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 if (Flags.isByVal()) {
1475 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1476 VA.getLocMemOffset(), isImmutable, false);
1477 return DAG.getFrameIndex(FI, getPointerTy());
1479 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1480 VA.getLocMemOffset(), isImmutable, false);
1481 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1482 return DAG.getLoad(ValVT, dl, Chain, FIN,
1483 PseudoSourceValue::getFixedStack(FI), 0);
1488 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1489 CallingConv::ID CallConv,
1491 const SmallVectorImpl<ISD::InputArg> &Ins,
1494 SmallVectorImpl<SDValue> &InVals) {
1496 MachineFunction &MF = DAG.getMachineFunction();
1497 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1499 const Function* Fn = MF.getFunction();
1500 if (Fn->hasExternalLinkage() &&
1501 Subtarget->isTargetCygMing() &&
1502 Fn->getName() == "main")
1503 FuncInfo->setForceFramePointer(true);
1505 // Decorate the function name.
1506 FuncInfo->setDecorationStyle(NameDecorationForCallConv(CallConv));
1508 MachineFrameInfo *MFI = MF.getFrameInfo();
1509 bool Is64Bit = Subtarget->is64Bit();
1510 bool IsWin64 = Subtarget->isTargetWin64();
1512 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1513 "Var args not supported with calling convention fastcc");
1515 // Assign locations to all of the incoming arguments.
1516 SmallVector<CCValAssign, 16> ArgLocs;
1517 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1518 ArgLocs, *DAG.getContext());
1519 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1521 unsigned LastVal = ~0U;
1523 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1524 CCValAssign &VA = ArgLocs[i];
1525 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1527 assert(VA.getValNo() != LastVal &&
1528 "Don't support value assigned to multiple locs yet");
1529 LastVal = VA.getValNo();
1531 if (VA.isRegLoc()) {
1532 EVT RegVT = VA.getLocVT();
1533 TargetRegisterClass *RC = NULL;
1534 if (RegVT == MVT::i32)
1535 RC = X86::GR32RegisterClass;
1536 else if (Is64Bit && RegVT == MVT::i64)
1537 RC = X86::GR64RegisterClass;
1538 else if (RegVT == MVT::f32)
1539 RC = X86::FR32RegisterClass;
1540 else if (RegVT == MVT::f64)
1541 RC = X86::FR64RegisterClass;
1542 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1543 RC = X86::VR128RegisterClass;
1544 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1545 RC = X86::VR64RegisterClass;
1547 llvm_unreachable("Unknown argument type!");
1549 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1550 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1552 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1553 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1555 if (VA.getLocInfo() == CCValAssign::SExt)
1556 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1557 DAG.getValueType(VA.getValVT()));
1558 else if (VA.getLocInfo() == CCValAssign::ZExt)
1559 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1560 DAG.getValueType(VA.getValVT()));
1561 else if (VA.getLocInfo() == CCValAssign::BCvt)
1562 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1564 if (VA.isExtInLoc()) {
1565 // Handle MMX values passed in XMM regs.
1566 if (RegVT.isVector()) {
1567 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1568 ArgValue, DAG.getConstant(0, MVT::i64));
1569 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1571 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1574 assert(VA.isMemLoc());
1575 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1578 // If value is passed via pointer - do a load.
1579 if (VA.getLocInfo() == CCValAssign::Indirect)
1580 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0);
1582 InVals.push_back(ArgValue);
1585 // The x86-64 ABI for returning structs by value requires that we copy
1586 // the sret argument into %rax for the return. Save the argument into
1587 // a virtual register so that we can access it from the return points.
1588 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1589 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1590 unsigned Reg = FuncInfo->getSRetReturnReg();
1592 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1593 FuncInfo->setSRetReturnReg(Reg);
1595 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1596 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1599 unsigned StackSize = CCInfo.getNextStackOffset();
1600 // Align stack specially for tail calls.
1601 if (FuncIsMadeTailCallSafe(CallConv))
1602 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1604 // If the function takes variable number of arguments, make a frame index for
1605 // the start of the first vararg value... for expansion of llvm.va_start.
1607 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1608 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize, true, false);
1611 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1613 // FIXME: We should really autogenerate these arrays
1614 static const unsigned GPR64ArgRegsWin64[] = {
1615 X86::RCX, X86::RDX, X86::R8, X86::R9
1617 static const unsigned XMMArgRegsWin64[] = {
1618 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1620 static const unsigned GPR64ArgRegs64Bit[] = {
1621 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1623 static const unsigned XMMArgRegs64Bit[] = {
1624 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1625 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1627 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1630 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1631 GPR64ArgRegs = GPR64ArgRegsWin64;
1632 XMMArgRegs = XMMArgRegsWin64;
1634 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1635 GPR64ArgRegs = GPR64ArgRegs64Bit;
1636 XMMArgRegs = XMMArgRegs64Bit;
1638 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1640 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1643 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1644 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1645 "SSE register cannot be used when SSE is disabled!");
1646 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1647 "SSE register cannot be used when SSE is disabled!");
1648 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1649 // Kernel mode asks for SSE to be disabled, so don't push them
1651 TotalNumXMMRegs = 0;
1653 // For X86-64, if there are vararg parameters that are passed via
1654 // registers, then we must store them to their spots on the stack so they
1655 // may be loaded by deferencing the result of va_next.
1656 VarArgsGPOffset = NumIntRegs * 8;
1657 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1658 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1659 TotalNumXMMRegs * 16, 16,
1662 // Store the integer parameter registers.
1663 SmallVector<SDValue, 8> MemOps;
1664 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1665 unsigned Offset = VarArgsGPOffset;
1666 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1667 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1668 DAG.getIntPtrConstant(Offset));
1669 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1670 X86::GR64RegisterClass);
1671 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1673 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1674 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1676 MemOps.push_back(Store);
1680 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1681 // Now store the XMM (fp + vector) parameter registers.
1682 SmallVector<SDValue, 11> SaveXMMOps;
1683 SaveXMMOps.push_back(Chain);
1685 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1686 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1687 SaveXMMOps.push_back(ALVal);
1689 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1690 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1692 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1693 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1694 X86::VR128RegisterClass);
1695 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1696 SaveXMMOps.push_back(Val);
1698 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1700 &SaveXMMOps[0], SaveXMMOps.size()));
1703 if (!MemOps.empty())
1704 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1705 &MemOps[0], MemOps.size());
1709 // Some CCs need callee pop.
1710 if (IsCalleePop(isVarArg, CallConv)) {
1711 BytesToPopOnReturn = StackSize; // Callee pops everything.
1713 BytesToPopOnReturn = 0; // Callee pops nothing.
1714 // If this is an sret function, the return should pop the hidden pointer.
1715 if (!Is64Bit && CallConv != CallingConv::Fast && ArgsAreStructReturn(Ins))
1716 BytesToPopOnReturn = 4;
1720 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1721 if (CallConv == CallingConv::X86_FastCall)
1722 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1725 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1731 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1732 SDValue StackPtr, SDValue Arg,
1733 DebugLoc dl, SelectionDAG &DAG,
1734 const CCValAssign &VA,
1735 ISD::ArgFlagsTy Flags) {
1736 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1737 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1738 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1739 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1740 if (Flags.isByVal()) {
1741 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1743 return DAG.getStore(Chain, dl, Arg, PtrOff,
1744 PseudoSourceValue::getStack(), LocMemOffset);
1747 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1748 /// optimization is performed and it is required.
1750 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1751 SDValue &OutRetAddr, SDValue Chain,
1752 bool IsTailCall, bool Is64Bit,
1753 int FPDiff, DebugLoc dl) {
1754 if (!IsTailCall || FPDiff==0) return Chain;
1756 // Adjust the Return address stack slot.
1757 EVT VT = getPointerTy();
1758 OutRetAddr = getReturnAddressFrameIndex(DAG);
1760 // Load the "old" Return address.
1761 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1762 return SDValue(OutRetAddr.getNode(), 1);
1765 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1766 /// optimization is performed and it is required (FPDiff!=0).
1768 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1769 SDValue Chain, SDValue RetAddrFrIdx,
1770 bool Is64Bit, int FPDiff, DebugLoc dl) {
1771 // Store the return address to the appropriate stack slot.
1772 if (!FPDiff) return Chain;
1773 // Calculate the new stack slot for the return address.
1774 int SlotSize = Is64Bit ? 8 : 4;
1775 int NewReturnAddrFI =
1776 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, true,false);
1777 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1778 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1779 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1780 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1785 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1786 CallingConv::ID CallConv, bool isVarArg,
1788 const SmallVectorImpl<ISD::OutputArg> &Outs,
1789 const SmallVectorImpl<ISD::InputArg> &Ins,
1790 DebugLoc dl, SelectionDAG &DAG,
1791 SmallVectorImpl<SDValue> &InVals) {
1792 MachineFunction &MF = DAG.getMachineFunction();
1793 bool Is64Bit = Subtarget->is64Bit();
1794 bool IsStructRet = CallIsStructReturn(Outs);
1797 // Check if it's really possible to do a tail call.
1798 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
1801 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1802 "Var args not supported with calling convention fastcc");
1804 // Analyze operands of the call, assigning locations to each operand.
1805 SmallVector<CCValAssign, 16> ArgLocs;
1806 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1807 ArgLocs, *DAG.getContext());
1808 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1810 // Get a count of how many bytes are to be pushed on the stack.
1811 unsigned NumBytes = CCInfo.getNextStackOffset();
1812 if (FuncIsMadeTailCallSafe(CallConv))
1813 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1814 else if (isTailCall && !PerformTailCallOpt)
1815 // This is a sibcall. The memory operands are available in caller's
1816 // own caller's stack.
1823 // Lower arguments at fp - stackoffset + fpdiff.
1824 unsigned NumBytesCallerPushed =
1825 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1826 FPDiff = NumBytesCallerPushed - NumBytes;
1828 // Set the delta of movement of the returnaddr stackslot.
1829 // But only set if delta is greater than previous delta.
1830 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1831 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1834 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1836 SDValue RetAddrFrIdx;
1837 // Load return adress for tail calls.
1838 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall, Is64Bit,
1841 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1842 SmallVector<SDValue, 8> MemOpChains;
1845 // Walk the register/memloc assignments, inserting copies/loads. In the case
1846 // of tail call optimization arguments are handle later.
1847 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1848 CCValAssign &VA = ArgLocs[i];
1849 EVT RegVT = VA.getLocVT();
1850 SDValue Arg = Outs[i].Val;
1851 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1852 bool isByVal = Flags.isByVal();
1854 // Promote the value if needed.
1855 switch (VA.getLocInfo()) {
1856 default: llvm_unreachable("Unknown loc info!");
1857 case CCValAssign::Full: break;
1858 case CCValAssign::SExt:
1859 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1861 case CCValAssign::ZExt:
1862 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1864 case CCValAssign::AExt:
1865 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1866 // Special case: passing MMX values in XMM registers.
1867 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1868 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1869 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1871 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1873 case CCValAssign::BCvt:
1874 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1876 case CCValAssign::Indirect: {
1877 // Store the argument.
1878 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1879 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1880 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1881 PseudoSourceValue::getFixedStack(FI), 0);
1887 if (VA.isRegLoc()) {
1888 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1890 if (!isTailCall || (isTailCall && isByVal)) {
1891 assert(VA.isMemLoc());
1892 if (StackPtr.getNode() == 0)
1893 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1895 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1896 dl, DAG, VA, Flags));
1901 if (!MemOpChains.empty())
1902 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1903 &MemOpChains[0], MemOpChains.size());
1905 // Build a sequence of copy-to-reg nodes chained together with token chain
1906 // and flag operands which copy the outgoing args into registers.
1908 // Tail call byval lowering might overwrite argument registers so in case of
1909 // tail call optimization the copies to registers are lowered later.
1911 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1912 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1913 RegsToPass[i].second, InFlag);
1914 InFlag = Chain.getValue(1);
1918 if (Subtarget->isPICStyleGOT()) {
1919 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1922 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1923 DAG.getNode(X86ISD::GlobalBaseReg,
1924 DebugLoc::getUnknownLoc(),
1927 InFlag = Chain.getValue(1);
1929 // If we are tail calling and generating PIC/GOT style code load the
1930 // address of the callee into ECX. The value in ecx is used as target of
1931 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1932 // for tail calls on PIC/GOT architectures. Normally we would just put the
1933 // address of GOT into ebx and then call target@PLT. But for tail calls
1934 // ebx would be restored (since ebx is callee saved) before jumping to the
1937 // Note: The actual moving to ECX is done further down.
1938 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1939 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1940 !G->getGlobal()->hasProtectedVisibility())
1941 Callee = LowerGlobalAddress(Callee, DAG);
1942 else if (isa<ExternalSymbolSDNode>(Callee))
1943 Callee = LowerExternalSymbol(Callee, DAG);
1947 if (Is64Bit && isVarArg) {
1948 // From AMD64 ABI document:
1949 // For calls that may call functions that use varargs or stdargs
1950 // (prototype-less calls or calls to functions containing ellipsis (...) in
1951 // the declaration) %al is used as hidden argument to specify the number
1952 // of SSE registers used. The contents of %al do not need to match exactly
1953 // the number of registers, but must be an ubound on the number of SSE
1954 // registers used and is in the range 0 - 8 inclusive.
1956 // FIXME: Verify this on Win64
1957 // Count the number of XMM registers allocated.
1958 static const unsigned XMMArgRegs[] = {
1959 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1960 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1962 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1963 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1964 && "SSE registers cannot be used when SSE is disabled");
1966 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1967 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1968 InFlag = Chain.getValue(1);
1972 // For tail calls lower the arguments to the 'real' stack slot.
1974 // Force all the incoming stack arguments to be loaded from the stack
1975 // before any new outgoing arguments are stored to the stack, because the
1976 // outgoing stack slots may alias the incoming argument stack slots, and
1977 // the alias isn't otherwise explicit. This is slightly more conservative
1978 // than necessary, because it means that each store effectively depends
1979 // on every argument instead of just those arguments it would clobber.
1980 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1982 SmallVector<SDValue, 8> MemOpChains2;
1985 // Do not flag preceeding copytoreg stuff together with the following stuff.
1987 if (PerformTailCallOpt) {
1988 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1989 CCValAssign &VA = ArgLocs[i];
1992 assert(VA.isMemLoc());
1993 SDValue Arg = Outs[i].Val;
1994 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1995 // Create frame index.
1996 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1997 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1998 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
1999 FIN = DAG.getFrameIndex(FI, getPointerTy());
2001 if (Flags.isByVal()) {
2002 // Copy relative to framepointer.
2003 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2004 if (StackPtr.getNode() == 0)
2005 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2007 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2009 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2013 // Store relative to framepointer.
2014 MemOpChains2.push_back(
2015 DAG.getStore(ArgChain, dl, Arg, FIN,
2016 PseudoSourceValue::getFixedStack(FI), 0));
2021 if (!MemOpChains2.empty())
2022 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2023 &MemOpChains2[0], MemOpChains2.size());
2025 // Copy arguments to their registers.
2026 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2027 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2028 RegsToPass[i].second, InFlag);
2029 InFlag = Chain.getValue(1);
2033 // Store the return address to the appropriate stack slot.
2034 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2038 bool WasGlobalOrExternal = false;
2039 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2040 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2041 // In the 64-bit large code model, we have to make all calls
2042 // through a register, since the call instruction's 32-bit
2043 // pc-relative offset may not be large enough to hold the whole
2045 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2046 WasGlobalOrExternal = true;
2047 // If the callee is a GlobalAddress node (quite common, every direct call
2048 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2051 // We should use extra load for direct calls to dllimported functions in
2053 GlobalValue *GV = G->getGlobal();
2054 if (!GV->hasDLLImportLinkage()) {
2055 unsigned char OpFlags = 0;
2057 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2058 // external symbols most go through the PLT in PIC mode. If the symbol
2059 // has hidden or protected visibility, or if it is static or local, then
2060 // we don't need to use the PLT - we can directly call it.
2061 if (Subtarget->isTargetELF() &&
2062 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2063 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2064 OpFlags = X86II::MO_PLT;
2065 } else if (Subtarget->isPICStyleStubAny() &&
2066 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2067 Subtarget->getDarwinVers() < 9) {
2068 // PC-relative references to external symbols should go through $stub,
2069 // unless we're building with the leopard linker or later, which
2070 // automatically synthesizes these stubs.
2071 OpFlags = X86II::MO_DARWIN_STUB;
2074 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2075 G->getOffset(), OpFlags);
2077 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2078 WasGlobalOrExternal = true;
2079 unsigned char OpFlags = 0;
2081 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2082 // symbols should go through the PLT.
2083 if (Subtarget->isTargetELF() &&
2084 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2085 OpFlags = X86II::MO_PLT;
2086 } else if (Subtarget->isPICStyleStubAny() &&
2087 Subtarget->getDarwinVers() < 9) {
2088 // PC-relative references to external symbols should go through $stub,
2089 // unless we're building with the leopard linker or later, which
2090 // automatically synthesizes these stubs.
2091 OpFlags = X86II::MO_DARWIN_STUB;
2094 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2098 if (isTailCall && !WasGlobalOrExternal) {
2099 unsigned Opc = Is64Bit ? X86::R11 : X86::EAX;
2101 Chain = DAG.getCopyToReg(Chain, dl,
2102 DAG.getRegister(Opc, getPointerTy()),
2104 Callee = DAG.getRegister(Opc, getPointerTy());
2105 // Add register as live out.
2106 MF.getRegInfo().addLiveOut(Opc);
2109 // Returns a chain & a flag for retval copy to use.
2110 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2111 SmallVector<SDValue, 8> Ops;
2114 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2115 DAG.getIntPtrConstant(0, true), InFlag);
2116 InFlag = Chain.getValue(1);
2119 Ops.push_back(Chain);
2120 Ops.push_back(Callee);
2123 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2125 // Add argument registers to the end of the list so that they are known live
2127 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2128 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2129 RegsToPass[i].second.getValueType()));
2131 // Add an implicit use GOT pointer in EBX.
2132 if (!isTailCall && Subtarget->isPICStyleGOT())
2133 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2135 // Add an implicit use of AL for x86 vararg functions.
2136 if (Is64Bit && isVarArg)
2137 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2139 if (InFlag.getNode())
2140 Ops.push_back(InFlag);
2143 // If this is the first return lowered for this function, add the regs
2144 // to the liveout set for the function.
2145 if (MF.getRegInfo().liveout_empty()) {
2146 SmallVector<CCValAssign, 16> RVLocs;
2147 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2149 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2150 for (unsigned i = 0; i != RVLocs.size(); ++i)
2151 if (RVLocs[i].isRegLoc())
2152 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2155 assert(((Callee.getOpcode() == ISD::Register &&
2156 (cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
2157 cast<RegisterSDNode>(Callee)->getReg() == X86::R11)) ||
2158 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2159 Callee.getOpcode() == ISD::TargetGlobalAddress) &&
2160 "Expecting a global address, external symbol, or scratch register");
2162 return DAG.getNode(X86ISD::TC_RETURN, dl,
2163 NodeTys, &Ops[0], Ops.size());
2166 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2167 InFlag = Chain.getValue(1);
2169 // Create the CALLSEQ_END node.
2170 unsigned NumBytesForCalleeToPush;
2171 if (IsCalleePop(isVarArg, CallConv))
2172 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2173 else if (!Is64Bit && CallConv != CallingConv::Fast && IsStructRet)
2174 // If this is is a call to a struct-return function, the callee
2175 // pops the hidden struct pointer, so we have to push it back.
2176 // This is common for Darwin/X86, Linux & Mingw32 targets.
2177 NumBytesForCalleeToPush = 4;
2179 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2181 // Returns a flag for retval copy to use.
2182 Chain = DAG.getCALLSEQ_END(Chain,
2183 DAG.getIntPtrConstant(NumBytes, true),
2184 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2187 InFlag = Chain.getValue(1);
2189 // Handle result values, copying them out of physregs into vregs that we
2191 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2192 Ins, dl, DAG, InVals);
2196 //===----------------------------------------------------------------------===//
2197 // Fast Calling Convention (tail call) implementation
2198 //===----------------------------------------------------------------------===//
2200 // Like std call, callee cleans arguments, convention except that ECX is
2201 // reserved for storing the tail called function address. Only 2 registers are
2202 // free for argument passing (inreg). Tail call optimization is performed
2204 // * tailcallopt is enabled
2205 // * caller/callee are fastcc
2206 // On X86_64 architecture with GOT-style position independent code only local
2207 // (within module) calls are supported at the moment.
2208 // To keep the stack aligned according to platform abi the function
2209 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2210 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2211 // If a tail called function callee has more arguments than the caller the
2212 // caller needs to make sure that there is room to move the RETADDR to. This is
2213 // achieved by reserving an area the size of the argument delta right after the
2214 // original REtADDR, but before the saved framepointer or the spilled registers
2215 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2227 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2228 /// for a 16 byte align requirement.
2229 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2230 SelectionDAG& DAG) {
2231 MachineFunction &MF = DAG.getMachineFunction();
2232 const TargetMachine &TM = MF.getTarget();
2233 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2234 unsigned StackAlignment = TFI.getStackAlignment();
2235 uint64_t AlignMask = StackAlignment - 1;
2236 int64_t Offset = StackSize;
2237 uint64_t SlotSize = TD->getPointerSize();
2238 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2239 // Number smaller than 12 so just add the difference.
2240 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2242 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2243 Offset = ((~AlignMask) & Offset) + StackAlignment +
2244 (StackAlignment-SlotSize);
2249 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2250 /// for tail call optimization. Targets which want to do tail call
2251 /// optimization should implement this function.
2253 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2254 CallingConv::ID CalleeCC,
2256 const SmallVectorImpl<ISD::OutputArg> &Outs,
2257 const SmallVectorImpl<ISD::InputArg> &Ins,
2258 SelectionDAG& DAG) const {
2259 if (CalleeCC != CallingConv::Fast &&
2260 CalleeCC != CallingConv::C)
2263 // If -tailcallopt is specified, make fastcc functions tail-callable.
2264 const Function *CallerF = DAG.getMachineFunction().getFunction();
2265 if (PerformTailCallOpt) {
2266 if (CalleeCC == CallingConv::Fast &&
2267 CallerF->getCallingConv() == CalleeCC)
2272 // Look for obvious safe cases to perform tail call optimization that does not
2273 // requite ABI changes. This is what gcc calls sibcall.
2275 // Do not tail call optimize vararg calls for now.
2279 // If the callee takes no arguments then go on to check the results of the
2281 if (!Outs.empty()) {
2282 // Check if stack adjustment is needed. For now, do not do this if any
2283 // argument is passed on the stack.
2284 SmallVector<CCValAssign, 16> ArgLocs;
2285 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2286 ArgLocs, *DAG.getContext());
2287 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2288 if (CCInfo.getNextStackOffset()) {
2289 MachineFunction &MF = DAG.getMachineFunction();
2290 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2292 if (Subtarget->isTargetWin64())
2293 // Win64 ABI has additional complications.
2296 // Check if the arguments are already laid out in the right way as
2297 // the caller's fixed stack objects.
2298 MachineFrameInfo *MFI = MF.getFrameInfo();
2299 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2300 CCValAssign &VA = ArgLocs[i];
2301 EVT RegVT = VA.getLocVT();
2302 SDValue Arg = Outs[i].Val;
2303 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2304 if (Flags.isByVal())
2305 return false; // TODO
2306 if (VA.getLocInfo() == CCValAssign::Indirect)
2308 if (!VA.isRegLoc()) {
2309 LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg);
2312 SDValue Ptr = Ld->getBasePtr();
2313 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2316 int FI = FINode->getIndex();
2317 if (!MFI->isFixedObjectIndex(FI))
2319 if (VA.getLocMemOffset() != MFI->getObjectOffset(FI))
2330 X86TargetLowering::createFastISel(MachineFunction &mf, MachineModuleInfo *mmo,
2332 DenseMap<const Value *, unsigned> &vm,
2333 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2334 DenseMap<const AllocaInst *, int> &am
2336 , SmallSet<Instruction*, 8> &cil
2339 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2347 //===----------------------------------------------------------------------===//
2348 // Other Lowering Hooks
2349 //===----------------------------------------------------------------------===//
2352 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2353 MachineFunction &MF = DAG.getMachineFunction();
2354 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2355 int ReturnAddrIndex = FuncInfo->getRAIndex();
2357 if (ReturnAddrIndex == 0) {
2358 // Set up a frame object for the return address.
2359 uint64_t SlotSize = TD->getPointerSize();
2360 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2362 FuncInfo->setRAIndex(ReturnAddrIndex);
2365 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2369 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2370 bool hasSymbolicDisplacement) {
2371 // Offset should fit into 32 bit immediate field.
2372 if (!isInt32(Offset))
2375 // If we don't have a symbolic displacement - we don't have any extra
2377 if (!hasSymbolicDisplacement)
2380 // FIXME: Some tweaks might be needed for medium code model.
2381 if (M != CodeModel::Small && M != CodeModel::Kernel)
2384 // For small code model we assume that latest object is 16MB before end of 31
2385 // bits boundary. We may also accept pretty large negative constants knowing
2386 // that all objects are in the positive half of address space.
2387 if (M == CodeModel::Small && Offset < 16*1024*1024)
2390 // For kernel code model we know that all object resist in the negative half
2391 // of 32bits address space. We may not accept negative offsets, since they may
2392 // be just off and we may accept pretty large positive ones.
2393 if (M == CodeModel::Kernel && Offset > 0)
2399 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2400 /// specific condition code, returning the condition code and the LHS/RHS of the
2401 /// comparison to make.
2402 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2403 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2405 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2406 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2407 // X > -1 -> X == 0, jump !sign.
2408 RHS = DAG.getConstant(0, RHS.getValueType());
2409 return X86::COND_NS;
2410 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2411 // X < 0 -> X == 0, jump on sign.
2413 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2415 RHS = DAG.getConstant(0, RHS.getValueType());
2416 return X86::COND_LE;
2420 switch (SetCCOpcode) {
2421 default: llvm_unreachable("Invalid integer condition!");
2422 case ISD::SETEQ: return X86::COND_E;
2423 case ISD::SETGT: return X86::COND_G;
2424 case ISD::SETGE: return X86::COND_GE;
2425 case ISD::SETLT: return X86::COND_L;
2426 case ISD::SETLE: return X86::COND_LE;
2427 case ISD::SETNE: return X86::COND_NE;
2428 case ISD::SETULT: return X86::COND_B;
2429 case ISD::SETUGT: return X86::COND_A;
2430 case ISD::SETULE: return X86::COND_BE;
2431 case ISD::SETUGE: return X86::COND_AE;
2435 // First determine if it is required or is profitable to flip the operands.
2437 // If LHS is a foldable load, but RHS is not, flip the condition.
2438 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2439 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2440 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2441 std::swap(LHS, RHS);
2444 switch (SetCCOpcode) {
2450 std::swap(LHS, RHS);
2454 // On a floating point condition, the flags are set as follows:
2456 // 0 | 0 | 0 | X > Y
2457 // 0 | 0 | 1 | X < Y
2458 // 1 | 0 | 0 | X == Y
2459 // 1 | 1 | 1 | unordered
2460 switch (SetCCOpcode) {
2461 default: llvm_unreachable("Condcode should be pre-legalized away");
2463 case ISD::SETEQ: return X86::COND_E;
2464 case ISD::SETOLT: // flipped
2466 case ISD::SETGT: return X86::COND_A;
2467 case ISD::SETOLE: // flipped
2469 case ISD::SETGE: return X86::COND_AE;
2470 case ISD::SETUGT: // flipped
2472 case ISD::SETLT: return X86::COND_B;
2473 case ISD::SETUGE: // flipped
2475 case ISD::SETLE: return X86::COND_BE;
2477 case ISD::SETNE: return X86::COND_NE;
2478 case ISD::SETUO: return X86::COND_P;
2479 case ISD::SETO: return X86::COND_NP;
2481 case ISD::SETUNE: return X86::COND_INVALID;
2485 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2486 /// code. Current x86 isa includes the following FP cmov instructions:
2487 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2488 static bool hasFPCMov(unsigned X86CC) {
2504 /// isFPImmLegal - Returns true if the target can instruction select the
2505 /// specified FP immediate natively. If false, the legalizer will
2506 /// materialize the FP immediate as a load from a constant pool.
2507 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2508 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2509 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2515 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2516 /// the specified range (L, H].
2517 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2518 return (Val < 0) || (Val >= Low && Val < Hi);
2521 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2522 /// specified value.
2523 static bool isUndefOrEqual(int Val, int CmpVal) {
2524 if (Val < 0 || Val == CmpVal)
2529 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2530 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2531 /// the second operand.
2532 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2533 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2534 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2535 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2536 return (Mask[0] < 2 && Mask[1] < 2);
2540 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2541 SmallVector<int, 8> M;
2543 return ::isPSHUFDMask(M, N->getValueType(0));
2546 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2547 /// is suitable for input to PSHUFHW.
2548 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2549 if (VT != MVT::v8i16)
2552 // Lower quadword copied in order or undef.
2553 for (int i = 0; i != 4; ++i)
2554 if (Mask[i] >= 0 && Mask[i] != i)
2557 // Upper quadword shuffled.
2558 for (int i = 4; i != 8; ++i)
2559 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2565 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2566 SmallVector<int, 8> M;
2568 return ::isPSHUFHWMask(M, N->getValueType(0));
2571 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2572 /// is suitable for input to PSHUFLW.
2573 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2574 if (VT != MVT::v8i16)
2577 // Upper quadword copied in order.
2578 for (int i = 4; i != 8; ++i)
2579 if (Mask[i] >= 0 && Mask[i] != i)
2582 // Lower quadword shuffled.
2583 for (int i = 0; i != 4; ++i)
2590 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2591 SmallVector<int, 8> M;
2593 return ::isPSHUFLWMask(M, N->getValueType(0));
2596 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2597 /// is suitable for input to PALIGNR.
2598 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2600 int i, e = VT.getVectorNumElements();
2602 // Do not handle v2i64 / v2f64 shuffles with palignr.
2603 if (e < 4 || !hasSSSE3)
2606 for (i = 0; i != e; ++i)
2610 // All undef, not a palignr.
2614 // Determine if it's ok to perform a palignr with only the LHS, since we
2615 // don't have access to the actual shuffle elements to see if RHS is undef.
2616 bool Unary = Mask[i] < (int)e;
2617 bool NeedsUnary = false;
2619 int s = Mask[i] - i;
2621 // Check the rest of the elements to see if they are consecutive.
2622 for (++i; i != e; ++i) {
2627 Unary = Unary && (m < (int)e);
2628 NeedsUnary = NeedsUnary || (m < s);
2630 if (NeedsUnary && !Unary)
2632 if (Unary && m != ((s+i) & (e-1)))
2634 if (!Unary && m != (s+i))
2640 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2641 SmallVector<int, 8> M;
2643 return ::isPALIGNRMask(M, N->getValueType(0), true);
2646 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2647 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2648 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2649 int NumElems = VT.getVectorNumElements();
2650 if (NumElems != 2 && NumElems != 4)
2653 int Half = NumElems / 2;
2654 for (int i = 0; i < Half; ++i)
2655 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2657 for (int i = Half; i < NumElems; ++i)
2658 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2664 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2665 SmallVector<int, 8> M;
2667 return ::isSHUFPMask(M, N->getValueType(0));
2670 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2671 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2672 /// half elements to come from vector 1 (which would equal the dest.) and
2673 /// the upper half to come from vector 2.
2674 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2675 int NumElems = VT.getVectorNumElements();
2677 if (NumElems != 2 && NumElems != 4)
2680 int Half = NumElems / 2;
2681 for (int i = 0; i < Half; ++i)
2682 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2684 for (int i = Half; i < NumElems; ++i)
2685 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2690 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2691 SmallVector<int, 8> M;
2693 return isCommutedSHUFPMask(M, N->getValueType(0));
2696 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2697 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2698 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2699 if (N->getValueType(0).getVectorNumElements() != 4)
2702 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2703 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2704 isUndefOrEqual(N->getMaskElt(1), 7) &&
2705 isUndefOrEqual(N->getMaskElt(2), 2) &&
2706 isUndefOrEqual(N->getMaskElt(3), 3);
2709 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2710 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2712 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2713 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2718 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2719 isUndefOrEqual(N->getMaskElt(1), 3) &&
2720 isUndefOrEqual(N->getMaskElt(2), 2) &&
2721 isUndefOrEqual(N->getMaskElt(3), 3);
2724 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2725 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2726 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2727 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2729 if (NumElems != 2 && NumElems != 4)
2732 for (unsigned i = 0; i < NumElems/2; ++i)
2733 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2736 for (unsigned i = NumElems/2; i < NumElems; ++i)
2737 if (!isUndefOrEqual(N->getMaskElt(i), i))
2743 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2744 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2745 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2746 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2748 if (NumElems != 2 && NumElems != 4)
2751 for (unsigned i = 0; i < NumElems/2; ++i)
2752 if (!isUndefOrEqual(N->getMaskElt(i), i))
2755 for (unsigned i = 0; i < NumElems/2; ++i)
2756 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2762 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2763 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2764 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2765 bool V2IsSplat = false) {
2766 int NumElts = VT.getVectorNumElements();
2767 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2770 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2772 int BitI1 = Mask[i+1];
2773 if (!isUndefOrEqual(BitI, j))
2776 if (!isUndefOrEqual(BitI1, NumElts))
2779 if (!isUndefOrEqual(BitI1, j + NumElts))
2786 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2787 SmallVector<int, 8> M;
2789 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2792 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2793 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2794 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2795 bool V2IsSplat = false) {
2796 int NumElts = VT.getVectorNumElements();
2797 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2800 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2802 int BitI1 = Mask[i+1];
2803 if (!isUndefOrEqual(BitI, j + NumElts/2))
2806 if (isUndefOrEqual(BitI1, NumElts))
2809 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2816 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2817 SmallVector<int, 8> M;
2819 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2822 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2823 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2825 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2826 int NumElems = VT.getVectorNumElements();
2827 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2830 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2832 int BitI1 = Mask[i+1];
2833 if (!isUndefOrEqual(BitI, j))
2835 if (!isUndefOrEqual(BitI1, j))
2841 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2842 SmallVector<int, 8> M;
2844 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2847 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2848 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2850 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2851 int NumElems = VT.getVectorNumElements();
2852 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2855 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2857 int BitI1 = Mask[i+1];
2858 if (!isUndefOrEqual(BitI, j))
2860 if (!isUndefOrEqual(BitI1, j))
2866 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2867 SmallVector<int, 8> M;
2869 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2872 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2873 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2874 /// MOVSD, and MOVD, i.e. setting the lowest element.
2875 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2876 if (VT.getVectorElementType().getSizeInBits() < 32)
2879 int NumElts = VT.getVectorNumElements();
2881 if (!isUndefOrEqual(Mask[0], NumElts))
2884 for (int i = 1; i < NumElts; ++i)
2885 if (!isUndefOrEqual(Mask[i], i))
2891 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2892 SmallVector<int, 8> M;
2894 return ::isMOVLMask(M, N->getValueType(0));
2897 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2898 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2899 /// element of vector 2 and the other elements to come from vector 1 in order.
2900 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2901 bool V2IsSplat = false, bool V2IsUndef = false) {
2902 int NumOps = VT.getVectorNumElements();
2903 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2906 if (!isUndefOrEqual(Mask[0], 0))
2909 for (int i = 1; i < NumOps; ++i)
2910 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2911 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2912 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2918 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2919 bool V2IsUndef = false) {
2920 SmallVector<int, 8> M;
2922 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2925 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2926 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2927 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2928 if (N->getValueType(0).getVectorNumElements() != 4)
2931 // Expect 1, 1, 3, 3
2932 for (unsigned i = 0; i < 2; ++i) {
2933 int Elt = N->getMaskElt(i);
2934 if (Elt >= 0 && Elt != 1)
2939 for (unsigned i = 2; i < 4; ++i) {
2940 int Elt = N->getMaskElt(i);
2941 if (Elt >= 0 && Elt != 3)
2946 // Don't use movshdup if it can be done with a shufps.
2947 // FIXME: verify that matching u, u, 3, 3 is what we want.
2951 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2952 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2953 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2954 if (N->getValueType(0).getVectorNumElements() != 4)
2957 // Expect 0, 0, 2, 2
2958 for (unsigned i = 0; i < 2; ++i)
2959 if (N->getMaskElt(i) > 0)
2963 for (unsigned i = 2; i < 4; ++i) {
2964 int Elt = N->getMaskElt(i);
2965 if (Elt >= 0 && Elt != 2)
2970 // Don't use movsldup if it can be done with a shufps.
2974 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2975 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2976 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2977 int e = N->getValueType(0).getVectorNumElements() / 2;
2979 for (int i = 0; i < e; ++i)
2980 if (!isUndefOrEqual(N->getMaskElt(i), i))
2982 for (int i = 0; i < e; ++i)
2983 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2988 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2989 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
2990 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2991 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2992 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2994 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2996 for (int i = 0; i < NumOperands; ++i) {
2997 int Val = SVOp->getMaskElt(NumOperands-i-1);
2998 if (Val < 0) Val = 0;
2999 if (Val >= NumOperands) Val -= NumOperands;
3001 if (i != NumOperands - 1)
3007 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3008 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3009 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3010 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3012 // 8 nodes, but we only care about the last 4.
3013 for (unsigned i = 7; i >= 4; --i) {
3014 int Val = SVOp->getMaskElt(i);
3023 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3024 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3025 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3026 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3028 // 8 nodes, but we only care about the first 4.
3029 for (int i = 3; i >= 0; --i) {
3030 int Val = SVOp->getMaskElt(i);
3039 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3040 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3041 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3042 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3043 EVT VVT = N->getValueType(0);
3044 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3048 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3049 Val = SVOp->getMaskElt(i);
3053 return (Val - i) * EltSize;
3056 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3058 bool X86::isZeroNode(SDValue Elt) {
3059 return ((isa<ConstantSDNode>(Elt) &&
3060 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
3061 (isa<ConstantFPSDNode>(Elt) &&
3062 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3065 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3066 /// their permute mask.
3067 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3068 SelectionDAG &DAG) {
3069 EVT VT = SVOp->getValueType(0);
3070 unsigned NumElems = VT.getVectorNumElements();
3071 SmallVector<int, 8> MaskVec;
3073 for (unsigned i = 0; i != NumElems; ++i) {
3074 int idx = SVOp->getMaskElt(i);
3076 MaskVec.push_back(idx);
3077 else if (idx < (int)NumElems)
3078 MaskVec.push_back(idx + NumElems);
3080 MaskVec.push_back(idx - NumElems);
3082 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3083 SVOp->getOperand(0), &MaskVec[0]);
3086 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3087 /// the two vector operands have swapped position.
3088 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3089 unsigned NumElems = VT.getVectorNumElements();
3090 for (unsigned i = 0; i != NumElems; ++i) {
3094 else if (idx < (int)NumElems)
3095 Mask[i] = idx + NumElems;
3097 Mask[i] = idx - NumElems;
3101 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3102 /// match movhlps. The lower half elements should come from upper half of
3103 /// V1 (and in order), and the upper half elements should come from the upper
3104 /// half of V2 (and in order).
3105 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3106 if (Op->getValueType(0).getVectorNumElements() != 4)
3108 for (unsigned i = 0, e = 2; i != e; ++i)
3109 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3111 for (unsigned i = 2; i != 4; ++i)
3112 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3117 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3118 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3120 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3121 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3123 N = N->getOperand(0).getNode();
3124 if (!ISD::isNON_EXTLoad(N))
3127 *LD = cast<LoadSDNode>(N);
3131 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3132 /// match movlp{s|d}. The lower half elements should come from lower half of
3133 /// V1 (and in order), and the upper half elements should come from the upper
3134 /// half of V2 (and in order). And since V1 will become the source of the
3135 /// MOVLP, it must be either a vector load or a scalar load to vector.
3136 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3137 ShuffleVectorSDNode *Op) {
3138 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3140 // Is V2 is a vector load, don't do this transformation. We will try to use
3141 // load folding shufps op.
3142 if (ISD::isNON_EXTLoad(V2))
3145 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3147 if (NumElems != 2 && NumElems != 4)
3149 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3150 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3152 for (unsigned i = NumElems/2; i != NumElems; ++i)
3153 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3158 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3160 static bool isSplatVector(SDNode *N) {
3161 if (N->getOpcode() != ISD::BUILD_VECTOR)
3164 SDValue SplatValue = N->getOperand(0);
3165 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3166 if (N->getOperand(i) != SplatValue)
3171 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3172 /// to an zero vector.
3173 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3174 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3175 SDValue V1 = N->getOperand(0);
3176 SDValue V2 = N->getOperand(1);
3177 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3178 for (unsigned i = 0; i != NumElems; ++i) {
3179 int Idx = N->getMaskElt(i);
3180 if (Idx >= (int)NumElems) {
3181 unsigned Opc = V2.getOpcode();
3182 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3184 if (Opc != ISD::BUILD_VECTOR ||
3185 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3187 } else if (Idx >= 0) {
3188 unsigned Opc = V1.getOpcode();
3189 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3191 if (Opc != ISD::BUILD_VECTOR ||
3192 !X86::isZeroNode(V1.getOperand(Idx)))
3199 /// getZeroVector - Returns a vector of specified type with all zero elements.
3201 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3203 assert(VT.isVector() && "Expected a vector type");
3205 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3206 // type. This ensures they get CSE'd.
3208 if (VT.getSizeInBits() == 64) { // MMX
3209 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3210 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3211 } else if (HasSSE2) { // SSE2
3212 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3213 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3215 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3216 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3218 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3221 /// getOnesVector - Returns a vector of specified type with all bits set.
3223 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3224 assert(VT.isVector() && "Expected a vector type");
3226 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3227 // type. This ensures they get CSE'd.
3228 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3230 if (VT.getSizeInBits() == 64) // MMX
3231 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3233 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3234 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3238 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3239 /// that point to V2 points to its first element.
3240 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3241 EVT VT = SVOp->getValueType(0);
3242 unsigned NumElems = VT.getVectorNumElements();
3244 bool Changed = false;
3245 SmallVector<int, 8> MaskVec;
3246 SVOp->getMask(MaskVec);
3248 for (unsigned i = 0; i != NumElems; ++i) {
3249 if (MaskVec[i] > (int)NumElems) {
3250 MaskVec[i] = NumElems;
3255 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3256 SVOp->getOperand(1), &MaskVec[0]);
3257 return SDValue(SVOp, 0);
3260 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3261 /// operation of specified width.
3262 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3264 unsigned NumElems = VT.getVectorNumElements();
3265 SmallVector<int, 8> Mask;
3266 Mask.push_back(NumElems);
3267 for (unsigned i = 1; i != NumElems; ++i)
3269 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3272 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3273 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3275 unsigned NumElems = VT.getVectorNumElements();
3276 SmallVector<int, 8> Mask;
3277 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3279 Mask.push_back(i + NumElems);
3281 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3284 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3285 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3287 unsigned NumElems = VT.getVectorNumElements();
3288 unsigned Half = NumElems/2;
3289 SmallVector<int, 8> Mask;
3290 for (unsigned i = 0; i != Half; ++i) {
3291 Mask.push_back(i + Half);
3292 Mask.push_back(i + NumElems + Half);
3294 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3297 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3298 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3300 if (SV->getValueType(0).getVectorNumElements() <= 4)
3301 return SDValue(SV, 0);
3303 EVT PVT = MVT::v4f32;
3304 EVT VT = SV->getValueType(0);
3305 DebugLoc dl = SV->getDebugLoc();
3306 SDValue V1 = SV->getOperand(0);
3307 int NumElems = VT.getVectorNumElements();
3308 int EltNo = SV->getSplatIndex();
3310 // unpack elements to the correct location
3311 while (NumElems > 4) {
3312 if (EltNo < NumElems/2) {
3313 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3315 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3316 EltNo -= NumElems/2;
3321 // Perform the splat.
3322 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3323 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3324 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3325 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3328 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3329 /// vector of zero or undef vector. This produces a shuffle where the low
3330 /// element of V2 is swizzled into the zero/undef vector, landing at element
3331 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3332 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3333 bool isZero, bool HasSSE2,
3334 SelectionDAG &DAG) {
3335 EVT VT = V2.getValueType();
3337 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3338 unsigned NumElems = VT.getVectorNumElements();
3339 SmallVector<int, 16> MaskVec;
3340 for (unsigned i = 0; i != NumElems; ++i)
3341 // If this is the insertion idx, put the low elt of V2 here.
3342 MaskVec.push_back(i == Idx ? NumElems : i);
3343 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3346 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3347 /// a shuffle that is zero.
3349 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3350 bool Low, SelectionDAG &DAG) {
3351 unsigned NumZeros = 0;
3352 for (int i = 0; i < NumElems; ++i) {
3353 unsigned Index = Low ? i : NumElems-i-1;
3354 int Idx = SVOp->getMaskElt(Index);
3359 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3360 if (Elt.getNode() && X86::isZeroNode(Elt))
3368 /// isVectorShift - Returns true if the shuffle can be implemented as a
3369 /// logical left or right shift of a vector.
3370 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3371 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3372 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3373 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3376 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3379 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3383 bool SeenV1 = false;
3384 bool SeenV2 = false;
3385 for (int i = NumZeros; i < NumElems; ++i) {
3386 int Val = isLeft ? (i - NumZeros) : i;
3387 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3399 if (SeenV1 && SeenV2)
3402 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3408 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3410 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3411 unsigned NumNonZero, unsigned NumZero,
3412 SelectionDAG &DAG, TargetLowering &TLI) {
3416 DebugLoc dl = Op.getDebugLoc();
3419 for (unsigned i = 0; i < 16; ++i) {
3420 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3421 if (ThisIsNonZero && First) {
3423 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3425 V = DAG.getUNDEF(MVT::v8i16);
3430 SDValue ThisElt(0, 0), LastElt(0, 0);
3431 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3432 if (LastIsNonZero) {
3433 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3434 MVT::i16, Op.getOperand(i-1));
3436 if (ThisIsNonZero) {
3437 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3438 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3439 ThisElt, DAG.getConstant(8, MVT::i8));
3441 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3445 if (ThisElt.getNode())
3446 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3447 DAG.getIntPtrConstant(i/2));
3451 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3454 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3456 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3457 unsigned NumNonZero, unsigned NumZero,
3458 SelectionDAG &DAG, TargetLowering &TLI) {
3462 DebugLoc dl = Op.getDebugLoc();
3465 for (unsigned i = 0; i < 8; ++i) {
3466 bool isNonZero = (NonZeros & (1 << i)) != 0;
3470 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3472 V = DAG.getUNDEF(MVT::v8i16);
3475 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3476 MVT::v8i16, V, Op.getOperand(i),
3477 DAG.getIntPtrConstant(i));
3484 /// getVShift - Return a vector logical shift node.
3486 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3487 unsigned NumBits, SelectionDAG &DAG,
3488 const TargetLowering &TLI, DebugLoc dl) {
3489 bool isMMX = VT.getSizeInBits() == 64;
3490 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3491 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3492 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3493 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3494 DAG.getNode(Opc, dl, ShVT, SrcOp,
3495 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3499 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3500 SelectionDAG &DAG) {
3502 // Check if the scalar load can be widened into a vector load. And if
3503 // the address is "base + cst" see if the cst can be "absorbed" into
3504 // the shuffle mask.
3505 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3506 SDValue Ptr = LD->getBasePtr();
3507 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3509 EVT PVT = LD->getValueType(0);
3510 if (PVT != MVT::i32 && PVT != MVT::f32)
3515 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3516 FI = FINode->getIndex();
3518 } else if (Ptr.getOpcode() == ISD::ADD &&
3519 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3520 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3521 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3522 Offset = Ptr.getConstantOperandVal(1);
3523 Ptr = Ptr.getOperand(0);
3528 SDValue Chain = LD->getChain();
3529 // Make sure the stack object alignment is at least 16.
3530 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3531 if (DAG.InferPtrAlignment(Ptr) < 16) {
3532 if (MFI->isFixedObjectIndex(FI)) {
3533 // Can't change the alignment. FIXME: It's possible to compute
3534 // the exact stack offset and reference FI + adjust offset instead.
3535 // If someone *really* cares about this. That's the way to implement it.
3538 MFI->setObjectAlignment(FI, 16);
3542 // (Offset % 16) must be multiple of 4. Then address is then
3543 // Ptr + (Offset & ~15).
3546 if ((Offset % 16) & 3)
3548 int64_t StartOffset = Offset & ~15;
3550 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3551 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3553 int EltNo = (Offset - StartOffset) >> 2;
3554 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3555 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3556 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0);
3557 // Canonicalize it to a v4i32 shuffle.
3558 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3559 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3560 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3561 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3568 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3569 DebugLoc dl = Op.getDebugLoc();
3570 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3571 if (ISD::isBuildVectorAllZeros(Op.getNode())
3572 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3573 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3574 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3575 // eliminated on x86-32 hosts.
3576 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3579 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3580 return getOnesVector(Op.getValueType(), DAG, dl);
3581 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3584 EVT VT = Op.getValueType();
3585 EVT ExtVT = VT.getVectorElementType();
3586 unsigned EVTBits = ExtVT.getSizeInBits();
3588 unsigned NumElems = Op.getNumOperands();
3589 unsigned NumZero = 0;
3590 unsigned NumNonZero = 0;
3591 unsigned NonZeros = 0;
3592 bool IsAllConstants = true;
3593 SmallSet<SDValue, 8> Values;
3594 for (unsigned i = 0; i < NumElems; ++i) {
3595 SDValue Elt = Op.getOperand(i);
3596 if (Elt.getOpcode() == ISD::UNDEF)
3599 if (Elt.getOpcode() != ISD::Constant &&
3600 Elt.getOpcode() != ISD::ConstantFP)
3601 IsAllConstants = false;
3602 if (X86::isZeroNode(Elt))
3605 NonZeros |= (1 << i);
3610 if (NumNonZero == 0) {
3611 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3612 return DAG.getUNDEF(VT);
3615 // Special case for single non-zero, non-undef, element.
3616 if (NumNonZero == 1) {
3617 unsigned Idx = CountTrailingZeros_32(NonZeros);
3618 SDValue Item = Op.getOperand(Idx);
3620 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3621 // the value are obviously zero, truncate the value to i32 and do the
3622 // insertion that way. Only do this if the value is non-constant or if the
3623 // value is a constant being inserted into element 0. It is cheaper to do
3624 // a constant pool load than it is to do a movd + shuffle.
3625 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3626 (!IsAllConstants || Idx == 0)) {
3627 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3628 // Handle MMX and SSE both.
3629 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3630 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3632 // Truncate the value (which may itself be a constant) to i32, and
3633 // convert it to a vector with movd (S2V+shuffle to zero extend).
3634 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3635 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3636 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3637 Subtarget->hasSSE2(), DAG);
3639 // Now we have our 32-bit value zero extended in the low element of
3640 // a vector. If Idx != 0, swizzle it into place.
3642 SmallVector<int, 4> Mask;
3643 Mask.push_back(Idx);
3644 for (unsigned i = 1; i != VecElts; ++i)
3646 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3647 DAG.getUNDEF(Item.getValueType()),
3650 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3654 // If we have a constant or non-constant insertion into the low element of
3655 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3656 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3657 // depending on what the source datatype is.
3660 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3661 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3662 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3663 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3664 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3665 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3667 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3668 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3669 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3670 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3671 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3672 Subtarget->hasSSE2(), DAG);
3673 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3677 // Is it a vector logical left shift?
3678 if (NumElems == 2 && Idx == 1 &&
3679 X86::isZeroNode(Op.getOperand(0)) &&
3680 !X86::isZeroNode(Op.getOperand(1))) {
3681 unsigned NumBits = VT.getSizeInBits();
3682 return getVShift(true, VT,
3683 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3684 VT, Op.getOperand(1)),
3685 NumBits/2, DAG, *this, dl);
3688 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3691 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3692 // is a non-constant being inserted into an element other than the low one,
3693 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3694 // movd/movss) to move this into the low element, then shuffle it into
3696 if (EVTBits == 32) {
3697 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3699 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3700 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3701 Subtarget->hasSSE2(), DAG);
3702 SmallVector<int, 8> MaskVec;
3703 for (unsigned i = 0; i < NumElems; i++)
3704 MaskVec.push_back(i == Idx ? 0 : 1);
3705 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3709 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3710 if (Values.size() == 1) {
3711 if (EVTBits == 32) {
3712 // Instead of a shuffle like this:
3713 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3714 // Check if it's possible to issue this instead.
3715 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3716 unsigned Idx = CountTrailingZeros_32(NonZeros);
3717 SDValue Item = Op.getOperand(Idx);
3718 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3719 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3724 // A vector full of immediates; various special cases are already
3725 // handled, so this is best done with a single constant-pool load.
3729 // Let legalizer expand 2-wide build_vectors.
3730 if (EVTBits == 64) {
3731 if (NumNonZero == 1) {
3732 // One half is zero or undef.
3733 unsigned Idx = CountTrailingZeros_32(NonZeros);
3734 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3735 Op.getOperand(Idx));
3736 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3737 Subtarget->hasSSE2(), DAG);
3742 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3743 if (EVTBits == 8 && NumElems == 16) {
3744 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3746 if (V.getNode()) return V;
3749 if (EVTBits == 16 && NumElems == 8) {
3750 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3752 if (V.getNode()) return V;
3755 // If element VT is == 32 bits, turn it into a number of shuffles.
3756 SmallVector<SDValue, 8> V;
3758 if (NumElems == 4 && NumZero > 0) {
3759 for (unsigned i = 0; i < 4; ++i) {
3760 bool isZero = !(NonZeros & (1 << i));
3762 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3764 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3767 for (unsigned i = 0; i < 2; ++i) {
3768 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3771 V[i] = V[i*2]; // Must be a zero vector.
3774 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3777 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3780 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3785 SmallVector<int, 8> MaskVec;
3786 bool Reverse = (NonZeros & 0x3) == 2;
3787 for (unsigned i = 0; i < 2; ++i)
3788 MaskVec.push_back(Reverse ? 1-i : i);
3789 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3790 for (unsigned i = 0; i < 2; ++i)
3791 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3792 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3795 if (Values.size() > 2) {
3796 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3797 // values to be inserted is equal to the number of elements, in which case
3798 // use the unpack code below in the hopes of matching the consecutive elts
3799 // load merge pattern for shuffles.
3800 // FIXME: We could probably just check that here directly.
3801 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3802 getSubtarget()->hasSSE41()) {
3803 V[0] = DAG.getUNDEF(VT);
3804 for (unsigned i = 0; i < NumElems; ++i)
3805 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3806 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3807 Op.getOperand(i), DAG.getIntPtrConstant(i));
3810 // Expand into a number of unpckl*.
3812 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3813 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3814 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3815 for (unsigned i = 0; i < NumElems; ++i)
3816 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3818 while (NumElems != 0) {
3819 for (unsigned i = 0; i < NumElems; ++i)
3820 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3830 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3831 // We support concatenate two MMX registers and place them in a MMX
3832 // register. This is better than doing a stack convert.
3833 DebugLoc dl = Op.getDebugLoc();
3834 EVT ResVT = Op.getValueType();
3835 assert(Op.getNumOperands() == 2);
3836 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
3837 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
3839 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
3840 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3841 InVec = Op.getOperand(1);
3842 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3843 unsigned NumElts = ResVT.getVectorNumElements();
3844 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3845 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
3846 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
3848 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
3849 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3850 Mask[0] = 0; Mask[1] = 2;
3851 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
3853 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3856 // v8i16 shuffles - Prefer shuffles in the following order:
3857 // 1. [all] pshuflw, pshufhw, optional move
3858 // 2. [ssse3] 1 x pshufb
3859 // 3. [ssse3] 2 x pshufb + 1 x por
3860 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3862 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3863 SelectionDAG &DAG, X86TargetLowering &TLI) {
3864 SDValue V1 = SVOp->getOperand(0);
3865 SDValue V2 = SVOp->getOperand(1);
3866 DebugLoc dl = SVOp->getDebugLoc();
3867 SmallVector<int, 8> MaskVals;
3869 // Determine if more than 1 of the words in each of the low and high quadwords
3870 // of the result come from the same quadword of one of the two inputs. Undef
3871 // mask values count as coming from any quadword, for better codegen.
3872 SmallVector<unsigned, 4> LoQuad(4);
3873 SmallVector<unsigned, 4> HiQuad(4);
3874 BitVector InputQuads(4);
3875 for (unsigned i = 0; i < 8; ++i) {
3876 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3877 int EltIdx = SVOp->getMaskElt(i);
3878 MaskVals.push_back(EltIdx);
3887 InputQuads.set(EltIdx / 4);
3890 int BestLoQuad = -1;
3891 unsigned MaxQuad = 1;
3892 for (unsigned i = 0; i < 4; ++i) {
3893 if (LoQuad[i] > MaxQuad) {
3895 MaxQuad = LoQuad[i];
3899 int BestHiQuad = -1;
3901 for (unsigned i = 0; i < 4; ++i) {
3902 if (HiQuad[i] > MaxQuad) {
3904 MaxQuad = HiQuad[i];
3908 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3909 // of the two input vectors, shuffle them into one input vector so only a
3910 // single pshufb instruction is necessary. If There are more than 2 input
3911 // quads, disable the next transformation since it does not help SSSE3.
3912 bool V1Used = InputQuads[0] || InputQuads[1];
3913 bool V2Used = InputQuads[2] || InputQuads[3];
3914 if (TLI.getSubtarget()->hasSSSE3()) {
3915 if (InputQuads.count() == 2 && V1Used && V2Used) {
3916 BestLoQuad = InputQuads.find_first();
3917 BestHiQuad = InputQuads.find_next(BestLoQuad);
3919 if (InputQuads.count() > 2) {
3925 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3926 // the shuffle mask. If a quad is scored as -1, that means that it contains
3927 // words from all 4 input quadwords.
3929 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3930 SmallVector<int, 8> MaskV;
3931 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3932 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3933 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3934 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3935 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3936 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3938 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3939 // source words for the shuffle, to aid later transformations.
3940 bool AllWordsInNewV = true;
3941 bool InOrder[2] = { true, true };
3942 for (unsigned i = 0; i != 8; ++i) {
3943 int idx = MaskVals[i];
3945 InOrder[i/4] = false;
3946 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3948 AllWordsInNewV = false;
3952 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3953 if (AllWordsInNewV) {
3954 for (int i = 0; i != 8; ++i) {
3955 int idx = MaskVals[i];
3958 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3959 if ((idx != i) && idx < 4)
3961 if ((idx != i) && idx > 3)
3970 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3971 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3972 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3973 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3974 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3978 // If we have SSSE3, and all words of the result are from 1 input vector,
3979 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3980 // is present, fall back to case 4.
3981 if (TLI.getSubtarget()->hasSSSE3()) {
3982 SmallVector<SDValue,16> pshufbMask;
3984 // If we have elements from both input vectors, set the high bit of the
3985 // shuffle mask element to zero out elements that come from V2 in the V1
3986 // mask, and elements that come from V1 in the V2 mask, so that the two
3987 // results can be OR'd together.
3988 bool TwoInputs = V1Used && V2Used;
3989 for (unsigned i = 0; i != 8; ++i) {
3990 int EltIdx = MaskVals[i] * 2;
3991 if (TwoInputs && (EltIdx >= 16)) {
3992 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3993 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3996 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3997 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3999 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4000 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4001 DAG.getNode(ISD::BUILD_VECTOR, dl,
4002 MVT::v16i8, &pshufbMask[0], 16));
4004 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4006 // Calculate the shuffle mask for the second input, shuffle it, and
4007 // OR it with the first shuffled input.
4009 for (unsigned i = 0; i != 8; ++i) {
4010 int EltIdx = MaskVals[i] * 2;
4012 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4013 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4016 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4017 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4019 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4020 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4021 DAG.getNode(ISD::BUILD_VECTOR, dl,
4022 MVT::v16i8, &pshufbMask[0], 16));
4023 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4024 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4027 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4028 // and update MaskVals with new element order.
4029 BitVector InOrder(8);
4030 if (BestLoQuad >= 0) {
4031 SmallVector<int, 8> MaskV;
4032 for (int i = 0; i != 4; ++i) {
4033 int idx = MaskVals[i];
4035 MaskV.push_back(-1);
4037 } else if ((idx / 4) == BestLoQuad) {
4038 MaskV.push_back(idx & 3);
4041 MaskV.push_back(-1);
4044 for (unsigned i = 4; i != 8; ++i)
4046 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4050 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4051 // and update MaskVals with the new element order.
4052 if (BestHiQuad >= 0) {
4053 SmallVector<int, 8> MaskV;
4054 for (unsigned i = 0; i != 4; ++i)
4056 for (unsigned i = 4; i != 8; ++i) {
4057 int idx = MaskVals[i];
4059 MaskV.push_back(-1);
4061 } else if ((idx / 4) == BestHiQuad) {
4062 MaskV.push_back((idx & 3) + 4);
4065 MaskV.push_back(-1);
4068 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4072 // In case BestHi & BestLo were both -1, which means each quadword has a word
4073 // from each of the four input quadwords, calculate the InOrder bitvector now
4074 // before falling through to the insert/extract cleanup.
4075 if (BestLoQuad == -1 && BestHiQuad == -1) {
4077 for (int i = 0; i != 8; ++i)
4078 if (MaskVals[i] < 0 || MaskVals[i] == i)
4082 // The other elements are put in the right place using pextrw and pinsrw.
4083 for (unsigned i = 0; i != 8; ++i) {
4086 int EltIdx = MaskVals[i];
4089 SDValue ExtOp = (EltIdx < 8)
4090 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4091 DAG.getIntPtrConstant(EltIdx))
4092 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4093 DAG.getIntPtrConstant(EltIdx - 8));
4094 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4095 DAG.getIntPtrConstant(i));
4100 // v16i8 shuffles - Prefer shuffles in the following order:
4101 // 1. [ssse3] 1 x pshufb
4102 // 2. [ssse3] 2 x pshufb + 1 x por
4103 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4105 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4106 SelectionDAG &DAG, X86TargetLowering &TLI) {
4107 SDValue V1 = SVOp->getOperand(0);
4108 SDValue V2 = SVOp->getOperand(1);
4109 DebugLoc dl = SVOp->getDebugLoc();
4110 SmallVector<int, 16> MaskVals;
4111 SVOp->getMask(MaskVals);
4113 // If we have SSSE3, case 1 is generated when all result bytes come from
4114 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4115 // present, fall back to case 3.
4116 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4119 for (unsigned i = 0; i < 16; ++i) {
4120 int EltIdx = MaskVals[i];
4129 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4130 if (TLI.getSubtarget()->hasSSSE3()) {
4131 SmallVector<SDValue,16> pshufbMask;
4133 // If all result elements are from one input vector, then only translate
4134 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4136 // Otherwise, we have elements from both input vectors, and must zero out
4137 // elements that come from V2 in the first mask, and V1 in the second mask
4138 // so that we can OR them together.
4139 bool TwoInputs = !(V1Only || V2Only);
4140 for (unsigned i = 0; i != 16; ++i) {
4141 int EltIdx = MaskVals[i];
4142 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4143 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4146 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4148 // If all the elements are from V2, assign it to V1 and return after
4149 // building the first pshufb.
4152 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4153 DAG.getNode(ISD::BUILD_VECTOR, dl,
4154 MVT::v16i8, &pshufbMask[0], 16));
4158 // Calculate the shuffle mask for the second input, shuffle it, and
4159 // OR it with the first shuffled input.
4161 for (unsigned i = 0; i != 16; ++i) {
4162 int EltIdx = MaskVals[i];
4164 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4167 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4169 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4170 DAG.getNode(ISD::BUILD_VECTOR, dl,
4171 MVT::v16i8, &pshufbMask[0], 16));
4172 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4175 // No SSSE3 - Calculate in place words and then fix all out of place words
4176 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4177 // the 16 different words that comprise the two doublequadword input vectors.
4178 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4179 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4180 SDValue NewV = V2Only ? V2 : V1;
4181 for (int i = 0; i != 8; ++i) {
4182 int Elt0 = MaskVals[i*2];
4183 int Elt1 = MaskVals[i*2+1];
4185 // This word of the result is all undef, skip it.
4186 if (Elt0 < 0 && Elt1 < 0)
4189 // This word of the result is already in the correct place, skip it.
4190 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4192 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4195 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4196 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4199 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4200 // using a single extract together, load it and store it.
4201 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4202 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4203 DAG.getIntPtrConstant(Elt1 / 2));
4204 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4205 DAG.getIntPtrConstant(i));
4209 // If Elt1 is defined, extract it from the appropriate source. If the
4210 // source byte is not also odd, shift the extracted word left 8 bits
4211 // otherwise clear the bottom 8 bits if we need to do an or.
4213 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4214 DAG.getIntPtrConstant(Elt1 / 2));
4215 if ((Elt1 & 1) == 0)
4216 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4217 DAG.getConstant(8, TLI.getShiftAmountTy()));
4219 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4220 DAG.getConstant(0xFF00, MVT::i16));
4222 // If Elt0 is defined, extract it from the appropriate source. If the
4223 // source byte is not also even, shift the extracted word right 8 bits. If
4224 // Elt1 was also defined, OR the extracted values together before
4225 // inserting them in the result.
4227 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4228 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4229 if ((Elt0 & 1) != 0)
4230 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4231 DAG.getConstant(8, TLI.getShiftAmountTy()));
4233 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4234 DAG.getConstant(0x00FF, MVT::i16));
4235 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4238 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4239 DAG.getIntPtrConstant(i));
4241 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4244 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4245 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4246 /// done when every pair / quad of shuffle mask elements point to elements in
4247 /// the right sequence. e.g.
4248 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4250 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4252 TargetLowering &TLI, DebugLoc dl) {
4253 EVT VT = SVOp->getValueType(0);
4254 SDValue V1 = SVOp->getOperand(0);
4255 SDValue V2 = SVOp->getOperand(1);
4256 unsigned NumElems = VT.getVectorNumElements();
4257 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4258 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4259 EVT MaskEltVT = MaskVT.getVectorElementType();
4261 switch (VT.getSimpleVT().SimpleTy) {
4262 default: assert(false && "Unexpected!");
4263 case MVT::v4f32: NewVT = MVT::v2f64; break;
4264 case MVT::v4i32: NewVT = MVT::v2i64; break;
4265 case MVT::v8i16: NewVT = MVT::v4i32; break;
4266 case MVT::v16i8: NewVT = MVT::v4i32; break;
4269 if (NewWidth == 2) {
4275 int Scale = NumElems / NewWidth;
4276 SmallVector<int, 8> MaskVec;
4277 for (unsigned i = 0; i < NumElems; i += Scale) {
4279 for (int j = 0; j < Scale; ++j) {
4280 int EltIdx = SVOp->getMaskElt(i+j);
4284 StartIdx = EltIdx - (EltIdx % Scale);
4285 if (EltIdx != StartIdx + j)
4289 MaskVec.push_back(-1);
4291 MaskVec.push_back(StartIdx / Scale);
4294 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4295 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4296 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4299 /// getVZextMovL - Return a zero-extending vector move low node.
4301 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4302 SDValue SrcOp, SelectionDAG &DAG,
4303 const X86Subtarget *Subtarget, DebugLoc dl) {
4304 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4305 LoadSDNode *LD = NULL;
4306 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4307 LD = dyn_cast<LoadSDNode>(SrcOp);
4309 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4311 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4312 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4313 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4314 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4315 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4317 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4318 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4319 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4320 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4328 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4329 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4330 DAG.getNode(ISD::BIT_CONVERT, dl,
4334 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4337 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4338 SDValue V1 = SVOp->getOperand(0);
4339 SDValue V2 = SVOp->getOperand(1);
4340 DebugLoc dl = SVOp->getDebugLoc();
4341 EVT VT = SVOp->getValueType(0);
4343 SmallVector<std::pair<int, int>, 8> Locs;
4345 SmallVector<int, 8> Mask1(4U, -1);
4346 SmallVector<int, 8> PermMask;
4347 SVOp->getMask(PermMask);
4351 for (unsigned i = 0; i != 4; ++i) {
4352 int Idx = PermMask[i];
4354 Locs[i] = std::make_pair(-1, -1);
4356 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4358 Locs[i] = std::make_pair(0, NumLo);
4362 Locs[i] = std::make_pair(1, NumHi);
4364 Mask1[2+NumHi] = Idx;
4370 if (NumLo <= 2 && NumHi <= 2) {
4371 // If no more than two elements come from either vector. This can be
4372 // implemented with two shuffles. First shuffle gather the elements.
4373 // The second shuffle, which takes the first shuffle as both of its
4374 // vector operands, put the elements into the right order.
4375 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4377 SmallVector<int, 8> Mask2(4U, -1);
4379 for (unsigned i = 0; i != 4; ++i) {
4380 if (Locs[i].first == -1)
4383 unsigned Idx = (i < 2) ? 0 : 4;
4384 Idx += Locs[i].first * 2 + Locs[i].second;
4389 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4390 } else if (NumLo == 3 || NumHi == 3) {
4391 // Otherwise, we must have three elements from one vector, call it X, and
4392 // one element from the other, call it Y. First, use a shufps to build an
4393 // intermediate vector with the one element from Y and the element from X
4394 // that will be in the same half in the final destination (the indexes don't
4395 // matter). Then, use a shufps to build the final vector, taking the half
4396 // containing the element from Y from the intermediate, and the other half
4399 // Normalize it so the 3 elements come from V1.
4400 CommuteVectorShuffleMask(PermMask, VT);
4404 // Find the element from V2.
4406 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4407 int Val = PermMask[HiIndex];
4414 Mask1[0] = PermMask[HiIndex];
4416 Mask1[2] = PermMask[HiIndex^1];
4418 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4421 Mask1[0] = PermMask[0];
4422 Mask1[1] = PermMask[1];
4423 Mask1[2] = HiIndex & 1 ? 6 : 4;
4424 Mask1[3] = HiIndex & 1 ? 4 : 6;
4425 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4427 Mask1[0] = HiIndex & 1 ? 2 : 0;
4428 Mask1[1] = HiIndex & 1 ? 0 : 2;
4429 Mask1[2] = PermMask[2];
4430 Mask1[3] = PermMask[3];
4435 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4439 // Break it into (shuffle shuffle_hi, shuffle_lo).
4441 SmallVector<int,8> LoMask(4U, -1);
4442 SmallVector<int,8> HiMask(4U, -1);
4444 SmallVector<int,8> *MaskPtr = &LoMask;
4445 unsigned MaskIdx = 0;
4448 for (unsigned i = 0; i != 4; ++i) {
4455 int Idx = PermMask[i];
4457 Locs[i] = std::make_pair(-1, -1);
4458 } else if (Idx < 4) {
4459 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4460 (*MaskPtr)[LoIdx] = Idx;
4463 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4464 (*MaskPtr)[HiIdx] = Idx;
4469 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4470 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4471 SmallVector<int, 8> MaskOps;
4472 for (unsigned i = 0; i != 4; ++i) {
4473 if (Locs[i].first == -1) {
4474 MaskOps.push_back(-1);
4476 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4477 MaskOps.push_back(Idx);
4480 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4484 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4485 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4486 SDValue V1 = Op.getOperand(0);
4487 SDValue V2 = Op.getOperand(1);
4488 EVT VT = Op.getValueType();
4489 DebugLoc dl = Op.getDebugLoc();
4490 unsigned NumElems = VT.getVectorNumElements();
4491 bool isMMX = VT.getSizeInBits() == 64;
4492 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4493 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4494 bool V1IsSplat = false;
4495 bool V2IsSplat = false;
4497 if (isZeroShuffle(SVOp))
4498 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4500 // Promote splats to v4f32.
4501 if (SVOp->isSplat()) {
4502 if (isMMX || NumElems < 4)
4504 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4507 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4509 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4510 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4511 if (NewOp.getNode())
4512 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4513 LowerVECTOR_SHUFFLE(NewOp, DAG));
4514 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4515 // FIXME: Figure out a cleaner way to do this.
4516 // Try to make use of movq to zero out the top part.
4517 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4518 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4519 if (NewOp.getNode()) {
4520 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4521 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4522 DAG, Subtarget, dl);
4524 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4525 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4526 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4527 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4528 DAG, Subtarget, dl);
4532 if (X86::isPSHUFDMask(SVOp))
4535 // Check if this can be converted into a logical shift.
4536 bool isLeft = false;
4539 bool isShift = getSubtarget()->hasSSE2() &&
4540 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4541 if (isShift && ShVal.hasOneUse()) {
4542 // If the shifted value has multiple uses, it may be cheaper to use
4543 // v_set0 + movlhps or movhlps, etc.
4544 EVT EltVT = VT.getVectorElementType();
4545 ShAmt *= EltVT.getSizeInBits();
4546 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4549 if (X86::isMOVLMask(SVOp)) {
4552 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4553 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4558 // FIXME: fold these into legal mask.
4559 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4560 X86::isMOVSLDUPMask(SVOp) ||
4561 X86::isMOVHLPSMask(SVOp) ||
4562 X86::isMOVLHPSMask(SVOp) ||
4563 X86::isMOVLPMask(SVOp)))
4566 if (ShouldXformToMOVHLPS(SVOp) ||
4567 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4568 return CommuteVectorShuffle(SVOp, DAG);
4571 // No better options. Use a vshl / vsrl.
4572 EVT EltVT = VT.getVectorElementType();
4573 ShAmt *= EltVT.getSizeInBits();
4574 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4577 bool Commuted = false;
4578 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4579 // 1,1,1,1 -> v8i16 though.
4580 V1IsSplat = isSplatVector(V1.getNode());
4581 V2IsSplat = isSplatVector(V2.getNode());
4583 // Canonicalize the splat or undef, if present, to be on the RHS.
4584 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4585 Op = CommuteVectorShuffle(SVOp, DAG);
4586 SVOp = cast<ShuffleVectorSDNode>(Op);
4587 V1 = SVOp->getOperand(0);
4588 V2 = SVOp->getOperand(1);
4589 std::swap(V1IsSplat, V2IsSplat);
4590 std::swap(V1IsUndef, V2IsUndef);
4594 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4595 // Shuffling low element of v1 into undef, just return v1.
4598 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4599 // the instruction selector will not match, so get a canonical MOVL with
4600 // swapped operands to undo the commute.
4601 return getMOVL(DAG, dl, VT, V2, V1);
4604 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4605 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4606 X86::isUNPCKLMask(SVOp) ||
4607 X86::isUNPCKHMask(SVOp))
4611 // Normalize mask so all entries that point to V2 points to its first
4612 // element then try to match unpck{h|l} again. If match, return a
4613 // new vector_shuffle with the corrected mask.
4614 SDValue NewMask = NormalizeMask(SVOp, DAG);
4615 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4616 if (NSVOp != SVOp) {
4617 if (X86::isUNPCKLMask(NSVOp, true)) {
4619 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4626 // Commute is back and try unpck* again.
4627 // FIXME: this seems wrong.
4628 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4629 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4630 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4631 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4632 X86::isUNPCKLMask(NewSVOp) ||
4633 X86::isUNPCKHMask(NewSVOp))
4637 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4639 // Normalize the node to match x86 shuffle ops if needed
4640 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4641 return CommuteVectorShuffle(SVOp, DAG);
4643 // Check for legal shuffle and return?
4644 SmallVector<int, 16> PermMask;
4645 SVOp->getMask(PermMask);
4646 if (isShuffleMaskLegal(PermMask, VT))
4649 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4650 if (VT == MVT::v8i16) {
4651 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4652 if (NewOp.getNode())
4656 if (VT == MVT::v16i8) {
4657 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4658 if (NewOp.getNode())
4662 // Handle all 4 wide cases with a number of shuffles except for MMX.
4663 if (NumElems == 4 && !isMMX)
4664 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4670 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4671 SelectionDAG &DAG) {
4672 EVT VT = Op.getValueType();
4673 DebugLoc dl = Op.getDebugLoc();
4674 if (VT.getSizeInBits() == 8) {
4675 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4676 Op.getOperand(0), Op.getOperand(1));
4677 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4678 DAG.getValueType(VT));
4679 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4680 } else if (VT.getSizeInBits() == 16) {
4681 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4682 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4684 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4685 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4686 DAG.getNode(ISD::BIT_CONVERT, dl,
4690 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4691 Op.getOperand(0), Op.getOperand(1));
4692 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4693 DAG.getValueType(VT));
4694 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4695 } else if (VT == MVT::f32) {
4696 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4697 // the result back to FR32 register. It's only worth matching if the
4698 // result has a single use which is a store or a bitcast to i32. And in
4699 // the case of a store, it's not worth it if the index is a constant 0,
4700 // because a MOVSSmr can be used instead, which is smaller and faster.
4701 if (!Op.hasOneUse())
4703 SDNode *User = *Op.getNode()->use_begin();
4704 if ((User->getOpcode() != ISD::STORE ||
4705 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4706 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4707 (User->getOpcode() != ISD::BIT_CONVERT ||
4708 User->getValueType(0) != MVT::i32))
4710 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4711 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4714 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4715 } else if (VT == MVT::i32) {
4716 // ExtractPS works with constant index.
4717 if (isa<ConstantSDNode>(Op.getOperand(1)))
4725 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4726 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4729 if (Subtarget->hasSSE41()) {
4730 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4735 EVT VT = Op.getValueType();
4736 DebugLoc dl = Op.getDebugLoc();
4737 // TODO: handle v16i8.
4738 if (VT.getSizeInBits() == 16) {
4739 SDValue Vec = Op.getOperand(0);
4740 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4742 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4743 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4744 DAG.getNode(ISD::BIT_CONVERT, dl,
4747 // Transform it so it match pextrw which produces a 32-bit result.
4748 EVT EltVT = MVT::i32;
4749 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4750 Op.getOperand(0), Op.getOperand(1));
4751 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4752 DAG.getValueType(VT));
4753 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4754 } else if (VT.getSizeInBits() == 32) {
4755 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4759 // SHUFPS the element to the lowest double word, then movss.
4760 int Mask[4] = { Idx, -1, -1, -1 };
4761 EVT VVT = Op.getOperand(0).getValueType();
4762 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4763 DAG.getUNDEF(VVT), Mask);
4764 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4765 DAG.getIntPtrConstant(0));
4766 } else if (VT.getSizeInBits() == 64) {
4767 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4768 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4769 // to match extract_elt for f64.
4770 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4774 // UNPCKHPD the element to the lowest double word, then movsd.
4775 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4776 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4777 int Mask[2] = { 1, -1 };
4778 EVT VVT = Op.getOperand(0).getValueType();
4779 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4780 DAG.getUNDEF(VVT), Mask);
4781 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4782 DAG.getIntPtrConstant(0));
4789 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4790 EVT VT = Op.getValueType();
4791 EVT EltVT = VT.getVectorElementType();
4792 DebugLoc dl = Op.getDebugLoc();
4794 SDValue N0 = Op.getOperand(0);
4795 SDValue N1 = Op.getOperand(1);
4796 SDValue N2 = Op.getOperand(2);
4798 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4799 isa<ConstantSDNode>(N2)) {
4800 unsigned Opc = (EltVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4802 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4804 if (N1.getValueType() != MVT::i32)
4805 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4806 if (N2.getValueType() != MVT::i32)
4807 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4808 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4809 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4810 // Bits [7:6] of the constant are the source select. This will always be
4811 // zero here. The DAG Combiner may combine an extract_elt index into these
4812 // bits. For example (insert (extract, 3), 2) could be matched by putting
4813 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4814 // Bits [5:4] of the constant are the destination select. This is the
4815 // value of the incoming immediate.
4816 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4817 // combine either bitwise AND or insert of float 0.0 to set these bits.
4818 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4819 // Create this as a scalar to vector..
4820 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4821 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4822 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4823 // PINSR* works with constant index.
4830 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4831 EVT VT = Op.getValueType();
4832 EVT EltVT = VT.getVectorElementType();
4834 if (Subtarget->hasSSE41())
4835 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4837 if (EltVT == MVT::i8)
4840 DebugLoc dl = Op.getDebugLoc();
4841 SDValue N0 = Op.getOperand(0);
4842 SDValue N1 = Op.getOperand(1);
4843 SDValue N2 = Op.getOperand(2);
4845 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4846 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4847 // as its second argument.
4848 if (N1.getValueType() != MVT::i32)
4849 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4850 if (N2.getValueType() != MVT::i32)
4851 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4852 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4858 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4859 DebugLoc dl = Op.getDebugLoc();
4860 if (Op.getValueType() == MVT::v2f32)
4861 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4862 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4863 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4864 Op.getOperand(0))));
4866 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4867 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4869 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4870 EVT VT = MVT::v2i32;
4871 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4878 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4879 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4882 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4883 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4884 // one of the above mentioned nodes. It has to be wrapped because otherwise
4885 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4886 // be used to form addressing mode. These wrapped nodes will be selected
4889 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4890 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4892 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4894 unsigned char OpFlag = 0;
4895 unsigned WrapperKind = X86ISD::Wrapper;
4896 CodeModel::Model M = getTargetMachine().getCodeModel();
4898 if (Subtarget->isPICStyleRIPRel() &&
4899 (M == CodeModel::Small || M == CodeModel::Kernel))
4900 WrapperKind = X86ISD::WrapperRIP;
4901 else if (Subtarget->isPICStyleGOT())
4902 OpFlag = X86II::MO_GOTOFF;
4903 else if (Subtarget->isPICStyleStubPIC())
4904 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4906 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4908 CP->getOffset(), OpFlag);
4909 DebugLoc DL = CP->getDebugLoc();
4910 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4911 // With PIC, the address is actually $g + Offset.
4913 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4914 DAG.getNode(X86ISD::GlobalBaseReg,
4915 DebugLoc::getUnknownLoc(), getPointerTy()),
4922 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4923 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4925 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4927 unsigned char OpFlag = 0;
4928 unsigned WrapperKind = X86ISD::Wrapper;
4929 CodeModel::Model M = getTargetMachine().getCodeModel();
4931 if (Subtarget->isPICStyleRIPRel() &&
4932 (M == CodeModel::Small || M == CodeModel::Kernel))
4933 WrapperKind = X86ISD::WrapperRIP;
4934 else if (Subtarget->isPICStyleGOT())
4935 OpFlag = X86II::MO_GOTOFF;
4936 else if (Subtarget->isPICStyleStubPIC())
4937 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4939 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4941 DebugLoc DL = JT->getDebugLoc();
4942 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4944 // With PIC, the address is actually $g + Offset.
4946 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4947 DAG.getNode(X86ISD::GlobalBaseReg,
4948 DebugLoc::getUnknownLoc(), getPointerTy()),
4956 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4957 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4959 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4961 unsigned char OpFlag = 0;
4962 unsigned WrapperKind = X86ISD::Wrapper;
4963 CodeModel::Model M = getTargetMachine().getCodeModel();
4965 if (Subtarget->isPICStyleRIPRel() &&
4966 (M == CodeModel::Small || M == CodeModel::Kernel))
4967 WrapperKind = X86ISD::WrapperRIP;
4968 else if (Subtarget->isPICStyleGOT())
4969 OpFlag = X86II::MO_GOTOFF;
4970 else if (Subtarget->isPICStyleStubPIC())
4971 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4973 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4975 DebugLoc DL = Op.getDebugLoc();
4976 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4979 // With PIC, the address is actually $g + Offset.
4980 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4981 !Subtarget->is64Bit()) {
4982 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4983 DAG.getNode(X86ISD::GlobalBaseReg,
4984 DebugLoc::getUnknownLoc(),
4993 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
4994 // Create the TargetBlockAddressAddress node.
4995 unsigned char OpFlags =
4996 Subtarget->ClassifyBlockAddressReference();
4997 CodeModel::Model M = getTargetMachine().getCodeModel();
4998 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
4999 DebugLoc dl = Op.getDebugLoc();
5000 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5001 /*isTarget=*/true, OpFlags);
5003 if (Subtarget->isPICStyleRIPRel() &&
5004 (M == CodeModel::Small || M == CodeModel::Kernel))
5005 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5007 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5009 // With PIC, the address is actually $g + Offset.
5010 if (isGlobalRelativeToPICBase(OpFlags)) {
5011 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5012 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5020 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5022 SelectionDAG &DAG) const {
5023 // Create the TargetGlobalAddress node, folding in the constant
5024 // offset if it is legal.
5025 unsigned char OpFlags =
5026 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5027 CodeModel::Model M = getTargetMachine().getCodeModel();
5029 if (OpFlags == X86II::MO_NO_FLAG &&
5030 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5031 // A direct static reference to a global.
5032 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5035 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5038 if (Subtarget->isPICStyleRIPRel() &&
5039 (M == CodeModel::Small || M == CodeModel::Kernel))
5040 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5042 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5044 // With PIC, the address is actually $g + Offset.
5045 if (isGlobalRelativeToPICBase(OpFlags)) {
5046 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5047 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5051 // For globals that require a load from a stub to get the address, emit the
5053 if (isGlobalStubReference(OpFlags))
5054 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5055 PseudoSourceValue::getGOT(), 0);
5057 // If there was a non-zero offset that we didn't fold, create an explicit
5060 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5061 DAG.getConstant(Offset, getPointerTy()));
5067 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
5068 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5069 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5070 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5074 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5075 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5076 unsigned char OperandFlags) {
5077 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5078 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5079 DebugLoc dl = GA->getDebugLoc();
5080 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5081 GA->getValueType(0),
5085 SDValue Ops[] = { Chain, TGA, *InFlag };
5086 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5088 SDValue Ops[] = { Chain, TGA };
5089 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5092 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5093 MFI->setHasCalls(true);
5095 SDValue Flag = Chain.getValue(1);
5096 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5099 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5101 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5104 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5105 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5106 DAG.getNode(X86ISD::GlobalBaseReg,
5107 DebugLoc::getUnknownLoc(),
5109 InFlag = Chain.getValue(1);
5111 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5114 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5116 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5118 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5119 X86::RAX, X86II::MO_TLSGD);
5122 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5123 // "local exec" model.
5124 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5125 const EVT PtrVT, TLSModel::Model model,
5127 DebugLoc dl = GA->getDebugLoc();
5128 // Get the Thread Pointer
5129 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5130 DebugLoc::getUnknownLoc(), PtrVT,
5131 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5134 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5137 unsigned char OperandFlags = 0;
5138 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5140 unsigned WrapperKind = X86ISD::Wrapper;
5141 if (model == TLSModel::LocalExec) {
5142 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5143 } else if (is64Bit) {
5144 assert(model == TLSModel::InitialExec);
5145 OperandFlags = X86II::MO_GOTTPOFF;
5146 WrapperKind = X86ISD::WrapperRIP;
5148 assert(model == TLSModel::InitialExec);
5149 OperandFlags = X86II::MO_INDNTPOFF;
5152 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5154 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5155 GA->getOffset(), OperandFlags);
5156 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5158 if (model == TLSModel::InitialExec)
5159 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5160 PseudoSourceValue::getGOT(), 0);
5162 // The address of the thread local variable is the add of the thread
5163 // pointer with the offset of the variable.
5164 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5168 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
5169 // TODO: implement the "local dynamic" model
5170 // TODO: implement the "initial exec"model for pic executables
5171 assert(Subtarget->isTargetELF() &&
5172 "TLS not implemented for non-ELF targets");
5173 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5174 const GlobalValue *GV = GA->getGlobal();
5176 // If GV is an alias then use the aliasee for determining
5177 // thread-localness.
5178 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5179 GV = GA->resolveAliasedGlobal(false);
5181 TLSModel::Model model = getTLSModel(GV,
5182 getTargetMachine().getRelocationModel());
5185 case TLSModel::GeneralDynamic:
5186 case TLSModel::LocalDynamic: // not implemented
5187 if (Subtarget->is64Bit())
5188 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5189 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5191 case TLSModel::InitialExec:
5192 case TLSModel::LocalExec:
5193 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5194 Subtarget->is64Bit());
5197 llvm_unreachable("Unreachable");
5202 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5203 /// take a 2 x i32 value to shift plus a shift amount.
5204 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
5205 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5206 EVT VT = Op.getValueType();
5207 unsigned VTBits = VT.getSizeInBits();
5208 DebugLoc dl = Op.getDebugLoc();
5209 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5210 SDValue ShOpLo = Op.getOperand(0);
5211 SDValue ShOpHi = Op.getOperand(1);
5212 SDValue ShAmt = Op.getOperand(2);
5213 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5214 DAG.getConstant(VTBits - 1, MVT::i8))
5215 : DAG.getConstant(0, VT);
5218 if (Op.getOpcode() == ISD::SHL_PARTS) {
5219 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5220 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5222 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5223 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5226 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5227 DAG.getConstant(VTBits, MVT::i8));
5228 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
5229 AndNode, DAG.getConstant(0, MVT::i8));
5232 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5233 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5234 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5236 if (Op.getOpcode() == ISD::SHL_PARTS) {
5237 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5238 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5240 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5241 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5244 SDValue Ops[2] = { Lo, Hi };
5245 return DAG.getMergeValues(Ops, 2, dl);
5248 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5249 EVT SrcVT = Op.getOperand(0).getValueType();
5251 if (SrcVT.isVector()) {
5252 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5258 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5259 "Unknown SINT_TO_FP to lower!");
5261 // These are really Legal; return the operand so the caller accepts it as
5263 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5265 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5266 Subtarget->is64Bit()) {
5270 DebugLoc dl = Op.getDebugLoc();
5271 unsigned Size = SrcVT.getSizeInBits()/8;
5272 MachineFunction &MF = DAG.getMachineFunction();
5273 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5274 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5275 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5277 PseudoSourceValue::getFixedStack(SSFI), 0);
5278 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5281 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5283 SelectionDAG &DAG) {
5285 DebugLoc dl = Op.getDebugLoc();
5287 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5289 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5291 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5292 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5293 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5294 Tys, Ops, array_lengthof(Ops));
5297 Chain = Result.getValue(1);
5298 SDValue InFlag = Result.getValue(2);
5300 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5301 // shouldn't be necessary except that RFP cannot be live across
5302 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5303 MachineFunction &MF = DAG.getMachineFunction();
5304 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5305 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5306 Tys = DAG.getVTList(MVT::Other);
5308 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5310 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5311 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5312 PseudoSourceValue::getFixedStack(SSFI), 0);
5318 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5319 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5320 // This algorithm is not obvious. Here it is in C code, more or less:
5322 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5323 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5324 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5326 // Copy ints to xmm registers.
5327 __m128i xh = _mm_cvtsi32_si128( hi );
5328 __m128i xl = _mm_cvtsi32_si128( lo );
5330 // Combine into low half of a single xmm register.
5331 __m128i x = _mm_unpacklo_epi32( xh, xl );
5335 // Merge in appropriate exponents to give the integer bits the right
5337 x = _mm_unpacklo_epi32( x, exp );
5339 // Subtract away the biases to deal with the IEEE-754 double precision
5341 d = _mm_sub_pd( (__m128d) x, bias );
5343 // All conversions up to here are exact. The correctly rounded result is
5344 // calculated using the current rounding mode using the following
5346 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5347 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5348 // store doesn't really need to be here (except
5349 // maybe to zero the other double)
5354 DebugLoc dl = Op.getDebugLoc();
5355 LLVMContext *Context = DAG.getContext();
5357 // Build some magic constants.
5358 std::vector<Constant*> CV0;
5359 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5360 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5361 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5362 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5363 Constant *C0 = ConstantVector::get(CV0);
5364 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5366 std::vector<Constant*> CV1;
5368 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5370 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5371 Constant *C1 = ConstantVector::get(CV1);
5372 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5374 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5375 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5377 DAG.getIntPtrConstant(1)));
5378 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5379 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5381 DAG.getIntPtrConstant(0)));
5382 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5383 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5384 PseudoSourceValue::getConstantPool(), 0,
5386 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5387 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5388 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5389 PseudoSourceValue::getConstantPool(), 0,
5391 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5393 // Add the halves; easiest way is to swap them into another reg first.
5394 int ShufMask[2] = { 1, -1 };
5395 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5396 DAG.getUNDEF(MVT::v2f64), ShufMask);
5397 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5398 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5399 DAG.getIntPtrConstant(0));
5402 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5403 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5404 DebugLoc dl = Op.getDebugLoc();
5405 // FP constant to bias correct the final result.
5406 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5409 // Load the 32-bit value into an XMM register.
5410 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5411 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5413 DAG.getIntPtrConstant(0)));
5415 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5416 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5417 DAG.getIntPtrConstant(0));
5419 // Or the load with the bias.
5420 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5421 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5422 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5424 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5425 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5426 MVT::v2f64, Bias)));
5427 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5428 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5429 DAG.getIntPtrConstant(0));
5431 // Subtract the bias.
5432 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5434 // Handle final rounding.
5435 EVT DestVT = Op.getValueType();
5437 if (DestVT.bitsLT(MVT::f64)) {
5438 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5439 DAG.getIntPtrConstant(0));
5440 } else if (DestVT.bitsGT(MVT::f64)) {
5441 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5444 // Handle final rounding.
5448 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5449 SDValue N0 = Op.getOperand(0);
5450 DebugLoc dl = Op.getDebugLoc();
5452 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5453 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5454 // the optimization here.
5455 if (DAG.SignBitIsZero(N0))
5456 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5458 EVT SrcVT = N0.getValueType();
5459 if (SrcVT == MVT::i64) {
5460 // We only handle SSE2 f64 target here; caller can expand the rest.
5461 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5464 return LowerUINT_TO_FP_i64(Op, DAG);
5465 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5466 return LowerUINT_TO_FP_i32(Op, DAG);
5469 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5471 // Make a 64-bit buffer, and use it to build an FILD.
5472 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5473 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5474 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5475 getPointerTy(), StackSlot, WordOff);
5476 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5477 StackSlot, NULL, 0);
5478 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5479 OffsetSlot, NULL, 0);
5480 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5483 std::pair<SDValue,SDValue> X86TargetLowering::
5484 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5485 DebugLoc dl = Op.getDebugLoc();
5487 EVT DstTy = Op.getValueType();
5490 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5494 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5495 DstTy.getSimpleVT() >= MVT::i16 &&
5496 "Unknown FP_TO_SINT to lower!");
5498 // These are really Legal.
5499 if (DstTy == MVT::i32 &&
5500 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5501 return std::make_pair(SDValue(), SDValue());
5502 if (Subtarget->is64Bit() &&
5503 DstTy == MVT::i64 &&
5504 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5505 return std::make_pair(SDValue(), SDValue());
5507 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5509 MachineFunction &MF = DAG.getMachineFunction();
5510 unsigned MemSize = DstTy.getSizeInBits()/8;
5511 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5512 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5515 switch (DstTy.getSimpleVT().SimpleTy) {
5516 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5517 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5518 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5519 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5522 SDValue Chain = DAG.getEntryNode();
5523 SDValue Value = Op.getOperand(0);
5524 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5525 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5526 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5527 PseudoSourceValue::getFixedStack(SSFI), 0);
5528 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5530 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5532 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5533 Chain = Value.getValue(1);
5534 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5535 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5538 // Build the FP_TO_INT*_IN_MEM
5539 SDValue Ops[] = { Chain, Value, StackSlot };
5540 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5542 return std::make_pair(FIST, StackSlot);
5545 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5546 if (Op.getValueType().isVector()) {
5547 if (Op.getValueType() == MVT::v2i32 &&
5548 Op.getOperand(0).getValueType() == MVT::v2f64) {
5554 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5555 SDValue FIST = Vals.first, StackSlot = Vals.second;
5556 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5557 if (FIST.getNode() == 0) return Op;
5560 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5561 FIST, StackSlot, NULL, 0);
5564 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5565 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5566 SDValue FIST = Vals.first, StackSlot = Vals.second;
5567 assert(FIST.getNode() && "Unexpected failure");
5570 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5571 FIST, StackSlot, NULL, 0);
5574 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5575 LLVMContext *Context = DAG.getContext();
5576 DebugLoc dl = Op.getDebugLoc();
5577 EVT VT = Op.getValueType();
5580 EltVT = VT.getVectorElementType();
5581 std::vector<Constant*> CV;
5582 if (EltVT == MVT::f64) {
5583 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5587 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5593 Constant *C = ConstantVector::get(CV);
5594 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5595 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5596 PseudoSourceValue::getConstantPool(), 0,
5598 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5601 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5602 LLVMContext *Context = DAG.getContext();
5603 DebugLoc dl = Op.getDebugLoc();
5604 EVT VT = Op.getValueType();
5607 EltVT = VT.getVectorElementType();
5608 std::vector<Constant*> CV;
5609 if (EltVT == MVT::f64) {
5610 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5614 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5620 Constant *C = ConstantVector::get(CV);
5621 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5622 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5623 PseudoSourceValue::getConstantPool(), 0,
5625 if (VT.isVector()) {
5626 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5627 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5628 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5630 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5632 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5636 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5637 LLVMContext *Context = DAG.getContext();
5638 SDValue Op0 = Op.getOperand(0);
5639 SDValue Op1 = Op.getOperand(1);
5640 DebugLoc dl = Op.getDebugLoc();
5641 EVT VT = Op.getValueType();
5642 EVT SrcVT = Op1.getValueType();
5644 // If second operand is smaller, extend it first.
5645 if (SrcVT.bitsLT(VT)) {
5646 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5649 // And if it is bigger, shrink it first.
5650 if (SrcVT.bitsGT(VT)) {
5651 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5655 // At this point the operands and the result should have the same
5656 // type, and that won't be f80 since that is not custom lowered.
5658 // First get the sign bit of second operand.
5659 std::vector<Constant*> CV;
5660 if (SrcVT == MVT::f64) {
5661 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5662 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5664 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5665 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5666 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5667 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5669 Constant *C = ConstantVector::get(CV);
5670 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5671 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5672 PseudoSourceValue::getConstantPool(), 0,
5674 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5676 // Shift sign bit right or left if the two operands have different types.
5677 if (SrcVT.bitsGT(VT)) {
5678 // Op0 is MVT::f32, Op1 is MVT::f64.
5679 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5680 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5681 DAG.getConstant(32, MVT::i32));
5682 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5683 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5684 DAG.getIntPtrConstant(0));
5687 // Clear first operand sign bit.
5689 if (VT == MVT::f64) {
5690 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5691 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5693 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5694 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5695 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5696 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5698 C = ConstantVector::get(CV);
5699 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5700 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5701 PseudoSourceValue::getConstantPool(), 0,
5703 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5705 // Or the value with the sign bit.
5706 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5709 /// Emit nodes that will be selected as "test Op0,Op0", or something
5711 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5712 SelectionDAG &DAG) {
5713 DebugLoc dl = Op.getDebugLoc();
5715 // CF and OF aren't always set the way we want. Determine which
5716 // of these we need.
5717 bool NeedCF = false;
5718 bool NeedOF = false;
5720 case X86::COND_A: case X86::COND_AE:
5721 case X86::COND_B: case X86::COND_BE:
5724 case X86::COND_G: case X86::COND_GE:
5725 case X86::COND_L: case X86::COND_LE:
5726 case X86::COND_O: case X86::COND_NO:
5732 // See if we can use the EFLAGS value from the operand instead of
5733 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5734 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5735 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5736 unsigned Opcode = 0;
5737 unsigned NumOperands = 0;
5738 switch (Op.getNode()->getOpcode()) {
5740 // Due to an isel shortcoming, be conservative if this add is likely to
5741 // be selected as part of a load-modify-store instruction. When the root
5742 // node in a match is a store, isel doesn't know how to remap non-chain
5743 // non-flag uses of other nodes in the match, such as the ADD in this
5744 // case. This leads to the ADD being left around and reselected, with
5745 // the result being two adds in the output.
5746 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5747 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5748 if (UI->getOpcode() == ISD::STORE)
5750 if (ConstantSDNode *C =
5751 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5752 // An add of one will be selected as an INC.
5753 if (C->getAPIntValue() == 1) {
5754 Opcode = X86ISD::INC;
5758 // An add of negative one (subtract of one) will be selected as a DEC.
5759 if (C->getAPIntValue().isAllOnesValue()) {
5760 Opcode = X86ISD::DEC;
5765 // Otherwise use a regular EFLAGS-setting add.
5766 Opcode = X86ISD::ADD;
5770 // If the primary and result isn't used, don't bother using X86ISD::AND,
5771 // because a TEST instruction will be better.
5772 bool NonFlagUse = false;
5773 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5774 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
5776 unsigned UOpNo = UI.getOperandNo();
5777 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
5778 // Look pass truncate.
5779 UOpNo = User->use_begin().getOperandNo();
5780 User = *User->use_begin();
5782 if (User->getOpcode() != ISD::BRCOND &&
5783 User->getOpcode() != ISD::SETCC &&
5784 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
5796 // Due to the ISEL shortcoming noted above, be conservative if this op is
5797 // likely to be selected as part of a load-modify-store instruction.
5798 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5799 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5800 if (UI->getOpcode() == ISD::STORE)
5802 // Otherwise use a regular EFLAGS-setting instruction.
5803 switch (Op.getNode()->getOpcode()) {
5804 case ISD::SUB: Opcode = X86ISD::SUB; break;
5805 case ISD::OR: Opcode = X86ISD::OR; break;
5806 case ISD::XOR: Opcode = X86ISD::XOR; break;
5807 case ISD::AND: Opcode = X86ISD::AND; break;
5808 default: llvm_unreachable("unexpected operator!");
5819 return SDValue(Op.getNode(), 1);
5825 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5826 SmallVector<SDValue, 4> Ops;
5827 for (unsigned i = 0; i != NumOperands; ++i)
5828 Ops.push_back(Op.getOperand(i));
5829 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5830 DAG.ReplaceAllUsesWith(Op, New);
5831 return SDValue(New.getNode(), 1);
5835 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5836 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5837 DAG.getConstant(0, Op.getValueType()));
5840 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5842 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5843 SelectionDAG &DAG) {
5844 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5845 if (C->getAPIntValue() == 0)
5846 return EmitTest(Op0, X86CC, DAG);
5848 DebugLoc dl = Op0.getDebugLoc();
5849 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5852 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
5853 /// if it's possible.
5854 static SDValue LowerToBT(SDValue Op0, ISD::CondCode CC,
5855 DebugLoc dl, SelectionDAG &DAG) {
5857 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5858 if (ConstantSDNode *Op010C =
5859 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5860 if (Op010C->getZExtValue() == 1) {
5861 LHS = Op0.getOperand(0);
5862 RHS = Op0.getOperand(1).getOperand(1);
5864 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5865 if (ConstantSDNode *Op000C =
5866 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5867 if (Op000C->getZExtValue() == 1) {
5868 LHS = Op0.getOperand(1);
5869 RHS = Op0.getOperand(0).getOperand(1);
5871 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5872 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5873 SDValue AndLHS = Op0.getOperand(0);
5874 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5875 LHS = AndLHS.getOperand(0);
5876 RHS = AndLHS.getOperand(1);
5880 if (LHS.getNode()) {
5881 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5882 // instruction. Since the shift amount is in-range-or-undefined, we know
5883 // that doing a bittest on the i16 value is ok. We extend to i32 because
5884 // the encoding for the i16 version is larger than the i32 version.
5885 if (LHS.getValueType() == MVT::i8)
5886 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5888 // If the operand types disagree, extend the shift amount to match. Since
5889 // BT ignores high bits (like shifts) we can use anyextend.
5890 if (LHS.getValueType() != RHS.getValueType())
5891 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5893 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5894 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5895 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5896 DAG.getConstant(Cond, MVT::i8), BT);
5902 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5903 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5904 SDValue Op0 = Op.getOperand(0);
5905 SDValue Op1 = Op.getOperand(1);
5906 DebugLoc dl = Op.getDebugLoc();
5907 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5909 // Optimize to BT if possible.
5910 // Lower (X & (1 << N)) == 0 to BT(X, N).
5911 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5912 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5913 if (Op0.getOpcode() == ISD::AND &&
5915 Op1.getOpcode() == ISD::Constant &&
5916 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5917 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5918 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
5919 if (NewSetCC.getNode())
5923 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5924 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5925 if (X86CC == X86::COND_INVALID)
5928 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5930 // Use sbb x, x to materialize carry bit into a GPR.
5931 if (X86CC == X86::COND_B)
5932 return DAG.getNode(ISD::AND, dl, MVT::i8,
5933 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
5934 DAG.getConstant(X86CC, MVT::i8), Cond),
5935 DAG.getConstant(1, MVT::i8));
5937 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5938 DAG.getConstant(X86CC, MVT::i8), Cond);
5941 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5943 SDValue Op0 = Op.getOperand(0);
5944 SDValue Op1 = Op.getOperand(1);
5945 SDValue CC = Op.getOperand(2);
5946 EVT VT = Op.getValueType();
5947 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5948 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5949 DebugLoc dl = Op.getDebugLoc();
5953 EVT VT0 = Op0.getValueType();
5954 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5955 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5958 switch (SetCCOpcode) {
5961 case ISD::SETEQ: SSECC = 0; break;
5963 case ISD::SETGT: Swap = true; // Fallthrough
5965 case ISD::SETOLT: SSECC = 1; break;
5967 case ISD::SETGE: Swap = true; // Fallthrough
5969 case ISD::SETOLE: SSECC = 2; break;
5970 case ISD::SETUO: SSECC = 3; break;
5972 case ISD::SETNE: SSECC = 4; break;
5973 case ISD::SETULE: Swap = true;
5974 case ISD::SETUGE: SSECC = 5; break;
5975 case ISD::SETULT: Swap = true;
5976 case ISD::SETUGT: SSECC = 6; break;
5977 case ISD::SETO: SSECC = 7; break;
5980 std::swap(Op0, Op1);
5982 // In the two special cases we can't handle, emit two comparisons.
5984 if (SetCCOpcode == ISD::SETUEQ) {
5986 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5987 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5988 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5990 else if (SetCCOpcode == ISD::SETONE) {
5992 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5993 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5994 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5996 llvm_unreachable("Illegal FP comparison");
5998 // Handle all other FP comparisons here.
5999 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6002 // We are handling one of the integer comparisons here. Since SSE only has
6003 // GT and EQ comparisons for integer, swapping operands and multiple
6004 // operations may be required for some comparisons.
6005 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6006 bool Swap = false, Invert = false, FlipSigns = false;
6008 switch (VT.getSimpleVT().SimpleTy) {
6011 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6013 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6015 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6016 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6019 switch (SetCCOpcode) {
6021 case ISD::SETNE: Invert = true;
6022 case ISD::SETEQ: Opc = EQOpc; break;
6023 case ISD::SETLT: Swap = true;
6024 case ISD::SETGT: Opc = GTOpc; break;
6025 case ISD::SETGE: Swap = true;
6026 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6027 case ISD::SETULT: Swap = true;
6028 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6029 case ISD::SETUGE: Swap = true;
6030 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6033 std::swap(Op0, Op1);
6035 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6036 // bits of the inputs before performing those operations.
6038 EVT EltVT = VT.getVectorElementType();
6039 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6041 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6042 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6044 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6045 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6048 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6050 // If the logical-not of the result is required, perform that now.
6052 Result = DAG.getNOT(dl, Result, VT);
6057 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6058 static bool isX86LogicalCmp(SDValue Op) {
6059 unsigned Opc = Op.getNode()->getOpcode();
6060 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6062 if (Op.getResNo() == 1 &&
6063 (Opc == X86ISD::ADD ||
6064 Opc == X86ISD::SUB ||
6065 Opc == X86ISD::SMUL ||
6066 Opc == X86ISD::UMUL ||
6067 Opc == X86ISD::INC ||
6068 Opc == X86ISD::DEC ||
6069 Opc == X86ISD::OR ||
6070 Opc == X86ISD::XOR ||
6071 Opc == X86ISD::AND))
6077 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
6078 bool addTest = true;
6079 SDValue Cond = Op.getOperand(0);
6080 DebugLoc dl = Op.getDebugLoc();
6083 if (Cond.getOpcode() == ISD::SETCC) {
6084 SDValue NewCond = LowerSETCC(Cond, DAG);
6085 if (NewCond.getNode())
6089 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6090 SDValue Op1 = Op.getOperand(1);
6091 SDValue Op2 = Op.getOperand(2);
6092 if (Cond.getOpcode() == X86ISD::SETCC &&
6093 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6094 SDValue Cmp = Cond.getOperand(1);
6095 if (Cmp.getOpcode() == X86ISD::CMP) {
6096 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6097 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6098 ConstantSDNode *RHSC =
6099 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6100 if (N1C && N1C->isAllOnesValue() &&
6101 N2C && N2C->isNullValue() &&
6102 RHSC && RHSC->isNullValue()) {
6103 SDValue CmpOp0 = Cmp.getOperand(0);
6104 Cmp = DAG.getNode(X86ISD::CMP, dl, CmpOp0.getValueType(),
6105 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6106 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6107 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6112 // Look pass (and (setcc_carry (cmp ...)), 1).
6113 if (Cond.getOpcode() == ISD::AND &&
6114 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6115 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6116 if (C && C->getAPIntValue() == 1)
6117 Cond = Cond.getOperand(0);
6120 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6121 // setting operand in place of the X86ISD::SETCC.
6122 if (Cond.getOpcode() == X86ISD::SETCC ||
6123 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6124 CC = Cond.getOperand(0);
6126 SDValue Cmp = Cond.getOperand(1);
6127 unsigned Opc = Cmp.getOpcode();
6128 EVT VT = Op.getValueType();
6130 bool IllegalFPCMov = false;
6131 if (VT.isFloatingPoint() && !VT.isVector() &&
6132 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6133 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6135 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6136 Opc == X86ISD::BT) { // FIXME
6143 // Look pass the truncate.
6144 if (Cond.getOpcode() == ISD::TRUNCATE)
6145 Cond = Cond.getOperand(0);
6147 // We know the result of AND is compared against zero. Try to match
6149 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6150 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6151 if (NewSetCC.getNode()) {
6152 CC = NewSetCC.getOperand(0);
6153 Cond = NewSetCC.getOperand(1);
6160 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6161 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6164 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6165 // condition is true.
6166 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6167 SDValue Ops[] = { Op2, Op1, CC, Cond };
6168 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6171 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6172 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6173 // from the AND / OR.
6174 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6175 Opc = Op.getOpcode();
6176 if (Opc != ISD::OR && Opc != ISD::AND)
6178 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6179 Op.getOperand(0).hasOneUse() &&
6180 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6181 Op.getOperand(1).hasOneUse());
6184 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6185 // 1 and that the SETCC node has a single use.
6186 static bool isXor1OfSetCC(SDValue Op) {
6187 if (Op.getOpcode() != ISD::XOR)
6189 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6190 if (N1C && N1C->getAPIntValue() == 1) {
6191 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6192 Op.getOperand(0).hasOneUse();
6197 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
6198 bool addTest = true;
6199 SDValue Chain = Op.getOperand(0);
6200 SDValue Cond = Op.getOperand(1);
6201 SDValue Dest = Op.getOperand(2);
6202 DebugLoc dl = Op.getDebugLoc();
6205 if (Cond.getOpcode() == ISD::SETCC) {
6206 SDValue NewCond = LowerSETCC(Cond, DAG);
6207 if (NewCond.getNode())
6211 // FIXME: LowerXALUO doesn't handle these!!
6212 else if (Cond.getOpcode() == X86ISD::ADD ||
6213 Cond.getOpcode() == X86ISD::SUB ||
6214 Cond.getOpcode() == X86ISD::SMUL ||
6215 Cond.getOpcode() == X86ISD::UMUL)
6216 Cond = LowerXALUO(Cond, DAG);
6219 // Look pass (and (setcc_carry (cmp ...)), 1).
6220 if (Cond.getOpcode() == ISD::AND &&
6221 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6222 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6223 if (C && C->getAPIntValue() == 1)
6224 Cond = Cond.getOperand(0);
6227 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6228 // setting operand in place of the X86ISD::SETCC.
6229 if (Cond.getOpcode() == X86ISD::SETCC ||
6230 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6231 CC = Cond.getOperand(0);
6233 SDValue Cmp = Cond.getOperand(1);
6234 unsigned Opc = Cmp.getOpcode();
6235 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6236 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6240 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6244 // These can only come from an arithmetic instruction with overflow,
6245 // e.g. SADDO, UADDO.
6246 Cond = Cond.getNode()->getOperand(1);
6253 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6254 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6255 if (CondOpc == ISD::OR) {
6256 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6257 // two branches instead of an explicit OR instruction with a
6259 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6260 isX86LogicalCmp(Cmp)) {
6261 CC = Cond.getOperand(0).getOperand(0);
6262 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6263 Chain, Dest, CC, Cmp);
6264 CC = Cond.getOperand(1).getOperand(0);
6268 } else { // ISD::AND
6269 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6270 // two branches instead of an explicit AND instruction with a
6271 // separate test. However, we only do this if this block doesn't
6272 // have a fall-through edge, because this requires an explicit
6273 // jmp when the condition is false.
6274 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6275 isX86LogicalCmp(Cmp) &&
6276 Op.getNode()->hasOneUse()) {
6277 X86::CondCode CCode =
6278 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6279 CCode = X86::GetOppositeBranchCondition(CCode);
6280 CC = DAG.getConstant(CCode, MVT::i8);
6281 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6282 // Look for an unconditional branch following this conditional branch.
6283 // We need this because we need to reverse the successors in order
6284 // to implement FCMP_OEQ.
6285 if (User.getOpcode() == ISD::BR) {
6286 SDValue FalseBB = User.getOperand(1);
6288 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6289 assert(NewBR == User);
6292 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6293 Chain, Dest, CC, Cmp);
6294 X86::CondCode CCode =
6295 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6296 CCode = X86::GetOppositeBranchCondition(CCode);
6297 CC = DAG.getConstant(CCode, MVT::i8);
6303 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6304 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6305 // It should be transformed during dag combiner except when the condition
6306 // is set by a arithmetics with overflow node.
6307 X86::CondCode CCode =
6308 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6309 CCode = X86::GetOppositeBranchCondition(CCode);
6310 CC = DAG.getConstant(CCode, MVT::i8);
6311 Cond = Cond.getOperand(0).getOperand(1);
6317 // Look pass the truncate.
6318 if (Cond.getOpcode() == ISD::TRUNCATE)
6319 Cond = Cond.getOperand(0);
6321 // We know the result of AND is compared against zero. Try to match
6323 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6324 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6325 if (NewSetCC.getNode()) {
6326 CC = NewSetCC.getOperand(0);
6327 Cond = NewSetCC.getOperand(1);
6334 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6335 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6337 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6338 Chain, Dest, CC, Cond);
6342 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6343 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6344 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6345 // that the guard pages used by the OS virtual memory manager are allocated in
6346 // correct sequence.
6348 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6349 SelectionDAG &DAG) {
6350 assert(Subtarget->isTargetCygMing() &&
6351 "This should be used only on Cygwin/Mingw targets");
6352 DebugLoc dl = Op.getDebugLoc();
6355 SDValue Chain = Op.getOperand(0);
6356 SDValue Size = Op.getOperand(1);
6357 // FIXME: Ensure alignment here
6361 EVT IntPtr = getPointerTy();
6362 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6364 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
6366 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6367 Flag = Chain.getValue(1);
6369 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6370 SDValue Ops[] = { Chain,
6371 DAG.getTargetExternalSymbol("_alloca", IntPtr),
6372 DAG.getRegister(X86::EAX, IntPtr),
6373 DAG.getRegister(X86StackPtr, SPTy),
6375 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
6376 Flag = Chain.getValue(1);
6378 Chain = DAG.getCALLSEQ_END(Chain,
6379 DAG.getIntPtrConstant(0, true),
6380 DAG.getIntPtrConstant(0, true),
6383 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6385 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6386 return DAG.getMergeValues(Ops1, 2, dl);
6390 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6392 SDValue Dst, SDValue Src,
6393 SDValue Size, unsigned Align,
6395 uint64_t DstSVOff) {
6396 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6398 // If not DWORD aligned or size is more than the threshold, call the library.
6399 // The libc version is likely to be faster for these cases. It can use the
6400 // address value and run time information about the CPU.
6401 if ((Align & 3) != 0 ||
6403 ConstantSize->getZExtValue() >
6404 getSubtarget()->getMaxInlineSizeThreshold()) {
6405 SDValue InFlag(0, 0);
6407 // Check to see if there is a specialized entry-point for memory zeroing.
6408 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6410 if (const char *bzeroEntry = V &&
6411 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6412 EVT IntPtr = getPointerTy();
6413 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6414 TargetLowering::ArgListTy Args;
6415 TargetLowering::ArgListEntry Entry;
6417 Entry.Ty = IntPtrTy;
6418 Args.push_back(Entry);
6420 Args.push_back(Entry);
6421 std::pair<SDValue,SDValue> CallResult =
6422 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6423 false, false, false, false,
6424 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6425 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl,
6426 DAG.GetOrdering(Chain.getNode()));
6427 return CallResult.second;
6430 // Otherwise have the target-independent code call memset.
6434 uint64_t SizeVal = ConstantSize->getZExtValue();
6435 SDValue InFlag(0, 0);
6438 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6439 unsigned BytesLeft = 0;
6440 bool TwoRepStos = false;
6443 uint64_t Val = ValC->getZExtValue() & 255;
6445 // If the value is a constant, then we can potentially use larger sets.
6446 switch (Align & 3) {
6447 case 2: // WORD aligned
6450 Val = (Val << 8) | Val;
6452 case 0: // DWORD aligned
6455 Val = (Val << 8) | Val;
6456 Val = (Val << 16) | Val;
6457 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6460 Val = (Val << 32) | Val;
6463 default: // Byte aligned
6466 Count = DAG.getIntPtrConstant(SizeVal);
6470 if (AVT.bitsGT(MVT::i8)) {
6471 unsigned UBytes = AVT.getSizeInBits() / 8;
6472 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6473 BytesLeft = SizeVal % UBytes;
6476 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6478 InFlag = Chain.getValue(1);
6481 Count = DAG.getIntPtrConstant(SizeVal);
6482 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6483 InFlag = Chain.getValue(1);
6486 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6489 InFlag = Chain.getValue(1);
6490 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6493 InFlag = Chain.getValue(1);
6495 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6496 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6497 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6500 InFlag = Chain.getValue(1);
6502 EVT CVT = Count.getValueType();
6503 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6504 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6505 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6508 InFlag = Chain.getValue(1);
6509 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6510 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
6511 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6512 } else if (BytesLeft) {
6513 // Handle the last 1 - 7 bytes.
6514 unsigned Offset = SizeVal - BytesLeft;
6515 EVT AddrVT = Dst.getValueType();
6516 EVT SizeVT = Size.getValueType();
6518 Chain = DAG.getMemset(Chain, dl,
6519 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6520 DAG.getConstant(Offset, AddrVT)),
6522 DAG.getConstant(BytesLeft, SizeVT),
6523 Align, DstSV, DstSVOff + Offset);
6526 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6531 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6532 SDValue Chain, SDValue Dst, SDValue Src,
6533 SDValue Size, unsigned Align,
6535 const Value *DstSV, uint64_t DstSVOff,
6536 const Value *SrcSV, uint64_t SrcSVOff) {
6537 // This requires the copy size to be a constant, preferrably
6538 // within a subtarget-specific limit.
6539 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6542 uint64_t SizeVal = ConstantSize->getZExtValue();
6543 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6546 /// If not DWORD aligned, call the library.
6547 if ((Align & 3) != 0)
6552 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6555 unsigned UBytes = AVT.getSizeInBits() / 8;
6556 unsigned CountVal = SizeVal / UBytes;
6557 SDValue Count = DAG.getIntPtrConstant(CountVal);
6558 unsigned BytesLeft = SizeVal % UBytes;
6560 SDValue InFlag(0, 0);
6561 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6564 InFlag = Chain.getValue(1);
6565 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6568 InFlag = Chain.getValue(1);
6569 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6572 InFlag = Chain.getValue(1);
6574 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6575 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6576 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
6577 array_lengthof(Ops));
6579 SmallVector<SDValue, 4> Results;
6580 Results.push_back(RepMovs);
6582 // Handle the last 1 - 7 bytes.
6583 unsigned Offset = SizeVal - BytesLeft;
6584 EVT DstVT = Dst.getValueType();
6585 EVT SrcVT = Src.getValueType();
6586 EVT SizeVT = Size.getValueType();
6587 Results.push_back(DAG.getMemcpy(Chain, dl,
6588 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6589 DAG.getConstant(Offset, DstVT)),
6590 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6591 DAG.getConstant(Offset, SrcVT)),
6592 DAG.getConstant(BytesLeft, SizeVT),
6593 Align, AlwaysInline,
6594 DstSV, DstSVOff + Offset,
6595 SrcSV, SrcSVOff + Offset));
6598 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6599 &Results[0], Results.size());
6602 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6603 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6604 DebugLoc dl = Op.getDebugLoc();
6606 if (!Subtarget->is64Bit()) {
6607 // vastart just stores the address of the VarArgsFrameIndex slot into the
6608 // memory location argument.
6609 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6610 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6614 // gp_offset (0 - 6 * 8)
6615 // fp_offset (48 - 48 + 8 * 16)
6616 // overflow_arg_area (point to parameters coming in memory).
6618 SmallVector<SDValue, 8> MemOps;
6619 SDValue FIN = Op.getOperand(1);
6621 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6622 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6624 MemOps.push_back(Store);
6627 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6628 FIN, DAG.getIntPtrConstant(4));
6629 Store = DAG.getStore(Op.getOperand(0), dl,
6630 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6632 MemOps.push_back(Store);
6634 // Store ptr to overflow_arg_area
6635 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6636 FIN, DAG.getIntPtrConstant(4));
6637 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6638 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6639 MemOps.push_back(Store);
6641 // Store ptr to reg_save_area.
6642 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6643 FIN, DAG.getIntPtrConstant(8));
6644 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6645 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6646 MemOps.push_back(Store);
6647 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6648 &MemOps[0], MemOps.size());
6651 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6652 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6653 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6654 SDValue Chain = Op.getOperand(0);
6655 SDValue SrcPtr = Op.getOperand(1);
6656 SDValue SrcSV = Op.getOperand(2);
6658 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6662 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6663 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6664 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6665 SDValue Chain = Op.getOperand(0);
6666 SDValue DstPtr = Op.getOperand(1);
6667 SDValue SrcPtr = Op.getOperand(2);
6668 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6669 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6670 DebugLoc dl = Op.getDebugLoc();
6672 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6673 DAG.getIntPtrConstant(24), 8, false,
6674 DstSV, 0, SrcSV, 0);
6678 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6679 DebugLoc dl = Op.getDebugLoc();
6680 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6682 default: return SDValue(); // Don't custom lower most intrinsics.
6683 // Comparison intrinsics.
6684 case Intrinsic::x86_sse_comieq_ss:
6685 case Intrinsic::x86_sse_comilt_ss:
6686 case Intrinsic::x86_sse_comile_ss:
6687 case Intrinsic::x86_sse_comigt_ss:
6688 case Intrinsic::x86_sse_comige_ss:
6689 case Intrinsic::x86_sse_comineq_ss:
6690 case Intrinsic::x86_sse_ucomieq_ss:
6691 case Intrinsic::x86_sse_ucomilt_ss:
6692 case Intrinsic::x86_sse_ucomile_ss:
6693 case Intrinsic::x86_sse_ucomigt_ss:
6694 case Intrinsic::x86_sse_ucomige_ss:
6695 case Intrinsic::x86_sse_ucomineq_ss:
6696 case Intrinsic::x86_sse2_comieq_sd:
6697 case Intrinsic::x86_sse2_comilt_sd:
6698 case Intrinsic::x86_sse2_comile_sd:
6699 case Intrinsic::x86_sse2_comigt_sd:
6700 case Intrinsic::x86_sse2_comige_sd:
6701 case Intrinsic::x86_sse2_comineq_sd:
6702 case Intrinsic::x86_sse2_ucomieq_sd:
6703 case Intrinsic::x86_sse2_ucomilt_sd:
6704 case Intrinsic::x86_sse2_ucomile_sd:
6705 case Intrinsic::x86_sse2_ucomigt_sd:
6706 case Intrinsic::x86_sse2_ucomige_sd:
6707 case Intrinsic::x86_sse2_ucomineq_sd: {
6709 ISD::CondCode CC = ISD::SETCC_INVALID;
6712 case Intrinsic::x86_sse_comieq_ss:
6713 case Intrinsic::x86_sse2_comieq_sd:
6717 case Intrinsic::x86_sse_comilt_ss:
6718 case Intrinsic::x86_sse2_comilt_sd:
6722 case Intrinsic::x86_sse_comile_ss:
6723 case Intrinsic::x86_sse2_comile_sd:
6727 case Intrinsic::x86_sse_comigt_ss:
6728 case Intrinsic::x86_sse2_comigt_sd:
6732 case Intrinsic::x86_sse_comige_ss:
6733 case Intrinsic::x86_sse2_comige_sd:
6737 case Intrinsic::x86_sse_comineq_ss:
6738 case Intrinsic::x86_sse2_comineq_sd:
6742 case Intrinsic::x86_sse_ucomieq_ss:
6743 case Intrinsic::x86_sse2_ucomieq_sd:
6744 Opc = X86ISD::UCOMI;
6747 case Intrinsic::x86_sse_ucomilt_ss:
6748 case Intrinsic::x86_sse2_ucomilt_sd:
6749 Opc = X86ISD::UCOMI;
6752 case Intrinsic::x86_sse_ucomile_ss:
6753 case Intrinsic::x86_sse2_ucomile_sd:
6754 Opc = X86ISD::UCOMI;
6757 case Intrinsic::x86_sse_ucomigt_ss:
6758 case Intrinsic::x86_sse2_ucomigt_sd:
6759 Opc = X86ISD::UCOMI;
6762 case Intrinsic::x86_sse_ucomige_ss:
6763 case Intrinsic::x86_sse2_ucomige_sd:
6764 Opc = X86ISD::UCOMI;
6767 case Intrinsic::x86_sse_ucomineq_ss:
6768 case Intrinsic::x86_sse2_ucomineq_sd:
6769 Opc = X86ISD::UCOMI;
6774 SDValue LHS = Op.getOperand(1);
6775 SDValue RHS = Op.getOperand(2);
6776 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6777 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6778 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6779 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6780 DAG.getConstant(X86CC, MVT::i8), Cond);
6781 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6783 // ptest intrinsics. The intrinsic these come from are designed to return
6784 // an integer value, not just an instruction so lower it to the ptest
6785 // pattern and a setcc for the result.
6786 case Intrinsic::x86_sse41_ptestz:
6787 case Intrinsic::x86_sse41_ptestc:
6788 case Intrinsic::x86_sse41_ptestnzc:{
6791 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6792 case Intrinsic::x86_sse41_ptestz:
6794 X86CC = X86::COND_E;
6796 case Intrinsic::x86_sse41_ptestc:
6798 X86CC = X86::COND_B;
6800 case Intrinsic::x86_sse41_ptestnzc:
6802 X86CC = X86::COND_A;
6806 SDValue LHS = Op.getOperand(1);
6807 SDValue RHS = Op.getOperand(2);
6808 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6809 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6810 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6811 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6814 // Fix vector shift instructions where the last operand is a non-immediate
6816 case Intrinsic::x86_sse2_pslli_w:
6817 case Intrinsic::x86_sse2_pslli_d:
6818 case Intrinsic::x86_sse2_pslli_q:
6819 case Intrinsic::x86_sse2_psrli_w:
6820 case Intrinsic::x86_sse2_psrli_d:
6821 case Intrinsic::x86_sse2_psrli_q:
6822 case Intrinsic::x86_sse2_psrai_w:
6823 case Intrinsic::x86_sse2_psrai_d:
6824 case Intrinsic::x86_mmx_pslli_w:
6825 case Intrinsic::x86_mmx_pslli_d:
6826 case Intrinsic::x86_mmx_pslli_q:
6827 case Intrinsic::x86_mmx_psrli_w:
6828 case Intrinsic::x86_mmx_psrli_d:
6829 case Intrinsic::x86_mmx_psrli_q:
6830 case Intrinsic::x86_mmx_psrai_w:
6831 case Intrinsic::x86_mmx_psrai_d: {
6832 SDValue ShAmt = Op.getOperand(2);
6833 if (isa<ConstantSDNode>(ShAmt))
6836 unsigned NewIntNo = 0;
6837 EVT ShAmtVT = MVT::v4i32;
6839 case Intrinsic::x86_sse2_pslli_w:
6840 NewIntNo = Intrinsic::x86_sse2_psll_w;
6842 case Intrinsic::x86_sse2_pslli_d:
6843 NewIntNo = Intrinsic::x86_sse2_psll_d;
6845 case Intrinsic::x86_sse2_pslli_q:
6846 NewIntNo = Intrinsic::x86_sse2_psll_q;
6848 case Intrinsic::x86_sse2_psrli_w:
6849 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6851 case Intrinsic::x86_sse2_psrli_d:
6852 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6854 case Intrinsic::x86_sse2_psrli_q:
6855 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6857 case Intrinsic::x86_sse2_psrai_w:
6858 NewIntNo = Intrinsic::x86_sse2_psra_w;
6860 case Intrinsic::x86_sse2_psrai_d:
6861 NewIntNo = Intrinsic::x86_sse2_psra_d;
6864 ShAmtVT = MVT::v2i32;
6866 case Intrinsic::x86_mmx_pslli_w:
6867 NewIntNo = Intrinsic::x86_mmx_psll_w;
6869 case Intrinsic::x86_mmx_pslli_d:
6870 NewIntNo = Intrinsic::x86_mmx_psll_d;
6872 case Intrinsic::x86_mmx_pslli_q:
6873 NewIntNo = Intrinsic::x86_mmx_psll_q;
6875 case Intrinsic::x86_mmx_psrli_w:
6876 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6878 case Intrinsic::x86_mmx_psrli_d:
6879 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6881 case Intrinsic::x86_mmx_psrli_q:
6882 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6884 case Intrinsic::x86_mmx_psrai_w:
6885 NewIntNo = Intrinsic::x86_mmx_psra_w;
6887 case Intrinsic::x86_mmx_psrai_d:
6888 NewIntNo = Intrinsic::x86_mmx_psra_d;
6890 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6896 // The vector shift intrinsics with scalars uses 32b shift amounts but
6897 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
6901 ShOps[1] = DAG.getConstant(0, MVT::i32);
6902 if (ShAmtVT == MVT::v4i32) {
6903 ShOps[2] = DAG.getUNDEF(MVT::i32);
6904 ShOps[3] = DAG.getUNDEF(MVT::i32);
6905 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
6907 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
6910 EVT VT = Op.getValueType();
6911 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
6912 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6913 DAG.getConstant(NewIntNo, MVT::i32),
6914 Op.getOperand(1), ShAmt);
6919 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6920 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6921 DebugLoc dl = Op.getDebugLoc();
6924 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6926 DAG.getConstant(TD->getPointerSize(),
6927 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6928 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6929 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6934 // Just load the return address.
6935 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6936 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6937 RetAddrFI, NULL, 0);
6940 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6941 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6942 MFI->setFrameAddressIsTaken(true);
6943 EVT VT = Op.getValueType();
6944 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6945 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6946 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6947 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6949 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6953 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6954 SelectionDAG &DAG) {
6955 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6958 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6960 MachineFunction &MF = DAG.getMachineFunction();
6961 SDValue Chain = Op.getOperand(0);
6962 SDValue Offset = Op.getOperand(1);
6963 SDValue Handler = Op.getOperand(2);
6964 DebugLoc dl = Op.getDebugLoc();
6966 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6968 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6970 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6971 DAG.getIntPtrConstant(-TD->getPointerSize()));
6972 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6973 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6974 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6975 MF.getRegInfo().addLiveOut(StoreAddrReg);
6977 return DAG.getNode(X86ISD::EH_RETURN, dl,
6979 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6982 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6983 SelectionDAG &DAG) {
6984 SDValue Root = Op.getOperand(0);
6985 SDValue Trmp = Op.getOperand(1); // trampoline
6986 SDValue FPtr = Op.getOperand(2); // nested function
6987 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6988 DebugLoc dl = Op.getDebugLoc();
6990 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6992 const X86InstrInfo *TII =
6993 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6995 if (Subtarget->is64Bit()) {
6996 SDValue OutChains[6];
6998 // Large code-model.
7000 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
7001 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
7003 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7004 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7006 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7008 // Load the pointer to the nested function into R11.
7009 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7010 SDValue Addr = Trmp;
7011 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7014 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7015 DAG.getConstant(2, MVT::i64));
7016 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
7018 // Load the 'nest' parameter value into R10.
7019 // R10 is specified in X86CallingConv.td
7020 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7021 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7022 DAG.getConstant(10, MVT::i64));
7023 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7024 Addr, TrmpAddr, 10);
7026 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7027 DAG.getConstant(12, MVT::i64));
7028 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
7030 // Jump to the nested function.
7031 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7032 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7033 DAG.getConstant(20, MVT::i64));
7034 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7035 Addr, TrmpAddr, 20);
7037 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7038 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7039 DAG.getConstant(22, MVT::i64));
7040 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7044 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7045 return DAG.getMergeValues(Ops, 2, dl);
7047 const Function *Func =
7048 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7049 CallingConv::ID CC = Func->getCallingConv();
7054 llvm_unreachable("Unsupported calling convention");
7055 case CallingConv::C:
7056 case CallingConv::X86_StdCall: {
7057 // Pass 'nest' parameter in ECX.
7058 // Must be kept in sync with X86CallingConv.td
7061 // Check that ECX wasn't needed by an 'inreg' parameter.
7062 const FunctionType *FTy = Func->getFunctionType();
7063 const AttrListPtr &Attrs = Func->getAttributes();
7065 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7066 unsigned InRegCount = 0;
7069 for (FunctionType::param_iterator I = FTy->param_begin(),
7070 E = FTy->param_end(); I != E; ++I, ++Idx)
7071 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7072 // FIXME: should only count parameters that are lowered to integers.
7073 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7075 if (InRegCount > 2) {
7076 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
7081 case CallingConv::X86_FastCall:
7082 case CallingConv::Fast:
7083 // Pass 'nest' parameter in EAX.
7084 // Must be kept in sync with X86CallingConv.td
7089 SDValue OutChains[4];
7092 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7093 DAG.getConstant(10, MVT::i32));
7094 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7096 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
7097 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7098 OutChains[0] = DAG.getStore(Root, dl,
7099 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7102 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7103 DAG.getConstant(1, MVT::i32));
7104 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
7106 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
7107 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7108 DAG.getConstant(5, MVT::i32));
7109 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7110 TrmpAddr, 5, false, 1);
7112 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7113 DAG.getConstant(6, MVT::i32));
7114 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
7117 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7118 return DAG.getMergeValues(Ops, 2, dl);
7122 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
7124 The rounding mode is in bits 11:10 of FPSR, and has the following
7131 FLT_ROUNDS, on the other hand, expects the following:
7138 To perform the conversion, we do:
7139 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7142 MachineFunction &MF = DAG.getMachineFunction();
7143 const TargetMachine &TM = MF.getTarget();
7144 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7145 unsigned StackAlignment = TFI.getStackAlignment();
7146 EVT VT = Op.getValueType();
7147 DebugLoc dl = Op.getDebugLoc();
7149 // Save FP Control Word to stack slot
7150 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7151 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7153 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7154 DAG.getEntryNode(), StackSlot);
7156 // Load FP Control Word from stack slot
7157 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
7159 // Transform as necessary
7161 DAG.getNode(ISD::SRL, dl, MVT::i16,
7162 DAG.getNode(ISD::AND, dl, MVT::i16,
7163 CWD, DAG.getConstant(0x800, MVT::i16)),
7164 DAG.getConstant(11, MVT::i8));
7166 DAG.getNode(ISD::SRL, dl, MVT::i16,
7167 DAG.getNode(ISD::AND, dl, MVT::i16,
7168 CWD, DAG.getConstant(0x400, MVT::i16)),
7169 DAG.getConstant(9, MVT::i8));
7172 DAG.getNode(ISD::AND, dl, MVT::i16,
7173 DAG.getNode(ISD::ADD, dl, MVT::i16,
7174 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7175 DAG.getConstant(1, MVT::i16)),
7176 DAG.getConstant(3, MVT::i16));
7179 return DAG.getNode((VT.getSizeInBits() < 16 ?
7180 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7183 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
7184 EVT VT = Op.getValueType();
7186 unsigned NumBits = VT.getSizeInBits();
7187 DebugLoc dl = Op.getDebugLoc();
7189 Op = Op.getOperand(0);
7190 if (VT == MVT::i8) {
7191 // Zero extend to i32 since there is not an i8 bsr.
7193 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7196 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7197 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7198 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7200 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7203 DAG.getConstant(NumBits+NumBits-1, OpVT),
7204 DAG.getConstant(X86::COND_E, MVT::i8),
7207 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7209 // Finally xor with NumBits-1.
7210 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7213 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7217 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
7218 EVT VT = Op.getValueType();
7220 unsigned NumBits = VT.getSizeInBits();
7221 DebugLoc dl = Op.getDebugLoc();
7223 Op = Op.getOperand(0);
7224 if (VT == MVT::i8) {
7226 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7229 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7230 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7231 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7233 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7236 DAG.getConstant(NumBits, OpVT),
7237 DAG.getConstant(X86::COND_E, MVT::i8),
7240 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7243 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7247 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
7248 EVT VT = Op.getValueType();
7249 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7250 DebugLoc dl = Op.getDebugLoc();
7252 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7253 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7254 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7255 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7256 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7258 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7259 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7260 // return AloBlo + AloBhi + AhiBlo;
7262 SDValue A = Op.getOperand(0);
7263 SDValue B = Op.getOperand(1);
7265 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7266 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7267 A, DAG.getConstant(32, MVT::i32));
7268 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7269 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7270 B, DAG.getConstant(32, MVT::i32));
7271 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7272 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7274 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7275 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7277 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7278 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7280 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7281 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7282 AloBhi, DAG.getConstant(32, MVT::i32));
7283 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7284 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7285 AhiBlo, DAG.getConstant(32, MVT::i32));
7286 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7287 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7292 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
7293 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7294 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7295 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7296 // has only one use.
7297 SDNode *N = Op.getNode();
7298 SDValue LHS = N->getOperand(0);
7299 SDValue RHS = N->getOperand(1);
7300 unsigned BaseOp = 0;
7302 DebugLoc dl = Op.getDebugLoc();
7304 switch (Op.getOpcode()) {
7305 default: llvm_unreachable("Unknown ovf instruction!");
7307 // A subtract of one will be selected as a INC. Note that INC doesn't
7308 // set CF, so we can't do this for UADDO.
7309 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7310 if (C->getAPIntValue() == 1) {
7311 BaseOp = X86ISD::INC;
7315 BaseOp = X86ISD::ADD;
7319 BaseOp = X86ISD::ADD;
7323 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7324 // set CF, so we can't do this for USUBO.
7325 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7326 if (C->getAPIntValue() == 1) {
7327 BaseOp = X86ISD::DEC;
7331 BaseOp = X86ISD::SUB;
7335 BaseOp = X86ISD::SUB;
7339 BaseOp = X86ISD::SMUL;
7343 BaseOp = X86ISD::UMUL;
7348 // Also sets EFLAGS.
7349 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7350 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7353 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7354 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7356 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7360 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
7361 EVT T = Op.getValueType();
7362 DebugLoc dl = Op.getDebugLoc();
7365 switch(T.getSimpleVT().SimpleTy) {
7367 assert(false && "Invalid value type!");
7368 case MVT::i8: Reg = X86::AL; size = 1; break;
7369 case MVT::i16: Reg = X86::AX; size = 2; break;
7370 case MVT::i32: Reg = X86::EAX; size = 4; break;
7372 assert(Subtarget->is64Bit() && "Node not type legal!");
7373 Reg = X86::RAX; size = 8;
7376 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7377 Op.getOperand(2), SDValue());
7378 SDValue Ops[] = { cpIn.getValue(0),
7381 DAG.getTargetConstant(size, MVT::i8),
7383 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7384 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7386 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7390 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7391 SelectionDAG &DAG) {
7392 assert(Subtarget->is64Bit() && "Result not type legalized?");
7393 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7394 SDValue TheChain = Op.getOperand(0);
7395 DebugLoc dl = Op.getDebugLoc();
7396 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7397 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7398 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7400 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7401 DAG.getConstant(32, MVT::i8));
7403 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7406 return DAG.getMergeValues(Ops, 2, dl);
7409 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7410 SDNode *Node = Op.getNode();
7411 DebugLoc dl = Node->getDebugLoc();
7412 EVT T = Node->getValueType(0);
7413 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7414 DAG.getConstant(0, T), Node->getOperand(2));
7415 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7416 cast<AtomicSDNode>(Node)->getMemoryVT(),
7417 Node->getOperand(0),
7418 Node->getOperand(1), negOp,
7419 cast<AtomicSDNode>(Node)->getSrcValue(),
7420 cast<AtomicSDNode>(Node)->getAlignment());
7423 /// LowerOperation - Provide custom lowering hooks for some operations.
7425 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7426 switch (Op.getOpcode()) {
7427 default: llvm_unreachable("Should not custom lower this!");
7428 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7429 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7430 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7431 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7432 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7433 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7434 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7435 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7436 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7437 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7438 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7439 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7440 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7441 case ISD::SHL_PARTS:
7442 case ISD::SRA_PARTS:
7443 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7444 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7445 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7446 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7447 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7448 case ISD::FABS: return LowerFABS(Op, DAG);
7449 case ISD::FNEG: return LowerFNEG(Op, DAG);
7450 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7451 case ISD::SETCC: return LowerSETCC(Op, DAG);
7452 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7453 case ISD::SELECT: return LowerSELECT(Op, DAG);
7454 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7455 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7456 case ISD::VASTART: return LowerVASTART(Op, DAG);
7457 case ISD::VAARG: return LowerVAARG(Op, DAG);
7458 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7459 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7460 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7461 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7462 case ISD::FRAME_TO_ARGS_OFFSET:
7463 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7464 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7465 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7466 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7467 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7468 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7469 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7470 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7476 case ISD::UMULO: return LowerXALUO(Op, DAG);
7477 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7481 void X86TargetLowering::
7482 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7483 SelectionDAG &DAG, unsigned NewOp) {
7484 EVT T = Node->getValueType(0);
7485 DebugLoc dl = Node->getDebugLoc();
7486 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7488 SDValue Chain = Node->getOperand(0);
7489 SDValue In1 = Node->getOperand(1);
7490 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7491 Node->getOperand(2), DAG.getIntPtrConstant(0));
7492 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7493 Node->getOperand(2), DAG.getIntPtrConstant(1));
7494 SDValue Ops[] = { Chain, In1, In2L, In2H };
7495 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7497 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7498 cast<MemSDNode>(Node)->getMemOperand());
7499 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7500 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7501 Results.push_back(Result.getValue(2));
7504 /// ReplaceNodeResults - Replace a node with an illegal result type
7505 /// with a new node built out of custom code.
7506 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7507 SmallVectorImpl<SDValue>&Results,
7508 SelectionDAG &DAG) {
7509 DebugLoc dl = N->getDebugLoc();
7510 switch (N->getOpcode()) {
7512 assert(false && "Do not know how to custom type legalize this operation!");
7514 case ISD::FP_TO_SINT: {
7515 std::pair<SDValue,SDValue> Vals =
7516 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7517 SDValue FIST = Vals.first, StackSlot = Vals.second;
7518 if (FIST.getNode() != 0) {
7519 EVT VT = N->getValueType(0);
7520 // Return a load from the stack slot.
7521 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
7525 case ISD::READCYCLECOUNTER: {
7526 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7527 SDValue TheChain = N->getOperand(0);
7528 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7529 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7531 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7533 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7534 SDValue Ops[] = { eax, edx };
7535 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7536 Results.push_back(edx.getValue(1));
7543 EVT WidenVT = getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
7544 Results.push_back(DAG.UnrollVectorOp(N, WidenVT.getVectorNumElements()));
7547 case ISD::ATOMIC_CMP_SWAP: {
7548 EVT T = N->getValueType(0);
7549 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7550 SDValue cpInL, cpInH;
7551 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7552 DAG.getConstant(0, MVT::i32));
7553 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7554 DAG.getConstant(1, MVT::i32));
7555 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7556 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7558 SDValue swapInL, swapInH;
7559 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7560 DAG.getConstant(0, MVT::i32));
7561 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7562 DAG.getConstant(1, MVT::i32));
7563 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7565 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7566 swapInL.getValue(1));
7567 SDValue Ops[] = { swapInH.getValue(0),
7569 swapInH.getValue(1) };
7570 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7571 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7572 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7573 MVT::i32, Result.getValue(1));
7574 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7575 MVT::i32, cpOutL.getValue(2));
7576 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7577 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7578 Results.push_back(cpOutH.getValue(1));
7581 case ISD::ATOMIC_LOAD_ADD:
7582 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7584 case ISD::ATOMIC_LOAD_AND:
7585 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7587 case ISD::ATOMIC_LOAD_NAND:
7588 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7590 case ISD::ATOMIC_LOAD_OR:
7591 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7593 case ISD::ATOMIC_LOAD_SUB:
7594 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7596 case ISD::ATOMIC_LOAD_XOR:
7597 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7599 case ISD::ATOMIC_SWAP:
7600 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7605 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7607 default: return NULL;
7608 case X86ISD::BSF: return "X86ISD::BSF";
7609 case X86ISD::BSR: return "X86ISD::BSR";
7610 case X86ISD::SHLD: return "X86ISD::SHLD";
7611 case X86ISD::SHRD: return "X86ISD::SHRD";
7612 case X86ISD::FAND: return "X86ISD::FAND";
7613 case X86ISD::FOR: return "X86ISD::FOR";
7614 case X86ISD::FXOR: return "X86ISD::FXOR";
7615 case X86ISD::FSRL: return "X86ISD::FSRL";
7616 case X86ISD::FILD: return "X86ISD::FILD";
7617 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7618 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7619 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7620 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7621 case X86ISD::FLD: return "X86ISD::FLD";
7622 case X86ISD::FST: return "X86ISD::FST";
7623 case X86ISD::CALL: return "X86ISD::CALL";
7624 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7625 case X86ISD::BT: return "X86ISD::BT";
7626 case X86ISD::CMP: return "X86ISD::CMP";
7627 case X86ISD::COMI: return "X86ISD::COMI";
7628 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7629 case X86ISD::SETCC: return "X86ISD::SETCC";
7630 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7631 case X86ISD::CMOV: return "X86ISD::CMOV";
7632 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7633 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7634 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7635 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7636 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7637 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7638 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7639 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7640 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7641 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7642 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7643 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7644 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7645 case X86ISD::FMAX: return "X86ISD::FMAX";
7646 case X86ISD::FMIN: return "X86ISD::FMIN";
7647 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7648 case X86ISD::FRCP: return "X86ISD::FRCP";
7649 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7650 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7651 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7652 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7653 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7654 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7655 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7656 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7657 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7658 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7659 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7660 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7661 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7662 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7663 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7664 case X86ISD::VSHL: return "X86ISD::VSHL";
7665 case X86ISD::VSRL: return "X86ISD::VSRL";
7666 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7667 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7668 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7669 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7670 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7671 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7672 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7673 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7674 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7675 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7676 case X86ISD::ADD: return "X86ISD::ADD";
7677 case X86ISD::SUB: return "X86ISD::SUB";
7678 case X86ISD::SMUL: return "X86ISD::SMUL";
7679 case X86ISD::UMUL: return "X86ISD::UMUL";
7680 case X86ISD::INC: return "X86ISD::INC";
7681 case X86ISD::DEC: return "X86ISD::DEC";
7682 case X86ISD::OR: return "X86ISD::OR";
7683 case X86ISD::XOR: return "X86ISD::XOR";
7684 case X86ISD::AND: return "X86ISD::AND";
7685 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7686 case X86ISD::PTEST: return "X86ISD::PTEST";
7687 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7691 // isLegalAddressingMode - Return true if the addressing mode represented
7692 // by AM is legal for this target, for a load/store of the specified type.
7693 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7694 const Type *Ty) const {
7695 // X86 supports extremely general addressing modes.
7696 CodeModel::Model M = getTargetMachine().getCodeModel();
7698 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7699 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7704 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7706 // If a reference to this global requires an extra load, we can't fold it.
7707 if (isGlobalStubReference(GVFlags))
7710 // If BaseGV requires a register for the PIC base, we cannot also have a
7711 // BaseReg specified.
7712 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7715 // If lower 4G is not available, then we must use rip-relative addressing.
7716 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7726 // These scales always work.
7731 // These scales are formed with basereg+scalereg. Only accept if there is
7736 default: // Other stuff never works.
7744 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7745 if (!Ty1->isInteger() || !Ty2->isInteger())
7747 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7748 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7749 if (NumBits1 <= NumBits2)
7751 return Subtarget->is64Bit() || NumBits1 < 64;
7754 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7755 if (!VT1.isInteger() || !VT2.isInteger())
7757 unsigned NumBits1 = VT1.getSizeInBits();
7758 unsigned NumBits2 = VT2.getSizeInBits();
7759 if (NumBits1 <= NumBits2)
7761 return Subtarget->is64Bit() || NumBits1 < 64;
7764 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7765 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7766 return Ty1->isInteger(32) && Ty2->isInteger(64) && Subtarget->is64Bit();
7769 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7770 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7771 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7774 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7775 // i16 instructions are longer (0x66 prefix) and potentially slower.
7776 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7779 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7780 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7781 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7782 /// are assumed to be legal.
7784 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7786 // Only do shuffles on 128-bit vector types for now.
7787 if (VT.getSizeInBits() == 64)
7790 // FIXME: pshufb, blends, shifts.
7791 return (VT.getVectorNumElements() == 2 ||
7792 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7793 isMOVLMask(M, VT) ||
7794 isSHUFPMask(M, VT) ||
7795 isPSHUFDMask(M, VT) ||
7796 isPSHUFHWMask(M, VT) ||
7797 isPSHUFLWMask(M, VT) ||
7798 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7799 isUNPCKLMask(M, VT) ||
7800 isUNPCKHMask(M, VT) ||
7801 isUNPCKL_v_undef_Mask(M, VT) ||
7802 isUNPCKH_v_undef_Mask(M, VT));
7806 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7808 unsigned NumElts = VT.getVectorNumElements();
7809 // FIXME: This collection of masks seems suspect.
7812 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7813 return (isMOVLMask(Mask, VT) ||
7814 isCommutedMOVLMask(Mask, VT, true) ||
7815 isSHUFPMask(Mask, VT) ||
7816 isCommutedSHUFPMask(Mask, VT));
7821 //===----------------------------------------------------------------------===//
7822 // X86 Scheduler Hooks
7823 //===----------------------------------------------------------------------===//
7825 // private utility function
7827 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7828 MachineBasicBlock *MBB,
7836 TargetRegisterClass *RC,
7837 bool invSrc) const {
7838 // For the atomic bitwise operator, we generate
7841 // ld t1 = [bitinstr.addr]
7842 // op t2 = t1, [bitinstr.val]
7844 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7846 // fallthrough -->nextMBB
7847 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7848 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7849 MachineFunction::iterator MBBIter = MBB;
7852 /// First build the CFG
7853 MachineFunction *F = MBB->getParent();
7854 MachineBasicBlock *thisMBB = MBB;
7855 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7856 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7857 F->insert(MBBIter, newMBB);
7858 F->insert(MBBIter, nextMBB);
7860 // Move all successors to thisMBB to nextMBB
7861 nextMBB->transferSuccessors(thisMBB);
7863 // Update thisMBB to fall through to newMBB
7864 thisMBB->addSuccessor(newMBB);
7866 // newMBB jumps to itself and fall through to nextMBB
7867 newMBB->addSuccessor(nextMBB);
7868 newMBB->addSuccessor(newMBB);
7870 // Insert instructions into newMBB based on incoming instruction
7871 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7872 "unexpected number of operands");
7873 DebugLoc dl = bInstr->getDebugLoc();
7874 MachineOperand& destOper = bInstr->getOperand(0);
7875 MachineOperand* argOpers[2 + X86AddrNumOperands];
7876 int numArgs = bInstr->getNumOperands() - 1;
7877 for (int i=0; i < numArgs; ++i)
7878 argOpers[i] = &bInstr->getOperand(i+1);
7880 // x86 address has 4 operands: base, index, scale, and displacement
7881 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7882 int valArgIndx = lastAddrIndx + 1;
7884 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7885 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7886 for (int i=0; i <= lastAddrIndx; ++i)
7887 (*MIB).addOperand(*argOpers[i]);
7889 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7891 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7896 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7897 assert((argOpers[valArgIndx]->isReg() ||
7898 argOpers[valArgIndx]->isImm()) &&
7900 if (argOpers[valArgIndx]->isReg())
7901 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7903 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7905 (*MIB).addOperand(*argOpers[valArgIndx]);
7907 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7910 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7911 for (int i=0; i <= lastAddrIndx; ++i)
7912 (*MIB).addOperand(*argOpers[i]);
7914 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7915 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7916 bInstr->memoperands_end());
7918 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7922 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7924 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7928 // private utility function: 64 bit atomics on 32 bit host.
7930 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7931 MachineBasicBlock *MBB,
7936 bool invSrc) const {
7937 // For the atomic bitwise operator, we generate
7938 // thisMBB (instructions are in pairs, except cmpxchg8b)
7939 // ld t1,t2 = [bitinstr.addr]
7941 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7942 // op t5, t6 <- out1, out2, [bitinstr.val]
7943 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7944 // mov ECX, EBX <- t5, t6
7945 // mov EAX, EDX <- t1, t2
7946 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7947 // mov t3, t4 <- EAX, EDX
7949 // result in out1, out2
7950 // fallthrough -->nextMBB
7952 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7953 const unsigned LoadOpc = X86::MOV32rm;
7954 const unsigned copyOpc = X86::MOV32rr;
7955 const unsigned NotOpc = X86::NOT32r;
7956 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7957 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7958 MachineFunction::iterator MBBIter = MBB;
7961 /// First build the CFG
7962 MachineFunction *F = MBB->getParent();
7963 MachineBasicBlock *thisMBB = MBB;
7964 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7965 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7966 F->insert(MBBIter, newMBB);
7967 F->insert(MBBIter, nextMBB);
7969 // Move all successors to thisMBB to nextMBB
7970 nextMBB->transferSuccessors(thisMBB);
7972 // Update thisMBB to fall through to newMBB
7973 thisMBB->addSuccessor(newMBB);
7975 // newMBB jumps to itself and fall through to nextMBB
7976 newMBB->addSuccessor(nextMBB);
7977 newMBB->addSuccessor(newMBB);
7979 DebugLoc dl = bInstr->getDebugLoc();
7980 // Insert instructions into newMBB based on incoming instruction
7981 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7982 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7983 "unexpected number of operands");
7984 MachineOperand& dest1Oper = bInstr->getOperand(0);
7985 MachineOperand& dest2Oper = bInstr->getOperand(1);
7986 MachineOperand* argOpers[2 + X86AddrNumOperands];
7987 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7988 argOpers[i] = &bInstr->getOperand(i+2);
7990 // x86 address has 5 operands: base, index, scale, displacement, and segment.
7991 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7993 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7994 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7995 for (int i=0; i <= lastAddrIndx; ++i)
7996 (*MIB).addOperand(*argOpers[i]);
7997 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7998 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7999 // add 4 to displacement.
8000 for (int i=0; i <= lastAddrIndx-2; ++i)
8001 (*MIB).addOperand(*argOpers[i]);
8002 MachineOperand newOp3 = *(argOpers[3]);
8004 newOp3.setImm(newOp3.getImm()+4);
8006 newOp3.setOffset(newOp3.getOffset()+4);
8007 (*MIB).addOperand(newOp3);
8008 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8010 // t3/4 are defined later, at the bottom of the loop
8011 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8012 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8013 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8014 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8015 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8016 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8018 // The subsequent operations should be using the destination registers of
8019 //the PHI instructions.
8021 t1 = F->getRegInfo().createVirtualRegister(RC);
8022 t2 = F->getRegInfo().createVirtualRegister(RC);
8023 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8024 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8026 t1 = dest1Oper.getReg();
8027 t2 = dest2Oper.getReg();
8030 int valArgIndx = lastAddrIndx + 1;
8031 assert((argOpers[valArgIndx]->isReg() ||
8032 argOpers[valArgIndx]->isImm()) &&
8034 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8035 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8036 if (argOpers[valArgIndx]->isReg())
8037 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8039 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8040 if (regOpcL != X86::MOV32rr)
8042 (*MIB).addOperand(*argOpers[valArgIndx]);
8043 assert(argOpers[valArgIndx + 1]->isReg() ==
8044 argOpers[valArgIndx]->isReg());
8045 assert(argOpers[valArgIndx + 1]->isImm() ==
8046 argOpers[valArgIndx]->isImm());
8047 if (argOpers[valArgIndx + 1]->isReg())
8048 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8050 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8051 if (regOpcH != X86::MOV32rr)
8053 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8055 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8057 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8060 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8062 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8065 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8066 for (int i=0; i <= lastAddrIndx; ++i)
8067 (*MIB).addOperand(*argOpers[i]);
8069 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8070 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8071 bInstr->memoperands_end());
8073 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8074 MIB.addReg(X86::EAX);
8075 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8076 MIB.addReg(X86::EDX);
8079 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
8081 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8085 // private utility function
8087 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8088 MachineBasicBlock *MBB,
8089 unsigned cmovOpc) const {
8090 // For the atomic min/max operator, we generate
8093 // ld t1 = [min/max.addr]
8094 // mov t2 = [min/max.val]
8096 // cmov[cond] t2 = t1
8098 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8100 // fallthrough -->nextMBB
8102 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8103 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8104 MachineFunction::iterator MBBIter = MBB;
8107 /// First build the CFG
8108 MachineFunction *F = MBB->getParent();
8109 MachineBasicBlock *thisMBB = MBB;
8110 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8111 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8112 F->insert(MBBIter, newMBB);
8113 F->insert(MBBIter, nextMBB);
8115 // Move all successors of thisMBB to nextMBB
8116 nextMBB->transferSuccessors(thisMBB);
8118 // Update thisMBB to fall through to newMBB
8119 thisMBB->addSuccessor(newMBB);
8121 // newMBB jumps to newMBB and fall through to nextMBB
8122 newMBB->addSuccessor(nextMBB);
8123 newMBB->addSuccessor(newMBB);
8125 DebugLoc dl = mInstr->getDebugLoc();
8126 // Insert instructions into newMBB based on incoming instruction
8127 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8128 "unexpected number of operands");
8129 MachineOperand& destOper = mInstr->getOperand(0);
8130 MachineOperand* argOpers[2 + X86AddrNumOperands];
8131 int numArgs = mInstr->getNumOperands() - 1;
8132 for (int i=0; i < numArgs; ++i)
8133 argOpers[i] = &mInstr->getOperand(i+1);
8135 // x86 address has 4 operands: base, index, scale, and displacement
8136 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8137 int valArgIndx = lastAddrIndx + 1;
8139 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8140 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8141 for (int i=0; i <= lastAddrIndx; ++i)
8142 (*MIB).addOperand(*argOpers[i]);
8144 // We only support register and immediate values
8145 assert((argOpers[valArgIndx]->isReg() ||
8146 argOpers[valArgIndx]->isImm()) &&
8149 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8150 if (argOpers[valArgIndx]->isReg())
8151 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8153 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8154 (*MIB).addOperand(*argOpers[valArgIndx]);
8156 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8159 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8164 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8165 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8169 // Cmp and exchange if none has modified the memory location
8170 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8171 for (int i=0; i <= lastAddrIndx; ++i)
8172 (*MIB).addOperand(*argOpers[i]);
8174 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8175 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8176 mInstr->memoperands_end());
8178 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8179 MIB.addReg(X86::EAX);
8182 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
8184 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8188 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8189 // all of this code can be replaced with that in the .td file.
8191 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8192 unsigned numArgs, bool memArg) const {
8194 MachineFunction *F = BB->getParent();
8195 DebugLoc dl = MI->getDebugLoc();
8196 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8200 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8202 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8204 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8206 for (unsigned i = 0; i < numArgs; ++i) {
8207 MachineOperand &Op = MI->getOperand(i+1);
8209 if (!(Op.isReg() && Op.isImplicit()))
8213 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8216 F->DeleteMachineInstr(MI);
8222 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8224 MachineBasicBlock *MBB) const {
8225 // Emit code to save XMM registers to the stack. The ABI says that the
8226 // number of registers to save is given in %al, so it's theoretically
8227 // possible to do an indirect jump trick to avoid saving all of them,
8228 // however this code takes a simpler approach and just executes all
8229 // of the stores if %al is non-zero. It's less code, and it's probably
8230 // easier on the hardware branch predictor, and stores aren't all that
8231 // expensive anyway.
8233 // Create the new basic blocks. One block contains all the XMM stores,
8234 // and one block is the final destination regardless of whether any
8235 // stores were performed.
8236 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8237 MachineFunction *F = MBB->getParent();
8238 MachineFunction::iterator MBBIter = MBB;
8240 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8241 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8242 F->insert(MBBIter, XMMSaveMBB);
8243 F->insert(MBBIter, EndMBB);
8246 // Move any original successors of MBB to the end block.
8247 EndMBB->transferSuccessors(MBB);
8248 // The original block will now fall through to the XMM save block.
8249 MBB->addSuccessor(XMMSaveMBB);
8250 // The XMMSaveMBB will fall through to the end block.
8251 XMMSaveMBB->addSuccessor(EndMBB);
8253 // Now add the instructions.
8254 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8255 DebugLoc DL = MI->getDebugLoc();
8257 unsigned CountReg = MI->getOperand(0).getReg();
8258 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8259 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8261 if (!Subtarget->isTargetWin64()) {
8262 // If %al is 0, branch around the XMM save block.
8263 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8264 BuildMI(MBB, DL, TII->get(X86::JE)).addMBB(EndMBB);
8265 MBB->addSuccessor(EndMBB);
8268 // In the XMM save block, save all the XMM argument registers.
8269 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8270 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8271 MachineMemOperand *MMO =
8272 F->getMachineMemOperand(
8273 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8274 MachineMemOperand::MOStore, Offset,
8275 /*Size=*/16, /*Align=*/16);
8276 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8277 .addFrameIndex(RegSaveFrameIndex)
8278 .addImm(/*Scale=*/1)
8279 .addReg(/*IndexReg=*/0)
8280 .addImm(/*Disp=*/Offset)
8281 .addReg(/*Segment=*/0)
8282 .addReg(MI->getOperand(i).getReg())
8283 .addMemOperand(MMO);
8286 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8292 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8293 MachineBasicBlock *BB,
8294 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8295 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8296 DebugLoc DL = MI->getDebugLoc();
8298 // To "insert" a SELECT_CC instruction, we actually have to insert the
8299 // diamond control-flow pattern. The incoming instruction knows the
8300 // destination vreg to set, the condition code register to branch on, the
8301 // true/false values to select between, and a branch opcode to use.
8302 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8303 MachineFunction::iterator It = BB;
8309 // cmpTY ccX, r1, r2
8311 // fallthrough --> copy0MBB
8312 MachineBasicBlock *thisMBB = BB;
8313 MachineFunction *F = BB->getParent();
8314 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8315 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8317 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8318 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8319 F->insert(It, copy0MBB);
8320 F->insert(It, sinkMBB);
8321 // Update machine-CFG edges by first adding all successors of the current
8322 // block to the new block which will contain the Phi node for the select.
8323 // Also inform sdisel of the edge changes.
8324 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8325 E = BB->succ_end(); I != E; ++I) {
8326 EM->insert(std::make_pair(*I, sinkMBB));
8327 sinkMBB->addSuccessor(*I);
8329 // Next, remove all successors of the current block, and add the true
8330 // and fallthrough blocks as its successors.
8331 while (!BB->succ_empty())
8332 BB->removeSuccessor(BB->succ_begin());
8333 // Add the true and fallthrough blocks as its successors.
8334 BB->addSuccessor(copy0MBB);
8335 BB->addSuccessor(sinkMBB);
8338 // %FalseValue = ...
8339 // # fallthrough to sinkMBB
8342 // Update machine-CFG edges
8343 BB->addSuccessor(sinkMBB);
8346 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8349 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8350 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8351 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8353 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8359 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8360 MachineBasicBlock *BB,
8361 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8362 switch (MI->getOpcode()) {
8363 default: assert(false && "Unexpected instr type to insert");
8365 case X86::CMOV_V1I64:
8366 case X86::CMOV_FR32:
8367 case X86::CMOV_FR64:
8368 case X86::CMOV_V4F32:
8369 case X86::CMOV_V2F64:
8370 case X86::CMOV_V2I64:
8371 return EmitLoweredSelect(MI, BB, EM);
8373 case X86::FP32_TO_INT16_IN_MEM:
8374 case X86::FP32_TO_INT32_IN_MEM:
8375 case X86::FP32_TO_INT64_IN_MEM:
8376 case X86::FP64_TO_INT16_IN_MEM:
8377 case X86::FP64_TO_INT32_IN_MEM:
8378 case X86::FP64_TO_INT64_IN_MEM:
8379 case X86::FP80_TO_INT16_IN_MEM:
8380 case X86::FP80_TO_INT32_IN_MEM:
8381 case X86::FP80_TO_INT64_IN_MEM: {
8382 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8383 DebugLoc DL = MI->getDebugLoc();
8385 // Change the floating point control register to use "round towards zero"
8386 // mode when truncating to an integer value.
8387 MachineFunction *F = BB->getParent();
8388 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8389 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8391 // Load the old value of the high byte of the control word...
8393 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8394 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8397 // Set the high part to be round to zero...
8398 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8401 // Reload the modified control word now...
8402 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8404 // Restore the memory image of control word to original value
8405 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8408 // Get the X86 opcode to use.
8410 switch (MI->getOpcode()) {
8411 default: llvm_unreachable("illegal opcode!");
8412 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8413 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8414 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8415 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8416 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8417 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8418 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8419 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8420 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8424 MachineOperand &Op = MI->getOperand(0);
8426 AM.BaseType = X86AddressMode::RegBase;
8427 AM.Base.Reg = Op.getReg();
8429 AM.BaseType = X86AddressMode::FrameIndexBase;
8430 AM.Base.FrameIndex = Op.getIndex();
8432 Op = MI->getOperand(1);
8434 AM.Scale = Op.getImm();
8435 Op = MI->getOperand(2);
8437 AM.IndexReg = Op.getImm();
8438 Op = MI->getOperand(3);
8439 if (Op.isGlobal()) {
8440 AM.GV = Op.getGlobal();
8442 AM.Disp = Op.getImm();
8444 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8445 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8447 // Reload the original control word now.
8448 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8450 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8453 // String/text processing lowering.
8454 case X86::PCMPISTRM128REG:
8455 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8456 case X86::PCMPISTRM128MEM:
8457 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8458 case X86::PCMPESTRM128REG:
8459 return EmitPCMP(MI, BB, 5, false /* in mem */);
8460 case X86::PCMPESTRM128MEM:
8461 return EmitPCMP(MI, BB, 5, true /* in mem */);
8464 case X86::ATOMAND32:
8465 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8466 X86::AND32ri, X86::MOV32rm,
8467 X86::LCMPXCHG32, X86::MOV32rr,
8468 X86::NOT32r, X86::EAX,
8469 X86::GR32RegisterClass);
8471 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8472 X86::OR32ri, X86::MOV32rm,
8473 X86::LCMPXCHG32, X86::MOV32rr,
8474 X86::NOT32r, X86::EAX,
8475 X86::GR32RegisterClass);
8476 case X86::ATOMXOR32:
8477 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8478 X86::XOR32ri, X86::MOV32rm,
8479 X86::LCMPXCHG32, X86::MOV32rr,
8480 X86::NOT32r, X86::EAX,
8481 X86::GR32RegisterClass);
8482 case X86::ATOMNAND32:
8483 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8484 X86::AND32ri, X86::MOV32rm,
8485 X86::LCMPXCHG32, X86::MOV32rr,
8486 X86::NOT32r, X86::EAX,
8487 X86::GR32RegisterClass, true);
8488 case X86::ATOMMIN32:
8489 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8490 case X86::ATOMMAX32:
8491 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8492 case X86::ATOMUMIN32:
8493 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8494 case X86::ATOMUMAX32:
8495 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8497 case X86::ATOMAND16:
8498 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8499 X86::AND16ri, X86::MOV16rm,
8500 X86::LCMPXCHG16, X86::MOV16rr,
8501 X86::NOT16r, X86::AX,
8502 X86::GR16RegisterClass);
8504 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8505 X86::OR16ri, X86::MOV16rm,
8506 X86::LCMPXCHG16, X86::MOV16rr,
8507 X86::NOT16r, X86::AX,
8508 X86::GR16RegisterClass);
8509 case X86::ATOMXOR16:
8510 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8511 X86::XOR16ri, X86::MOV16rm,
8512 X86::LCMPXCHG16, X86::MOV16rr,
8513 X86::NOT16r, X86::AX,
8514 X86::GR16RegisterClass);
8515 case X86::ATOMNAND16:
8516 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8517 X86::AND16ri, X86::MOV16rm,
8518 X86::LCMPXCHG16, X86::MOV16rr,
8519 X86::NOT16r, X86::AX,
8520 X86::GR16RegisterClass, true);
8521 case X86::ATOMMIN16:
8522 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8523 case X86::ATOMMAX16:
8524 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8525 case X86::ATOMUMIN16:
8526 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8527 case X86::ATOMUMAX16:
8528 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8531 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8532 X86::AND8ri, X86::MOV8rm,
8533 X86::LCMPXCHG8, X86::MOV8rr,
8534 X86::NOT8r, X86::AL,
8535 X86::GR8RegisterClass);
8537 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8538 X86::OR8ri, X86::MOV8rm,
8539 X86::LCMPXCHG8, X86::MOV8rr,
8540 X86::NOT8r, X86::AL,
8541 X86::GR8RegisterClass);
8543 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8544 X86::XOR8ri, X86::MOV8rm,
8545 X86::LCMPXCHG8, X86::MOV8rr,
8546 X86::NOT8r, X86::AL,
8547 X86::GR8RegisterClass);
8548 case X86::ATOMNAND8:
8549 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8550 X86::AND8ri, X86::MOV8rm,
8551 X86::LCMPXCHG8, X86::MOV8rr,
8552 X86::NOT8r, X86::AL,
8553 X86::GR8RegisterClass, true);
8554 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8555 // This group is for 64-bit host.
8556 case X86::ATOMAND64:
8557 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8558 X86::AND64ri32, X86::MOV64rm,
8559 X86::LCMPXCHG64, X86::MOV64rr,
8560 X86::NOT64r, X86::RAX,
8561 X86::GR64RegisterClass);
8563 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8564 X86::OR64ri32, X86::MOV64rm,
8565 X86::LCMPXCHG64, X86::MOV64rr,
8566 X86::NOT64r, X86::RAX,
8567 X86::GR64RegisterClass);
8568 case X86::ATOMXOR64:
8569 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8570 X86::XOR64ri32, X86::MOV64rm,
8571 X86::LCMPXCHG64, X86::MOV64rr,
8572 X86::NOT64r, X86::RAX,
8573 X86::GR64RegisterClass);
8574 case X86::ATOMNAND64:
8575 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8576 X86::AND64ri32, X86::MOV64rm,
8577 X86::LCMPXCHG64, X86::MOV64rr,
8578 X86::NOT64r, X86::RAX,
8579 X86::GR64RegisterClass, true);
8580 case X86::ATOMMIN64:
8581 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8582 case X86::ATOMMAX64:
8583 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8584 case X86::ATOMUMIN64:
8585 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8586 case X86::ATOMUMAX64:
8587 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8589 // This group does 64-bit operations on a 32-bit host.
8590 case X86::ATOMAND6432:
8591 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8592 X86::AND32rr, X86::AND32rr,
8593 X86::AND32ri, X86::AND32ri,
8595 case X86::ATOMOR6432:
8596 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8597 X86::OR32rr, X86::OR32rr,
8598 X86::OR32ri, X86::OR32ri,
8600 case X86::ATOMXOR6432:
8601 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8602 X86::XOR32rr, X86::XOR32rr,
8603 X86::XOR32ri, X86::XOR32ri,
8605 case X86::ATOMNAND6432:
8606 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8607 X86::AND32rr, X86::AND32rr,
8608 X86::AND32ri, X86::AND32ri,
8610 case X86::ATOMADD6432:
8611 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8612 X86::ADD32rr, X86::ADC32rr,
8613 X86::ADD32ri, X86::ADC32ri,
8615 case X86::ATOMSUB6432:
8616 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8617 X86::SUB32rr, X86::SBB32rr,
8618 X86::SUB32ri, X86::SBB32ri,
8620 case X86::ATOMSWAP6432:
8621 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8622 X86::MOV32rr, X86::MOV32rr,
8623 X86::MOV32ri, X86::MOV32ri,
8625 case X86::VASTART_SAVE_XMM_REGS:
8626 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8630 //===----------------------------------------------------------------------===//
8631 // X86 Optimization Hooks
8632 //===----------------------------------------------------------------------===//
8634 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8638 const SelectionDAG &DAG,
8639 unsigned Depth) const {
8640 unsigned Opc = Op.getOpcode();
8641 assert((Opc >= ISD::BUILTIN_OP_END ||
8642 Opc == ISD::INTRINSIC_WO_CHAIN ||
8643 Opc == ISD::INTRINSIC_W_CHAIN ||
8644 Opc == ISD::INTRINSIC_VOID) &&
8645 "Should use MaskedValueIsZero if you don't know whether Op"
8646 " is a target node!");
8648 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8660 // These nodes' second result is a boolean.
8661 if (Op.getResNo() == 0)
8665 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8666 Mask.getBitWidth() - 1);
8671 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8672 /// node is a GlobalAddress + offset.
8673 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8674 GlobalValue* &GA, int64_t &Offset) const{
8675 if (N->getOpcode() == X86ISD::Wrapper) {
8676 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8677 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8678 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8682 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8685 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8686 EVT EltVT, LoadSDNode *&LDBase,
8687 unsigned &LastLoadedElt,
8688 SelectionDAG &DAG, MachineFrameInfo *MFI,
8689 const TargetLowering &TLI) {
8691 LastLoadedElt = -1U;
8692 for (unsigned i = 0; i < NumElems; ++i) {
8693 if (N->getMaskElt(i) < 0) {
8699 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8700 if (!Elt.getNode() ||
8701 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8704 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8706 LDBase = cast<LoadSDNode>(Elt.getNode());
8710 if (Elt.getOpcode() == ISD::UNDEF)
8713 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8714 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
8721 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8722 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8723 /// if the load addresses are consecutive, non-overlapping, and in the right
8724 /// order. In the case of v2i64, it will see if it can rewrite the
8725 /// shuffle to be an appropriate build vector so it can take advantage of
8726 // performBuildVectorCombine.
8727 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8728 const TargetLowering &TLI) {
8729 DebugLoc dl = N->getDebugLoc();
8730 EVT VT = N->getValueType(0);
8731 EVT EltVT = VT.getVectorElementType();
8732 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8733 unsigned NumElems = VT.getVectorNumElements();
8735 if (VT.getSizeInBits() != 128)
8738 // Try to combine a vector_shuffle into a 128-bit load.
8739 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8740 LoadSDNode *LD = NULL;
8741 unsigned LastLoadedElt;
8742 if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
8746 if (LastLoadedElt == NumElems - 1) {
8747 if (DAG.InferPtrAlignment(LD->getBasePtr()) >= 16)
8748 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8749 LD->getSrcValue(), LD->getSrcValueOffset(),
8751 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8752 LD->getSrcValue(), LD->getSrcValueOffset(),
8753 LD->isVolatile(), LD->getAlignment());
8754 } else if (NumElems == 4 && LastLoadedElt == 1) {
8755 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8756 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8757 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8758 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8763 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8764 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8765 const X86Subtarget *Subtarget) {
8766 DebugLoc DL = N->getDebugLoc();
8767 SDValue Cond = N->getOperand(0);
8768 // Get the LHS/RHS of the select.
8769 SDValue LHS = N->getOperand(1);
8770 SDValue RHS = N->getOperand(2);
8772 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8773 // instructions have the peculiarity that if either operand is a NaN,
8774 // they chose what we call the RHS operand (and as such are not symmetric).
8775 // It happens that this matches the semantics of the common C idiom
8776 // x<y?x:y and related forms, so we can recognize these cases.
8777 if (Subtarget->hasSSE2() &&
8778 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8779 Cond.getOpcode() == ISD::SETCC) {
8780 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8782 unsigned Opcode = 0;
8783 // Check for x CC y ? x : y.
8784 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8788 // This can be a min if we can prove that at least one of the operands
8790 if (!FiniteOnlyFPMath()) {
8791 if (DAG.isKnownNeverNaN(RHS)) {
8792 // Put the potential NaN in the RHS so that SSE will preserve it.
8793 std::swap(LHS, RHS);
8794 } else if (!DAG.isKnownNeverNaN(LHS))
8797 Opcode = X86ISD::FMIN;
8800 // This can be a min if we can prove that at least one of the operands
8802 if (!FiniteOnlyFPMath()) {
8803 if (DAG.isKnownNeverNaN(LHS)) {
8804 // Put the potential NaN in the RHS so that SSE will preserve it.
8805 std::swap(LHS, RHS);
8806 } else if (!DAG.isKnownNeverNaN(RHS))
8809 Opcode = X86ISD::FMIN;
8812 // This can be a min, but if either operand is a NaN we need it to
8813 // preserve the original LHS.
8814 std::swap(LHS, RHS);
8818 Opcode = X86ISD::FMIN;
8822 // This can be a max if we can prove that at least one of the operands
8824 if (!FiniteOnlyFPMath()) {
8825 if (DAG.isKnownNeverNaN(LHS)) {
8826 // Put the potential NaN in the RHS so that SSE will preserve it.
8827 std::swap(LHS, RHS);
8828 } else if (!DAG.isKnownNeverNaN(RHS))
8831 Opcode = X86ISD::FMAX;
8834 // This can be a max if we can prove that at least one of the operands
8836 if (!FiniteOnlyFPMath()) {
8837 if (DAG.isKnownNeverNaN(RHS)) {
8838 // Put the potential NaN in the RHS so that SSE will preserve it.
8839 std::swap(LHS, RHS);
8840 } else if (!DAG.isKnownNeverNaN(LHS))
8843 Opcode = X86ISD::FMAX;
8846 // This can be a max, but if either operand is a NaN we need it to
8847 // preserve the original LHS.
8848 std::swap(LHS, RHS);
8852 Opcode = X86ISD::FMAX;
8855 // Check for x CC y ? y : x -- a min/max with reversed arms.
8856 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8860 // This can be a min if we can prove that at least one of the operands
8862 if (!FiniteOnlyFPMath()) {
8863 if (DAG.isKnownNeverNaN(RHS)) {
8864 // Put the potential NaN in the RHS so that SSE will preserve it.
8865 std::swap(LHS, RHS);
8866 } else if (!DAG.isKnownNeverNaN(LHS))
8869 Opcode = X86ISD::FMIN;
8872 // This can be a min if we can prove that at least one of the operands
8874 if (!FiniteOnlyFPMath()) {
8875 if (DAG.isKnownNeverNaN(LHS)) {
8876 // Put the potential NaN in the RHS so that SSE will preserve it.
8877 std::swap(LHS, RHS);
8878 } else if (!DAG.isKnownNeverNaN(RHS))
8881 Opcode = X86ISD::FMIN;
8884 // This can be a min, but if either operand is a NaN we need it to
8885 // preserve the original LHS.
8886 std::swap(LHS, RHS);
8890 Opcode = X86ISD::FMIN;
8894 // This can be a max if we can prove that at least one of the operands
8896 if (!FiniteOnlyFPMath()) {
8897 if (DAG.isKnownNeverNaN(LHS)) {
8898 // Put the potential NaN in the RHS so that SSE will preserve it.
8899 std::swap(LHS, RHS);
8900 } else if (!DAG.isKnownNeverNaN(RHS))
8903 Opcode = X86ISD::FMAX;
8906 // This can be a max if we can prove that at least one of the operands
8908 if (!FiniteOnlyFPMath()) {
8909 if (DAG.isKnownNeverNaN(RHS)) {
8910 // Put the potential NaN in the RHS so that SSE will preserve it.
8911 std::swap(LHS, RHS);
8912 } else if (!DAG.isKnownNeverNaN(LHS))
8915 Opcode = X86ISD::FMAX;
8918 // This can be a max, but if either operand is a NaN we need it to
8919 // preserve the original LHS.
8920 std::swap(LHS, RHS);
8924 Opcode = X86ISD::FMAX;
8930 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8933 // If this is a select between two integer constants, try to do some
8935 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8936 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8937 // Don't do this for crazy integer types.
8938 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8939 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8940 // so that TrueC (the true value) is larger than FalseC.
8941 bool NeedsCondInvert = false;
8943 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8944 // Efficiently invertible.
8945 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8946 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8947 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8948 NeedsCondInvert = true;
8949 std::swap(TrueC, FalseC);
8952 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8953 if (FalseC->getAPIntValue() == 0 &&
8954 TrueC->getAPIntValue().isPowerOf2()) {
8955 if (NeedsCondInvert) // Invert the condition if needed.
8956 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8957 DAG.getConstant(1, Cond.getValueType()));
8959 // Zero extend the condition if needed.
8960 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8962 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8963 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8964 DAG.getConstant(ShAmt, MVT::i8));
8967 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8968 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8969 if (NeedsCondInvert) // Invert the condition if needed.
8970 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8971 DAG.getConstant(1, Cond.getValueType()));
8973 // Zero extend the condition if needed.
8974 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8975 FalseC->getValueType(0), Cond);
8976 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8977 SDValue(FalseC, 0));
8980 // Optimize cases that will turn into an LEA instruction. This requires
8981 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8982 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8983 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8984 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8986 bool isFastMultiplier = false;
8988 switch ((unsigned char)Diff) {
8990 case 1: // result = add base, cond
8991 case 2: // result = lea base( , cond*2)
8992 case 3: // result = lea base(cond, cond*2)
8993 case 4: // result = lea base( , cond*4)
8994 case 5: // result = lea base(cond, cond*4)
8995 case 8: // result = lea base( , cond*8)
8996 case 9: // result = lea base(cond, cond*8)
8997 isFastMultiplier = true;
9002 if (isFastMultiplier) {
9003 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9004 if (NeedsCondInvert) // Invert the condition if needed.
9005 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9006 DAG.getConstant(1, Cond.getValueType()));
9008 // Zero extend the condition if needed.
9009 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9011 // Scale the condition by the difference.
9013 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9014 DAG.getConstant(Diff, Cond.getValueType()));
9016 // Add the base if non-zero.
9017 if (FalseC->getAPIntValue() != 0)
9018 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9019 SDValue(FalseC, 0));
9029 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9030 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9031 TargetLowering::DAGCombinerInfo &DCI) {
9032 DebugLoc DL = N->getDebugLoc();
9034 // If the flag operand isn't dead, don't touch this CMOV.
9035 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9038 // If this is a select between two integer constants, try to do some
9039 // optimizations. Note that the operands are ordered the opposite of SELECT
9041 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9042 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9043 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9044 // larger than FalseC (the false value).
9045 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9047 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9048 CC = X86::GetOppositeBranchCondition(CC);
9049 std::swap(TrueC, FalseC);
9052 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9053 // This is efficient for any integer data type (including i8/i16) and
9055 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9056 SDValue Cond = N->getOperand(3);
9057 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9058 DAG.getConstant(CC, MVT::i8), Cond);
9060 // Zero extend the condition if needed.
9061 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9063 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9064 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9065 DAG.getConstant(ShAmt, MVT::i8));
9066 if (N->getNumValues() == 2) // Dead flag value?
9067 return DCI.CombineTo(N, Cond, SDValue());
9071 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9072 // for any integer data type, including i8/i16.
9073 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9074 SDValue Cond = N->getOperand(3);
9075 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9076 DAG.getConstant(CC, MVT::i8), Cond);
9078 // Zero extend the condition if needed.
9079 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9080 FalseC->getValueType(0), Cond);
9081 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9082 SDValue(FalseC, 0));
9084 if (N->getNumValues() == 2) // Dead flag value?
9085 return DCI.CombineTo(N, Cond, SDValue());
9089 // Optimize cases that will turn into an LEA instruction. This requires
9090 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9091 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9092 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9093 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9095 bool isFastMultiplier = false;
9097 switch ((unsigned char)Diff) {
9099 case 1: // result = add base, cond
9100 case 2: // result = lea base( , cond*2)
9101 case 3: // result = lea base(cond, cond*2)
9102 case 4: // result = lea base( , cond*4)
9103 case 5: // result = lea base(cond, cond*4)
9104 case 8: // result = lea base( , cond*8)
9105 case 9: // result = lea base(cond, cond*8)
9106 isFastMultiplier = true;
9111 if (isFastMultiplier) {
9112 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9113 SDValue Cond = N->getOperand(3);
9114 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9115 DAG.getConstant(CC, MVT::i8), Cond);
9116 // Zero extend the condition if needed.
9117 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9119 // Scale the condition by the difference.
9121 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9122 DAG.getConstant(Diff, Cond.getValueType()));
9124 // Add the base if non-zero.
9125 if (FalseC->getAPIntValue() != 0)
9126 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9127 SDValue(FalseC, 0));
9128 if (N->getNumValues() == 2) // Dead flag value?
9129 return DCI.CombineTo(N, Cond, SDValue());
9139 /// PerformMulCombine - Optimize a single multiply with constant into two
9140 /// in order to implement it with two cheaper instructions, e.g.
9141 /// LEA + SHL, LEA + LEA.
9142 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9143 TargetLowering::DAGCombinerInfo &DCI) {
9144 if (DAG.getMachineFunction().
9145 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
9148 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9151 EVT VT = N->getValueType(0);
9155 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9158 uint64_t MulAmt = C->getZExtValue();
9159 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9162 uint64_t MulAmt1 = 0;
9163 uint64_t MulAmt2 = 0;
9164 if ((MulAmt % 9) == 0) {
9166 MulAmt2 = MulAmt / 9;
9167 } else if ((MulAmt % 5) == 0) {
9169 MulAmt2 = MulAmt / 5;
9170 } else if ((MulAmt % 3) == 0) {
9172 MulAmt2 = MulAmt / 3;
9175 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9176 DebugLoc DL = N->getDebugLoc();
9178 if (isPowerOf2_64(MulAmt2) &&
9179 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9180 // If second multiplifer is pow2, issue it first. We want the multiply by
9181 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9183 std::swap(MulAmt1, MulAmt2);
9186 if (isPowerOf2_64(MulAmt1))
9187 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9188 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9190 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9191 DAG.getConstant(MulAmt1, VT));
9193 if (isPowerOf2_64(MulAmt2))
9194 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9195 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9197 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9198 DAG.getConstant(MulAmt2, VT));
9200 // Do not add new nodes to DAG combiner worklist.
9201 DCI.CombineTo(N, NewMul, false);
9206 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9207 SDValue N0 = N->getOperand(0);
9208 SDValue N1 = N->getOperand(1);
9209 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9210 EVT VT = N0.getValueType();
9212 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9213 // since the result of setcc_c is all zero's or all ones.
9214 if (N1C && N0.getOpcode() == ISD::AND &&
9215 N0.getOperand(1).getOpcode() == ISD::Constant) {
9216 SDValue N00 = N0.getOperand(0);
9217 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9218 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9219 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9220 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9221 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9222 APInt ShAmt = N1C->getAPIntValue();
9223 Mask = Mask.shl(ShAmt);
9225 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9226 N00, DAG.getConstant(Mask, VT));
9233 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9235 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9236 const X86Subtarget *Subtarget) {
9237 EVT VT = N->getValueType(0);
9238 if (!VT.isVector() && VT.isInteger() &&
9239 N->getOpcode() == ISD::SHL)
9240 return PerformSHLCombine(N, DAG);
9242 // On X86 with SSE2 support, we can transform this to a vector shift if
9243 // all elements are shifted by the same amount. We can't do this in legalize
9244 // because the a constant vector is typically transformed to a constant pool
9245 // so we have no knowledge of the shift amount.
9246 if (!Subtarget->hasSSE2())
9249 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9252 SDValue ShAmtOp = N->getOperand(1);
9253 EVT EltVT = VT.getVectorElementType();
9254 DebugLoc DL = N->getDebugLoc();
9255 SDValue BaseShAmt = SDValue();
9256 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9257 unsigned NumElts = VT.getVectorNumElements();
9259 for (; i != NumElts; ++i) {
9260 SDValue Arg = ShAmtOp.getOperand(i);
9261 if (Arg.getOpcode() == ISD::UNDEF) continue;
9265 for (; i != NumElts; ++i) {
9266 SDValue Arg = ShAmtOp.getOperand(i);
9267 if (Arg.getOpcode() == ISD::UNDEF) continue;
9268 if (Arg != BaseShAmt) {
9272 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9273 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9274 SDValue InVec = ShAmtOp.getOperand(0);
9275 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9276 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9278 for (; i != NumElts; ++i) {
9279 SDValue Arg = InVec.getOperand(i);
9280 if (Arg.getOpcode() == ISD::UNDEF) continue;
9284 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9285 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9286 unsigned SplatIdx = cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9287 if (C->getZExtValue() == SplatIdx)
9288 BaseShAmt = InVec.getOperand(1);
9291 if (BaseShAmt.getNode() == 0)
9292 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9293 DAG.getIntPtrConstant(0));
9297 // The shift amount is an i32.
9298 if (EltVT.bitsGT(MVT::i32))
9299 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9300 else if (EltVT.bitsLT(MVT::i32))
9301 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9303 // The shift amount is identical so we can do a vector shift.
9304 SDValue ValOp = N->getOperand(0);
9305 switch (N->getOpcode()) {
9307 llvm_unreachable("Unknown shift opcode!");
9310 if (VT == MVT::v2i64)
9311 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9312 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9314 if (VT == MVT::v4i32)
9315 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9316 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9318 if (VT == MVT::v8i16)
9319 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9320 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9324 if (VT == MVT::v4i32)
9325 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9326 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9328 if (VT == MVT::v8i16)
9329 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9330 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9334 if (VT == MVT::v2i64)
9335 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9336 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9338 if (VT == MVT::v4i32)
9339 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9340 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9342 if (VT == MVT::v8i16)
9343 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9344 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9351 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9352 const X86Subtarget *Subtarget) {
9353 EVT VT = N->getValueType(0);
9354 if (VT != MVT::i64 || !Subtarget->is64Bit())
9357 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9358 SDValue N0 = N->getOperand(0);
9359 SDValue N1 = N->getOperand(1);
9360 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9362 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9365 SDValue ShAmt0 = N0.getOperand(1);
9366 if (ShAmt0.getValueType() != MVT::i8)
9368 SDValue ShAmt1 = N1.getOperand(1);
9369 if (ShAmt1.getValueType() != MVT::i8)
9371 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9372 ShAmt0 = ShAmt0.getOperand(0);
9373 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9374 ShAmt1 = ShAmt1.getOperand(0);
9376 DebugLoc DL = N->getDebugLoc();
9377 unsigned Opc = X86ISD::SHLD;
9378 SDValue Op0 = N0.getOperand(0);
9379 SDValue Op1 = N1.getOperand(0);
9380 if (ShAmt0.getOpcode() == ISD::SUB) {
9382 std::swap(Op0, Op1);
9383 std::swap(ShAmt0, ShAmt1);
9386 if (ShAmt1.getOpcode() == ISD::SUB) {
9387 SDValue Sum = ShAmt1.getOperand(0);
9388 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9389 if (SumC->getSExtValue() == 64 &&
9390 ShAmt1.getOperand(1) == ShAmt0)
9391 return DAG.getNode(Opc, DL, VT,
9393 DAG.getNode(ISD::TRUNCATE, DL,
9396 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9397 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9399 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
9400 return DAG.getNode(Opc, DL, VT,
9401 N0.getOperand(0), N1.getOperand(0),
9402 DAG.getNode(ISD::TRUNCATE, DL,
9409 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9410 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9411 const X86Subtarget *Subtarget) {
9412 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9413 // the FP state in cases where an emms may be missing.
9414 // A preferable solution to the general problem is to figure out the right
9415 // places to insert EMMS. This qualifies as a quick hack.
9417 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9418 StoreSDNode *St = cast<StoreSDNode>(N);
9419 EVT VT = St->getValue().getValueType();
9420 if (VT.getSizeInBits() != 64)
9423 const Function *F = DAG.getMachineFunction().getFunction();
9424 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9425 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9426 && Subtarget->hasSSE2();
9427 if ((VT.isVector() ||
9428 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9429 isa<LoadSDNode>(St->getValue()) &&
9430 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9431 St->getChain().hasOneUse() && !St->isVolatile()) {
9432 SDNode* LdVal = St->getValue().getNode();
9434 int TokenFactorIndex = -1;
9435 SmallVector<SDValue, 8> Ops;
9436 SDNode* ChainVal = St->getChain().getNode();
9437 // Must be a store of a load. We currently handle two cases: the load
9438 // is a direct child, and it's under an intervening TokenFactor. It is
9439 // possible to dig deeper under nested TokenFactors.
9440 if (ChainVal == LdVal)
9441 Ld = cast<LoadSDNode>(St->getChain());
9442 else if (St->getValue().hasOneUse() &&
9443 ChainVal->getOpcode() == ISD::TokenFactor) {
9444 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9445 if (ChainVal->getOperand(i).getNode() == LdVal) {
9446 TokenFactorIndex = i;
9447 Ld = cast<LoadSDNode>(St->getValue());
9449 Ops.push_back(ChainVal->getOperand(i));
9453 if (!Ld || !ISD::isNormalLoad(Ld))
9456 // If this is not the MMX case, i.e. we are just turning i64 load/store
9457 // into f64 load/store, avoid the transformation if there are multiple
9458 // uses of the loaded value.
9459 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9462 DebugLoc LdDL = Ld->getDebugLoc();
9463 DebugLoc StDL = N->getDebugLoc();
9464 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9465 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9467 if (Subtarget->is64Bit() || F64IsLegal) {
9468 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9469 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9470 Ld->getBasePtr(), Ld->getSrcValue(),
9471 Ld->getSrcValueOffset(), Ld->isVolatile(),
9472 Ld->getAlignment());
9473 SDValue NewChain = NewLd.getValue(1);
9474 if (TokenFactorIndex != -1) {
9475 Ops.push_back(NewChain);
9476 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9479 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9480 St->getSrcValue(), St->getSrcValueOffset(),
9481 St->isVolatile(), St->getAlignment());
9484 // Otherwise, lower to two pairs of 32-bit loads / stores.
9485 SDValue LoAddr = Ld->getBasePtr();
9486 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9487 DAG.getConstant(4, MVT::i32));
9489 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9490 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9491 Ld->isVolatile(), Ld->getAlignment());
9492 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9493 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9495 MinAlign(Ld->getAlignment(), 4));
9497 SDValue NewChain = LoLd.getValue(1);
9498 if (TokenFactorIndex != -1) {
9499 Ops.push_back(LoLd);
9500 Ops.push_back(HiLd);
9501 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9505 LoAddr = St->getBasePtr();
9506 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9507 DAG.getConstant(4, MVT::i32));
9509 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9510 St->getSrcValue(), St->getSrcValueOffset(),
9511 St->isVolatile(), St->getAlignment());
9512 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9514 St->getSrcValueOffset() + 4,
9516 MinAlign(St->getAlignment(), 4));
9517 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9522 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9523 /// X86ISD::FXOR nodes.
9524 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9525 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9526 // F[X]OR(0.0, x) -> x
9527 // F[X]OR(x, 0.0) -> x
9528 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9529 if (C->getValueAPF().isPosZero())
9530 return N->getOperand(1);
9531 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9532 if (C->getValueAPF().isPosZero())
9533 return N->getOperand(0);
9537 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9538 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9539 // FAND(0.0, x) -> 0.0
9540 // FAND(x, 0.0) -> 0.0
9541 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9542 if (C->getValueAPF().isPosZero())
9543 return N->getOperand(0);
9544 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9545 if (C->getValueAPF().isPosZero())
9546 return N->getOperand(1);
9550 static SDValue PerformBTCombine(SDNode *N,
9552 TargetLowering::DAGCombinerInfo &DCI) {
9553 // BT ignores high bits in the bit index operand.
9554 SDValue Op1 = N->getOperand(1);
9555 if (Op1.hasOneUse()) {
9556 unsigned BitWidth = Op1.getValueSizeInBits();
9557 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9558 APInt KnownZero, KnownOne;
9559 TargetLowering::TargetLoweringOpt TLO(DAG);
9560 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9561 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9562 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9563 DCI.CommitTargetLoweringOpt(TLO);
9568 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9569 SDValue Op = N->getOperand(0);
9570 if (Op.getOpcode() == ISD::BIT_CONVERT)
9571 Op = Op.getOperand(0);
9572 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9573 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9574 VT.getVectorElementType().getSizeInBits() ==
9575 OpVT.getVectorElementType().getSizeInBits()) {
9576 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9581 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9582 // Locked instructions, in turn, have implicit fence semantics (all memory
9583 // operations are flushed before issuing the locked instruction, and the
9584 // are not buffered), so we can fold away the common pattern of
9585 // fence-atomic-fence.
9586 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9587 SDValue atomic = N->getOperand(0);
9588 switch (atomic.getOpcode()) {
9589 case ISD::ATOMIC_CMP_SWAP:
9590 case ISD::ATOMIC_SWAP:
9591 case ISD::ATOMIC_LOAD_ADD:
9592 case ISD::ATOMIC_LOAD_SUB:
9593 case ISD::ATOMIC_LOAD_AND:
9594 case ISD::ATOMIC_LOAD_OR:
9595 case ISD::ATOMIC_LOAD_XOR:
9596 case ISD::ATOMIC_LOAD_NAND:
9597 case ISD::ATOMIC_LOAD_MIN:
9598 case ISD::ATOMIC_LOAD_MAX:
9599 case ISD::ATOMIC_LOAD_UMIN:
9600 case ISD::ATOMIC_LOAD_UMAX:
9606 SDValue fence = atomic.getOperand(0);
9607 if (fence.getOpcode() != ISD::MEMBARRIER)
9610 switch (atomic.getOpcode()) {
9611 case ISD::ATOMIC_CMP_SWAP:
9612 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9613 atomic.getOperand(1), atomic.getOperand(2),
9614 atomic.getOperand(3));
9615 case ISD::ATOMIC_SWAP:
9616 case ISD::ATOMIC_LOAD_ADD:
9617 case ISD::ATOMIC_LOAD_SUB:
9618 case ISD::ATOMIC_LOAD_AND:
9619 case ISD::ATOMIC_LOAD_OR:
9620 case ISD::ATOMIC_LOAD_XOR:
9621 case ISD::ATOMIC_LOAD_NAND:
9622 case ISD::ATOMIC_LOAD_MIN:
9623 case ISD::ATOMIC_LOAD_MAX:
9624 case ISD::ATOMIC_LOAD_UMIN:
9625 case ISD::ATOMIC_LOAD_UMAX:
9626 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9627 atomic.getOperand(1), atomic.getOperand(2));
9633 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9634 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9635 // (and (i32 x86isd::setcc_carry), 1)
9636 // This eliminates the zext. This transformation is necessary because
9637 // ISD::SETCC is always legalized to i8.
9638 DebugLoc dl = N->getDebugLoc();
9639 SDValue N0 = N->getOperand(0);
9640 EVT VT = N->getValueType(0);
9641 if (N0.getOpcode() == ISD::AND &&
9643 N0.getOperand(0).hasOneUse()) {
9644 SDValue N00 = N0.getOperand(0);
9645 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9647 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9648 if (!C || C->getZExtValue() != 1)
9650 return DAG.getNode(ISD::AND, dl, VT,
9651 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9652 N00.getOperand(0), N00.getOperand(1)),
9653 DAG.getConstant(1, VT));
9659 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9660 DAGCombinerInfo &DCI) const {
9661 SelectionDAG &DAG = DCI.DAG;
9662 switch (N->getOpcode()) {
9664 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9665 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9666 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9667 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9670 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9671 case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
9672 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9674 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9675 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9676 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9677 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9678 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9679 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9685 //===----------------------------------------------------------------------===//
9686 // X86 Inline Assembly Support
9687 //===----------------------------------------------------------------------===//
9689 static bool LowerToBSwap(CallInst *CI) {
9690 // FIXME: this should verify that we are targetting a 486 or better. If not,
9691 // we will turn this bswap into something that will be lowered to logical ops
9692 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9693 // so don't worry about this.
9695 // Verify this is a simple bswap.
9696 if (CI->getNumOperands() != 2 ||
9697 CI->getType() != CI->getOperand(1)->getType() ||
9698 !CI->getType()->isInteger())
9701 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9702 if (!Ty || Ty->getBitWidth() % 16 != 0)
9705 // Okay, we can do this xform, do so now.
9706 const Type *Tys[] = { Ty };
9707 Module *M = CI->getParent()->getParent()->getParent();
9708 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9710 Value *Op = CI->getOperand(1);
9711 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9713 CI->replaceAllUsesWith(Op);
9714 CI->eraseFromParent();
9718 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9719 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9720 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9722 std::string AsmStr = IA->getAsmString();
9724 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9725 SmallVector<StringRef, 4> AsmPieces;
9726 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9728 switch (AsmPieces.size()) {
9729 default: return false;
9731 AsmStr = AsmPieces[0];
9733 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9736 if (AsmPieces.size() == 2 &&
9737 (AsmPieces[0] == "bswap" ||
9738 AsmPieces[0] == "bswapq" ||
9739 AsmPieces[0] == "bswapl") &&
9740 (AsmPieces[1] == "$0" ||
9741 AsmPieces[1] == "${0:q}")) {
9742 // No need to check constraints, nothing other than the equivalent of
9743 // "=r,0" would be valid here.
9744 return LowerToBSwap(CI);
9746 // rorw $$8, ${0:w} --> llvm.bswap.i16
9747 if (CI->getType()->isInteger(16) &&
9748 AsmPieces.size() == 3 &&
9749 AsmPieces[0] == "rorw" &&
9750 AsmPieces[1] == "$$8," &&
9751 AsmPieces[2] == "${0:w}" &&
9752 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
9753 return LowerToBSwap(CI);
9757 if (CI->getType()->isInteger(64) &&
9758 Constraints.size() >= 2 &&
9759 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9760 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9761 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9762 SmallVector<StringRef, 4> Words;
9763 SplitString(AsmPieces[0], Words, " \t");
9764 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9766 SplitString(AsmPieces[1], Words, " \t");
9767 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9769 SplitString(AsmPieces[2], Words, " \t,");
9770 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9771 Words[2] == "%edx") {
9772 return LowerToBSwap(CI);
9784 /// getConstraintType - Given a constraint letter, return the type of
9785 /// constraint it is for this target.
9786 X86TargetLowering::ConstraintType
9787 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
9788 if (Constraint.size() == 1) {
9789 switch (Constraint[0]) {
9801 return C_RegisterClass;
9809 return TargetLowering::getConstraintType(Constraint);
9812 /// LowerXConstraint - try to replace an X constraint, which matches anything,
9813 /// with another that has more specific requirements based on the type of the
9814 /// corresponding operand.
9815 const char *X86TargetLowering::
9816 LowerXConstraint(EVT ConstraintVT) const {
9817 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
9818 // 'f' like normal targets.
9819 if (ConstraintVT.isFloatingPoint()) {
9820 if (Subtarget->hasSSE2())
9822 if (Subtarget->hasSSE1())
9826 return TargetLowering::LowerXConstraint(ConstraintVT);
9829 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9830 /// vector. If it is invalid, don't add anything to Ops.
9831 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9834 std::vector<SDValue>&Ops,
9835 SelectionDAG &DAG) const {
9836 SDValue Result(0, 0);
9838 switch (Constraint) {
9841 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9842 if (C->getZExtValue() <= 31) {
9843 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9849 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9850 if (C->getZExtValue() <= 63) {
9851 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9857 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9858 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9859 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9865 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9866 if (C->getZExtValue() <= 255) {
9867 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9873 // 32-bit signed value
9874 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9875 const ConstantInt *CI = C->getConstantIntValue();
9876 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9877 C->getSExtValue())) {
9878 // Widen to 64 bits here to get it sign extended.
9879 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9882 // FIXME gcc accepts some relocatable values here too, but only in certain
9883 // memory models; it's complicated.
9888 // 32-bit unsigned value
9889 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9890 const ConstantInt *CI = C->getConstantIntValue();
9891 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9892 C->getZExtValue())) {
9893 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9897 // FIXME gcc accepts some relocatable values here too, but only in certain
9898 // memory models; it's complicated.
9902 // Literal immediates are always ok.
9903 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
9904 // Widen to 64 bits here to get it sign extended.
9905 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
9909 // If we are in non-pic codegen mode, we allow the address of a global (with
9910 // an optional displacement) to be used with 'i'.
9911 GlobalAddressSDNode *GA = 0;
9914 // Match either (GA), (GA+C), (GA+C1+C2), etc.
9916 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
9917 Offset += GA->getOffset();
9919 } else if (Op.getOpcode() == ISD::ADD) {
9920 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9921 Offset += C->getZExtValue();
9922 Op = Op.getOperand(0);
9925 } else if (Op.getOpcode() == ISD::SUB) {
9926 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9927 Offset += -C->getZExtValue();
9928 Op = Op.getOperand(0);
9933 // Otherwise, this isn't something we can handle, reject it.
9937 GlobalValue *GV = GA->getGlobal();
9938 // If we require an extra load to get this address, as in PIC mode, we
9940 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
9941 getTargetMachine())))
9945 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
9947 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
9953 if (Result.getNode()) {
9954 Ops.push_back(Result);
9957 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
9961 std::vector<unsigned> X86TargetLowering::
9962 getRegClassForInlineAsmConstraint(const std::string &Constraint,
9964 if (Constraint.size() == 1) {
9965 // FIXME: not handling fp-stack yet!
9966 switch (Constraint[0]) { // GCC X86 Constraint Letters
9967 default: break; // Unknown constraint letter
9968 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
9969 if (Subtarget->is64Bit()) {
9971 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
9972 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
9973 X86::R10D,X86::R11D,X86::R12D,
9974 X86::R13D,X86::R14D,X86::R15D,
9975 X86::EBP, X86::ESP, 0);
9976 else if (VT == MVT::i16)
9977 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
9978 X86::SI, X86::DI, X86::R8W,X86::R9W,
9979 X86::R10W,X86::R11W,X86::R12W,
9980 X86::R13W,X86::R14W,X86::R15W,
9981 X86::BP, X86::SP, 0);
9982 else if (VT == MVT::i8)
9983 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
9984 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
9985 X86::R10B,X86::R11B,X86::R12B,
9986 X86::R13B,X86::R14B,X86::R15B,
9987 X86::BPL, X86::SPL, 0);
9989 else if (VT == MVT::i64)
9990 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
9991 X86::RSI, X86::RDI, X86::R8, X86::R9,
9992 X86::R10, X86::R11, X86::R12,
9993 X86::R13, X86::R14, X86::R15,
9994 X86::RBP, X86::RSP, 0);
9998 // 32-bit fallthrough
10000 if (VT == MVT::i32)
10001 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10002 else if (VT == MVT::i16)
10003 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10004 else if (VT == MVT::i8)
10005 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10006 else if (VT == MVT::i64)
10007 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10012 return std::vector<unsigned>();
10015 std::pair<unsigned, const TargetRegisterClass*>
10016 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10018 // First, see if this is a constraint that directly corresponds to an LLVM
10020 if (Constraint.size() == 1) {
10021 // GCC Constraint Letters
10022 switch (Constraint[0]) {
10024 case 'r': // GENERAL_REGS
10025 case 'l': // INDEX_REGS
10027 return std::make_pair(0U, X86::GR8RegisterClass);
10028 if (VT == MVT::i16)
10029 return std::make_pair(0U, X86::GR16RegisterClass);
10030 if (VT == MVT::i32 || !Subtarget->is64Bit())
10031 return std::make_pair(0U, X86::GR32RegisterClass);
10032 return std::make_pair(0U, X86::GR64RegisterClass);
10033 case 'R': // LEGACY_REGS
10035 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10036 if (VT == MVT::i16)
10037 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10038 if (VT == MVT::i32 || !Subtarget->is64Bit())
10039 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10040 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10041 case 'f': // FP Stack registers.
10042 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10043 // value to the correct fpstack register class.
10044 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10045 return std::make_pair(0U, X86::RFP32RegisterClass);
10046 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10047 return std::make_pair(0U, X86::RFP64RegisterClass);
10048 return std::make_pair(0U, X86::RFP80RegisterClass);
10049 case 'y': // MMX_REGS if MMX allowed.
10050 if (!Subtarget->hasMMX()) break;
10051 return std::make_pair(0U, X86::VR64RegisterClass);
10052 case 'Y': // SSE_REGS if SSE2 allowed
10053 if (!Subtarget->hasSSE2()) break;
10055 case 'x': // SSE_REGS if SSE1 allowed
10056 if (!Subtarget->hasSSE1()) break;
10058 switch (VT.getSimpleVT().SimpleTy) {
10060 // Scalar SSE types.
10063 return std::make_pair(0U, X86::FR32RegisterClass);
10066 return std::make_pair(0U, X86::FR64RegisterClass);
10074 return std::make_pair(0U, X86::VR128RegisterClass);
10080 // Use the default implementation in TargetLowering to convert the register
10081 // constraint into a member of a register class.
10082 std::pair<unsigned, const TargetRegisterClass*> Res;
10083 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10085 // Not found as a standard register?
10086 if (Res.second == 0) {
10087 // Map st(0) -> st(7) -> ST0
10088 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10089 tolower(Constraint[1]) == 's' &&
10090 tolower(Constraint[2]) == 't' &&
10091 Constraint[3] == '(' &&
10092 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10093 Constraint[5] == ')' &&
10094 Constraint[6] == '}') {
10096 Res.first = X86::ST0+Constraint[4]-'0';
10097 Res.second = X86::RFP80RegisterClass;
10101 // GCC allows "st(0)" to be called just plain "st".
10102 if (StringRef("{st}").equals_lower(Constraint)) {
10103 Res.first = X86::ST0;
10104 Res.second = X86::RFP80RegisterClass;
10109 if (StringRef("{flags}").equals_lower(Constraint)) {
10110 Res.first = X86::EFLAGS;
10111 Res.second = X86::CCRRegisterClass;
10115 // 'A' means EAX + EDX.
10116 if (Constraint == "A") {
10117 Res.first = X86::EAX;
10118 Res.second = X86::GR32_ADRegisterClass;
10124 // Otherwise, check to see if this is a register class of the wrong value
10125 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10126 // turn into {ax},{dx}.
10127 if (Res.second->hasType(VT))
10128 return Res; // Correct type already, nothing to do.
10130 // All of the single-register GCC register classes map their values onto
10131 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10132 // really want an 8-bit or 32-bit register, map to the appropriate register
10133 // class and return the appropriate register.
10134 if (Res.second == X86::GR16RegisterClass) {
10135 if (VT == MVT::i8) {
10136 unsigned DestReg = 0;
10137 switch (Res.first) {
10139 case X86::AX: DestReg = X86::AL; break;
10140 case X86::DX: DestReg = X86::DL; break;
10141 case X86::CX: DestReg = X86::CL; break;
10142 case X86::BX: DestReg = X86::BL; break;
10145 Res.first = DestReg;
10146 Res.second = X86::GR8RegisterClass;
10148 } else if (VT == MVT::i32) {
10149 unsigned DestReg = 0;
10150 switch (Res.first) {
10152 case X86::AX: DestReg = X86::EAX; break;
10153 case X86::DX: DestReg = X86::EDX; break;
10154 case X86::CX: DestReg = X86::ECX; break;
10155 case X86::BX: DestReg = X86::EBX; break;
10156 case X86::SI: DestReg = X86::ESI; break;
10157 case X86::DI: DestReg = X86::EDI; break;
10158 case X86::BP: DestReg = X86::EBP; break;
10159 case X86::SP: DestReg = X86::ESP; break;
10162 Res.first = DestReg;
10163 Res.second = X86::GR32RegisterClass;
10165 } else if (VT == MVT::i64) {
10166 unsigned DestReg = 0;
10167 switch (Res.first) {
10169 case X86::AX: DestReg = X86::RAX; break;
10170 case X86::DX: DestReg = X86::RDX; break;
10171 case X86::CX: DestReg = X86::RCX; break;
10172 case X86::BX: DestReg = X86::RBX; break;
10173 case X86::SI: DestReg = X86::RSI; break;
10174 case X86::DI: DestReg = X86::RDI; break;
10175 case X86::BP: DestReg = X86::RBP; break;
10176 case X86::SP: DestReg = X86::RSP; break;
10179 Res.first = DestReg;
10180 Res.second = X86::GR64RegisterClass;
10183 } else if (Res.second == X86::FR32RegisterClass ||
10184 Res.second == X86::FR64RegisterClass ||
10185 Res.second == X86::VR128RegisterClass) {
10186 // Handle references to XMM physical registers that got mapped into the
10187 // wrong class. This can happen with constraints like {xmm0} where the
10188 // target independent register mapper will just pick the first match it can
10189 // find, ignoring the required type.
10190 if (VT == MVT::f32)
10191 Res.second = X86::FR32RegisterClass;
10192 else if (VT == MVT::f64)
10193 Res.second = X86::FR64RegisterClass;
10194 else if (X86::VR128RegisterClass->hasType(VT))
10195 Res.second = X86::VR128RegisterClass;
10201 //===----------------------------------------------------------------------===//
10202 // X86 Widen vector type
10203 //===----------------------------------------------------------------------===//
10205 /// getWidenVectorType: given a vector type, returns the type to widen
10206 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
10207 /// If there is no vector type that we want to widen to, returns MVT::Other
10208 /// When and where to widen is target dependent based on the cost of
10209 /// scalarizing vs using the wider vector type.
10211 EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
10212 assert(VT.isVector());
10213 if (isTypeLegal(VT))
10216 // TODO: In computeRegisterProperty, we can compute the list of legal vector
10217 // type based on element type. This would speed up our search (though
10218 // it may not be worth it since the size of the list is relatively
10220 EVT EltVT = VT.getVectorElementType();
10221 unsigned NElts = VT.getVectorNumElements();
10223 // On X86, it make sense to widen any vector wider than 1
10227 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
10228 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
10229 EVT SVT = (MVT::SimpleValueType)nVT;
10231 if (isTypeLegal(SVT) &&
10232 SVT.getVectorElementType() == EltVT &&
10233 SVT.getVectorNumElements() > NElts)