConstant(0, MVT::i32);
6862 if (ShAmtVT == MVT::v4i32) {
6863 ShOps[2] = DAG.getUNDEF(MVT::i32);
6864 ShOps[3] = DAG.getUNDEF(MVT::i32);
6865 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
6867 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
6870 EVT VT = Op.getValueType();
6871 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
6872 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6873 DAG.getConstant(NewIntNo, MVT::i32),
6874 Op.getOperand(1), ShAmt);
6879 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6880 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6881 DebugLoc dl = Op.getDebugLoc();
6884 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6886 DAG.getConstant(TD->getPointerSize(),
6887 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6888 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6889 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6894 // Just load the return address.
6895 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6896 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6897 RetAddrFI, NULL, 0);
6900 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6901 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6902 MFI->setFrameAddressIsTaken(true);
6903 EVT VT = Op.getValueType();
6904 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6905 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6906 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6907 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6909 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6913 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6914 SelectionDAG &DAG) {
6915 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6918 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6920 MachineFunction &MF = DAG.getMachineFunction();
6921 SDValue Chain = Op.getOperand(0);
6922 SDValue Offset = Op.getOperand(1);
6923 SDValue Handler = Op.getOperand(2);
6924 DebugLoc dl = Op.getDebugLoc();
6926 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6928 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6930 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6931 DAG.getIntPtrConstant(-TD->getPointerSize()));
6932 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6933 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6934 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6935 MF.getRegInfo().addLiveOut(StoreAddrReg);
6937 return DAG.getNode(X86ISD::EH_RETURN, dl,
6939 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6942 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6943 SelectionDAG &DAG) {
6944 SDValue Root = Op.getOperand(0);
6945 SDValue Trmp = Op.getOperand(1); // trampoline
6946 SDValue FPtr = Op.getOperand(2); // nested function
6947 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6948 DebugLoc dl = Op.getDebugLoc();
6950 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6952 const X86InstrInfo *TII =
6953 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6955 if (Subtarget->is64Bit()) {
6956 SDValue OutChains[6];
6958 // Large code-model.
6960 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6961 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6963 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6964 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6966 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6968 // Load the pointer to the nested function into R11.
6969 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6970 SDValue Addr = Trmp;
6971 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6974 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6975 DAG.getConstant(2, MVT::i64));
6976 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6978 // Load the 'nest' parameter value into R10.
6979 // R10 is specified in X86CallingConv.td
6980 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6981 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6982 DAG.getConstant(10, MVT::i64));
6983 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6984 Addr, TrmpAddr, 10);
6986 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6987 DAG.getConstant(12, MVT::i64));
6988 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6990 // Jump to the nested function.
6991 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6992 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6993 DAG.getConstant(20, MVT::i64));
6994 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6995 Addr, TrmpAddr, 20);
6997 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6998 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6999 DAG.getConstant(22, MVT::i64));
7000 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7004 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7005 return DAG.getMergeValues(Ops, 2, dl);
7007 const Function *Func =
7008 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7009 CallingConv::ID CC = Func->getCallingConv();
7014 llvm_unreachable("Unsupported calling convention");
7015 case CallingConv::C:
7016 case CallingConv::X86_StdCall: {
7017 // Pass 'nest' parameter in ECX.
7018 // Must be kept in sync with X86CallingConv.td
7021 // Check that ECX wasn't needed by an 'inreg' parameter.
7022 const FunctionType *FTy = Func->getFunctionType();
7023 const AttrListPtr &Attrs = Func->getAttributes();
7025 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7026 unsigned InRegCount = 0;
7029 for (FunctionType::param_iterator I = FTy->param_begin(),
7030 E = FTy->param_end(); I != E; ++I, ++Idx)
7031 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7032 // FIXME: should only count parameters that are lowered to integers.
7033 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7035 if (InRegCount > 2) {
7036 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
7041 case CallingConv::X86_FastCall:
7042 case CallingConv::Fast:
7043 // Pass 'nest' parameter in EAX.
7044 // Must be kept in sync with X86CallingConv.td
7049 SDValue OutChains[4];
7052 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7053 DAG.getConstant(10, MVT::i32));
7054 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7056 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
7057 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7058 OutChains[0] = DAG.getStore(Root, dl,
7059 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7062 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7063 DAG.getConstant(1, MVT::i32));
7064 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
7066 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
7067 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7068 DAG.getConstant(5, MVT::i32));
7069 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7070 TrmpAddr, 5, false, 1);
7072 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7073 DAG.getConstant(6, MVT::i32));
7074 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
7077 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7078 return DAG.getMergeValues(Ops, 2, dl);
7082 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
7084 The rounding mode is in bits 11:10 of FPSR, and has the following
7091 FLT_ROUNDS, on the other hand, expects the following:
7098 To perform the conversion, we do:
7099 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7102 MachineFunction &MF = DAG.getMachineFunction();
7103 const TargetMachine &TM = MF.getTarget();
7104 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7105 unsigned StackAlignment = TFI.getStackAlignment();
7106 EVT VT = Op.getValueType();
7107 DebugLoc dl = Op.getDebugLoc();
7109 // Save FP Control Word to stack slot
7110 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7111 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7113 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7114 DAG.getEntryNode(), StackSlot);
7116 // Load FP Control Word from stack slot
7117 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
7119 // Transform as necessary
7121 DAG.getNode(ISD::SRL, dl, MVT::i16,
7122 DAG.getNode(ISD::AND, dl, MVT::i16,
7123 CWD, DAG.getConstant(0x800, MVT::i16)),
7124 DAG.getConstant(11, MVT::i8));
7126 DAG.getNode(ISD::SRL, dl, MVT::i16,
7127 DAG.getNode(ISD::AND, dl, MVT::i16,
7128 CWD, DAG.getConstant(0x400, MVT::i16)),
7129 DAG.getConstant(9, MVT::i8));
7132 DAG.getNode(ISD::AND, dl, MVT::i16,
7133 DAG.getNode(ISD::ADD, dl, MVT::i16,
7134 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7135 DAG.getConstant(1, MVT::i16)),
7136 DAG.getConstant(3, MVT::i16));
7139 return DAG.getNode((VT.getSizeInBits() < 16 ?
7140 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7143 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
7144 EVT VT = Op.getValueType();
7146 unsigned NumBits = VT.getSizeInBits();
7147 DebugLoc dl = Op.getDebugLoc();
7149 Op = Op.getOperand(0);
7150 if (VT == MVT::i8) {
7151 // Zero extend to i32 since there is not an i8 bsr.
7153 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7156 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7157 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7158 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7160 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7163 DAG.getConstant(NumBits+NumBits-1, OpVT),
7164 DAG.getConstant(X86::COND_E, MVT::i8),
7167 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7169 // Finally xor with NumBits-1.
7170 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7173 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7177 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
7178 EVT VT = Op.getValueType();
7180 unsigned NumBits = VT.getSizeInBits();
7181 DebugLoc dl = Op.getDebugLoc();
7183 Op = Op.getOperand(0);
7184 if (VT == MVT::i8) {
7186 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7189 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7190 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7191 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7193 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7196 DAG.getConstant(NumBits, OpVT),
7197 DAG.getConstant(X86::COND_E, MVT::i8),
7200 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7203 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7207 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
7208 EVT VT = Op.getValueType();
7209 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7210 DebugLoc dl = Op.getDebugLoc();
7212 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7213 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7214 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7215 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7216 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7218 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7219 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7220 // return AloBlo + AloBhi + AhiBlo;
7222 SDValue A = Op.getOperand(0);
7223 SDValue B = Op.getOperand(1);
7225 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7226 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7227 A, DAG.getConstant(32, MVT::i32));
7228 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7229 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7230 B, DAG.getConstant(32, MVT::i32));
7231 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7232 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7234 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7235 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7237 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7238 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7240 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7241 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7242 AloBhi, DAG.getConstant(32, MVT::i32));
7243 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7244 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7245 AhiBlo, DAG.getConstant(32, MVT::i32));
7246 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7247 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7252 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
7253 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7254 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7255 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7256 // has only one use.
7257 SDNode *N = Op.getNode();
7258 SDValue LHS = N->getOperand(0);
7259 SDValue RHS = N->getOperand(1);
7260 unsigned BaseOp = 0;
7262 DebugLoc dl = Op.getDebugLoc();
7264 switch (Op.getOpcode()) {
7265 default: llvm_unreachable("Unknown ovf instruction!");
7267 // A subtract of one will be selected as a INC. Note that INC doesn't
7268 // set CF, so we can't do this for UADDO.
7269 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7270 if (C->getAPIntValue() == 1) {
7271 BaseOp = X86ISD::INC;
7275 BaseOp = X86ISD::ADD;
7279 BaseOp = X86ISD::ADD;
7283 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7284 // set CF, so we can't do this for USUBO.
7285 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7286 if (C->getAPIntValue() == 1) {
7287 BaseOp = X86ISD::DEC;
7291 BaseOp = X86ISD::SUB;
7295 BaseOp = X86ISD::SUB;
7299 BaseOp = X86ISD::SMUL;
7303 BaseOp = X86ISD::UMUL;
7308 // Also sets EFLAGS.
7309 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7310 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7313 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7314 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7316 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7320 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
7321 EVT T = Op.getValueType();
7322 DebugLoc dl = Op.getDebugLoc();
7325 switch(T.getSimpleVT().SimpleTy) {
7327 assert(false && "Invalid value type!");
7328 case MVT::i8: Reg = X86::AL; size = 1; break;
7329 case MVT::i16: Reg = X86::AX; size = 2; break;
7330 case MVT::i32: Reg = X86::EAX; size = 4; break;
7332 assert(Subtarget->is64Bit() && "Node not type legal!");
7333 Reg = X86::RAX; size = 8;
7336 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7337 Op.getOperand(2), SDValue());
7338 SDValue Ops[] = { cpIn.getValue(0),
7341 DAG.getTargetConstant(size, MVT::i8),
7343 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7344 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7346 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7350 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7351 SelectionDAG &DAG) {
7352 assert(Subtarget->is64Bit() && "Result not type legalized?");
7353 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7354 SDValue TheChain = Op.getOperand(0);
7355 DebugLoc dl = Op.getDebugLoc();
7356 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7357 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7358 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7360 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7361 DAG.getConstant(32, MVT::i8));
7363 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7366 return DAG.getMergeValues(Ops, 2, dl);
7369 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7370 SDNode *Node = Op.getNode();
7371 DebugLoc dl = Node->getDebugLoc();
7372 EVT T = Node->getValueType(0);
7373 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7374 DAG.getConstant(0, T), Node->getOperand(2));
7375 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7376 cast<AtomicSDNode>(Node)->getMemoryVT(),
7377 Node->getOperand(0),
7378 Node->getOperand(1), negOp,
7379 cast<AtomicSDNode>(Node)->getSrcValue(),
7380 cast<AtomicSDNode>(Node)->getAlignment());
7383 /// LowerOperation - Provide custom lowering hooks for some operations.
7385 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7386 switch (Op.getOpcode()) {
7387 default: llvm_unreachable("Should not custom lower this!");
7388 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7389 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7390 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7391 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7392 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7393 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7394 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7395 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7396 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7397 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7398 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7399 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7400 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7401 case ISD::SHL_PARTS:
7402 case ISD::SRA_PARTS:
7403 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7404 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7405 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7406 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7407 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7408 case ISD::FABS: return LowerFABS(Op, DAG);
7409 case ISD::FNEG: return LowerFNEG(Op, DAG);
7410 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7411 case ISD::SETCC: return LowerSETCC(Op, DAG);
7412 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7413 case ISD::SELECT: return LowerSELECT(Op, DAG);
7414 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7415 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7416 case ISD::VASTART: return LowerVASTART(Op, DAG);
7417 case ISD::VAARG: return LowerVAARG(Op, DAG);
7418 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7419 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7420 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7421 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7422 case ISD::FRAME_TO_ARGS_OFFSET:
7423 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7424 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7425 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7426 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7427 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7428 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7429 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7430 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7436 case ISD::UMULO: return LowerXALUO(Op, DAG);
7437 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7441 void X86TargetLowering::
7442 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7443 SelectionDAG &DAG, unsigned NewOp) {
7444 EVT T = Node->getValueType(0);
7445 DebugLoc dl = Node->getDebugLoc();
7446 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7448 SDValue Chain = Node->getOperand(0);
7449 SDValue In1 = Node->getOperand(1);
7450 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7451 Node->getOperand(2), DAG.getIntPtrConstant(0));
7452 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7453 Node->getOperand(2), DAG.getIntPtrConstant(1));
7454 SDValue Ops[] = { Chain, In1, In2L, In2H };
7455 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7457 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7458 cast<MemSDNode>(Node)->getMemOperand());
7459 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7460 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7461 Results.push_back(Result.getValue(2));
7464 /// ReplaceNodeResults - Replace a node with an illegal result type
7465 /// with a new node built out of custom code.
7466 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7467 SmallVectorImpl<SDValue>&Results,
7468 SelectionDAG &DAG) {
7469 DebugLoc dl = N->getDebugLoc();
7470 switch (N->getOpcode()) {
7472 assert(false && "Do not know how to custom type legalize this operation!");
7474 case ISD::FP_TO_SINT: {
7475 std::pair<SDValue,SDValue> Vals =
7476 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7477 SDValue FIST = Vals.first, StackSlot = Vals.second;
7478 if (FIST.getNode() != 0) {
7479 EVT VT = N->getValueType(0);
7480 // Return a load from the stack slot.
7481 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
7485 case ISD::READCYCLECOUNTER: {
7486 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7487 SDValue TheChain = N->getOperand(0);
7488 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7489 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7491 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7493 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7494 SDValue Ops[] = { eax, edx };
7495 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7496 Results.push_back(edx.getValue(1));
7503 EVT WidenVT = getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
7504 Results.push_back(DAG.UnrollVectorOp(N, WidenVT.getVectorNumElements()));
7507 case ISD::ATOMIC_CMP_SWAP: {
7508 EVT T = N->getValueType(0);
7509 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7510 SDValue cpInL, cpInH;
7511 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7512 DAG.getConstant(0, MVT::i32));
7513 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7514 DAG.getConstant(1, MVT::i32));
7515 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7516 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7518 SDValue swapInL, swapInH;
7519 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7520 DAG.getConstant(0, MVT::i32));
7521 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7522 DAG.getConstant(1, MVT::i32));
7523 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7525 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7526 swapInL.getValue(1));
7527 SDValue Ops[] = { swapInH.getValue(0),
7529 swapInH.getValue(1) };
7530 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7531 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7532 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7533 MVT::i32, Result.getValue(1));
7534 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7535 MVT::i32, cpOutL.getValue(2));
7536 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7537 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7538 Results.push_back(cpOutH.getValue(1));
7541 case ISD::ATOMIC_LOAD_ADD:
7542 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7544 case ISD::ATOMIC_LOAD_AND:
7545 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7547 case ISD::ATOMIC_LOAD_NAND:
7548 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7550 case ISD::ATOMIC_LOAD_OR:
7551 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7553 case ISD::ATOMIC_LOAD_SUB:
7554 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7556 case ISD::ATOMIC_LOAD_XOR:
7557 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7559 case ISD::ATOMIC_SWAP:
7560 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7565 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7567 default: return NULL;
7568 case X86ISD::BSF: return "X86ISD::BSF";
7569 case X86ISD::BSR: return "X86ISD::BSR";
7570 case X86ISD::SHLD: return "X86ISD::SHLD";
7571 case X86ISD::SHRD: return "X86ISD::SHRD";
7572 case X86ISD::FAND: return "X86ISD::FAND";
7573 case X86ISD::FOR: return "X86ISD::FOR";
7574 case X86ISD::FXOR: return "X86ISD::FXOR";
7575 case X86ISD::FSRL: return "X86ISD::FSRL";
7576 case X86ISD::FILD: return "X86ISD::FILD";
7577 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7578 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7579 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7580 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7581 case X86ISD::FLD: return "X86ISD::FLD";
7582 case X86ISD::FST: return "X86ISD::FST";
7583 case X86ISD::CALL: return "X86ISD::CALL";
7584 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7585 case X86ISD::BT: return "X86ISD::BT";
7586 case X86ISD::CMP: return "X86ISD::CMP";
7587 case X86ISD::COMI: return "X86ISD::COMI";
7588 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7589 case X86ISD::SETCC: return "X86ISD::SETCC";
7590 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7591 case X86ISD::CMOV: return "X86ISD::CMOV";
7592 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7593 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7594 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7595 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7596 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7597 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7598 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7599 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7600 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7601 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7602 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7603 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7604 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7605 case X86ISD::FMAX: return "X86ISD::FMAX";
7606 case X86ISD::FMIN: return "X86ISD::FMIN";
7607 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7608 case X86ISD::FRCP: return "X86ISD::FRCP";
7609 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7610 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7611 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7612 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7613 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7614 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7615 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7616 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7617 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7618 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7619 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7620 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7621 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7622 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7623 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7624 case X86ISD::VSHL: return "X86ISD::VSHL";
7625 case X86ISD::VSRL: return "X86ISD::VSRL";
7626 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7627 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7628 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7629 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7630 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7631 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7632 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7633 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7634 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7635 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7636 case X86ISD::ADD: return "X86ISD::ADD";
7637 case X86ISD::SUB: return "X86ISD::SUB";
7638 case X86ISD::SMUL: return "X86ISD::SMUL";
7639 case X86ISD::UMUL: return "X86ISD::UMUL";
7640 case X86ISD::INC: return "X86ISD::INC";
7641 case X86ISD::DEC: return "X86ISD::DEC";
7642 case X86ISD::OR: return "X86ISD::OR";
7643 case X86ISD::XOR: return "X86ISD::XOR";
7644 case X86ISD::AND: return "X86ISD::AND";
7645 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7646 case X86ISD::PTEST: return "X86ISD::PTEST";
7647 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7651 // isLegalAddressingMode - Return true if the addressing mode represented
7652 // by AM is legal for this target, for a load/store of the specified type.
7653 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7654 const Type *Ty) const {
7655 // X86 supports extremely general addressing modes.
7656 CodeModel::Model M = getTargetMachine().getCodeModel();
7658 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7659 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7664 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7666 // If a reference to this global requires an extra load, we can't fold it.
7667 if (isGlobalStubReference(GVFlags))
7670 // If BaseGV requires a register for the PIC base, we cannot also have a
7671 // BaseReg specified.
7672 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7675 // If lower 4G is not available, then we must use rip-relative addressing.
7676 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7686 // These scales always work.
7691 // These scales are formed with basereg+scalereg. Only accept if there is
7696 default: // Other stuff never works.
7704 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7705 if (!Ty1->isInteger() || !Ty2->isInteger())
7707 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7708 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7709 if (NumBits1 <= NumBits2)
7711 return Subtarget->is64Bit() || NumBits1 < 64;
7714 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7715 if (!VT1.isInteger() || !VT2.isInteger())
7717 unsigned NumBits1 = VT1.getSizeInBits();
7718 unsigned NumBits2 = VT2.getSizeInBits();
7719 if (NumBits1 <= NumBits2)
7721 return Subtarget->is64Bit() || NumBits1 < 64;
7724 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7725 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7726 return Ty1->isInteger(32) && Ty2->isInteger(64) && Subtarget->is64Bit();
7729 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7730 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7731 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7734 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7735 // i16 instructions are longer (0x66 prefix) and potentially slower.
7736 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7739 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7740 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7741 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7742 /// are assumed to be legal.
7744 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7746 // Only do shuffles on 128-bit vector types for now.
7747 if (VT.getSizeInBits() == 64)
7750 // FIXME: pshufb, blends, shifts.
7751 return (VT.getVectorNumElements() == 2 ||
7752 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7753 isMOVLMask(M, VT) ||
7754 isSHUFPMask(M, VT) ||
7755 isPSHUFDMask(M, VT) ||
7756 isPSHUFHWMask(M, VT) ||
7757 isPSHUFLWMask(M, VT) ||
7758 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7759 isUNPCKLMask(M, VT) ||
7760 isUNPCKHMask(M, VT) ||
7761 isUNPCKL_v_undef_Mask(M, VT) ||
7762 isUNPCKH_v_undef_Mask(M, VT));
7766 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7768 unsigned NumElts = VT.getVectorNumElements();
7769 // FIXME: This collection of masks seems suspect.
7772 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7773 return (isMOVLMask(Mask, VT) ||
7774 isCommutedMOVLMask(Mask, VT, true) ||
7775 isSHUFPMask(Mask, VT) ||
7776 isCommutedSHUFPMask(Mask, VT));
7781 //===----------------------------------------------------------------------===//
7782 // X86 Scheduler Hooks
7783 //===----------------------------------------------------------------------===//
7785 // private utility function
7787 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7788 MachineBasicBlock *MBB,
7796 TargetRegisterClass *RC,
7797 bool invSrc) const {
7798 // For the atomic bitwise operator, we generate
7801 // ld t1 = [bitinstr.addr]
7802 // op t2 = t1, [bitinstr.val]
7804 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7806 // fallthrough -->nextMBB
7807 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7808 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7809 MachineFunction::iterator MBBIter = MBB;
7812 /// First build the CFG
7813 MachineFunction *F = MBB->getParent();
7814 MachineBasicBlock *thisMBB = MBB;
7815 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7816 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7817 F->insert(MBBIter, newMBB);
7818 F->insert(MBBIter, nextMBB);
7820 // Move all successors to thisMBB to nextMBB
7821 nextMBB->transferSuccessors(thisMBB);
7823 // Update thisMBB to fall through to newMBB
7824 thisMBB->addSuccessor(newMBB);
7826 // newMBB jumps to itself and fall through to nextMBB
7827 newMBB->addSuccessor(nextMBB);
7828 newMBB->addSuccessor(newMBB);
7830 // Insert instructions into newMBB based on incoming instruction
7831 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7832 "unexpected number of operands");
7833 DebugLoc dl = bInstr->getDebugLoc();
7834 MachineOperand& destOper = bInstr->getOperand(0);
7835 MachineOperand* argOpers[2 + X86AddrNumOperands];
7836 int numArgs = bInstr->getNumOperands() - 1;
7837 for (int i=0; i < numArgs; ++i)
7838 argOpers[i] = &bInstr->getOperand(i+1);
7840 // x86 address has 4 operands: base, index, scale, and displacement
7841 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7842 int valArgIndx = lastAddrIndx + 1;
7844 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7845 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7846 for (int i=0; i <= lastAddrIndx; ++i)
7847 (*MIB).addOperand(*argOpers[i]);
7849 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7851 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7856 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7857 assert((argOpers[valArgIndx]->isReg() ||
7858 argOpers[valArgIndx]->isImm()) &&
7860 if (argOpers[valArgIndx]->isReg())
7861 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7863 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7865 (*MIB).addOperand(*argOpers[valArgIndx]);
7867 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7870 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7871 for (int i=0; i <= lastAddrIndx; ++i)
7872 (*MIB).addOperand(*argOpers[i]);
7874 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7875 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7876 bInstr->memoperands_end());
7878 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7882 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7884 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7888 // private utility function: 64 bit atomics on 32 bit host.
7890 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7891 MachineBasicBlock *MBB,
7896 bool invSrc) const {
7897 // For the atomic bitwise operator, we generate
7898 // thisMBB (instructions are in pairs, except cmpxchg8b)
7899 // ld t1,t2 = [bitinstr.addr]
7901 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7902 // op t5, t6 <- out1, out2, [bitinstr.val]
7903 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7904 // mov ECX, EBX <- t5, t6
7905 // mov EAX, EDX <- t1, t2
7906 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7907 // mov t3, t4 <- EAX, EDX
7909 // result in out1, out2
7910 // fallthrough -->nextMBB
7912 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7913 const unsigned LoadOpc = X86::MOV32rm;
7914 const unsigned copyOpc = X86::MOV32rr;
7915 const unsigned NotOpc = X86::NOT32r;
7916 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7917 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7918 MachineFunction::iterator MBBIter = MBB;
7921 /// First build the CFG
7922 MachineFunction *F = MBB->getParent();
7923 MachineBasicBlock *thisMBB = MBB;
7924 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7925 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7926 F->insert(MBBIter, newMBB);
7927 F->insert(MBBIter, nextMBB);
7929 // Move all successors to thisMBB to nextMBB
7930 nextMBB->transferSuccessors(thisMBB);
7932 // Update thisMBB to fall through to newMBB
7933 thisMBB->addSuccessor(newMBB);
7935 // newMBB jumps to itself and fall through to nextMBB
7936 newMBB->addSuccessor(nextMBB);
7937 newMBB->addSuccessor(newMBB);
7939 DebugLoc dl = bInstr->getDebugLoc();
7940 // Insert instructions into newMBB based on incoming instruction
7941 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7942 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7943 "unexpected number of operands");
7944 MachineOperand& dest1Oper = bInstr->getOperand(0);
7945 MachineOperand& dest2Oper = bInstr->getOperand(1);
7946 MachineOperand* argOpers[2 + X86AddrNumOperands];
7947 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7948 argOpers[i] = &bInstr->getOperand(i+2);
7950 // x86 address has 5 operands: base, index, scale, displacement, and segment.
7951 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7953 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7954 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7955 for (int i=0; i <= lastAddrIndx; ++i)
7956 (*MIB).addOperand(*argOpers[i]);
7957 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7958 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7959 // add 4 to displacement.
7960 for (int i=0; i <= lastAddrIndx-2; ++i)
7961 (*MIB).addOperand(*argOpers[i]);
7962 MachineOperand newOp3 = *(argOpers[3]);
7964 newOp3.setImm(newOp3.getImm()+4);
7966 newOp3.setOffset(newOp3.getOffset()+4);
7967 (*MIB).addOperand(newOp3);
7968 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7970 // t3/4 are defined later, at the bottom of the loop
7971 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7972 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7973 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7974 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7975 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7976 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7978 // The subsequent operations should be using the destination registers of
7979 //the PHI instructions.
7981 t1 = F->getRegInfo().createVirtualRegister(RC);
7982 t2 = F->getRegInfo().createVirtualRegister(RC);
7983 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
7984 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
7986 t1 = dest1Oper.getReg();
7987 t2 = dest2Oper.getReg();
7990 int valArgIndx = lastAddrIndx + 1;
7991 assert((argOpers[valArgIndx]->isReg() ||
7992 argOpers[valArgIndx]->isImm()) &&
7994 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7995 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7996 if (argOpers[valArgIndx]->isReg())
7997 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7999 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8000 if (regOpcL != X86::MOV32rr)
8002 (*MIB).addOperand(*argOpers[valArgIndx]);
8003 assert(argOpers[valArgIndx + 1]->isReg() ==
8004 argOpers[valArgIndx]->isReg());
8005 assert(argOpers[valArgIndx + 1]->isImm() ==
8006 argOpers[valArgIndx]->isImm());
8007 if (argOpers[valArgIndx + 1]->isReg())
8008 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8010 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8011 if (regOpcH != X86::MOV32rr)
8013 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8015 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8017 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8020 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8022 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8025 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8026 for (int i=0; i <= lastAddrIndx; ++i)
8027 (*MIB).addOperand(*argOpers[i]);
8029 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8030 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8031 bInstr->memoperands_end());
8033 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8034 MIB.addReg(X86::EAX);
8035 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8036 MIB.addReg(X86::EDX);
8039 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
8041 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8045 // private utility function
8047 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8048 MachineBasicBlock *MBB,
8049 unsigned cmovOpc) const {
8050 // For the atomic min/max operator, we generate
8053 // ld t1 = [min/max.addr]
8054 // mov t2 = [min/max.val]
8056 // cmov[cond] t2 = t1
8058 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8060 // fallthrough -->nextMBB
8062 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8063 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8064 MachineFunction::iterator MBBIter = MBB;
8067 /// First build the CFG
8068 MachineFunction *F = MBB->getParent();
8069 MachineBasicBlock *thisMBB = MBB;
8070 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8071 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8072 F->insert(MBBIter, newMBB);
8073 F->insert(MBBIter, nextMBB);
8075 // Move all successors of thisMBB to nextMBB
8076 nextMBB->transferSuccessors(thisMBB);
8078 // Update thisMBB to fall through to newMBB
8079 thisMBB->addSuccessor(newMBB);
8081 // newMBB jumps to newMBB and fall through to nextMBB
8082 newMBB->addSuccessor(nextMBB);
8083 newMBB->addSuccessor(newMBB);
8085 DebugLoc dl = mInstr->getDebugLoc();
8086 // Insert instructions into newMBB based on incoming instruction
8087 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8088 "unexpected number of operands");
8089 MachineOperand& destOper = mInstr->getOperand(0);
8090 MachineOperand* argOpers[2 + X86AddrNumOperands];
8091 int numArgs = mInstr->getNumOperands() - 1;
8092 for (int i=0; i < numArgs; ++i)
8093 argOpers[i] = &mInstr->getOperand(i+1);
8095 // x86 address has 4 operands: base, index, scale, and displacement
8096 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8097 int valArgIndx = lastAddrIndx + 1;
8099 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8100 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8101 for (int i=0; i <= lastAddrIndx; ++i)
8102 (*MIB).addOperand(*argOpers[i]);
8104 // We only support register and immediate values
8105 assert((argOpers[valArgIndx]->isReg() ||
8106 argOpers[valArgIndx]->isImm()) &&
8109 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8110 if (argOpers[valArgIndx]->isReg())
8111 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8113 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8114 (*MIB).addOperand(*argOpers[valArgIndx]);
8116 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8119 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8124 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8125 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8129 // Cmp and exchange if none has modified the memory location
8130 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8131 for (int i=0; i <= lastAddrIndx; ++i)
8132 (*MIB).addOperand(*argOpers[i]);
8134 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8135 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8136 mInstr->memoperands_end());
8138 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8139 MIB.addReg(X86::EAX);
8142 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
8144 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8148 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8149 // all of this code can be replaced with that in the .td file.
8151 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8152 unsigned numArgs, bool memArg) const {
8154 MachineFunction *F = BB->getParent();
8155 DebugLoc dl = MI->getDebugLoc();
8156 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8160 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8162 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8164 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8166 for (unsigned i = 0; i < numArgs; ++i) {
8167 MachineOperand &Op = MI->getOperand(i+1);
8169 if (!(Op.isReg() && Op.isImplicit()))
8173 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8176 F->DeleteMachineInstr(MI);
8182 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8184 MachineBasicBlock *MBB) const {
8185 // Emit code to save XMM registers to the stack. The ABI says that the
8186 // number of registers to save is given in %al, so it's theoretically
8187 // possible to do an indirect jump trick to avoid saving all of them,
8188 // however this code takes a simpler approach and just executes all
8189 // of the stores if %al is non-zero. It's less code, and it's probably
8190 // easier on the hardware branch predictor, and stores aren't all that
8191 // expensive anyway.
8193 // Create the new basic blocks. One block contains all the XMM stores,
8194 // and one block is the final destination regardless of whether any
8195 // stores were performed.
8196 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8197 MachineFunction *F = MBB->getParent();
8198 MachineFunction::iterator MBBIter = MBB;
8200 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8201 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8202 F->insert(MBBIter, XMMSaveMBB);
8203 F->insert(MBBIter, EndMBB);
8206 // Move any original successors of MBB to the end block.
8207 EndMBB->transferSuccessors(MBB);
8208 // The original block will now fall through to the XMM save block.
8209 MBB->addSuccessor(XMMSaveMBB);
8210 // The XMMSaveMBB will fall through to the end block.
8211 XMMSaveMBB->addSuccessor(EndMBB);
8213 // Now add the instructions.
8214 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8215 DebugLoc DL = MI->getDebugLoc();
8217 unsigned CountReg = MI->getOperand(0).getReg();
8218 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8219 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8221 if (!Subtarget->isTargetWin64()) {
8222 // If %al is 0, branch around the XMM save block.
8223 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8224 BuildMI(MBB, DL, TII->get(X86::JE)).addMBB(EndMBB);
8225 MBB->addSuccessor(EndMBB);
8228 // In the XMM save block, save all the XMM argument registers.
8229 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8230 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8231 MachineMemOperand *MMO =
8232 F->getMachineMemOperand(
8233 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8234 MachineMemOperand::MOStore, Offset,
8235 /*Size=*/16, /*Align=*/16);
8236 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8237 .addFrameIndex(RegSaveFrameIndex)
8238 .addImm(/*Scale=*/1)
8239 .addReg(/*IndexReg=*/0)
8240 .addImm(/*Disp=*/Offset)
8241 .addReg(/*Segment=*/0)
8242 .addReg(MI->getOperand(i).getReg())
8243 .addMemOperand(MMO);
8246 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8252 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8253 MachineBasicBlock *BB,
8254 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8255 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8256 DebugLoc DL = MI->getDebugLoc();
8258 // To "insert" a SELECT_CC instruction, we actually have to insert the
8259 // diamond control-flow pattern. The incoming instruction knows the
8260 // destination vreg to set, the condition code register to branch on, the
8261 // true/false values to select between, and a branch opcode to use.
8262 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8263 MachineFunction::iterator It = BB;
8269 // cmpTY ccX, r1, r2
8271 // fallthrough --> copy0MBB
8272 MachineBasicBlock *thisMBB = BB;
8273 MachineFunction *F = BB->getParent();
8274 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8275 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8277 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8278 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8279 F->insert(It, copy0MBB);
8280 F->insert(It, sinkMBB);
8281 // Update machine-CFG edges by first adding all successors of the current
8282 // block to the new block which will contain the Phi node for the select.
8283 // Also inform sdisel of the edge changes.
8284 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8285 E = BB->succ_end(); I != E; ++I) {
8286 EM->insert(std::make_pair(*I, sinkMBB));
8287 sinkMBB->addSuccessor(*I);
8289 // Next, remove all successors of the current block, and add the true
8290 // and fallthrough blocks as its successors.
8291 while (!BB->succ_empty())
8292 BB->removeSuccessor(BB->succ_begin());
8293 // Add the true and fallthrough blocks as its successors.
8294 BB->addSuccessor(copy0MBB);
8295 BB->addSuccessor(sinkMBB);
8298 // %FalseValue = ...
8299 // # fallthrough to sinkMBB
8302 // Update machine-CFG edges
8303 BB->addSuccessor(sinkMBB);
8306 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8309 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8310 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8311 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8313 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8319 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8320 MachineBasicBlock *BB,
8321 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8322 switch (MI->getOpcode()) {
8323 default: assert(false && "Unexpected instr type to insert");
8325 case X86::CMOV_V1I64:
8326 case X86::CMOV_FR32:
8327 case X86::CMOV_FR64:
8328 case X86::CMOV_V4F32:
8329 case X86::CMOV_V2F64:
8330 case X86::CMOV_V2I64:
8331 return EmitLoweredSelect(MI, BB, EM);
8333 case X86::FP32_TO_INT16_IN_MEM:
8334 case X86::FP32_TO_INT32_IN_MEM:
8335 case X86::FP32_TO_INT64_IN_MEM:
8336 case X86::FP64_TO_INT16_IN_MEM:
8337 case X86::FP64_TO_INT32_IN_MEM:
8338 case X86::FP64_TO_INT64_IN_MEM:
8339 case X86::FP80_TO_INT16_IN_MEM:
8340 case X86::FP80_TO_INT32_IN_MEM:
8341 case X86::FP80_TO_INT64_IN_MEM: {
8342 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8343 DebugLoc DL = MI->getDebugLoc();
8345 // Change the floating point control register to use "round towards zero"
8346 // mode when truncating to an integer value.
8347 MachineFunction *F = BB->getParent();
8348 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8349 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8351 // Load the old value of the high byte of the control word...
8353 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8354 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8357 // Set the high part to be round to zero...
8358 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8361 // Reload the modified control word now...
8362 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8364 // Restore the memory image of control word to original value
8365 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8368 // Get the X86 opcode to use.
8370 switch (MI->getOpcode()) {
8371 default: llvm_unreachable("illegal opcode!");
8372 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8373 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8374 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8375 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8376 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8377 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8378 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8379 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8380 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8384 MachineOperand &Op = MI->getOperand(0);
8386 AM.BaseType = X86AddressMode::RegBase;
8387 AM.Base.Reg = Op.getReg();
8389 AM.BaseType = X86AddressMode::FrameIndexBase;
8390 AM.Base.FrameIndex = Op.getIndex();
8392 Op = MI->getOperand(1);
8394 AM.Scale = Op.getImm();
8395 Op = MI->getOperand(2);
8397 AM.IndexReg = Op.getImm();
8398 Op = MI->getOperand(3);
8399 if (Op.isGlobal()) {
8400 AM.GV = Op.getGlobal();
8402 AM.Disp = Op.getImm();
8404 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8405 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8407 // Reload the original control word now.
8408 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8410 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8413 // String/text processing lowering.
8414 case X86::PCMPISTRM128REG:
8415 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8416 case X86::PCMPISTRM128MEM:
8417 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8418 case X86::PCMPESTRM128REG:
8419 return EmitPCMP(MI, BB, 5, false /* in mem */);
8420 case X86::PCMPESTRM128MEM:
8421 return EmitPCMP(MI, BB, 5, true /* in mem */);
8424 case X86::ATOMAND32:
8425 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8426 X86::AND32ri, X86::MOV32rm,
8427 X86::LCMPXCHG32, X86::MOV32rr,
8428 X86::NOT32r, X86::EAX,
8429 X86::GR32RegisterClass);
8431 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8432 X86::OR32ri, X86::MOV32rm,
8433 X86::LCMPXCHG32, X86::MOV32rr,
8434 X86::NOT32r, X86::EAX,
8435 X86::GR32RegisterClass);
8436 case X86::ATOMXOR32:
8437 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8438 X86::XOR32ri, X86::MOV32rm,
8439 X86::LCMPXCHG32, X86::MOV32rr,
8440 X86::NOT32r, X86::EAX,
8441 X86::GR32RegisterClass);
8442 case X86::ATOMNAND32:
8443 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8444 X86::AND32ri, X86::MOV32rm,
8445 X86::LCMPXCHG32, X86::MOV32rr,
8446 X86::NOT32r, X86::EAX,
8447 X86::GR32RegisterClass, true);
8448 case X86::ATOMMIN32:
8449 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8450 case X86::ATOMMAX32:
8451 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8452 case X86::ATOMUMIN32:
8453 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8454 case X86::ATOMUMAX32:
8455 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8457 case X86::ATOMAND16:
8458 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8459 X86::AND16ri, X86::MOV16rm,
8460 X86::LCMPXCHG16, X86::MOV16rr,
8461 X86::NOT16r, X86::AX,
8462 X86::GR16RegisterClass);
8464 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8465 X86::OR16ri, X86::MOV16rm,
8466 X86::LCMPXCHG16, X86::MOV16rr,
8467 X86::NOT16r, X86::AX,
8468 X86::GR16RegisterClass);
8469 case X86::ATOMXOR16:
8470 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8471 X86::XOR16ri, X86::MOV16rm,
8472 X86::LCMPXCHG16, X86::MOV16rr,
8473 X86::NOT16r, X86::AX,
8474 X86::GR16RegisterClass);
8475 case X86::ATOMNAND16:
8476 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8477 X86::AND16ri, X86::MOV16rm,
8478 X86::LCMPXCHG16, X86::MOV16rr,
8479 X86::NOT16r, X86::AX,
8480 X86::GR16RegisterClass, true);
8481 case X86::ATOMMIN16:
8482 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8483 case X86::ATOMMAX16:
8484 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8485 case X86::ATOMUMIN16:
8486 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8487 case X86::ATOMUMAX16:
8488 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8491 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8492 X86::AND8ri, X86::MOV8rm,
8493 X86::LCMPXCHG8, X86::MOV8rr,
8494 X86::NOT8r, X86::AL,
8495 X86::GR8RegisterClass);
8497 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8498 X86::OR8ri, X86::MOV8rm,
8499 X86::LCMPXCHG8, X86::MOV8rr,
8500 X86::NOT8r, X86::AL,
8501 X86::GR8RegisterClass);
8503 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8504 X86::XOR8ri, X86::MOV8rm,
8505 X86::LCMPXCHG8, X86::MOV8rr,
8506 X86::NOT8r, X86::AL,
8507 X86::GR8RegisterClass);
8508 case X86::ATOMNAND8:
8509 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8510 X86::AND8ri, X86::MOV8rm,
8511 X86::LCMPXCHG8, X86::MOV8rr,
8512 X86::NOT8r, X86::AL,
8513 X86::GR8RegisterClass, true);
8514 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8515 // This group is for 64-bit host.
8516 case X86::ATOMAND64:
8517 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8518 X86::AND64ri32, X86::MOV64rm,
8519 X86::LCMPXCHG64, X86::MOV64rr,
8520 X86::NOT64r, X86::RAX,
8521 X86::GR64RegisterClass);
8523 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8524 X86::OR64ri32, X86::MOV64rm,
8525 X86::LCMPXCHG64, X86::MOV64rr,
8526 X86::NOT64r, X86::RAX,
8527 X86::GR64RegisterClass);
8528 case X86::ATOMXOR64:
8529 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8530 X86::XOR64ri32, X86::MOV64rm,
8531 X86::LCMPXCHG64, X86::MOV64rr,
8532 X86::NOT64r, X86::RAX,
8533 X86::GR64RegisterClass);
8534 case X86::ATOMNAND64:
8535 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8536 X86::AND64ri32, X86::MOV64rm,
8537 X86::LCMPXCHG64, X86::MOV64rr,
8538 X86::NOT64r, X86::RAX,
8539 X86::GR64RegisterClass, true);
8540 case X86::ATOMMIN64:
8541 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8542 case X86::ATOMMAX64:
8543 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8544 case X86::ATOMUMIN64:
8545 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8546 case X86::ATOMUMAX64:
8547 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8549 // This group does 64-bit operations on a 32-bit host.
8550 case X86::ATOMAND6432:
8551 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8552 X86::AND32rr, X86::AND32rr,
8553 X86::AND32ri, X86::AND32ri,
8555 case X86::ATOMOR6432:
8556 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8557 X86::OR32rr, X86::OR32rr,
8558 X86::OR32ri, X86::OR32ri,
8560 case X86::ATOMXOR6432:
8561 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8562 X86::XOR32rr, X86::XOR32rr,
8563 X86::XOR32ri, X86::XOR32ri,
8565 case X86::ATOMNAND6432:
8566 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8567 X86::AND32rr, X86::AND32rr,
8568 X86::AND32ri, X86::AND32ri,
8570 case X86::ATOMADD6432:
8571 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8572 X86::ADD32rr, X86::ADC32rr,
8573 X86::ADD32ri, X86::ADC32ri,
8575 case X86::ATOMSUB6432:
8576 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8577 X86::SUB32rr, X86::SBB32rr,
8578 X86::SUB32ri, X86::SBB32ri,
8580 case X86::ATOMSWAP6432:
8581 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8582 X86::MOV32rr, X86::MOV32rr,
8583 X86::MOV32ri, X86::MOV32ri,
8585 case X86::VASTART_SAVE_XMM_REGS:
8586 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8590 //===----------------------------------------------------------------------===//
8591 // X86 Optimization Hooks
8592 //===----------------------------------------------------------------------===//
8594 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8598 const SelectionDAG &DAG,
8599 unsigned Depth) const {
8600 unsigned Opc = Op.getOpcode();
8601 assert((Opc >= ISD::BUILTIN_OP_END ||
8602 Opc == ISD::INTRINSIC_WO_CHAIN ||
8603 Opc == ISD::INTRINSIC_W_CHAIN ||
8604 Opc == ISD::INTRINSIC_VOID) &&
8605 "Should use MaskedValueIsZero if you don't know whether Op"
8606 " is a target node!");
8608 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8620 // These nodes' second result is a boolean.
8621 if (Op.getResNo() == 0)
8625 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8626 Mask.getBitWidth() - 1);
8631 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8632 /// node is a GlobalAddress + offset.
8633 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8634 GlobalValue* &GA, int64_t &Offset) const{
8635 if (N->getOpcode() == X86ISD::Wrapper) {
8636 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8637 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8638 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8642 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8645 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8646 EVT EltVT, LoadSDNode *&LDBase,
8647 unsigned &LastLoadedElt,
8648 SelectionDAG &DAG, MachineFrameInfo *MFI,
8649 const TargetLowering &TLI) {
8651 LastLoadedElt = -1U;
8652 for (unsigned i = 0; i < NumElems; ++i) {
8653 if (N->getMaskElt(i) < 0) {
8659 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8660 if (!Elt.getNode() ||
8661 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8664 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8666 LDBase = cast<LoadSDNode>(Elt.getNode());
8670 if (Elt.getOpcode() == ISD::UNDEF)
8673 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8674 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
8681 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8682 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8683 /// if the load addresses are consecutive, non-overlapping, and in the right
8684 /// order. In the case of v2i64, it will see if it can rewrite the
8685 /// shuffle to be an appropriate build vector so it can take advantage of
8686 // performBuildVectorCombine.
8687 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8688 const TargetLowering &TLI) {
8689 DebugLoc dl = N->getDebugLoc();
8690 EVT VT = N->getValueType(0);
8691 EVT EltVT = VT.getVectorElementType();
8692 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8693 unsigned NumElems = VT.getVectorNumElements();
8695 if (VT.getSizeInBits() != 128)
8698 // Try to combine a vector_shuffle into a 128-bit load.
8699 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8700 LoadSDNode *LD = NULL;
8701 unsigned LastLoadedElt;
8702 if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
8706 if (LastLoadedElt == NumElems - 1) {
8707 if (DAG.InferPtrAlignment(LD->getBasePtr()) >= 16)
8708 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8709 LD->getSrcValue(), LD->getSrcValueOffset(),
8711 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8712 LD->getSrcValue(), LD->getSrcValueOffset(),
8713 LD->isVolatile(), LD->getAlignment());
8714 } else if (NumElems == 4 && LastLoadedElt == 1) {
8715 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8716 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8717 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8718 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8723 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8724 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8725 const X86Subtarget *Subtarget) {
8726 DebugLoc DL = N->getDebugLoc();
8727 SDValue Cond = N->getOperand(0);
8728 // Get the LHS/RHS of the select.
8729 SDValue LHS = N->getOperand(1);
8730 SDValue RHS = N->getOperand(2);
8732 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8733 // instructions have the peculiarity that if either operand is a NaN,
8734 // they chose what we call the RHS operand (and as such are not symmetric).
8735 // It happens that this matches the semantics of the common C idiom
8736 // x<y?x:y and related forms, so we can recognize these cases.
8737 if (Subtarget->hasSSE2() &&
8738 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8739 Cond.getOpcode() == ISD::SETCC) {
8740 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8742 unsigned Opcode = 0;
8743 // Check for x CC y ? x : y.
8744 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8748 // This can be a min if we can prove that at least one of the operands
8750 if (!FiniteOnlyFPMath()) {
8751 if (DAG.isKnownNeverNaN(RHS)) {
8752 // Put the potential NaN in the RHS so that SSE will preserve it.
8753 std::swap(LHS, RHS);
8754 } else if (!DAG.isKnownNeverNaN(LHS))
8757 Opcode = X86ISD::FMIN;
8760 // This can be a min if we can prove that at least one of the operands
8762 if (!FiniteOnlyFPMath()) {
8763 if (DAG.isKnownNeverNaN(LHS)) {
8764 // Put the potential NaN in the RHS so that SSE will preserve it.
8765 std::swap(LHS, RHS);
8766 } else if (!DAG.isKnownNeverNaN(RHS))
8769 Opcode = X86ISD::FMIN;
8772 // This can be a min, but if either operand is a NaN we need it to
8773 // preserve the original LHS.
8774 std::swap(LHS, RHS);
8778 Opcode = X86ISD::FMIN;
8782 // This can be a max if we can prove that at least one of the operands
8784 if (!FiniteOnlyFPMath()) {
8785 if (DAG.isKnownNeverNaN(LHS)) {
8786 // Put the potential NaN in the RHS so that SSE will preserve it.
8787 std::swap(LHS, RHS);
8788 } else if (!DAG.isKnownNeverNaN(RHS))
8791 Opcode = X86ISD::FMAX;
8794 // This can be a max if we can prove that at least one of the operands
8796 if (!FiniteOnlyFPMath()) {
8797 if (DAG.isKnownNeverNaN(RHS)) {
8798 // Put the potential NaN in the RHS so that SSE will preserve it.
8799 std::swap(LHS, RHS);
8800 } else if (!DAG.isKnownNeverNaN(LHS))
8803 Opcode = X86ISD::FMAX;
8806 // This can be a max, but if either operand is a NaN we need it to
8807 // preserve the original LHS.
8808 std::swap(LHS, RHS);
8812 Opcode = X86ISD::FMAX;
8815 // Check for x CC y ? y : x -- a min/max with reversed arms.
8816 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8820 // This can be a min if we can prove that at least one of the operands
8822 if (!FiniteOnlyFPMath()) {
8823 if (DAG.isKnownNeverNaN(RHS)) {
8824 // Put the potential NaN in the RHS so that SSE will preserve it.
8825 std::swap(LHS, RHS);
8826 } else if (!DAG.isKnownNeverNaN(LHS))
8829 Opcode = X86ISD::FMIN;
8832 // This can be a min if we can prove that at least one of the operands
8834 if (!FiniteOnlyFPMath()) {
8835 if (DAG.isKnownNeverNaN(LHS)) {
8836 // Put the potential NaN in the RHS so that SSE will preserve it.
8837 std::swap(LHS, RHS);
8838 } else if (!DAG.isKnownNeverNaN(RHS))
8841 Opcode = X86ISD::FMIN;
8844 // This can be a min, but if either operand is a NaN we need it to
8845 // preserve the original LHS.
8846 std::swap(LHS, RHS);
8850 Opcode = X86ISD::FMIN;
8854 // This can be a max if we can prove that at least one of the operands
8856 if (!FiniteOnlyFPMath()) {
8857 if (DAG.isKnownNeverNaN(LHS)) {
8858 // Put the potential NaN in the RHS so that SSE will preserve it.
8859 std::swap(LHS, RHS);
8860 } else if (!DAG.isKnownNeverNaN(RHS))
8863 Opcode = X86ISD::FMAX;
8866 // This can be a max if we can prove that at least one of the operands
8868 if (!FiniteOnlyFPMath()) {
8869 if (DAG.isKnownNeverNaN(RHS)) {
8870 // Put the potential NaN in the RHS so that SSE will preserve it.
8871 std::swap(LHS, RHS);
8872 } else if (!DAG.isKnownNeverNaN(LHS))
8875 Opcode = X86ISD::FMAX;
8878 // This can be a max, but if either operand is a NaN we need it to
8879 // preserve the original LHS.
8880 std::swap(LHS, RHS);
8884 Opcode = X86ISD::FMAX;
8890 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8893 // If this is a select between two integer constants, try to do some
8895 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8896 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8897 // Don't do this for crazy integer types.
8898 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8899 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8900 // so that TrueC (the true value) is larger than FalseC.
8901 bool NeedsCondInvert = false;
8903 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8904 // Efficiently invertible.
8905 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8906 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8907 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8908 NeedsCondInvert = true;
8909 std::swap(TrueC, FalseC);
8912 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8913 if (FalseC->getAPIntValue() == 0 &&
8914 TrueC->getAPIntValue().isPowerOf2()) {
8915 if (NeedsCondInvert) // Invert the condition if needed.
8916 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8917 DAG.getConstant(1, Cond.getValueType()));
8919 // Zero extend the condition if needed.
8920 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8922 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8923 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8924 DAG.getConstant(ShAmt, MVT::i8));
8927 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8928 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8929 if (NeedsCondInvert) // Invert the condition if needed.
8930 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8931 DAG.getConstant(1, Cond.getValueType()));
8933 // Zero extend the condition if needed.
8934 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8935 FalseC->getValueType(0), Cond);
8936 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8937 SDValue(FalseC, 0));
8940 // Optimize cases that will turn into an LEA instruction. This requires
8941 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8942 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8943 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8944 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8946 bool isFastMultiplier = false;
8948 switch ((unsigned char)Diff) {
8950 case 1: // result = add base, cond
8951 case 2: // result = lea base( , cond*2)
8952 case 3: // result = lea base(cond, cond*2)
8953 case 4: // result = lea base( , cond*4)
8954 case 5: // result = lea base(cond, cond*4)
8955 case 8: // result = lea base( , cond*8)
8956 case 9: // result = lea base(cond, cond*8)
8957 isFastMultiplier = true;
8962 if (isFastMultiplier) {
8963 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8964 if (NeedsCondInvert) // Invert the condition if needed.
8965 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8966 DAG.getConstant(1, Cond.getValueType()));
8968 // Zero extend the condition if needed.
8969 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8971 // Scale the condition by the difference.
8973 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8974 DAG.getConstant(Diff, Cond.getValueType()));
8976 // Add the base if non-zero.
8977 if (FalseC->getAPIntValue() != 0)
8978 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8979 SDValue(FalseC, 0));
8989 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
8990 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8991 TargetLowering::DAGCombinerInfo &DCI) {
8992 DebugLoc DL = N->getDebugLoc();
8994 // If the flag operand isn't dead, don't touch this CMOV.
8995 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8998 // If this is a select between two integer constants, try to do some
8999 // optimizations. Note that the operands are ordered the opposite of SELECT
9001 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9002 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9003 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9004 // larger than FalseC (the false value).
9005 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9007 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9008 CC = X86::GetOppositeBranchCondition(CC);
9009 std::swap(TrueC, FalseC);
9012 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9013 // This is efficient for any integer data type (including i8/i16) and
9015 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9016 SDValue Cond = N->getOperand(3);
9017 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9018 DAG.getConstant(CC, MVT::i8), Cond);
9020 // Zero extend the condition if needed.
9021 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9023 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9024 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9025 DAG.getConstant(ShAmt, MVT::i8));
9026 if (N->getNumValues() == 2) // Dead flag value?
9027 return DCI.CombineTo(N, Cond, SDValue());
9031 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9032 // for any integer data type, including i8/i16.
9033 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9034 SDValue Cond = N->getOperand(3);
9035 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9036 DAG.getConstant(CC, MVT::i8), Cond);
9038 // Zero extend the condition if needed.
9039 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9040 FalseC->getValueType(0), Cond);
9041 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9042 SDValue(FalseC, 0));
9044 if (N->getNumValues() == 2) // Dead flag value?
9045 return DCI.CombineTo(N, Cond, SDValue());
9049 // Optimize cases that will turn into an LEA instruction. This requires
9050 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9051 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9052 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9053 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9055 bool isFastMultiplier = false;
9057 switch ((unsigned char)Diff) {
9059 case 1: // result = add base, cond
9060 case 2: // result = lea base( , cond*2)
9061 case 3: // result = lea base(cond, cond*2)
9062 case 4: // result = lea base( , cond*4)
9063 case 5: // result = lea base(cond, cond*4)
9064 case 8: // result = lea base( , cond*8)
9065 case 9: // result = lea base(cond, cond*8)
9066 isFastMultiplier = true;
9071 if (isFastMultiplier) {
9072 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9073 SDValue Cond = N->getOperand(3);
9074 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9075 DAG.getConstant(CC, MVT::i8), Cond);
9076 // Zero extend the condition if needed.
9077 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9079 // Scale the condition by the difference.
9081 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9082 DAG.getConstant(Diff, Cond.getValueType()));
9084 // Add the base if non-zero.
9085 if (FalseC->getAPIntValue() != 0)
9086 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9087 SDValue(FalseC, 0));
9088 if (N->getNumValues() == 2) // Dead flag value?
9089 return DCI.CombineTo(N, Cond, SDValue());
9099 /// PerformMulCombine - Optimize a single multiply with constant into two
9100 /// in order to implement it with two cheaper instructions, e.g.
9101 /// LEA + SHL, LEA + LEA.
9102 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9103 TargetLowering::DAGCombinerInfo &DCI) {
9104 if (DAG.getMachineFunction().
9105 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
9108 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9111 EVT VT = N->getValueType(0);
9115 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9118 uint64_t MulAmt = C->getZExtValue();
9119 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9122 uint64_t MulAmt1 = 0;
9123 uint64_t MulAmt2 = 0;
9124 if ((MulAmt % 9) == 0) {
9126 MulAmt2 = MulAmt / 9;
9127 } else if ((MulAmt % 5) == 0) {
9129 MulAmt2 = MulAmt / 5;
9130 } else if ((MulAmt % 3) == 0) {
9132 MulAmt2 = MulAmt / 3;
9135 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9136 DebugLoc DL = N->getDebugLoc();
9138 if (isPowerOf2_64(MulAmt2) &&
9139 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9140 // If second multiplifer is pow2, issue it first. We want the multiply by
9141 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9143 std::swap(MulAmt1, MulAmt2);
9146 if (isPowerOf2_64(MulAmt1))
9147 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9148 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9150 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9151 DAG.getConstant(MulAmt1, VT));
9153 if (isPowerOf2_64(MulAmt2))
9154 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9155 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9157 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9158 DAG.getConstant(MulAmt2, VT));
9160 // Do not add new nodes to DAG combiner worklist.
9161 DCI.CombineTo(N, NewMul, false);
9166 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9167 SDValue N0 = N->getOperand(0);
9168 SDValue N1 = N->getOperand(1);
9169 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9170 EVT VT = N0.getValueType();
9172 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9173 // since the result of setcc_c is all zero's or all ones.
9174 if (N1C && N0.getOpcode() == ISD::AND &&
9175 N0.getOperand(1).getOpcode() == ISD::Constant) {
9176 SDValue N00 = N0.getOperand(0);
9177 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9178 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9179 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9180 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9181 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9182 APInt ShAmt = N1C->getAPIntValue();
9183 Mask = Mask.shl(ShAmt);
9185 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9186 N00, DAG.getConstant(Mask, VT));
9193 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9195 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9196 const X86Subtarget *Subtarget) {
9197 EVT VT = N->getValueType(0);
9198 if (!VT.isVector() && VT.isInteger() &&
9199 N->getOpcode() == ISD::SHL)
9200 return PerformSHLCombine(N, DAG);
9202 // On X86 with SSE2 support, we can transform this to a vector shift if
9203 // all elements are shifted by the same amount. We can't do this in legalize
9204 // because the a constant vector is typically transformed to a constant pool
9205 // so we have no knowledge of the shift amount.
9206 if (!Subtarget->hasSSE2())
9209 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9212 SDValue ShAmtOp = N->getOperand(1);
9213 EVT EltVT = VT.getVectorElementType();
9214 DebugLoc DL = N->getDebugLoc();
9215 SDValue BaseShAmt = SDValue();
9216 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9217 unsigned NumElts = VT.getVectorNumElements();
9219 for (; i != NumElts; ++i) {
9220 SDValue Arg = ShAmtOp.getOperand(i);
9221 if (Arg.getOpcode() == ISD::UNDEF) continue;
9225 for (; i != NumElts; ++i) {
9226 SDValue Arg = ShAmtOp.getOperand(i);
9227 if (Arg.getOpcode() == ISD::UNDEF) continue;
9228 if (Arg != BaseShAmt) {
9232 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9233 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9234 SDValue InVec = ShAmtOp.getOperand(0);
9235 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9236 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9238 for (; i != NumElts; ++i) {
9239 SDValue Arg = InVec.getOperand(i);
9240 if (Arg.getOpcode() == ISD::UNDEF) continue;
9244 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9245 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9246 unsigned SplatIdx = cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9247 if (C->getZExtValue() == SplatIdx)
9248 BaseShAmt = InVec.getOperand(1);
9251 if (BaseShAmt.getNode() == 0)
9252 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9253 DAG.getIntPtrConstant(0));
9257 // The shift amount is an i32.
9258 if (EltVT.bitsGT(MVT::i32))
9259 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9260 else if (EltVT.bitsLT(MVT::i32))
9261 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9263 // The shift amount is identical so we can do a vector shift.
9264 SDValue ValOp = N->getOperand(0);
9265 switch (N->getOpcode()) {
9267 llvm_unreachable("Unknown shift opcode!");
9270 if (VT == MVT::v2i64)
9271 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9272 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9274 if (VT == MVT::v4i32)
9275 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9276 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9278 if (VT == MVT::v8i16)
9279 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9280 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9284 if (VT == MVT::v4i32)
9285 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9286 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9288 if (VT == MVT::v8i16)
9289 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9290 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9294 if (VT == MVT::v2i64)
9295 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9296 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9298 if (VT == MVT::v4i32)
9299 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9300 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9302 if (VT == MVT::v8i16)
9303 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9304 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9311 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9312 const X86Subtarget *Subtarget) {
9313 EVT VT = N->getValueType(0);
9314 if (VT != MVT::i64 || !Subtarget->is64Bit())
9317 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9318 SDValue N0 = N->getOperand(0);
9319 SDValue N1 = N->getOperand(1);
9320 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9322 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9325 SDValue ShAmt0 = N0.getOperand(1);
9326 if (ShAmt0.getValueType() != MVT::i8)
9328 SDValue ShAmt1 = N1.getOperand(1);
9329 if (ShAmt1.getValueType() != MVT::i8)
9331 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9332 ShAmt0 = ShAmt0.getOperand(0);
9333 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9334 ShAmt1 = ShAmt1.getOperand(0);
9336 DebugLoc DL = N->getDebugLoc();
9337 unsigned Opc = X86ISD::SHLD;
9338 SDValue Op0 = N0.getOperand(0);
9339 SDValue Op1 = N1.getOperand(0);
9340 if (ShAmt0.getOpcode() == ISD::SUB) {
9342 std::swap(Op0, Op1);
9343 std::swap(ShAmt0, ShAmt1);
9346 if (ShAmt1.getOpcode() == ISD::SUB) {
9347 SDValue Sum = ShAmt1.getOperand(0);
9348 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9349 if (SumC->getSExtValue() == 64 &&
9350 ShAmt1.getOperand(1) == ShAmt0)
9351 return DAG.getNode(Opc, DL, VT,
9353 DAG.getNode(ISD::TRUNCATE, DL,
9356 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9357 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9359 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
9360 return DAG.getNode(Opc, DL, VT,
9361 N0.getOperand(0), N1.getOperand(0),
9362 DAG.getNode(ISD::TRUNCATE, DL,
9369 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9370 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9371 const X86Subtarget *Subtarget) {
9372 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9373 // the FP state in cases where an emms may be missing.
9374 // A preferable solution to the general problem is to figure out the right
9375 // places to insert EMMS. This qualifies as a quick hack.
9377 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9378 StoreSDNode *St = cast<StoreSDNode>(N);
9379 EVT VT = St->getValue().getValueType();
9380 if (VT.getSizeInBits() != 64)
9383 const Function *F = DAG.getMachineFunction().getFunction();
9384 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9385 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9386 && Subtarget->hasSSE2();
9387 if ((VT.isVector() ||
9388 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9389 isa<LoadSDNode>(St->getValue()) &&
9390 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9391 St->getChain().hasOneUse() && !St->isVolatile()) {
9392 SDNode* LdVal = St->getValue().getNode();
9394 int TokenFactorIndex = -1;
9395 SmallVector<SDValue, 8> Ops;
9396 SDNode* ChainVal = St->getChain().getNode();
9397 // Must be a store of a load. We currently handle two cases: the load
9398 // is a direct child, and it's under an intervening TokenFactor. It is
9399 // possible to dig deeper under nested TokenFactors.
9400 if (ChainVal == LdVal)
9401 Ld = cast<LoadSDNode>(St->getChain());
9402 else if (St->getValue().hasOneUse() &&
9403 ChainVal->getOpcode() == ISD::TokenFactor) {
9404 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9405 if (ChainVal->getOperand(i).getNode() == LdVal) {
9406 TokenFactorIndex = i;
9407 Ld = cast<LoadSDNode>(St->getValue());
9409 Ops.push_back(ChainVal->getOperand(i));
9413 if (!Ld || !ISD::isNormalLoad(Ld))
9416 // If this is not the MMX case, i.e. we are just turning i64 load/store
9417 // into f64 load/store, avoid the transformation if there are multiple
9418 // uses of the loaded value.
9419 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9422 DebugLoc LdDL = Ld->getDebugLoc();
9423 DebugLoc StDL = N->getDebugLoc();
9424 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9425 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9427 if (Subtarget->is64Bit() || F64IsLegal) {
9428 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9429 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9430 Ld->getBasePtr(), Ld->getSrcValue(),
9431 Ld->getSrcValueOffset(), Ld->isVolatile(),
9432 Ld->getAlignment());
9433 SDValue NewChain = NewLd.getValue(1);
9434 if (TokenFactorIndex != -1) {
9435 Ops.push_back(NewChain);
9436 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9439 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9440 St->getSrcValue(), St->getSrcValueOffset(),
9441 St->isVolatile(), St->getAlignment());
9444 // Otherwise, lower to two pairs of 32-bit loads / stores.
9445 SDValue LoAddr = Ld->getBasePtr();
9446 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9447 DAG.getConstant(4, MVT::i32));
9449 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9450 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9451 Ld->isVolatile(), Ld->getAlignment());
9452 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9453 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9455 MinAlign(Ld->getAlignment(), 4));
9457 SDValue NewChain = LoLd.getValue(1);
9458 if (TokenFactorIndex != -1) {
9459 Ops.push_back(LoLd);
9460 Ops.push_back(HiLd);
9461 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9465 LoAddr = St->getBasePtr();
9466 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9467 DAG.getConstant(4, MVT::i32));
9469 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9470 St->getSrcValue(), St->getSrcValueOffset(),
9471 St->isVolatile(), St->getAlignment());
9472 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9474 St->getSrcValueOffset() + 4,
9476 MinAlign(St->getAlignment(), 4));
9477 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9482 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9483 /// X86ISD::FXOR nodes.
9484 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9485 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9486 // F[X]OR(0.0, x) -> x
9487 // F[X]OR(x, 0.0) -> x
9488 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9489 if (C->getValueAPF().isPosZero())
9490 return N->getOperand(1);
9491 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9492 if (C->getValueAPF().isPosZero())
9493 return N->getOperand(0);
9497 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9498 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9499 // FAND(0.0, x) -> 0.0
9500 // FAND(x, 0.0) -> 0.0
9501 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9502 if (C->getValueAPF().isPosZero())
9503 return N->getOperand(0);
9504 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9505 if (C->getValueAPF().isPosZero())
9506 return N->getOperand(1);
9510 static SDValue PerformBTCombine(SDNode *N,
9512 TargetLowering::DAGCombinerInfo &DCI) {
9513 // BT ignores high bits in the bit index operand.
9514 SDValue Op1 = N->getOperand(1);
9515 if (Op1.hasOneUse()) {
9516 unsigned BitWidth = Op1.getValueSizeInBits();
9517 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9518 APInt KnownZero, KnownOne;
9519 TargetLowering::TargetLoweringOpt TLO(DAG);
9520 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9521 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9522 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9523 DCI.CommitTargetLoweringOpt(TLO);
9528 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9529 SDValue Op = N->getOperand(0);
9530 if (Op.getOpcode() == ISD::BIT_CONVERT)
9531 Op = Op.getOperand(0);
9532 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9533 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9534 VT.getVectorElementType().getSizeInBits() ==
9535 OpVT.getVectorElementType().getSizeInBits()) {
9536 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9541 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9542 // Locked instructions, in turn, have implicit fence semantics (all memory
9543 // operations are flushed before issuing the locked instruction, and the
9544 // are not buffered), so we can fold away the common pattern of
9545 // fence-atomic-fence.
9546 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9547 SDValue atomic = N->getOperand(0);
9548 switch (atomic.getOpcode()) {
9549 case ISD::ATOMIC_CMP_SWAP:
9550 case ISD::ATOMIC_SWAP:
9551 case ISD::ATOMIC_LOAD_ADD:
9552 case ISD::ATOMIC_LOAD_SUB:
9553 case ISD::ATOMIC_LOAD_AND:
9554 case ISD::ATOMIC_LOAD_OR:
9555 case ISD::ATOMIC_LOAD_XOR:
9556 case ISD::ATOMIC_LOAD_NAND:
9557 case ISD::ATOMIC_LOAD_MIN:
9558 case ISD::ATOMIC_LOAD_MAX:
9559 case ISD::ATOMIC_LOAD_UMIN:
9560 case ISD::ATOMIC_LOAD_UMAX:
9566 SDValue fence = atomic.getOperand(0);
9567 if (fence.getOpcode() != ISD::MEMBARRIER)
9570 switch (atomic.getOpcode()) {
9571 case ISD::ATOMIC_CMP_SWAP:
9572 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9573 atomic.getOperand(1), atomic.getOperand(2),
9574 atomic.getOperand(3));
9575 case ISD::ATOMIC_SWAP:
9576 case ISD::ATOMIC_LOAD_ADD:
9577 case ISD::ATOMIC_LOAD_SUB:
9578 case ISD::ATOMIC_LOAD_AND:
9579 case ISD::ATOMIC_LOAD_OR:
9580 case ISD::ATOMIC_LOAD_XOR:
9581 case ISD::ATOMIC_LOAD_NAND:
9582 case ISD::ATOMIC_LOAD_MIN:
9583 case ISD::ATOMIC_LOAD_MAX:
9584 case ISD::ATOMIC_LOAD_UMIN:
9585 case ISD::ATOMIC_LOAD_UMAX:
9586 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9587 atomic.getOperand(1), atomic.getOperand(2));
9593 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9594 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9595 // (and (i32 x86isd::setcc_carry), 1)
9596 // This eliminates the zext. This transformation is necessary because
9597 // ISD::SETCC is always legalized to i8.
9598 DebugLoc dl = N->getDebugLoc();
9599 SDValue N0 = N->getOperand(0);
9600 EVT VT = N->getValueType(0);
9601 if (N0.getOpcode() == ISD::AND &&
9603 N0.getOperand(0).hasOneUse()) {
9604 SDValue N00 = N0.getOperand(0);
9605 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9607 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9608 if (!C || C->getZExtValue() != 1)
9610 return DAG.getNode(ISD::AND, dl, VT,
9611 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9612 N00.getOperand(0), N00.getOperand(1)),
9613 DAG.getConstant(1, VT));
9619 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9620 DAGCombinerInfo &DCI) const {
9621 SelectionDAG &DAG = DCI.DAG;
9622 switch (N->getOpcode()) {
9624 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9625 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9626 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9627 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9630 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9631 case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
9632 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9634 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9635 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9636 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9637 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9638 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9639 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9645 //===----------------------------------------------------------------------===//
9646 // X86 Inline Assembly Support
9647 //===----------------------------------------------------------------------===//
9649 static bool LowerToBSwap(CallInst *CI) {
9650 // FIXME: this should verify that we are targetting a 486 or better. If not,
9651 // we will turn this bswap into something that will be lowered to logical ops
9652 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9653 // so don't worry about this.
9655 // Verify this is a simple bswap.
9656 if (CI->getNumOperands() != 2 ||
9657 CI->getType() != CI->getOperand(1)->getType() ||
9658 !CI->getType()->isInteger())
9661 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9662 if (!Ty || Ty->getBitWidth() % 16 != 0)
9665 // Okay, we can do this xform, do so now.
9666 const Type *Tys[] = { Ty };
9667 Module *M = CI->getParent()->getParent()->getParent();
9668 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9670 Value *Op = CI->getOperand(1);
9671 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9673 CI->replaceAllUsesWith(Op);
9674 CI->eraseFromParent();
9678 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9679 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9680 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9682 std::string AsmStr = IA->getAsmString();
9684 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9685 SmallVector<StringRef, 4> AsmPieces;
9686 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9688 switch (AsmPieces.size()) {
9689 default: return false;
9691 AsmStr = AsmPieces[0];
9693 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9696 if (AsmPieces.size() == 2 &&
9697 (AsmPieces[0] == "bswap" ||
9698 AsmPieces[0] == "bswapq" ||
9699 AsmPieces[0] == "bswapl") &&
9700 (AsmPieces[1] == "$0" ||
9701 AsmPieces[1] == "${0:q}")) {
9702 // No need to check constraints, nothing other than the equivalent of
9703 // "=r,0" would be valid here.
9704 return LowerToBSwap(CI);
9706 // rorw $$8, ${0:w} --> llvm.bswap.i16
9707 if (CI->getType()->isInteger(16) &&
9708 AsmPieces.size() == 3 &&
9709 AsmPieces[0] == "rorw" &&
9710 AsmPieces[1] == "$$8," &&
9711 AsmPieces[2] == "${0:w}" &&
9712 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
9713 return LowerToBSwap(CI);
9717 if (CI->getType()->isInteger(64) &&
9718 Constraints.size() >= 2 &&
9719 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9720 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9721 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9722 SmallVector<StringRef, 4> Words;
9723 SplitString(AsmPieces[0], Words, " \t");
9724 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9726 SplitString(AsmPieces[1], Words, " \t");
9727 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9729 SplitString(AsmPieces[2], Words, " \t,");
9730 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9731 Words[2] == "%edx") {
9732 return LowerToBSwap(CI);
9744 /// getConstraintType - Given a constraint letter, return the type of
9745 /// constraint it is for this target.
9746 X86TargetLowering::ConstraintType
9747 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
9748 if (Constraint.size() == 1) {
9749 switch (Constraint[0]) {
9761 return C_RegisterClass;
9769 return TargetLowering::getConstraintType(Constraint);
9772 /// LowerXConstraint - try to replace an X constraint, which matches anything,
9773 /// with another that has more specific requirements based on the type of the
9774 /// corresponding operand.
9775 const char *X86TargetLowering::
9776 LowerXConstraint(EVT ConstraintVT) const {
9777 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
9778 // 'f' like normal targets.
9779 if (ConstraintVT.isFloatingPoint()) {
9780 if (Subtarget->hasSSE2())
9782 if (Subtarget->hasSSE1())
9786 return TargetLowering::LowerXConstraint(ConstraintVT);
9789 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9790 /// vector. If it is invalid, don't add anything to Ops.
9791 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9794 std::vector<SDValue>&Ops,
9795 SelectionDAG &DAG) const {
9796 SDValue Result(0, 0);
9798 switch (Constraint) {
9801 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9802 if (C->getZExtValue() <= 31) {
9803 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9809 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9810 if (C->getZExtValue() <= 63) {
9811 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9817 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9818 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9819 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9825 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9826 if (C->getZExtValue() <= 255) {
9827 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9833 // 32-bit signed value
9834 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9835 const ConstantInt *CI = C->getConstantIntValue();
9836 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9837 C->getSExtValue())) {
9838 // Widen to 64 bits here to get it sign extended.
9839 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9842 // FIXME gcc accepts some relocatable values here too, but only in certain
9843 // memory models; it's complicated.
9848 // 32-bit unsigned value
9849 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9850 const ConstantInt *CI = C->getConstantIntValue();
9851 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9852 C->getZExtValue())) {
9853 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9857 // FIXME gcc accepts some relocatable values here too, but only in certain
9858 // memory models; it's complicated.
9862 // Literal immediates are always ok.
9863 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
9864 // Widen to 64 bits here to get it sign extended.
9865 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
9869 // If we are in non-pic codegen mode, we allow the address of a global (with
9870 // an optional displacement) to be used with 'i'.
9871 GlobalAddressSDNode *GA = 0;
9874 // Match either (GA), (GA+C), (GA+C1+C2), etc.
9876 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
9877 Offset += GA->getOffset();
9879 } else if (Op.getOpcode() == ISD::ADD) {
9880 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9881 Offset += C->getZExtValue();
9882 Op = Op.getOperand(0);
9885 } else if (Op.getOpcode() == ISD::SUB) {
9886 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9887 Offset += -C->getZExtValue();
9888 Op = Op.getOperand(0);
9893 // Otherwise, this isn't something we can handle, reject it.
9897 GlobalValue *GV = GA->getGlobal();
9898 // If we require an extra load to get this address, as in PIC mode, we
9900 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
9901 getTargetMachine())))
9905 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
9907 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
9913 if (Result.getNode()) {
9914 Ops.push_back(Result);
9917 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
9921 std::vector<unsigned> X86TargetLowering::
9922 getRegClassForInlineAsmConstraint(const std::string &Constraint,
9924 if (Constraint.size() == 1) {
9925 // FIXME: not handling fp-stack yet!
9926 switch (Constraint[0]) { // GCC X86 Constraint Letters
9927 default: break; // Unknown constraint letter
9928 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
9929 if (Subtarget->is64Bit()) {
9931 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
9932 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
9933 X86::R10D,X86::R11D,X86::R12D,
9934 X86::R13D,X86::R14D,X86::R15D,
9935 X86::EBP, X86::ESP, 0);
9936 else if (VT == MVT::i16)
9937 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
9938 X86::SI, X86::DI, X86::R8W,X86::R9W,
9939 X86::R10W,X86::R11W,X86::R12W,
9940 X86::R13W,X86::R14W,X86::R15W,
9941 X86::BP, X86::SP, 0);
9942 else if (VT == MVT::i8)
9943 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
9944 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
9945 X86::R10B,X86::R11B,X86::R12B,
9946 X86::R13B,X86::R14B,X86::R15B,
9947 X86::BPL, X86::SPL, 0);
9949 else if (VT == MVT::i64)
9950 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
9951 X86::RSI, X86::RDI, X86::R8, X86::R9,
9952 X86::R10, X86::R11, X86::R12,
9953 X86::R13, X86::R14, X86::R15,
9954 X86::RBP, X86::RSP, 0);
9958 // 32-bit fallthrough
9961 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
9962 else if (VT == MVT::i16)
9963 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
9964 else if (VT == MVT::i8)
9965 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
9966 else if (VT == MVT::i64)
9967 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
9972 return std::vector<unsigned>();
9975 std::pair<unsigned, const TargetRegisterClass*>
9976 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9978 // First, see if this is a constraint that directly corresponds to an LLVM
9980 if (Constraint.size() == 1) {
9981 // GCC Constraint Letters
9982 switch (Constraint[0]) {
9984 case 'r': // GENERAL_REGS
9985 case 'l': // INDEX_REGS
9987 return std::make_pair(0U, X86::GR8RegisterClass);
9989 return std::make_pair(0U, X86::GR16RegisterClass);
9990 if (VT == MVT::i32 || !Subtarget->is64Bit())
9991 return std::make_pair(0U, X86::GR32RegisterClass);
9992 return std::make_pair(0U, X86::GR64RegisterClass);
9993 case 'R': // LEGACY_REGS
9995 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
9997 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
9998 if (VT == MVT::i32 || !Subtarget->is64Bit())
9999 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10000 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10001 case 'f': // FP Stack registers.
10002 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10003 // value to the correct fpstack register class.
10004 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10005 return std::make_pair(0U, X86::RFP32RegisterClass);
10006 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10007 return std::make_pair(0U, X86::RFP64RegisterClass);
10008 return std::make_pair(0U, X86::RFP80RegisterClass);
10009 case 'y': // MMX_REGS if MMX allowed.
10010 if (!Subtarget->hasMMX()) break;
10011 return std::make_pair(0U, X86::VR64RegisterClass);
10012 case 'Y': // SSE_REGS if SSE2 allowed
10013 if (!Subtarget->hasSSE2()) break;
10015 case 'x': // SSE_REGS if SSE1 allowed
10016 if (!Subtarget->hasSSE1()) break;
10018 switch (VT.getSimpleVT().SimpleTy) {
10020 // Scalar SSE types.
10023 return std::make_pair(0U, X86::FR32RegisterClass);
10026 return std::make_pair(0U, X86::FR64RegisterClass);
10034 return std::make_pair(0U, X86::VR128RegisterClass);
10040 // Use the default implementation in TargetLowering to convert the register
10041 // constraint into a member of a register class.
10042 std::pair<unsigned, const TargetRegisterClass*> Res;
10043 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10045 // Not found as a standard register?
10046 if (Res.second == 0) {
10047 // Map st(0) -> st(7) -> ST0
10048 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10049 tolower(Constraint[1]) == 's' &&
10050 tolower(Constraint[2]) == 't' &&
10051 Constraint[3] == '(' &&
10052 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10053 Constraint[5] == ')' &&
10054 Constraint[6] == '}') {
10056 Res.first = X86::ST0+Constraint[4]-'0';
10057 Res.second = X86::RFP80RegisterClass;
10061 // GCC allows "st(0)" to be called just plain "st".
10062 if (StringRef("{st}").equals_lower(Constraint)) {
10063 Res.first = X86::ST0;
10064 Res.second = X86::RFP80RegisterClass;
10069 if (StringRef("{flags}").equals_lower(Constraint)) {
10070 Res.first = X86::EFLAGS;
10071 Res.second = X86::CCRRegisterClass;
10075 // 'A' means EAX + EDX.
10076 if (Constraint == "A") {
10077 Res.first = X86::EAX;
10078 Res.second = X86::GR32_ADRegisterClass;
10084 // Otherwise, check to see if this is a register class of the wrong value
10085 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10086 // turn into {ax},{dx}.
10087 if (Res.second->hasType(VT))
10088 return Res; // Correct type already, nothing to do.
10090 // All of the single-register GCC register classes map their values onto
10091 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10092 // really want an 8-bit or 32-bit register, map to the appropriate register
10093 // class and return the appropriate register.
10094 if (Res.second == X86::GR16RegisterClass) {
10095 if (VT == MVT::i8) {
10096 unsigned DestReg = 0;
10097 switch (Res.first) {
10099 case X86::AX: DestReg = X86::AL; break;
10100 case X86::DX: DestReg = X86::DL; break;
10101 case X86::CX: DestReg = X86::CL; break;
10102 case X86::BX: DestReg = X86::BL; break;
10105 Res.first = DestReg;
10106 Res.second = X86::GR8RegisterClass;
10108 } else if (VT == MVT::i32) {
10109 unsigned DestReg = 0;
10110 switch (Res.first) {
10112 case X86::AX: DestReg = X86::EAX; break;
10113 case X86::DX: DestReg = X86::EDX; break;
10114 case X86::CX: DestReg = X86::ECX; break;
10115 case X86::BX: DestReg = X86::EBX; break;
10116 case X86::SI: DestReg = X86::ESI; break;
10117 case X86::DI: DestReg = X86::EDI; break;
10118 case X86::BP: DestReg = X86::EBP; break;
10119 case X86::SP: DestReg = X86::ESP; break;
10122 Res.first = DestReg;
10123 Res.second = X86::GR32RegisterClass;
10125 } else if (VT == MVT::i64) {
10126 unsigned DestReg = 0;
10127 switch (Res.first) {
10129 case X86::AX: DestReg = X86::RAX; break;
10130 case X86::DX: DestReg = X86::RDX; break;
10131 case X86::CX: DestReg = X86::RCX; break;
10132 case X86::BX: DestReg = X86::RBX; break;
10133 case X86::SI: DestReg = X86::RSI; break;
10134 case X86::DI: DestReg = X86::RDI; break;
10135 case X86::BP: DestReg = X86::RBP; break;
10136 case X86::SP: DestReg = X86::RSP; break;
10139 Res.first = DestReg;
10140 Res.second = X86::GR64RegisterClass;
10143 } else if (Res.second == X86::FR32RegisterClass ||
10144 Res.second == X86::FR64RegisterClass ||
10145 Res.second == X86::VR128RegisterClass) {
10146 // Handle references to XMM physical registers that got mapped into the
10147 // wrong class. This can happen with constraints like {xmm0} where the
10148 // target independent register mapper will just pick the first match it can
10149 // find, ignoring the required type.
10150 if (VT == MVT::f32)
10151 Res.second = X86::FR32RegisterClass;
10152 else if (VT == MVT::f64)
10153 Res.second = X86::FR64RegisterClass;
10154 else if (X86::VR128RegisterClass->hasType(VT))
10155 Res.second = X86::VR128RegisterClass;
10161 //===----------------------------------------------------------------------===//
10162 // X86 Widen vector type
10163 //===----------------------------------------------------------------------===//
10165 /// getWidenVectorType: given a vector type, returns the type to widen
10166 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
10167 /// If there is no vector type that we want to widen to, returns MVT::Other
10168 /// When and where to widen is target dependent based on the cost of
10169 /// scalarizing vs using the wider vector type.
10171 EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
10172 assert(VT.isVector());
10173 if (isTypeLegal(VT))
10176 // TODO: In computeRegisterProperty, we can compute the list of legal vector
10177 // type based on element type. This would speed up our search (though
10178 // it may not be worth it since the size of the list is relatively
10180 EVT EltVT = VT.getVectorElementType();
10181 unsigned NElts = VT.getVectorNumElements();
10183 // On X86, it make sense to widen any vector wider than 1
10187 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
10188 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
10189 EVT SVT = (MVT::SimpleValueType)nVT;
10191 if (isTypeLegal(SVT) &&
10192 SVT.getVectorElementType() == EltVT &&
10193 SVT.getVectorNumElements() > NElts)