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 "X86MCTargetExpr.h"
20 #include "X86TargetMachine.h"
21 #include "X86TargetObjectFile.h"
22 #include "llvm/CallingConv.h"
23 #include "llvm/Constants.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/GlobalAlias.h"
26 #include "llvm/GlobalVariable.h"
27 #include "llvm/Function.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/Intrinsics.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineJumpTableInfo.h"
35 #include "llvm/CodeGen/MachineModuleInfo.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/PseudoSourceValue.h"
38 #include "llvm/MC/MCAsmInfo.h"
39 #include "llvm/MC/MCContext.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 // FIXME: These should be based on subtarget info. Plus, the values should
1005 // be smaller when we are in optimizing for size mode.
1006 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1007 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
1008 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1009 setPrefLoopAlignment(16);
1010 benefitFromCodePlacementOpt = true;
1014 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1019 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1020 /// the desired ByVal argument alignment.
1021 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1024 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1025 if (VTy->getBitWidth() == 128)
1027 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1028 unsigned EltAlign = 0;
1029 getMaxByValAlign(ATy->getElementType(), EltAlign);
1030 if (EltAlign > MaxAlign)
1031 MaxAlign = EltAlign;
1032 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1033 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1034 unsigned EltAlign = 0;
1035 getMaxByValAlign(STy->getElementType(i), EltAlign);
1036 if (EltAlign > MaxAlign)
1037 MaxAlign = EltAlign;
1045 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1046 /// function arguments in the caller parameter area. For X86, aggregates
1047 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1048 /// are at 4-byte boundaries.
1049 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1050 if (Subtarget->is64Bit()) {
1051 // Max of 8 and alignment of type.
1052 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1059 if (Subtarget->hasSSE1())
1060 getMaxByValAlign(Ty, Align);
1064 /// getOptimalMemOpType - Returns the target specific optimal type for load
1065 /// and store operations as a result of memset, memcpy, and memmove
1066 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
1069 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1070 bool isSrcConst, bool isSrcStr,
1071 SelectionDAG &DAG) const {
1072 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1073 // linux. This is because the stack realignment code can't handle certain
1074 // cases like PR2962. This should be removed when PR2962 is fixed.
1075 const Function *F = DAG.getMachineFunction().getFunction();
1076 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1077 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1078 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1080 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1083 if (Subtarget->is64Bit() && Size >= 8)
1088 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1089 /// current function. The returned value is a member of the
1090 /// MachineJumpTableInfo::JTEntryKind enum.
1091 unsigned X86TargetLowering::getJumpTableEncoding() const {
1092 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1094 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1095 Subtarget->isPICStyleGOT())
1096 return MachineJumpTableInfo::EK_Custom32;
1098 // Otherwise, use the normal jump table encoding heuristics.
1099 return TargetLowering::getJumpTableEncoding();
1102 /// getPICBaseSymbol - Return the X86-32 PIC base.
1104 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1105 MCContext &Ctx) const {
1106 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1107 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1108 Twine(MF->getFunctionNumber())+"$pb");
1113 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1114 const MachineBasicBlock *MBB,
1115 unsigned uid,MCContext &Ctx) const{
1116 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1117 Subtarget->isPICStyleGOT());
1118 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1120 return X86MCTargetExpr::Create(MBB->getSymbol(Ctx),
1121 X86MCTargetExpr::GOTOFF, Ctx);
1124 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1126 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1127 SelectionDAG &DAG) const {
1128 if (!Subtarget->is64Bit())
1129 // This doesn't have DebugLoc associated with it, but is not really the
1130 // same as a Register.
1131 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1136 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1137 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1139 const MCExpr *X86TargetLowering::
1140 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1141 MCContext &Ctx) const {
1142 // X86-64 uses RIP relative addressing based on the jump table label.
1143 if (Subtarget->isPICStyleRIPRel())
1144 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1146 // Otherwise, the reference is relative to the PIC base.
1147 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1150 /// getFunctionAlignment - Return the Log2 alignment of this function.
1151 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1152 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1155 //===----------------------------------------------------------------------===//
1156 // Return Value Calling Convention Implementation
1157 //===----------------------------------------------------------------------===//
1159 #include "X86GenCallingConv.inc"
1162 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1163 const SmallVectorImpl<EVT> &OutTys,
1164 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1165 SelectionDAG &DAG) {
1166 SmallVector<CCValAssign, 16> RVLocs;
1167 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1168 RVLocs, *DAG.getContext());
1169 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1173 X86TargetLowering::LowerReturn(SDValue Chain,
1174 CallingConv::ID CallConv, bool isVarArg,
1175 const SmallVectorImpl<ISD::OutputArg> &Outs,
1176 DebugLoc dl, SelectionDAG &DAG) {
1178 SmallVector<CCValAssign, 16> RVLocs;
1179 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1180 RVLocs, *DAG.getContext());
1181 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1183 // Add the regs to the liveout set for the function.
1184 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1185 for (unsigned i = 0; i != RVLocs.size(); ++i)
1186 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1187 MRI.addLiveOut(RVLocs[i].getLocReg());
1191 SmallVector<SDValue, 6> RetOps;
1192 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1193 // Operand #1 = Bytes To Pop
1194 RetOps.push_back(DAG.getTargetConstant(getBytesToPopOnReturn(), MVT::i16));
1196 // Copy the result values into the output registers.
1197 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1198 CCValAssign &VA = RVLocs[i];
1199 assert(VA.isRegLoc() && "Can only return in registers!");
1200 SDValue ValToCopy = Outs[i].Val;
1202 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1203 // the RET instruction and handled by the FP Stackifier.
1204 if (VA.getLocReg() == X86::ST0 ||
1205 VA.getLocReg() == X86::ST1) {
1206 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1207 // change the value to the FP stack register class.
1208 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1209 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1210 RetOps.push_back(ValToCopy);
1211 // Don't emit a copytoreg.
1215 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1216 // which is returned in RAX / RDX.
1217 if (Subtarget->is64Bit()) {
1218 EVT ValVT = ValToCopy.getValueType();
1219 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1220 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1221 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1222 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1226 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1227 Flag = Chain.getValue(1);
1230 // The x86-64 ABI for returning structs by value requires that we copy
1231 // the sret argument into %rax for the return. We saved the argument into
1232 // a virtual register in the entry block, so now we copy the value out
1234 if (Subtarget->is64Bit() &&
1235 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1236 MachineFunction &MF = DAG.getMachineFunction();
1237 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1238 unsigned Reg = FuncInfo->getSRetReturnReg();
1240 Reg = MRI.createVirtualRegister(getRegClassFor(MVT::i64));
1241 FuncInfo->setSRetReturnReg(Reg);
1243 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1245 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1246 Flag = Chain.getValue(1);
1248 // RAX now acts like a return value.
1249 MRI.addLiveOut(X86::RAX);
1252 RetOps[0] = Chain; // Update chain.
1254 // Add the flag if we have it.
1256 RetOps.push_back(Flag);
1258 return DAG.getNode(X86ISD::RET_FLAG, dl,
1259 MVT::Other, &RetOps[0], RetOps.size());
1262 /// LowerCallResult - Lower the result values of a call into the
1263 /// appropriate copies out of appropriate physical registers.
1266 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1267 CallingConv::ID CallConv, bool isVarArg,
1268 const SmallVectorImpl<ISD::InputArg> &Ins,
1269 DebugLoc dl, SelectionDAG &DAG,
1270 SmallVectorImpl<SDValue> &InVals) {
1272 // Assign locations to each value returned by this call.
1273 SmallVector<CCValAssign, 16> RVLocs;
1274 bool Is64Bit = Subtarget->is64Bit();
1275 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1276 RVLocs, *DAG.getContext());
1277 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1279 // Copy all of the result registers out of their specified physreg.
1280 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1281 CCValAssign &VA = RVLocs[i];
1282 EVT CopyVT = VA.getValVT();
1284 // If this is x86-64, and we disabled SSE, we can't return FP values
1285 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1286 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1287 llvm_report_error("SSE register return with SSE disabled");
1290 // If this is a call to a function that returns an fp value on the floating
1291 // point stack, but where we prefer to use the value in xmm registers, copy
1292 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1293 if ((VA.getLocReg() == X86::ST0 ||
1294 VA.getLocReg() == X86::ST1) &&
1295 isScalarFPTypeInSSEReg(VA.getValVT())) {
1300 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1301 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1302 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1303 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1304 MVT::v2i64, InFlag).getValue(1);
1305 Val = Chain.getValue(0);
1306 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1307 Val, DAG.getConstant(0, MVT::i64));
1309 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1310 MVT::i64, InFlag).getValue(1);
1311 Val = Chain.getValue(0);
1313 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1315 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1316 CopyVT, InFlag).getValue(1);
1317 Val = Chain.getValue(0);
1319 InFlag = Chain.getValue(2);
1321 if (CopyVT != VA.getValVT()) {
1322 // Round the F80 the right size, which also moves to the appropriate xmm
1324 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1325 // This truncation won't change the value.
1326 DAG.getIntPtrConstant(1));
1329 InVals.push_back(Val);
1336 //===----------------------------------------------------------------------===//
1337 // C & StdCall & Fast Calling Convention implementation
1338 //===----------------------------------------------------------------------===//
1339 // StdCall calling convention seems to be standard for many Windows' API
1340 // routines and around. It differs from C calling convention just a little:
1341 // callee should clean up the stack, not caller. Symbols should be also
1342 // decorated in some fancy way :) It doesn't support any vector arguments.
1343 // For info on fast calling convention see Fast Calling Convention (tail call)
1344 // implementation LowerX86_32FastCCCallTo.
1346 /// CallIsStructReturn - Determines whether a call uses struct return
1348 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1352 return Outs[0].Flags.isSRet();
1355 /// ArgsAreStructReturn - Determines whether a function uses struct
1356 /// return semantics.
1358 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1362 return Ins[0].Flags.isSRet();
1365 /// IsCalleePop - Determines whether the callee is required to pop its
1366 /// own arguments. Callee pop is necessary to support tail calls.
1367 bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
1371 switch (CallingConv) {
1374 case CallingConv::X86_StdCall:
1375 return !Subtarget->is64Bit();
1376 case CallingConv::X86_FastCall:
1377 return !Subtarget->is64Bit();
1378 case CallingConv::Fast:
1379 return GuaranteedTailCallOpt;
1383 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1384 /// given CallingConvention value.
1385 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1386 if (Subtarget->is64Bit()) {
1387 if (Subtarget->isTargetWin64())
1388 return CC_X86_Win64_C;
1393 if (CC == CallingConv::X86_FastCall)
1394 return CC_X86_32_FastCall;
1395 else if (CC == CallingConv::Fast)
1396 return CC_X86_32_FastCC;
1401 /// NameDecorationForCallConv - Selects the appropriate decoration to
1402 /// apply to a MachineFunction containing a given calling convention.
1404 X86TargetLowering::NameDecorationForCallConv(CallingConv::ID CallConv) {
1405 if (CallConv == CallingConv::X86_FastCall)
1407 else if (CallConv == CallingConv::X86_StdCall)
1413 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1414 /// by "Src" to address "Dst" with size and alignment information specified by
1415 /// the specific parameter attribute. The copy will be passed as a byval
1416 /// function parameter.
1418 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1419 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1421 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1422 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1423 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1426 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1427 /// a tailcall target by changing its ABI.
1428 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1429 return GuaranteedTailCallOpt && CC == CallingConv::Fast;
1433 X86TargetLowering::LowerMemArgument(SDValue Chain,
1434 CallingConv::ID CallConv,
1435 const SmallVectorImpl<ISD::InputArg> &Ins,
1436 DebugLoc dl, SelectionDAG &DAG,
1437 const CCValAssign &VA,
1438 MachineFrameInfo *MFI,
1440 // Create the nodes corresponding to a load from this parameter slot.
1441 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1442 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1443 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1446 // If value is passed by pointer we have address passed instead of the value
1448 if (VA.getLocInfo() == CCValAssign::Indirect)
1449 ValVT = VA.getLocVT();
1451 ValVT = VA.getValVT();
1453 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1454 // changed with more analysis.
1455 // In case of tail call optimization mark all arguments mutable. Since they
1456 // could be overwritten by lowering of arguments in case of a tail call.
1457 if (Flags.isByVal()) {
1458 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1459 VA.getLocMemOffset(), isImmutable, false);
1460 return DAG.getFrameIndex(FI, getPointerTy());
1462 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1463 VA.getLocMemOffset(), isImmutable, false);
1464 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1465 return DAG.getLoad(ValVT, dl, Chain, FIN,
1466 PseudoSourceValue::getFixedStack(FI), 0);
1471 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1472 CallingConv::ID CallConv,
1474 const SmallVectorImpl<ISD::InputArg> &Ins,
1477 SmallVectorImpl<SDValue> &InVals) {
1479 MachineFunction &MF = DAG.getMachineFunction();
1480 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1482 const Function* Fn = MF.getFunction();
1483 if (Fn->hasExternalLinkage() &&
1484 Subtarget->isTargetCygMing() &&
1485 Fn->getName() == "main")
1486 FuncInfo->setForceFramePointer(true);
1488 // Decorate the function name.
1489 FuncInfo->setDecorationStyle(NameDecorationForCallConv(CallConv));
1491 MachineFrameInfo *MFI = MF.getFrameInfo();
1492 bool Is64Bit = Subtarget->is64Bit();
1493 bool IsWin64 = Subtarget->isTargetWin64();
1495 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1496 "Var args not supported with calling convention fastcc");
1498 // Assign locations to all of the incoming arguments.
1499 SmallVector<CCValAssign, 16> ArgLocs;
1500 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1501 ArgLocs, *DAG.getContext());
1502 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1504 unsigned LastVal = ~0U;
1506 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1507 CCValAssign &VA = ArgLocs[i];
1508 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1510 assert(VA.getValNo() != LastVal &&
1511 "Don't support value assigned to multiple locs yet");
1512 LastVal = VA.getValNo();
1514 if (VA.isRegLoc()) {
1515 EVT RegVT = VA.getLocVT();
1516 TargetRegisterClass *RC = NULL;
1517 if (RegVT == MVT::i32)
1518 RC = X86::GR32RegisterClass;
1519 else if (Is64Bit && RegVT == MVT::i64)
1520 RC = X86::GR64RegisterClass;
1521 else if (RegVT == MVT::f32)
1522 RC = X86::FR32RegisterClass;
1523 else if (RegVT == MVT::f64)
1524 RC = X86::FR64RegisterClass;
1525 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1526 RC = X86::VR128RegisterClass;
1527 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1528 RC = X86::VR64RegisterClass;
1530 llvm_unreachable("Unknown argument type!");
1532 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1533 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1535 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1536 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1538 if (VA.getLocInfo() == CCValAssign::SExt)
1539 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1540 DAG.getValueType(VA.getValVT()));
1541 else if (VA.getLocInfo() == CCValAssign::ZExt)
1542 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1543 DAG.getValueType(VA.getValVT()));
1544 else if (VA.getLocInfo() == CCValAssign::BCvt)
1545 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1547 if (VA.isExtInLoc()) {
1548 // Handle MMX values passed in XMM regs.
1549 if (RegVT.isVector()) {
1550 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1551 ArgValue, DAG.getConstant(0, MVT::i64));
1552 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1554 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1557 assert(VA.isMemLoc());
1558 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1561 // If value is passed via pointer - do a load.
1562 if (VA.getLocInfo() == CCValAssign::Indirect)
1563 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0);
1565 InVals.push_back(ArgValue);
1568 // The x86-64 ABI for returning structs by value requires that we copy
1569 // the sret argument into %rax for the return. Save the argument into
1570 // a virtual register so that we can access it from the return points.
1571 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1572 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1573 unsigned Reg = FuncInfo->getSRetReturnReg();
1575 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1576 FuncInfo->setSRetReturnReg(Reg);
1578 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1579 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1582 unsigned StackSize = CCInfo.getNextStackOffset();
1583 // Align stack specially for tail calls.
1584 if (FuncIsMadeTailCallSafe(CallConv))
1585 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1587 // If the function takes variable number of arguments, make a frame index for
1588 // the start of the first vararg value... for expansion of llvm.va_start.
1590 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1591 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize, true, false);
1594 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1596 // FIXME: We should really autogenerate these arrays
1597 static const unsigned GPR64ArgRegsWin64[] = {
1598 X86::RCX, X86::RDX, X86::R8, X86::R9
1600 static const unsigned XMMArgRegsWin64[] = {
1601 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1603 static const unsigned GPR64ArgRegs64Bit[] = {
1604 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1606 static const unsigned XMMArgRegs64Bit[] = {
1607 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1608 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1610 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1613 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1614 GPR64ArgRegs = GPR64ArgRegsWin64;
1615 XMMArgRegs = XMMArgRegsWin64;
1617 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1618 GPR64ArgRegs = GPR64ArgRegs64Bit;
1619 XMMArgRegs = XMMArgRegs64Bit;
1621 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1623 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1626 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1627 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1628 "SSE register cannot be used when SSE is disabled!");
1629 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1630 "SSE register cannot be used when SSE is disabled!");
1631 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1632 // Kernel mode asks for SSE to be disabled, so don't push them
1634 TotalNumXMMRegs = 0;
1636 // For X86-64, if there are vararg parameters that are passed via
1637 // registers, then we must store them to their spots on the stack so they
1638 // may be loaded by deferencing the result of va_next.
1639 VarArgsGPOffset = NumIntRegs * 8;
1640 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1641 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1642 TotalNumXMMRegs * 16, 16,
1645 // Store the integer parameter registers.
1646 SmallVector<SDValue, 8> MemOps;
1647 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1648 unsigned Offset = VarArgsGPOffset;
1649 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1650 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1651 DAG.getIntPtrConstant(Offset));
1652 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1653 X86::GR64RegisterClass);
1654 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1656 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1657 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1659 MemOps.push_back(Store);
1663 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1664 // Now store the XMM (fp + vector) parameter registers.
1665 SmallVector<SDValue, 11> SaveXMMOps;
1666 SaveXMMOps.push_back(Chain);
1668 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1669 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1670 SaveXMMOps.push_back(ALVal);
1672 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1673 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1675 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1676 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1677 X86::VR128RegisterClass);
1678 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1679 SaveXMMOps.push_back(Val);
1681 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1683 &SaveXMMOps[0], SaveXMMOps.size()));
1686 if (!MemOps.empty())
1687 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1688 &MemOps[0], MemOps.size());
1692 // Some CCs need callee pop.
1693 if (IsCalleePop(isVarArg, CallConv)) {
1694 BytesToPopOnReturn = StackSize; // Callee pops everything.
1696 BytesToPopOnReturn = 0; // Callee pops nothing.
1697 // If this is an sret function, the return should pop the hidden pointer.
1698 if (!Is64Bit && CallConv != CallingConv::Fast && ArgsAreStructReturn(Ins))
1699 BytesToPopOnReturn = 4;
1703 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1704 if (CallConv == CallingConv::X86_FastCall)
1705 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1708 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1714 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1715 SDValue StackPtr, SDValue Arg,
1716 DebugLoc dl, SelectionDAG &DAG,
1717 const CCValAssign &VA,
1718 ISD::ArgFlagsTy Flags) {
1719 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1720 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1721 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1722 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1723 if (Flags.isByVal()) {
1724 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1726 return DAG.getStore(Chain, dl, Arg, PtrOff,
1727 PseudoSourceValue::getStack(), LocMemOffset);
1730 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1731 /// optimization is performed and it is required.
1733 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1734 SDValue &OutRetAddr, SDValue Chain,
1735 bool IsTailCall, bool Is64Bit,
1736 int FPDiff, DebugLoc dl) {
1737 // Adjust the Return address stack slot.
1738 EVT VT = getPointerTy();
1739 OutRetAddr = getReturnAddressFrameIndex(DAG);
1741 // Load the "old" Return address.
1742 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1743 return SDValue(OutRetAddr.getNode(), 1);
1746 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1747 /// optimization is performed and it is required (FPDiff!=0).
1749 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1750 SDValue Chain, SDValue RetAddrFrIdx,
1751 bool Is64Bit, int FPDiff, DebugLoc dl) {
1752 // Store the return address to the appropriate stack slot.
1753 if (!FPDiff) return Chain;
1754 // Calculate the new stack slot for the return address.
1755 int SlotSize = Is64Bit ? 8 : 4;
1756 int NewReturnAddrFI =
1757 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, true,false);
1758 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1759 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1760 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1761 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1766 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1767 CallingConv::ID CallConv, bool isVarArg,
1769 const SmallVectorImpl<ISD::OutputArg> &Outs,
1770 const SmallVectorImpl<ISD::InputArg> &Ins,
1771 DebugLoc dl, SelectionDAG &DAG,
1772 SmallVectorImpl<SDValue> &InVals) {
1773 MachineFunction &MF = DAG.getMachineFunction();
1774 bool Is64Bit = Subtarget->is64Bit();
1775 bool IsStructRet = CallIsStructReturn(Outs);
1776 bool IsSibcall = false;
1779 // Check if it's really possible to do a tail call.
1780 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
1783 // Sibcalls are automatically detected tailcalls which do not require
1785 if (!GuaranteedTailCallOpt && isTailCall)
1792 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1793 "Var args not supported with calling convention fastcc");
1795 // Analyze operands of the call, assigning locations to each operand.
1796 SmallVector<CCValAssign, 16> ArgLocs;
1797 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1798 ArgLocs, *DAG.getContext());
1799 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1801 // Get a count of how many bytes are to be pushed on the stack.
1802 unsigned NumBytes = CCInfo.getNextStackOffset();
1804 // This is a sibcall. The memory operands are available in caller's
1805 // own caller's stack.
1807 else if (GuaranteedTailCallOpt && CallConv == CallingConv::Fast)
1808 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1811 if (isTailCall && !IsSibcall) {
1812 // Lower arguments at fp - stackoffset + fpdiff.
1813 unsigned NumBytesCallerPushed =
1814 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1815 FPDiff = NumBytesCallerPushed - NumBytes;
1817 // Set the delta of movement of the returnaddr stackslot.
1818 // But only set if delta is greater than previous delta.
1819 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1820 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1824 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1826 SDValue RetAddrFrIdx;
1827 // Load return adress for tail calls.
1828 if (isTailCall && FPDiff)
1829 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1830 Is64Bit, FPDiff, dl);
1832 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1833 SmallVector<SDValue, 8> MemOpChains;
1836 // Walk the register/memloc assignments, inserting copies/loads. In the case
1837 // of tail call optimization arguments are handle later.
1838 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1839 CCValAssign &VA = ArgLocs[i];
1840 EVT RegVT = VA.getLocVT();
1841 SDValue Arg = Outs[i].Val;
1842 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1843 bool isByVal = Flags.isByVal();
1845 // Promote the value if needed.
1846 switch (VA.getLocInfo()) {
1847 default: llvm_unreachable("Unknown loc info!");
1848 case CCValAssign::Full: break;
1849 case CCValAssign::SExt:
1850 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1852 case CCValAssign::ZExt:
1853 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1855 case CCValAssign::AExt:
1856 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1857 // Special case: passing MMX values in XMM registers.
1858 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1859 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1860 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1862 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1864 case CCValAssign::BCvt:
1865 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1867 case CCValAssign::Indirect: {
1868 // Store the argument.
1869 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1870 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1871 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1872 PseudoSourceValue::getFixedStack(FI), 0);
1878 if (VA.isRegLoc()) {
1879 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1880 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1881 assert(VA.isMemLoc());
1882 if (StackPtr.getNode() == 0)
1883 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1884 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1885 dl, DAG, VA, Flags));
1889 if (!MemOpChains.empty())
1890 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1891 &MemOpChains[0], MemOpChains.size());
1893 // Build a sequence of copy-to-reg nodes chained together with token chain
1894 // and flag operands which copy the outgoing args into registers.
1896 // Tail call byval lowering might overwrite argument registers so in case of
1897 // tail call optimization the copies to registers are lowered later.
1899 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1900 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1901 RegsToPass[i].second, InFlag);
1902 InFlag = Chain.getValue(1);
1905 if (Subtarget->isPICStyleGOT()) {
1906 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1909 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1910 DAG.getNode(X86ISD::GlobalBaseReg,
1911 DebugLoc::getUnknownLoc(),
1914 InFlag = Chain.getValue(1);
1916 // If we are tail calling and generating PIC/GOT style code load the
1917 // address of the callee into ECX. The value in ecx is used as target of
1918 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1919 // for tail calls on PIC/GOT architectures. Normally we would just put the
1920 // address of GOT into ebx and then call target@PLT. But for tail calls
1921 // ebx would be restored (since ebx is callee saved) before jumping to the
1924 // Note: The actual moving to ECX is done further down.
1925 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1926 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1927 !G->getGlobal()->hasProtectedVisibility())
1928 Callee = LowerGlobalAddress(Callee, DAG);
1929 else if (isa<ExternalSymbolSDNode>(Callee))
1930 Callee = LowerExternalSymbol(Callee, DAG);
1934 if (Is64Bit && isVarArg) {
1935 // From AMD64 ABI document:
1936 // For calls that may call functions that use varargs or stdargs
1937 // (prototype-less calls or calls to functions containing ellipsis (...) in
1938 // the declaration) %al is used as hidden argument to specify the number
1939 // of SSE registers used. The contents of %al do not need to match exactly
1940 // the number of registers, but must be an ubound on the number of SSE
1941 // registers used and is in the range 0 - 8 inclusive.
1943 // FIXME: Verify this on Win64
1944 // Count the number of XMM registers allocated.
1945 static const unsigned XMMArgRegs[] = {
1946 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1947 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1949 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1950 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1951 && "SSE registers cannot be used when SSE is disabled");
1953 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1954 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1955 InFlag = Chain.getValue(1);
1959 // For tail calls lower the arguments to the 'real' stack slot.
1961 // Force all the incoming stack arguments to be loaded from the stack
1962 // before any new outgoing arguments are stored to the stack, because the
1963 // outgoing stack slots may alias the incoming argument stack slots, and
1964 // the alias isn't otherwise explicit. This is slightly more conservative
1965 // than necessary, because it means that each store effectively depends
1966 // on every argument instead of just those arguments it would clobber.
1967 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1969 SmallVector<SDValue, 8> MemOpChains2;
1972 // Do not flag preceeding copytoreg stuff together with the following stuff.
1974 if (GuaranteedTailCallOpt) {
1975 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1976 CCValAssign &VA = ArgLocs[i];
1979 assert(VA.isMemLoc());
1980 SDValue Arg = Outs[i].Val;
1981 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1982 // Create frame index.
1983 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1984 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1985 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
1986 FIN = DAG.getFrameIndex(FI, getPointerTy());
1988 if (Flags.isByVal()) {
1989 // Copy relative to framepointer.
1990 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1991 if (StackPtr.getNode() == 0)
1992 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1994 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1996 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2000 // Store relative to framepointer.
2001 MemOpChains2.push_back(
2002 DAG.getStore(ArgChain, dl, Arg, FIN,
2003 PseudoSourceValue::getFixedStack(FI), 0));
2008 if (!MemOpChains2.empty())
2009 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2010 &MemOpChains2[0], MemOpChains2.size());
2012 // Copy arguments to their registers.
2013 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2014 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2015 RegsToPass[i].second, InFlag);
2016 InFlag = Chain.getValue(1);
2020 // Store the return address to the appropriate stack slot.
2021 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2025 bool WasGlobalOrExternal = false;
2026 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2027 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2028 // In the 64-bit large code model, we have to make all calls
2029 // through a register, since the call instruction's 32-bit
2030 // pc-relative offset may not be large enough to hold the whole
2032 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2033 WasGlobalOrExternal = true;
2034 // If the callee is a GlobalAddress node (quite common, every direct call
2035 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2038 // We should use extra load for direct calls to dllimported functions in
2040 GlobalValue *GV = G->getGlobal();
2041 if (!GV->hasDLLImportLinkage()) {
2042 unsigned char OpFlags = 0;
2044 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2045 // external symbols most go through the PLT in PIC mode. If the symbol
2046 // has hidden or protected visibility, or if it is static or local, then
2047 // we don't need to use the PLT - we can directly call it.
2048 if (Subtarget->isTargetELF() &&
2049 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2050 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2051 OpFlags = X86II::MO_PLT;
2052 } else if (Subtarget->isPICStyleStubAny() &&
2053 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2054 Subtarget->getDarwinVers() < 9) {
2055 // PC-relative references to external symbols should go through $stub,
2056 // unless we're building with the leopard linker or later, which
2057 // automatically synthesizes these stubs.
2058 OpFlags = X86II::MO_DARWIN_STUB;
2061 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2062 G->getOffset(), OpFlags);
2064 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2065 WasGlobalOrExternal = true;
2066 unsigned char OpFlags = 0;
2068 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2069 // symbols should go through the PLT.
2070 if (Subtarget->isTargetELF() &&
2071 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2072 OpFlags = X86II::MO_PLT;
2073 } else if (Subtarget->isPICStyleStubAny() &&
2074 Subtarget->getDarwinVers() < 9) {
2075 // PC-relative references to external symbols should go through $stub,
2076 // unless we're building with the leopard linker or later, which
2077 // automatically synthesizes these stubs.
2078 OpFlags = X86II::MO_DARWIN_STUB;
2081 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2085 if (isTailCall && !WasGlobalOrExternal) {
2086 // Force the address into a (call preserved) caller-saved register since
2087 // tailcall must happen after callee-saved registers are poped.
2088 // FIXME: Give it a special register class that contains caller-saved
2089 // register instead?
2090 unsigned TCReg = Is64Bit ? X86::R11 : X86::EAX;
2091 Chain = DAG.getCopyToReg(Chain, dl,
2092 DAG.getRegister(TCReg, getPointerTy()),
2094 Callee = DAG.getRegister(TCReg, getPointerTy());
2097 // Returns a chain & a flag for retval copy to use.
2098 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2099 SmallVector<SDValue, 8> Ops;
2101 if (!IsSibcall && isTailCall) {
2102 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2103 DAG.getIntPtrConstant(0, true), InFlag);
2104 InFlag = Chain.getValue(1);
2107 Ops.push_back(Chain);
2108 Ops.push_back(Callee);
2111 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2113 // Add argument registers to the end of the list so that they are known live
2115 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2116 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2117 RegsToPass[i].second.getValueType()));
2119 // Add an implicit use GOT pointer in EBX.
2120 if (!isTailCall && Subtarget->isPICStyleGOT())
2121 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2123 // Add an implicit use of AL for x86 vararg functions.
2124 if (Is64Bit && isVarArg)
2125 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2127 if (InFlag.getNode())
2128 Ops.push_back(InFlag);
2131 // If this is the first return lowered for this function, add the regs
2132 // to the liveout set for the function.
2133 if (MF.getRegInfo().liveout_empty()) {
2134 SmallVector<CCValAssign, 16> RVLocs;
2135 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2137 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2138 for (unsigned i = 0; i != RVLocs.size(); ++i)
2139 if (RVLocs[i].isRegLoc())
2140 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2143 assert(((Callee.getOpcode() == ISD::Register &&
2144 (cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
2145 cast<RegisterSDNode>(Callee)->getReg() == X86::R11)) ||
2146 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2147 Callee.getOpcode() == ISD::TargetGlobalAddress) &&
2148 "Expecting a global address, external symbol, or scratch register");
2150 return DAG.getNode(X86ISD::TC_RETURN, dl,
2151 NodeTys, &Ops[0], Ops.size());
2154 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2155 InFlag = Chain.getValue(1);
2157 // Create the CALLSEQ_END node.
2158 unsigned NumBytesForCalleeToPush;
2159 if (IsCalleePop(isVarArg, CallConv))
2160 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2161 else if (!Is64Bit && CallConv != CallingConv::Fast && IsStructRet)
2162 // If this is a call to a struct-return function, the callee
2163 // pops the hidden struct pointer, so we have to push it back.
2164 // This is common for Darwin/X86, Linux & Mingw32 targets.
2165 NumBytesForCalleeToPush = 4;
2167 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2169 // Returns a flag for retval copy to use.
2171 Chain = DAG.getCALLSEQ_END(Chain,
2172 DAG.getIntPtrConstant(NumBytes, true),
2173 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2176 InFlag = Chain.getValue(1);
2179 // Handle result values, copying them out of physregs into vregs that we
2181 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2182 Ins, dl, DAG, InVals);
2186 //===----------------------------------------------------------------------===//
2187 // Fast Calling Convention (tail call) implementation
2188 //===----------------------------------------------------------------------===//
2190 // Like std call, callee cleans arguments, convention except that ECX is
2191 // reserved for storing the tail called function address. Only 2 registers are
2192 // free for argument passing (inreg). Tail call optimization is performed
2194 // * tailcallopt is enabled
2195 // * caller/callee are fastcc
2196 // On X86_64 architecture with GOT-style position independent code only local
2197 // (within module) calls are supported at the moment.
2198 // To keep the stack aligned according to platform abi the function
2199 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2200 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2201 // If a tail called function callee has more arguments than the caller the
2202 // caller needs to make sure that there is room to move the RETADDR to. This is
2203 // achieved by reserving an area the size of the argument delta right after the
2204 // original REtADDR, but before the saved framepointer or the spilled registers
2205 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2217 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2218 /// for a 16 byte align requirement.
2219 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2220 SelectionDAG& DAG) {
2221 MachineFunction &MF = DAG.getMachineFunction();
2222 const TargetMachine &TM = MF.getTarget();
2223 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2224 unsigned StackAlignment = TFI.getStackAlignment();
2225 uint64_t AlignMask = StackAlignment - 1;
2226 int64_t Offset = StackSize;
2227 uint64_t SlotSize = TD->getPointerSize();
2228 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2229 // Number smaller than 12 so just add the difference.
2230 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2232 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2233 Offset = ((~AlignMask) & Offset) + StackAlignment +
2234 (StackAlignment-SlotSize);
2239 /// MatchingStackOffset - Return true if the given stack call argument is
2240 /// already available in the same position (relatively) of the caller's
2241 /// incoming argument stack.
2243 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2244 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2245 const X86InstrInfo *TII) {
2247 if (Arg.getOpcode() == ISD::CopyFromReg) {
2248 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2249 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2251 MachineInstr *Def = MRI->getVRegDef(VR);
2254 if (!Flags.isByVal()) {
2255 if (!TII->isLoadFromStackSlot(Def, FI))
2258 unsigned Opcode = Def->getOpcode();
2259 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2260 Def->getOperand(1).isFI()) {
2261 FI = Def->getOperand(1).getIndex();
2262 if (MFI->getObjectSize(FI) != Flags.getByValSize())
2268 LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg);
2271 SDValue Ptr = Ld->getBasePtr();
2272 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2275 FI = FINode->getIndex();
2278 if (!MFI->isFixedObjectIndex(FI))
2280 return Offset == MFI->getObjectOffset(FI);
2283 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2284 /// for tail call optimization. Targets which want to do tail call
2285 /// optimization should implement this function.
2287 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2288 CallingConv::ID CalleeCC,
2290 const SmallVectorImpl<ISD::OutputArg> &Outs,
2291 const SmallVectorImpl<ISD::InputArg> &Ins,
2292 SelectionDAG& DAG) const {
2293 if (CalleeCC != CallingConv::Fast &&
2294 CalleeCC != CallingConv::C)
2297 // If -tailcallopt is specified, make fastcc functions tail-callable.
2298 const Function *CallerF = DAG.getMachineFunction().getFunction();
2299 if (GuaranteedTailCallOpt) {
2300 if (CalleeCC == CallingConv::Fast &&
2301 CallerF->getCallingConv() == CalleeCC)
2306 // Look for obvious safe cases to perform tail call optimization that does not
2307 // requite ABI changes. This is what gcc calls sibcall.
2309 // Do not tail call optimize vararg calls for now.
2313 // If the callee takes no arguments then go on to check the results of the
2315 if (!Outs.empty()) {
2316 // Check if stack adjustment is needed. For now, do not do this if any
2317 // argument is passed on the stack.
2318 SmallVector<CCValAssign, 16> ArgLocs;
2319 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2320 ArgLocs, *DAG.getContext());
2321 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2322 if (CCInfo.getNextStackOffset()) {
2323 MachineFunction &MF = DAG.getMachineFunction();
2324 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2326 if (Subtarget->isTargetWin64())
2327 // Win64 ABI has additional complications.
2330 // Check if the arguments are already laid out in the right way as
2331 // the caller's fixed stack objects.
2332 MachineFrameInfo *MFI = MF.getFrameInfo();
2333 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2334 const X86InstrInfo *TII =
2335 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2336 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2337 CCValAssign &VA = ArgLocs[i];
2338 EVT RegVT = VA.getLocVT();
2339 SDValue Arg = Outs[i].Val;
2340 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2341 if (VA.getLocInfo() == CCValAssign::Indirect)
2343 if (!VA.isRegLoc()) {
2344 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2356 X86TargetLowering::createFastISel(MachineFunction &mf, MachineModuleInfo *mmo,
2358 DenseMap<const Value *, unsigned> &vm,
2359 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2360 DenseMap<const AllocaInst *, int> &am
2362 , SmallSet<Instruction*, 8> &cil
2365 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2373 //===----------------------------------------------------------------------===//
2374 // Other Lowering Hooks
2375 //===----------------------------------------------------------------------===//
2378 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2379 MachineFunction &MF = DAG.getMachineFunction();
2380 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2381 int ReturnAddrIndex = FuncInfo->getRAIndex();
2383 if (ReturnAddrIndex == 0) {
2384 // Set up a frame object for the return address.
2385 uint64_t SlotSize = TD->getPointerSize();
2386 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2388 FuncInfo->setRAIndex(ReturnAddrIndex);
2391 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2395 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2396 bool hasSymbolicDisplacement) {
2397 // Offset should fit into 32 bit immediate field.
2398 if (!isInt32(Offset))
2401 // If we don't have a symbolic displacement - we don't have any extra
2403 if (!hasSymbolicDisplacement)
2406 // FIXME: Some tweaks might be needed for medium code model.
2407 if (M != CodeModel::Small && M != CodeModel::Kernel)
2410 // For small code model we assume that latest object is 16MB before end of 31
2411 // bits boundary. We may also accept pretty large negative constants knowing
2412 // that all objects are in the positive half of address space.
2413 if (M == CodeModel::Small && Offset < 16*1024*1024)
2416 // For kernel code model we know that all object resist in the negative half
2417 // of 32bits address space. We may not accept negative offsets, since they may
2418 // be just off and we may accept pretty large positive ones.
2419 if (M == CodeModel::Kernel && Offset > 0)
2425 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2426 /// specific condition code, returning the condition code and the LHS/RHS of the
2427 /// comparison to make.
2428 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2429 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2431 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2432 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2433 // X > -1 -> X == 0, jump !sign.
2434 RHS = DAG.getConstant(0, RHS.getValueType());
2435 return X86::COND_NS;
2436 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2437 // X < 0 -> X == 0, jump on sign.
2439 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2441 RHS = DAG.getConstant(0, RHS.getValueType());
2442 return X86::COND_LE;
2446 switch (SetCCOpcode) {
2447 default: llvm_unreachable("Invalid integer condition!");
2448 case ISD::SETEQ: return X86::COND_E;
2449 case ISD::SETGT: return X86::COND_G;
2450 case ISD::SETGE: return X86::COND_GE;
2451 case ISD::SETLT: return X86::COND_L;
2452 case ISD::SETLE: return X86::COND_LE;
2453 case ISD::SETNE: return X86::COND_NE;
2454 case ISD::SETULT: return X86::COND_B;
2455 case ISD::SETUGT: return X86::COND_A;
2456 case ISD::SETULE: return X86::COND_BE;
2457 case ISD::SETUGE: return X86::COND_AE;
2461 // First determine if it is required or is profitable to flip the operands.
2463 // If LHS is a foldable load, but RHS is not, flip the condition.
2464 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2465 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2466 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2467 std::swap(LHS, RHS);
2470 switch (SetCCOpcode) {
2476 std::swap(LHS, RHS);
2480 // On a floating point condition, the flags are set as follows:
2482 // 0 | 0 | 0 | X > Y
2483 // 0 | 0 | 1 | X < Y
2484 // 1 | 0 | 0 | X == Y
2485 // 1 | 1 | 1 | unordered
2486 switch (SetCCOpcode) {
2487 default: llvm_unreachable("Condcode should be pre-legalized away");
2489 case ISD::SETEQ: return X86::COND_E;
2490 case ISD::SETOLT: // flipped
2492 case ISD::SETGT: return X86::COND_A;
2493 case ISD::SETOLE: // flipped
2495 case ISD::SETGE: return X86::COND_AE;
2496 case ISD::SETUGT: // flipped
2498 case ISD::SETLT: return X86::COND_B;
2499 case ISD::SETUGE: // flipped
2501 case ISD::SETLE: return X86::COND_BE;
2503 case ISD::SETNE: return X86::COND_NE;
2504 case ISD::SETUO: return X86::COND_P;
2505 case ISD::SETO: return X86::COND_NP;
2507 case ISD::SETUNE: return X86::COND_INVALID;
2511 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2512 /// code. Current x86 isa includes the following FP cmov instructions:
2513 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2514 static bool hasFPCMov(unsigned X86CC) {
2530 /// isFPImmLegal - Returns true if the target can instruction select the
2531 /// specified FP immediate natively. If false, the legalizer will
2532 /// materialize the FP immediate as a load from a constant pool.
2533 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2534 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2535 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2541 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2542 /// the specified range (L, H].
2543 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2544 return (Val < 0) || (Val >= Low && Val < Hi);
2547 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2548 /// specified value.
2549 static bool isUndefOrEqual(int Val, int CmpVal) {
2550 if (Val < 0 || Val == CmpVal)
2555 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2556 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2557 /// the second operand.
2558 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2559 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2560 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2561 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2562 return (Mask[0] < 2 && Mask[1] < 2);
2566 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2567 SmallVector<int, 8> M;
2569 return ::isPSHUFDMask(M, N->getValueType(0));
2572 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2573 /// is suitable for input to PSHUFHW.
2574 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2575 if (VT != MVT::v8i16)
2578 // Lower quadword copied in order or undef.
2579 for (int i = 0; i != 4; ++i)
2580 if (Mask[i] >= 0 && Mask[i] != i)
2583 // Upper quadword shuffled.
2584 for (int i = 4; i != 8; ++i)
2585 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2591 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2592 SmallVector<int, 8> M;
2594 return ::isPSHUFHWMask(M, N->getValueType(0));
2597 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2598 /// is suitable for input to PSHUFLW.
2599 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2600 if (VT != MVT::v8i16)
2603 // Upper quadword copied in order.
2604 for (int i = 4; i != 8; ++i)
2605 if (Mask[i] >= 0 && Mask[i] != i)
2608 // Lower quadword shuffled.
2609 for (int i = 0; i != 4; ++i)
2616 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2617 SmallVector<int, 8> M;
2619 return ::isPSHUFLWMask(M, N->getValueType(0));
2622 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2623 /// is suitable for input to PALIGNR.
2624 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2626 int i, e = VT.getVectorNumElements();
2628 // Do not handle v2i64 / v2f64 shuffles with palignr.
2629 if (e < 4 || !hasSSSE3)
2632 for (i = 0; i != e; ++i)
2636 // All undef, not a palignr.
2640 // Determine if it's ok to perform a palignr with only the LHS, since we
2641 // don't have access to the actual shuffle elements to see if RHS is undef.
2642 bool Unary = Mask[i] < (int)e;
2643 bool NeedsUnary = false;
2645 int s = Mask[i] - i;
2647 // Check the rest of the elements to see if they are consecutive.
2648 for (++i; i != e; ++i) {
2653 Unary = Unary && (m < (int)e);
2654 NeedsUnary = NeedsUnary || (m < s);
2656 if (NeedsUnary && !Unary)
2658 if (Unary && m != ((s+i) & (e-1)))
2660 if (!Unary && m != (s+i))
2666 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2667 SmallVector<int, 8> M;
2669 return ::isPALIGNRMask(M, N->getValueType(0), true);
2672 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2673 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2674 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2675 int NumElems = VT.getVectorNumElements();
2676 if (NumElems != 2 && NumElems != 4)
2679 int Half = NumElems / 2;
2680 for (int i = 0; i < Half; ++i)
2681 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2683 for (int i = Half; i < NumElems; ++i)
2684 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2690 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2691 SmallVector<int, 8> M;
2693 return ::isSHUFPMask(M, N->getValueType(0));
2696 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2697 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2698 /// half elements to come from vector 1 (which would equal the dest.) and
2699 /// the upper half to come from vector 2.
2700 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2701 int NumElems = VT.getVectorNumElements();
2703 if (NumElems != 2 && NumElems != 4)
2706 int Half = NumElems / 2;
2707 for (int i = 0; i < Half; ++i)
2708 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2710 for (int i = Half; i < NumElems; ++i)
2711 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2716 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2717 SmallVector<int, 8> M;
2719 return isCommutedSHUFPMask(M, N->getValueType(0));
2722 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2723 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2724 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2725 if (N->getValueType(0).getVectorNumElements() != 4)
2728 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2729 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2730 isUndefOrEqual(N->getMaskElt(1), 7) &&
2731 isUndefOrEqual(N->getMaskElt(2), 2) &&
2732 isUndefOrEqual(N->getMaskElt(3), 3);
2735 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2736 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2738 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2739 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2744 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2745 isUndefOrEqual(N->getMaskElt(1), 3) &&
2746 isUndefOrEqual(N->getMaskElt(2), 2) &&
2747 isUndefOrEqual(N->getMaskElt(3), 3);
2750 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2751 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2752 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2753 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2755 if (NumElems != 2 && NumElems != 4)
2758 for (unsigned i = 0; i < NumElems/2; ++i)
2759 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2762 for (unsigned i = NumElems/2; i < NumElems; ++i)
2763 if (!isUndefOrEqual(N->getMaskElt(i), i))
2769 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2770 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2771 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2772 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2774 if (NumElems != 2 && NumElems != 4)
2777 for (unsigned i = 0; i < NumElems/2; ++i)
2778 if (!isUndefOrEqual(N->getMaskElt(i), i))
2781 for (unsigned i = 0; i < NumElems/2; ++i)
2782 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2788 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2789 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2790 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2791 bool V2IsSplat = false) {
2792 int NumElts = VT.getVectorNumElements();
2793 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2796 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2798 int BitI1 = Mask[i+1];
2799 if (!isUndefOrEqual(BitI, j))
2802 if (!isUndefOrEqual(BitI1, NumElts))
2805 if (!isUndefOrEqual(BitI1, j + NumElts))
2812 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2813 SmallVector<int, 8> M;
2815 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2818 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2819 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2820 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2821 bool V2IsSplat = false) {
2822 int NumElts = VT.getVectorNumElements();
2823 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2826 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2828 int BitI1 = Mask[i+1];
2829 if (!isUndefOrEqual(BitI, j + NumElts/2))
2832 if (isUndefOrEqual(BitI1, NumElts))
2835 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2842 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2843 SmallVector<int, 8> M;
2845 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2848 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2849 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2851 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2852 int NumElems = VT.getVectorNumElements();
2853 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2856 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2858 int BitI1 = Mask[i+1];
2859 if (!isUndefOrEqual(BitI, j))
2861 if (!isUndefOrEqual(BitI1, j))
2867 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2868 SmallVector<int, 8> M;
2870 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2873 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2874 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2876 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2877 int NumElems = VT.getVectorNumElements();
2878 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2881 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2883 int BitI1 = Mask[i+1];
2884 if (!isUndefOrEqual(BitI, j))
2886 if (!isUndefOrEqual(BitI1, j))
2892 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2893 SmallVector<int, 8> M;
2895 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2898 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2899 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2900 /// MOVSD, and MOVD, i.e. setting the lowest element.
2901 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2902 if (VT.getVectorElementType().getSizeInBits() < 32)
2905 int NumElts = VT.getVectorNumElements();
2907 if (!isUndefOrEqual(Mask[0], NumElts))
2910 for (int i = 1; i < NumElts; ++i)
2911 if (!isUndefOrEqual(Mask[i], i))
2917 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2918 SmallVector<int, 8> M;
2920 return ::isMOVLMask(M, N->getValueType(0));
2923 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2924 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2925 /// element of vector 2 and the other elements to come from vector 1 in order.
2926 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2927 bool V2IsSplat = false, bool V2IsUndef = false) {
2928 int NumOps = VT.getVectorNumElements();
2929 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2932 if (!isUndefOrEqual(Mask[0], 0))
2935 for (int i = 1; i < NumOps; ++i)
2936 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2937 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2938 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2944 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2945 bool V2IsUndef = false) {
2946 SmallVector<int, 8> M;
2948 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2951 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2952 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2953 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2954 if (N->getValueType(0).getVectorNumElements() != 4)
2957 // Expect 1, 1, 3, 3
2958 for (unsigned i = 0; i < 2; ++i) {
2959 int Elt = N->getMaskElt(i);
2960 if (Elt >= 0 && Elt != 1)
2965 for (unsigned i = 2; i < 4; ++i) {
2966 int Elt = N->getMaskElt(i);
2967 if (Elt >= 0 && Elt != 3)
2972 // Don't use movshdup if it can be done with a shufps.
2973 // FIXME: verify that matching u, u, 3, 3 is what we want.
2977 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2978 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2979 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2980 if (N->getValueType(0).getVectorNumElements() != 4)
2983 // Expect 0, 0, 2, 2
2984 for (unsigned i = 0; i < 2; ++i)
2985 if (N->getMaskElt(i) > 0)
2989 for (unsigned i = 2; i < 4; ++i) {
2990 int Elt = N->getMaskElt(i);
2991 if (Elt >= 0 && Elt != 2)
2996 // Don't use movsldup if it can be done with a shufps.
3000 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3001 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3002 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3003 int e = N->getValueType(0).getVectorNumElements() / 2;
3005 for (int i = 0; i < e; ++i)
3006 if (!isUndefOrEqual(N->getMaskElt(i), i))
3008 for (int i = 0; i < e; ++i)
3009 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3014 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3015 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3016 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3017 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3018 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3020 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3022 for (int i = 0; i < NumOperands; ++i) {
3023 int Val = SVOp->getMaskElt(NumOperands-i-1);
3024 if (Val < 0) Val = 0;
3025 if (Val >= NumOperands) Val -= NumOperands;
3027 if (i != NumOperands - 1)
3033 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3034 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3035 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3036 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3038 // 8 nodes, but we only care about the last 4.
3039 for (unsigned i = 7; i >= 4; --i) {
3040 int Val = SVOp->getMaskElt(i);
3049 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3050 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3051 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3052 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3054 // 8 nodes, but we only care about the first 4.
3055 for (int i = 3; i >= 0; --i) {
3056 int Val = SVOp->getMaskElt(i);
3065 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3066 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3067 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3068 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3069 EVT VVT = N->getValueType(0);
3070 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3074 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3075 Val = SVOp->getMaskElt(i);
3079 return (Val - i) * EltSize;
3082 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3084 bool X86::isZeroNode(SDValue Elt) {
3085 return ((isa<ConstantSDNode>(Elt) &&
3086 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
3087 (isa<ConstantFPSDNode>(Elt) &&
3088 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3091 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3092 /// their permute mask.
3093 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3094 SelectionDAG &DAG) {
3095 EVT VT = SVOp->getValueType(0);
3096 unsigned NumElems = VT.getVectorNumElements();
3097 SmallVector<int, 8> MaskVec;
3099 for (unsigned i = 0; i != NumElems; ++i) {
3100 int idx = SVOp->getMaskElt(i);
3102 MaskVec.push_back(idx);
3103 else if (idx < (int)NumElems)
3104 MaskVec.push_back(idx + NumElems);
3106 MaskVec.push_back(idx - NumElems);
3108 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3109 SVOp->getOperand(0), &MaskVec[0]);
3112 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3113 /// the two vector operands have swapped position.
3114 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3115 unsigned NumElems = VT.getVectorNumElements();
3116 for (unsigned i = 0; i != NumElems; ++i) {
3120 else if (idx < (int)NumElems)
3121 Mask[i] = idx + NumElems;
3123 Mask[i] = idx - NumElems;
3127 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3128 /// match movhlps. The lower half elements should come from upper half of
3129 /// V1 (and in order), and the upper half elements should come from the upper
3130 /// half of V2 (and in order).
3131 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3132 if (Op->getValueType(0).getVectorNumElements() != 4)
3134 for (unsigned i = 0, e = 2; i != e; ++i)
3135 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3137 for (unsigned i = 2; i != 4; ++i)
3138 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3143 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3144 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3146 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3147 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3149 N = N->getOperand(0).getNode();
3150 if (!ISD::isNON_EXTLoad(N))
3153 *LD = cast<LoadSDNode>(N);
3157 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3158 /// match movlp{s|d}. The lower half elements should come from lower half of
3159 /// V1 (and in order), and the upper half elements should come from the upper
3160 /// half of V2 (and in order). And since V1 will become the source of the
3161 /// MOVLP, it must be either a vector load or a scalar load to vector.
3162 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3163 ShuffleVectorSDNode *Op) {
3164 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3166 // Is V2 is a vector load, don't do this transformation. We will try to use
3167 // load folding shufps op.
3168 if (ISD::isNON_EXTLoad(V2))
3171 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3173 if (NumElems != 2 && NumElems != 4)
3175 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3176 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3178 for (unsigned i = NumElems/2; i != NumElems; ++i)
3179 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3184 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3186 static bool isSplatVector(SDNode *N) {
3187 if (N->getOpcode() != ISD::BUILD_VECTOR)
3190 SDValue SplatValue = N->getOperand(0);
3191 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3192 if (N->getOperand(i) != SplatValue)
3197 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3198 /// to an zero vector.
3199 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3200 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3201 SDValue V1 = N->getOperand(0);
3202 SDValue V2 = N->getOperand(1);
3203 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3204 for (unsigned i = 0; i != NumElems; ++i) {
3205 int Idx = N->getMaskElt(i);
3206 if (Idx >= (int)NumElems) {
3207 unsigned Opc = V2.getOpcode();
3208 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3210 if (Opc != ISD::BUILD_VECTOR ||
3211 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3213 } else if (Idx >= 0) {
3214 unsigned Opc = V1.getOpcode();
3215 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3217 if (Opc != ISD::BUILD_VECTOR ||
3218 !X86::isZeroNode(V1.getOperand(Idx)))
3225 /// getZeroVector - Returns a vector of specified type with all zero elements.
3227 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3229 assert(VT.isVector() && "Expected a vector type");
3231 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3232 // type. This ensures they get CSE'd.
3234 if (VT.getSizeInBits() == 64) { // MMX
3235 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3236 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3237 } else if (HasSSE2) { // SSE2
3238 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3239 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3241 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3242 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3244 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3247 /// getOnesVector - Returns a vector of specified type with all bits set.
3249 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3250 assert(VT.isVector() && "Expected a vector type");
3252 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3253 // type. This ensures they get CSE'd.
3254 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3256 if (VT.getSizeInBits() == 64) // MMX
3257 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3259 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3260 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3264 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3265 /// that point to V2 points to its first element.
3266 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3267 EVT VT = SVOp->getValueType(0);
3268 unsigned NumElems = VT.getVectorNumElements();
3270 bool Changed = false;
3271 SmallVector<int, 8> MaskVec;
3272 SVOp->getMask(MaskVec);
3274 for (unsigned i = 0; i != NumElems; ++i) {
3275 if (MaskVec[i] > (int)NumElems) {
3276 MaskVec[i] = NumElems;
3281 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3282 SVOp->getOperand(1), &MaskVec[0]);
3283 return SDValue(SVOp, 0);
3286 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3287 /// operation of specified width.
3288 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3290 unsigned NumElems = VT.getVectorNumElements();
3291 SmallVector<int, 8> Mask;
3292 Mask.push_back(NumElems);
3293 for (unsigned i = 1; i != NumElems; ++i)
3295 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3298 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3299 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3301 unsigned NumElems = VT.getVectorNumElements();
3302 SmallVector<int, 8> Mask;
3303 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3305 Mask.push_back(i + NumElems);
3307 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3310 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3311 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3313 unsigned NumElems = VT.getVectorNumElements();
3314 unsigned Half = NumElems/2;
3315 SmallVector<int, 8> Mask;
3316 for (unsigned i = 0; i != Half; ++i) {
3317 Mask.push_back(i + Half);
3318 Mask.push_back(i + NumElems + Half);
3320 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3323 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3324 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3326 if (SV->getValueType(0).getVectorNumElements() <= 4)
3327 return SDValue(SV, 0);
3329 EVT PVT = MVT::v4f32;
3330 EVT VT = SV->getValueType(0);
3331 DebugLoc dl = SV->getDebugLoc();
3332 SDValue V1 = SV->getOperand(0);
3333 int NumElems = VT.getVectorNumElements();
3334 int EltNo = SV->getSplatIndex();
3336 // unpack elements to the correct location
3337 while (NumElems > 4) {
3338 if (EltNo < NumElems/2) {
3339 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3341 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3342 EltNo -= NumElems/2;
3347 // Perform the splat.
3348 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3349 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3350 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3351 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3354 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3355 /// vector of zero or undef vector. This produces a shuffle where the low
3356 /// element of V2 is swizzled into the zero/undef vector, landing at element
3357 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3358 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3359 bool isZero, bool HasSSE2,
3360 SelectionDAG &DAG) {
3361 EVT VT = V2.getValueType();
3363 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3364 unsigned NumElems = VT.getVectorNumElements();
3365 SmallVector<int, 16> MaskVec;
3366 for (unsigned i = 0; i != NumElems; ++i)
3367 // If this is the insertion idx, put the low elt of V2 here.
3368 MaskVec.push_back(i == Idx ? NumElems : i);
3369 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3372 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3373 /// a shuffle that is zero.
3375 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3376 bool Low, SelectionDAG &DAG) {
3377 unsigned NumZeros = 0;
3378 for (int i = 0; i < NumElems; ++i) {
3379 unsigned Index = Low ? i : NumElems-i-1;
3380 int Idx = SVOp->getMaskElt(Index);
3385 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3386 if (Elt.getNode() && X86::isZeroNode(Elt))
3394 /// isVectorShift - Returns true if the shuffle can be implemented as a
3395 /// logical left or right shift of a vector.
3396 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3397 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3398 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3399 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3402 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3405 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3409 bool SeenV1 = false;
3410 bool SeenV2 = false;
3411 for (int i = NumZeros; i < NumElems; ++i) {
3412 int Val = isLeft ? (i - NumZeros) : i;
3413 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3425 if (SeenV1 && SeenV2)
3428 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3434 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3436 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3437 unsigned NumNonZero, unsigned NumZero,
3438 SelectionDAG &DAG, TargetLowering &TLI) {
3442 DebugLoc dl = Op.getDebugLoc();
3445 for (unsigned i = 0; i < 16; ++i) {
3446 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3447 if (ThisIsNonZero && First) {
3449 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3451 V = DAG.getUNDEF(MVT::v8i16);
3456 SDValue ThisElt(0, 0), LastElt(0, 0);
3457 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3458 if (LastIsNonZero) {
3459 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3460 MVT::i16, Op.getOperand(i-1));
3462 if (ThisIsNonZero) {
3463 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3464 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3465 ThisElt, DAG.getConstant(8, MVT::i8));
3467 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3471 if (ThisElt.getNode())
3472 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3473 DAG.getIntPtrConstant(i/2));
3477 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3480 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3482 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3483 unsigned NumNonZero, unsigned NumZero,
3484 SelectionDAG &DAG, TargetLowering &TLI) {
3488 DebugLoc dl = Op.getDebugLoc();
3491 for (unsigned i = 0; i < 8; ++i) {
3492 bool isNonZero = (NonZeros & (1 << i)) != 0;
3496 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3498 V = DAG.getUNDEF(MVT::v8i16);
3501 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3502 MVT::v8i16, V, Op.getOperand(i),
3503 DAG.getIntPtrConstant(i));
3510 /// getVShift - Return a vector logical shift node.
3512 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3513 unsigned NumBits, SelectionDAG &DAG,
3514 const TargetLowering &TLI, DebugLoc dl) {
3515 bool isMMX = VT.getSizeInBits() == 64;
3516 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3517 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3518 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3519 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3520 DAG.getNode(Opc, dl, ShVT, SrcOp,
3521 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3525 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3526 SelectionDAG &DAG) {
3528 // Check if the scalar load can be widened into a vector load. And if
3529 // the address is "base + cst" see if the cst can be "absorbed" into
3530 // the shuffle mask.
3531 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3532 SDValue Ptr = LD->getBasePtr();
3533 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3535 EVT PVT = LD->getValueType(0);
3536 if (PVT != MVT::i32 && PVT != MVT::f32)
3541 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3542 FI = FINode->getIndex();
3544 } else if (Ptr.getOpcode() == ISD::ADD &&
3545 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3546 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3547 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3548 Offset = Ptr.getConstantOperandVal(1);
3549 Ptr = Ptr.getOperand(0);
3554 SDValue Chain = LD->getChain();
3555 // Make sure the stack object alignment is at least 16.
3556 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3557 if (DAG.InferPtrAlignment(Ptr) < 16) {
3558 if (MFI->isFixedObjectIndex(FI)) {
3559 // Can't change the alignment. FIXME: It's possible to compute
3560 // the exact stack offset and reference FI + adjust offset instead.
3561 // If someone *really* cares about this. That's the way to implement it.
3564 MFI->setObjectAlignment(FI, 16);
3568 // (Offset % 16) must be multiple of 4. Then address is then
3569 // Ptr + (Offset & ~15).
3572 if ((Offset % 16) & 3)
3574 int64_t StartOffset = Offset & ~15;
3576 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3577 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3579 int EltNo = (Offset - StartOffset) >> 2;
3580 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3581 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3582 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0);
3583 // Canonicalize it to a v4i32 shuffle.
3584 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3585 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3586 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3587 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3594 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3595 DebugLoc dl = Op.getDebugLoc();
3596 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3597 if (ISD::isBuildVectorAllZeros(Op.getNode())
3598 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3599 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3600 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3601 // eliminated on x86-32 hosts.
3602 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3605 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3606 return getOnesVector(Op.getValueType(), DAG, dl);
3607 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3610 EVT VT = Op.getValueType();
3611 EVT ExtVT = VT.getVectorElementType();
3612 unsigned EVTBits = ExtVT.getSizeInBits();
3614 unsigned NumElems = Op.getNumOperands();
3615 unsigned NumZero = 0;
3616 unsigned NumNonZero = 0;
3617 unsigned NonZeros = 0;
3618 bool IsAllConstants = true;
3619 SmallSet<SDValue, 8> Values;
3620 for (unsigned i = 0; i < NumElems; ++i) {
3621 SDValue Elt = Op.getOperand(i);
3622 if (Elt.getOpcode() == ISD::UNDEF)
3625 if (Elt.getOpcode() != ISD::Constant &&
3626 Elt.getOpcode() != ISD::ConstantFP)
3627 IsAllConstants = false;
3628 if (X86::isZeroNode(Elt))
3631 NonZeros |= (1 << i);
3636 if (NumNonZero == 0) {
3637 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3638 return DAG.getUNDEF(VT);
3641 // Special case for single non-zero, non-undef, element.
3642 if (NumNonZero == 1) {
3643 unsigned Idx = CountTrailingZeros_32(NonZeros);
3644 SDValue Item = Op.getOperand(Idx);
3646 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3647 // the value are obviously zero, truncate the value to i32 and do the
3648 // insertion that way. Only do this if the value is non-constant or if the
3649 // value is a constant being inserted into element 0. It is cheaper to do
3650 // a constant pool load than it is to do a movd + shuffle.
3651 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3652 (!IsAllConstants || Idx == 0)) {
3653 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3654 // Handle MMX and SSE both.
3655 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3656 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3658 // Truncate the value (which may itself be a constant) to i32, and
3659 // convert it to a vector with movd (S2V+shuffle to zero extend).
3660 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3661 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3662 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3663 Subtarget->hasSSE2(), DAG);
3665 // Now we have our 32-bit value zero extended in the low element of
3666 // a vector. If Idx != 0, swizzle it into place.
3668 SmallVector<int, 4> Mask;
3669 Mask.push_back(Idx);
3670 for (unsigned i = 1; i != VecElts; ++i)
3672 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3673 DAG.getUNDEF(Item.getValueType()),
3676 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3680 // If we have a constant or non-constant insertion into the low element of
3681 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3682 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3683 // depending on what the source datatype is.
3686 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3687 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3688 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3689 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3690 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3691 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3693 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3694 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3695 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3696 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3697 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3698 Subtarget->hasSSE2(), DAG);
3699 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3703 // Is it a vector logical left shift?
3704 if (NumElems == 2 && Idx == 1 &&
3705 X86::isZeroNode(Op.getOperand(0)) &&
3706 !X86::isZeroNode(Op.getOperand(1))) {
3707 unsigned NumBits = VT.getSizeInBits();
3708 return getVShift(true, VT,
3709 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3710 VT, Op.getOperand(1)),
3711 NumBits/2, DAG, *this, dl);
3714 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3717 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3718 // is a non-constant being inserted into an element other than the low one,
3719 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3720 // movd/movss) to move this into the low element, then shuffle it into
3722 if (EVTBits == 32) {
3723 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3725 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3726 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3727 Subtarget->hasSSE2(), DAG);
3728 SmallVector<int, 8> MaskVec;
3729 for (unsigned i = 0; i < NumElems; i++)
3730 MaskVec.push_back(i == Idx ? 0 : 1);
3731 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3735 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3736 if (Values.size() == 1) {
3737 if (EVTBits == 32) {
3738 // Instead of a shuffle like this:
3739 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3740 // Check if it's possible to issue this instead.
3741 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3742 unsigned Idx = CountTrailingZeros_32(NonZeros);
3743 SDValue Item = Op.getOperand(Idx);
3744 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3745 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3750 // A vector full of immediates; various special cases are already
3751 // handled, so this is best done with a single constant-pool load.
3755 // Let legalizer expand 2-wide build_vectors.
3756 if (EVTBits == 64) {
3757 if (NumNonZero == 1) {
3758 // One half is zero or undef.
3759 unsigned Idx = CountTrailingZeros_32(NonZeros);
3760 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3761 Op.getOperand(Idx));
3762 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3763 Subtarget->hasSSE2(), DAG);
3768 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3769 if (EVTBits == 8 && NumElems == 16) {
3770 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3772 if (V.getNode()) return V;
3775 if (EVTBits == 16 && NumElems == 8) {
3776 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3778 if (V.getNode()) return V;
3781 // If element VT is == 32 bits, turn it into a number of shuffles.
3782 SmallVector<SDValue, 8> V;
3784 if (NumElems == 4 && NumZero > 0) {
3785 for (unsigned i = 0; i < 4; ++i) {
3786 bool isZero = !(NonZeros & (1 << i));
3788 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3790 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3793 for (unsigned i = 0; i < 2; ++i) {
3794 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3797 V[i] = V[i*2]; // Must be a zero vector.
3800 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3803 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3806 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3811 SmallVector<int, 8> MaskVec;
3812 bool Reverse = (NonZeros & 0x3) == 2;
3813 for (unsigned i = 0; i < 2; ++i)
3814 MaskVec.push_back(Reverse ? 1-i : i);
3815 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3816 for (unsigned i = 0; i < 2; ++i)
3817 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3818 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3821 if (Values.size() > 2) {
3822 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3823 // values to be inserted is equal to the number of elements, in which case
3824 // use the unpack code below in the hopes of matching the consecutive elts
3825 // load merge pattern for shuffles.
3826 // FIXME: We could probably just check that here directly.
3827 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3828 getSubtarget()->hasSSE41()) {
3829 V[0] = DAG.getUNDEF(VT);
3830 for (unsigned i = 0; i < NumElems; ++i)
3831 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3832 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3833 Op.getOperand(i), DAG.getIntPtrConstant(i));
3836 // Expand into a number of unpckl*.
3838 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3839 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3840 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3841 for (unsigned i = 0; i < NumElems; ++i)
3842 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3844 while (NumElems != 0) {
3845 for (unsigned i = 0; i < NumElems; ++i)
3846 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3856 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3857 // We support concatenate two MMX registers and place them in a MMX
3858 // register. This is better than doing a stack convert.
3859 DebugLoc dl = Op.getDebugLoc();
3860 EVT ResVT = Op.getValueType();
3861 assert(Op.getNumOperands() == 2);
3862 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
3863 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
3865 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
3866 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3867 InVec = Op.getOperand(1);
3868 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3869 unsigned NumElts = ResVT.getVectorNumElements();
3870 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3871 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
3872 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
3874 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
3875 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3876 Mask[0] = 0; Mask[1] = 2;
3877 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
3879 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3882 // v8i16 shuffles - Prefer shuffles in the following order:
3883 // 1. [all] pshuflw, pshufhw, optional move
3884 // 2. [ssse3] 1 x pshufb
3885 // 3. [ssse3] 2 x pshufb + 1 x por
3886 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3888 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3889 SelectionDAG &DAG, X86TargetLowering &TLI) {
3890 SDValue V1 = SVOp->getOperand(0);
3891 SDValue V2 = SVOp->getOperand(1);
3892 DebugLoc dl = SVOp->getDebugLoc();
3893 SmallVector<int, 8> MaskVals;
3895 // Determine if more than 1 of the words in each of the low and high quadwords
3896 // of the result come from the same quadword of one of the two inputs. Undef
3897 // mask values count as coming from any quadword, for better codegen.
3898 SmallVector<unsigned, 4> LoQuad(4);
3899 SmallVector<unsigned, 4> HiQuad(4);
3900 BitVector InputQuads(4);
3901 for (unsigned i = 0; i < 8; ++i) {
3902 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3903 int EltIdx = SVOp->getMaskElt(i);
3904 MaskVals.push_back(EltIdx);
3913 InputQuads.set(EltIdx / 4);
3916 int BestLoQuad = -1;
3917 unsigned MaxQuad = 1;
3918 for (unsigned i = 0; i < 4; ++i) {
3919 if (LoQuad[i] > MaxQuad) {
3921 MaxQuad = LoQuad[i];
3925 int BestHiQuad = -1;
3927 for (unsigned i = 0; i < 4; ++i) {
3928 if (HiQuad[i] > MaxQuad) {
3930 MaxQuad = HiQuad[i];
3934 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3935 // of the two input vectors, shuffle them into one input vector so only a
3936 // single pshufb instruction is necessary. If There are more than 2 input
3937 // quads, disable the next transformation since it does not help SSSE3.
3938 bool V1Used = InputQuads[0] || InputQuads[1];
3939 bool V2Used = InputQuads[2] || InputQuads[3];
3940 if (TLI.getSubtarget()->hasSSSE3()) {
3941 if (InputQuads.count() == 2 && V1Used && V2Used) {
3942 BestLoQuad = InputQuads.find_first();
3943 BestHiQuad = InputQuads.find_next(BestLoQuad);
3945 if (InputQuads.count() > 2) {
3951 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3952 // the shuffle mask. If a quad is scored as -1, that means that it contains
3953 // words from all 4 input quadwords.
3955 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3956 SmallVector<int, 8> MaskV;
3957 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3958 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3959 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3960 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3961 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3962 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3964 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3965 // source words for the shuffle, to aid later transformations.
3966 bool AllWordsInNewV = true;
3967 bool InOrder[2] = { true, true };
3968 for (unsigned i = 0; i != 8; ++i) {
3969 int idx = MaskVals[i];
3971 InOrder[i/4] = false;
3972 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3974 AllWordsInNewV = false;
3978 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3979 if (AllWordsInNewV) {
3980 for (int i = 0; i != 8; ++i) {
3981 int idx = MaskVals[i];
3984 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3985 if ((idx != i) && idx < 4)
3987 if ((idx != i) && idx > 3)
3996 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3997 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3998 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3999 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4000 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4004 // If we have SSSE3, and all words of the result are from 1 input vector,
4005 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4006 // is present, fall back to case 4.
4007 if (TLI.getSubtarget()->hasSSSE3()) {
4008 SmallVector<SDValue,16> pshufbMask;
4010 // If we have elements from both input vectors, set the high bit of the
4011 // shuffle mask element to zero out elements that come from V2 in the V1
4012 // mask, and elements that come from V1 in the V2 mask, so that the two
4013 // results can be OR'd together.
4014 bool TwoInputs = V1Used && V2Used;
4015 for (unsigned i = 0; i != 8; ++i) {
4016 int EltIdx = MaskVals[i] * 2;
4017 if (TwoInputs && (EltIdx >= 16)) {
4018 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4019 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4022 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4023 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4025 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4026 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4027 DAG.getNode(ISD::BUILD_VECTOR, dl,
4028 MVT::v16i8, &pshufbMask[0], 16));
4030 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4032 // Calculate the shuffle mask for the second input, shuffle it, and
4033 // OR it with the first shuffled input.
4035 for (unsigned i = 0; i != 8; ++i) {
4036 int EltIdx = MaskVals[i] * 2;
4038 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4039 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4042 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4043 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4045 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4046 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4047 DAG.getNode(ISD::BUILD_VECTOR, dl,
4048 MVT::v16i8, &pshufbMask[0], 16));
4049 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4050 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4053 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4054 // and update MaskVals with new element order.
4055 BitVector InOrder(8);
4056 if (BestLoQuad >= 0) {
4057 SmallVector<int, 8> MaskV;
4058 for (int i = 0; i != 4; ++i) {
4059 int idx = MaskVals[i];
4061 MaskV.push_back(-1);
4063 } else if ((idx / 4) == BestLoQuad) {
4064 MaskV.push_back(idx & 3);
4067 MaskV.push_back(-1);
4070 for (unsigned i = 4; i != 8; ++i)
4072 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4076 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4077 // and update MaskVals with the new element order.
4078 if (BestHiQuad >= 0) {
4079 SmallVector<int, 8> MaskV;
4080 for (unsigned i = 0; i != 4; ++i)
4082 for (unsigned i = 4; i != 8; ++i) {
4083 int idx = MaskVals[i];
4085 MaskV.push_back(-1);
4087 } else if ((idx / 4) == BestHiQuad) {
4088 MaskV.push_back((idx & 3) + 4);
4091 MaskV.push_back(-1);
4094 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4098 // In case BestHi & BestLo were both -1, which means each quadword has a word
4099 // from each of the four input quadwords, calculate the InOrder bitvector now
4100 // before falling through to the insert/extract cleanup.
4101 if (BestLoQuad == -1 && BestHiQuad == -1) {
4103 for (int i = 0; i != 8; ++i)
4104 if (MaskVals[i] < 0 || MaskVals[i] == i)
4108 // The other elements are put in the right place using pextrw and pinsrw.
4109 for (unsigned i = 0; i != 8; ++i) {
4112 int EltIdx = MaskVals[i];
4115 SDValue ExtOp = (EltIdx < 8)
4116 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4117 DAG.getIntPtrConstant(EltIdx))
4118 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4119 DAG.getIntPtrConstant(EltIdx - 8));
4120 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4121 DAG.getIntPtrConstant(i));
4126 // v16i8 shuffles - Prefer shuffles in the following order:
4127 // 1. [ssse3] 1 x pshufb
4128 // 2. [ssse3] 2 x pshufb + 1 x por
4129 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4131 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4132 SelectionDAG &DAG, X86TargetLowering &TLI) {
4133 SDValue V1 = SVOp->getOperand(0);
4134 SDValue V2 = SVOp->getOperand(1);
4135 DebugLoc dl = SVOp->getDebugLoc();
4136 SmallVector<int, 16> MaskVals;
4137 SVOp->getMask(MaskVals);
4139 // If we have SSSE3, case 1 is generated when all result bytes come from
4140 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4141 // present, fall back to case 3.
4142 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4145 for (unsigned i = 0; i < 16; ++i) {
4146 int EltIdx = MaskVals[i];
4155 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4156 if (TLI.getSubtarget()->hasSSSE3()) {
4157 SmallVector<SDValue,16> pshufbMask;
4159 // If all result elements are from one input vector, then only translate
4160 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4162 // Otherwise, we have elements from both input vectors, and must zero out
4163 // elements that come from V2 in the first mask, and V1 in the second mask
4164 // so that we can OR them together.
4165 bool TwoInputs = !(V1Only || V2Only);
4166 for (unsigned i = 0; i != 16; ++i) {
4167 int EltIdx = MaskVals[i];
4168 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4169 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4172 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4174 // If all the elements are from V2, assign it to V1 and return after
4175 // building the first pshufb.
4178 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4179 DAG.getNode(ISD::BUILD_VECTOR, dl,
4180 MVT::v16i8, &pshufbMask[0], 16));
4184 // Calculate the shuffle mask for the second input, shuffle it, and
4185 // OR it with the first shuffled input.
4187 for (unsigned i = 0; i != 16; ++i) {
4188 int EltIdx = MaskVals[i];
4190 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4193 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4195 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4196 DAG.getNode(ISD::BUILD_VECTOR, dl,
4197 MVT::v16i8, &pshufbMask[0], 16));
4198 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4201 // No SSSE3 - Calculate in place words and then fix all out of place words
4202 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4203 // the 16 different words that comprise the two doublequadword input vectors.
4204 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4205 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4206 SDValue NewV = V2Only ? V2 : V1;
4207 for (int i = 0; i != 8; ++i) {
4208 int Elt0 = MaskVals[i*2];
4209 int Elt1 = MaskVals[i*2+1];
4211 // This word of the result is all undef, skip it.
4212 if (Elt0 < 0 && Elt1 < 0)
4215 // This word of the result is already in the correct place, skip it.
4216 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4218 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4221 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4222 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4225 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4226 // using a single extract together, load it and store it.
4227 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4228 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4229 DAG.getIntPtrConstant(Elt1 / 2));
4230 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4231 DAG.getIntPtrConstant(i));
4235 // If Elt1 is defined, extract it from the appropriate source. If the
4236 // source byte is not also odd, shift the extracted word left 8 bits
4237 // otherwise clear the bottom 8 bits if we need to do an or.
4239 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4240 DAG.getIntPtrConstant(Elt1 / 2));
4241 if ((Elt1 & 1) == 0)
4242 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4243 DAG.getConstant(8, TLI.getShiftAmountTy()));
4245 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4246 DAG.getConstant(0xFF00, MVT::i16));
4248 // If Elt0 is defined, extract it from the appropriate source. If the
4249 // source byte is not also even, shift the extracted word right 8 bits. If
4250 // Elt1 was also defined, OR the extracted values together before
4251 // inserting them in the result.
4253 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4254 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4255 if ((Elt0 & 1) != 0)
4256 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4257 DAG.getConstant(8, TLI.getShiftAmountTy()));
4259 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4260 DAG.getConstant(0x00FF, MVT::i16));
4261 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4264 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4265 DAG.getIntPtrConstant(i));
4267 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4270 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4271 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4272 /// done when every pair / quad of shuffle mask elements point to elements in
4273 /// the right sequence. e.g.
4274 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4276 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4278 TargetLowering &TLI, DebugLoc dl) {
4279 EVT VT = SVOp->getValueType(0);
4280 SDValue V1 = SVOp->getOperand(0);
4281 SDValue V2 = SVOp->getOperand(1);
4282 unsigned NumElems = VT.getVectorNumElements();
4283 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4284 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4285 EVT MaskEltVT = MaskVT.getVectorElementType();
4287 switch (VT.getSimpleVT().SimpleTy) {
4288 default: assert(false && "Unexpected!");
4289 case MVT::v4f32: NewVT = MVT::v2f64; break;
4290 case MVT::v4i32: NewVT = MVT::v2i64; break;
4291 case MVT::v8i16: NewVT = MVT::v4i32; break;
4292 case MVT::v16i8: NewVT = MVT::v4i32; break;
4295 if (NewWidth == 2) {
4301 int Scale = NumElems / NewWidth;
4302 SmallVector<int, 8> MaskVec;
4303 for (unsigned i = 0; i < NumElems; i += Scale) {
4305 for (int j = 0; j < Scale; ++j) {
4306 int EltIdx = SVOp->getMaskElt(i+j);
4310 StartIdx = EltIdx - (EltIdx % Scale);
4311 if (EltIdx != StartIdx + j)
4315 MaskVec.push_back(-1);
4317 MaskVec.push_back(StartIdx / Scale);
4320 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4321 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4322 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4325 /// getVZextMovL - Return a zero-extending vector move low node.
4327 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4328 SDValue SrcOp, SelectionDAG &DAG,
4329 const X86Subtarget *Subtarget, DebugLoc dl) {
4330 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4331 LoadSDNode *LD = NULL;
4332 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4333 LD = dyn_cast<LoadSDNode>(SrcOp);
4335 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4337 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4338 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4339 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4340 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4341 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4343 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4344 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4345 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4346 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4354 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4355 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4356 DAG.getNode(ISD::BIT_CONVERT, dl,
4360 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4363 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4364 SDValue V1 = SVOp->getOperand(0);
4365 SDValue V2 = SVOp->getOperand(1);
4366 DebugLoc dl = SVOp->getDebugLoc();
4367 EVT VT = SVOp->getValueType(0);
4369 SmallVector<std::pair<int, int>, 8> Locs;
4371 SmallVector<int, 8> Mask1(4U, -1);
4372 SmallVector<int, 8> PermMask;
4373 SVOp->getMask(PermMask);
4377 for (unsigned i = 0; i != 4; ++i) {
4378 int Idx = PermMask[i];
4380 Locs[i] = std::make_pair(-1, -1);
4382 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4384 Locs[i] = std::make_pair(0, NumLo);
4388 Locs[i] = std::make_pair(1, NumHi);
4390 Mask1[2+NumHi] = Idx;
4396 if (NumLo <= 2 && NumHi <= 2) {
4397 // If no more than two elements come from either vector. This can be
4398 // implemented with two shuffles. First shuffle gather the elements.
4399 // The second shuffle, which takes the first shuffle as both of its
4400 // vector operands, put the elements into the right order.
4401 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4403 SmallVector<int, 8> Mask2(4U, -1);
4405 for (unsigned i = 0; i != 4; ++i) {
4406 if (Locs[i].first == -1)
4409 unsigned Idx = (i < 2) ? 0 : 4;
4410 Idx += Locs[i].first * 2 + Locs[i].second;
4415 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4416 } else if (NumLo == 3 || NumHi == 3) {
4417 // Otherwise, we must have three elements from one vector, call it X, and
4418 // one element from the other, call it Y. First, use a shufps to build an
4419 // intermediate vector with the one element from Y and the element from X
4420 // that will be in the same half in the final destination (the indexes don't
4421 // matter). Then, use a shufps to build the final vector, taking the half
4422 // containing the element from Y from the intermediate, and the other half
4425 // Normalize it so the 3 elements come from V1.
4426 CommuteVectorShuffleMask(PermMask, VT);
4430 // Find the element from V2.
4432 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4433 int Val = PermMask[HiIndex];
4440 Mask1[0] = PermMask[HiIndex];
4442 Mask1[2] = PermMask[HiIndex^1];
4444 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4447 Mask1[0] = PermMask[0];
4448 Mask1[1] = PermMask[1];
4449 Mask1[2] = HiIndex & 1 ? 6 : 4;
4450 Mask1[3] = HiIndex & 1 ? 4 : 6;
4451 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4453 Mask1[0] = HiIndex & 1 ? 2 : 0;
4454 Mask1[1] = HiIndex & 1 ? 0 : 2;
4455 Mask1[2] = PermMask[2];
4456 Mask1[3] = PermMask[3];
4461 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4465 // Break it into (shuffle shuffle_hi, shuffle_lo).
4467 SmallVector<int,8> LoMask(4U, -1);
4468 SmallVector<int,8> HiMask(4U, -1);
4470 SmallVector<int,8> *MaskPtr = &LoMask;
4471 unsigned MaskIdx = 0;
4474 for (unsigned i = 0; i != 4; ++i) {
4481 int Idx = PermMask[i];
4483 Locs[i] = std::make_pair(-1, -1);
4484 } else if (Idx < 4) {
4485 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4486 (*MaskPtr)[LoIdx] = Idx;
4489 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4490 (*MaskPtr)[HiIdx] = Idx;
4495 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4496 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4497 SmallVector<int, 8> MaskOps;
4498 for (unsigned i = 0; i != 4; ++i) {
4499 if (Locs[i].first == -1) {
4500 MaskOps.push_back(-1);
4502 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4503 MaskOps.push_back(Idx);
4506 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4510 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4511 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4512 SDValue V1 = Op.getOperand(0);
4513 SDValue V2 = Op.getOperand(1);
4514 EVT VT = Op.getValueType();
4515 DebugLoc dl = Op.getDebugLoc();
4516 unsigned NumElems = VT.getVectorNumElements();
4517 bool isMMX = VT.getSizeInBits() == 64;
4518 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4519 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4520 bool V1IsSplat = false;
4521 bool V2IsSplat = false;
4523 if (isZeroShuffle(SVOp))
4524 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4526 // Promote splats to v4f32.
4527 if (SVOp->isSplat()) {
4528 if (isMMX || NumElems < 4)
4530 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4533 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4535 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4536 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4537 if (NewOp.getNode())
4538 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4539 LowerVECTOR_SHUFFLE(NewOp, DAG));
4540 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4541 // FIXME: Figure out a cleaner way to do this.
4542 // Try to make use of movq to zero out the top part.
4543 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4544 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4545 if (NewOp.getNode()) {
4546 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4547 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4548 DAG, Subtarget, dl);
4550 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4551 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4552 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4553 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4554 DAG, Subtarget, dl);
4558 if (X86::isPSHUFDMask(SVOp))
4561 // Check if this can be converted into a logical shift.
4562 bool isLeft = false;
4565 bool isShift = getSubtarget()->hasSSE2() &&
4566 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4567 if (isShift && ShVal.hasOneUse()) {
4568 // If the shifted value has multiple uses, it may be cheaper to use
4569 // v_set0 + movlhps or movhlps, etc.
4570 EVT EltVT = VT.getVectorElementType();
4571 ShAmt *= EltVT.getSizeInBits();
4572 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4575 if (X86::isMOVLMask(SVOp)) {
4578 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4579 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4584 // FIXME: fold these into legal mask.
4585 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4586 X86::isMOVSLDUPMask(SVOp) ||
4587 X86::isMOVHLPSMask(SVOp) ||
4588 X86::isMOVLHPSMask(SVOp) ||
4589 X86::isMOVLPMask(SVOp)))
4592 if (ShouldXformToMOVHLPS(SVOp) ||
4593 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4594 return CommuteVectorShuffle(SVOp, DAG);
4597 // No better options. Use a vshl / vsrl.
4598 EVT EltVT = VT.getVectorElementType();
4599 ShAmt *= EltVT.getSizeInBits();
4600 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4603 bool Commuted = false;
4604 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4605 // 1,1,1,1 -> v8i16 though.
4606 V1IsSplat = isSplatVector(V1.getNode());
4607 V2IsSplat = isSplatVector(V2.getNode());
4609 // Canonicalize the splat or undef, if present, to be on the RHS.
4610 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4611 Op = CommuteVectorShuffle(SVOp, DAG);
4612 SVOp = cast<ShuffleVectorSDNode>(Op);
4613 V1 = SVOp->getOperand(0);
4614 V2 = SVOp->getOperand(1);
4615 std::swap(V1IsSplat, V2IsSplat);
4616 std::swap(V1IsUndef, V2IsUndef);
4620 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4621 // Shuffling low element of v1 into undef, just return v1.
4624 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4625 // the instruction selector will not match, so get a canonical MOVL with
4626 // swapped operands to undo the commute.
4627 return getMOVL(DAG, dl, VT, V2, V1);
4630 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4631 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4632 X86::isUNPCKLMask(SVOp) ||
4633 X86::isUNPCKHMask(SVOp))
4637 // Normalize mask so all entries that point to V2 points to its first
4638 // element then try to match unpck{h|l} again. If match, return a
4639 // new vector_shuffle with the corrected mask.
4640 SDValue NewMask = NormalizeMask(SVOp, DAG);
4641 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4642 if (NSVOp != SVOp) {
4643 if (X86::isUNPCKLMask(NSVOp, true)) {
4645 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4652 // Commute is back and try unpck* again.
4653 // FIXME: this seems wrong.
4654 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4655 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4656 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4657 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4658 X86::isUNPCKLMask(NewSVOp) ||
4659 X86::isUNPCKHMask(NewSVOp))
4663 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4665 // Normalize the node to match x86 shuffle ops if needed
4666 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4667 return CommuteVectorShuffle(SVOp, DAG);
4669 // Check for legal shuffle and return?
4670 SmallVector<int, 16> PermMask;
4671 SVOp->getMask(PermMask);
4672 if (isShuffleMaskLegal(PermMask, VT))
4675 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4676 if (VT == MVT::v8i16) {
4677 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4678 if (NewOp.getNode())
4682 if (VT == MVT::v16i8) {
4683 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4684 if (NewOp.getNode())
4688 // Handle all 4 wide cases with a number of shuffles except for MMX.
4689 if (NumElems == 4 && !isMMX)
4690 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4696 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4697 SelectionDAG &DAG) {
4698 EVT VT = Op.getValueType();
4699 DebugLoc dl = Op.getDebugLoc();
4700 if (VT.getSizeInBits() == 8) {
4701 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4702 Op.getOperand(0), Op.getOperand(1));
4703 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4704 DAG.getValueType(VT));
4705 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4706 } else if (VT.getSizeInBits() == 16) {
4707 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4708 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4710 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4711 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4712 DAG.getNode(ISD::BIT_CONVERT, dl,
4716 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4717 Op.getOperand(0), Op.getOperand(1));
4718 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4719 DAG.getValueType(VT));
4720 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4721 } else if (VT == MVT::f32) {
4722 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4723 // the result back to FR32 register. It's only worth matching if the
4724 // result has a single use which is a store or a bitcast to i32. And in
4725 // the case of a store, it's not worth it if the index is a constant 0,
4726 // because a MOVSSmr can be used instead, which is smaller and faster.
4727 if (!Op.hasOneUse())
4729 SDNode *User = *Op.getNode()->use_begin();
4730 if ((User->getOpcode() != ISD::STORE ||
4731 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4732 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4733 (User->getOpcode() != ISD::BIT_CONVERT ||
4734 User->getValueType(0) != MVT::i32))
4736 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4737 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4740 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4741 } else if (VT == MVT::i32) {
4742 // ExtractPS works with constant index.
4743 if (isa<ConstantSDNode>(Op.getOperand(1)))
4751 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4752 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4755 if (Subtarget->hasSSE41()) {
4756 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4761 EVT VT = Op.getValueType();
4762 DebugLoc dl = Op.getDebugLoc();
4763 // TODO: handle v16i8.
4764 if (VT.getSizeInBits() == 16) {
4765 SDValue Vec = Op.getOperand(0);
4766 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4768 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4769 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4770 DAG.getNode(ISD::BIT_CONVERT, dl,
4773 // Transform it so it match pextrw which produces a 32-bit result.
4774 EVT EltVT = MVT::i32;
4775 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4776 Op.getOperand(0), Op.getOperand(1));
4777 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4778 DAG.getValueType(VT));
4779 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4780 } else if (VT.getSizeInBits() == 32) {
4781 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4785 // SHUFPS the element to the lowest double word, then movss.
4786 int Mask[4] = { Idx, -1, -1, -1 };
4787 EVT VVT = Op.getOperand(0).getValueType();
4788 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4789 DAG.getUNDEF(VVT), Mask);
4790 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4791 DAG.getIntPtrConstant(0));
4792 } else if (VT.getSizeInBits() == 64) {
4793 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4794 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4795 // to match extract_elt for f64.
4796 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4800 // UNPCKHPD the element to the lowest double word, then movsd.
4801 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4802 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4803 int Mask[2] = { 1, -1 };
4804 EVT VVT = Op.getOperand(0).getValueType();
4805 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4806 DAG.getUNDEF(VVT), Mask);
4807 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4808 DAG.getIntPtrConstant(0));
4815 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4816 EVT VT = Op.getValueType();
4817 EVT EltVT = VT.getVectorElementType();
4818 DebugLoc dl = Op.getDebugLoc();
4820 SDValue N0 = Op.getOperand(0);
4821 SDValue N1 = Op.getOperand(1);
4822 SDValue N2 = Op.getOperand(2);
4824 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4825 isa<ConstantSDNode>(N2)) {
4826 unsigned Opc = (EltVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4828 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4830 if (N1.getValueType() != MVT::i32)
4831 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4832 if (N2.getValueType() != MVT::i32)
4833 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4834 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4835 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4836 // Bits [7:6] of the constant are the source select. This will always be
4837 // zero here. The DAG Combiner may combine an extract_elt index into these
4838 // bits. For example (insert (extract, 3), 2) could be matched by putting
4839 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4840 // Bits [5:4] of the constant are the destination select. This is the
4841 // value of the incoming immediate.
4842 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4843 // combine either bitwise AND or insert of float 0.0 to set these bits.
4844 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4845 // Create this as a scalar to vector..
4846 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4847 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4848 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4849 // PINSR* works with constant index.
4856 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4857 EVT VT = Op.getValueType();
4858 EVT EltVT = VT.getVectorElementType();
4860 if (Subtarget->hasSSE41())
4861 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4863 if (EltVT == MVT::i8)
4866 DebugLoc dl = Op.getDebugLoc();
4867 SDValue N0 = Op.getOperand(0);
4868 SDValue N1 = Op.getOperand(1);
4869 SDValue N2 = Op.getOperand(2);
4871 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4872 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4873 // as its second argument.
4874 if (N1.getValueType() != MVT::i32)
4875 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4876 if (N2.getValueType() != MVT::i32)
4877 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4878 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4884 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4885 DebugLoc dl = Op.getDebugLoc();
4886 if (Op.getValueType() == MVT::v2f32)
4887 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4888 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4889 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4890 Op.getOperand(0))));
4892 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4893 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4895 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4896 EVT VT = MVT::v2i32;
4897 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4904 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4905 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4908 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4909 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4910 // one of the above mentioned nodes. It has to be wrapped because otherwise
4911 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4912 // be used to form addressing mode. These wrapped nodes will be selected
4915 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4916 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4918 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4920 unsigned char OpFlag = 0;
4921 unsigned WrapperKind = X86ISD::Wrapper;
4922 CodeModel::Model M = getTargetMachine().getCodeModel();
4924 if (Subtarget->isPICStyleRIPRel() &&
4925 (M == CodeModel::Small || M == CodeModel::Kernel))
4926 WrapperKind = X86ISD::WrapperRIP;
4927 else if (Subtarget->isPICStyleGOT())
4928 OpFlag = X86II::MO_GOTOFF;
4929 else if (Subtarget->isPICStyleStubPIC())
4930 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4932 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4934 CP->getOffset(), OpFlag);
4935 DebugLoc DL = CP->getDebugLoc();
4936 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4937 // With PIC, the address is actually $g + Offset.
4939 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4940 DAG.getNode(X86ISD::GlobalBaseReg,
4941 DebugLoc::getUnknownLoc(), getPointerTy()),
4948 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4949 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4951 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4953 unsigned char OpFlag = 0;
4954 unsigned WrapperKind = X86ISD::Wrapper;
4955 CodeModel::Model M = getTargetMachine().getCodeModel();
4957 if (Subtarget->isPICStyleRIPRel() &&
4958 (M == CodeModel::Small || M == CodeModel::Kernel))
4959 WrapperKind = X86ISD::WrapperRIP;
4960 else if (Subtarget->isPICStyleGOT())
4961 OpFlag = X86II::MO_GOTOFF;
4962 else if (Subtarget->isPICStyleStubPIC())
4963 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4965 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4967 DebugLoc DL = JT->getDebugLoc();
4968 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4970 // With PIC, the address is actually $g + Offset.
4972 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4973 DAG.getNode(X86ISD::GlobalBaseReg,
4974 DebugLoc::getUnknownLoc(), getPointerTy()),
4982 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4983 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4985 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4987 unsigned char OpFlag = 0;
4988 unsigned WrapperKind = X86ISD::Wrapper;
4989 CodeModel::Model M = getTargetMachine().getCodeModel();
4991 if (Subtarget->isPICStyleRIPRel() &&
4992 (M == CodeModel::Small || M == CodeModel::Kernel))
4993 WrapperKind = X86ISD::WrapperRIP;
4994 else if (Subtarget->isPICStyleGOT())
4995 OpFlag = X86II::MO_GOTOFF;
4996 else if (Subtarget->isPICStyleStubPIC())
4997 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4999 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5001 DebugLoc DL = Op.getDebugLoc();
5002 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5005 // With PIC, the address is actually $g + Offset.
5006 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5007 !Subtarget->is64Bit()) {
5008 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5009 DAG.getNode(X86ISD::GlobalBaseReg,
5010 DebugLoc::getUnknownLoc(),
5019 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
5020 // Create the TargetBlockAddressAddress node.
5021 unsigned char OpFlags =
5022 Subtarget->ClassifyBlockAddressReference();
5023 CodeModel::Model M = getTargetMachine().getCodeModel();
5024 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5025 DebugLoc dl = Op.getDebugLoc();
5026 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5027 /*isTarget=*/true, OpFlags);
5029 if (Subtarget->isPICStyleRIPRel() &&
5030 (M == CodeModel::Small || M == CodeModel::Kernel))
5031 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5033 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5035 // With PIC, the address is actually $g + Offset.
5036 if (isGlobalRelativeToPICBase(OpFlags)) {
5037 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5038 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5046 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5048 SelectionDAG &DAG) const {
5049 // Create the TargetGlobalAddress node, folding in the constant
5050 // offset if it is legal.
5051 unsigned char OpFlags =
5052 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5053 CodeModel::Model M = getTargetMachine().getCodeModel();
5055 if (OpFlags == X86II::MO_NO_FLAG &&
5056 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5057 // A direct static reference to a global.
5058 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5061 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5064 if (Subtarget->isPICStyleRIPRel() &&
5065 (M == CodeModel::Small || M == CodeModel::Kernel))
5066 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5068 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5070 // With PIC, the address is actually $g + Offset.
5071 if (isGlobalRelativeToPICBase(OpFlags)) {
5072 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5073 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5077 // For globals that require a load from a stub to get the address, emit the
5079 if (isGlobalStubReference(OpFlags))
5080 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5081 PseudoSourceValue::getGOT(), 0);
5083 // If there was a non-zero offset that we didn't fold, create an explicit
5086 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5087 DAG.getConstant(Offset, getPointerTy()));
5093 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
5094 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5095 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5096 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5100 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5101 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5102 unsigned char OperandFlags) {
5103 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5104 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5105 DebugLoc dl = GA->getDebugLoc();
5106 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5107 GA->getValueType(0),
5111 SDValue Ops[] = { Chain, TGA, *InFlag };
5112 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5114 SDValue Ops[] = { Chain, TGA };
5115 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5118 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5119 MFI->setHasCalls(true);
5121 SDValue Flag = Chain.getValue(1);
5122 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5125 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5127 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5130 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5131 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5132 DAG.getNode(X86ISD::GlobalBaseReg,
5133 DebugLoc::getUnknownLoc(),
5135 InFlag = Chain.getValue(1);
5137 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5140 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5142 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5144 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5145 X86::RAX, X86II::MO_TLSGD);
5148 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5149 // "local exec" model.
5150 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5151 const EVT PtrVT, TLSModel::Model model,
5153 DebugLoc dl = GA->getDebugLoc();
5154 // Get the Thread Pointer
5155 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5156 DebugLoc::getUnknownLoc(), PtrVT,
5157 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5160 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5163 unsigned char OperandFlags = 0;
5164 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5166 unsigned WrapperKind = X86ISD::Wrapper;
5167 if (model == TLSModel::LocalExec) {
5168 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5169 } else if (is64Bit) {
5170 assert(model == TLSModel::InitialExec);
5171 OperandFlags = X86II::MO_GOTTPOFF;
5172 WrapperKind = X86ISD::WrapperRIP;
5174 assert(model == TLSModel::InitialExec);
5175 OperandFlags = X86II::MO_INDNTPOFF;
5178 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5180 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5181 GA->getOffset(), OperandFlags);
5182 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5184 if (model == TLSModel::InitialExec)
5185 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5186 PseudoSourceValue::getGOT(), 0);
5188 // The address of the thread local variable is the add of the thread
5189 // pointer with the offset of the variable.
5190 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5194 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
5195 // TODO: implement the "local dynamic" model
5196 // TODO: implement the "initial exec"model for pic executables
5197 assert(Subtarget->isTargetELF() &&
5198 "TLS not implemented for non-ELF targets");
5199 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5200 const GlobalValue *GV = GA->getGlobal();
5202 // If GV is an alias then use the aliasee for determining
5203 // thread-localness.
5204 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5205 GV = GA->resolveAliasedGlobal(false);
5207 TLSModel::Model model = getTLSModel(GV,
5208 getTargetMachine().getRelocationModel());
5211 case TLSModel::GeneralDynamic:
5212 case TLSModel::LocalDynamic: // not implemented
5213 if (Subtarget->is64Bit())
5214 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5215 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5217 case TLSModel::InitialExec:
5218 case TLSModel::LocalExec:
5219 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5220 Subtarget->is64Bit());
5223 llvm_unreachable("Unreachable");
5228 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5229 /// take a 2 x i32 value to shift plus a shift amount.
5230 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
5231 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5232 EVT VT = Op.getValueType();
5233 unsigned VTBits = VT.getSizeInBits();
5234 DebugLoc dl = Op.getDebugLoc();
5235 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5236 SDValue ShOpLo = Op.getOperand(0);
5237 SDValue ShOpHi = Op.getOperand(1);
5238 SDValue ShAmt = Op.getOperand(2);
5239 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5240 DAG.getConstant(VTBits - 1, MVT::i8))
5241 : DAG.getConstant(0, VT);
5244 if (Op.getOpcode() == ISD::SHL_PARTS) {
5245 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5246 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5248 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5249 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5252 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5253 DAG.getConstant(VTBits, MVT::i8));
5254 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
5255 AndNode, DAG.getConstant(0, MVT::i8));
5258 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5259 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5260 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5262 if (Op.getOpcode() == ISD::SHL_PARTS) {
5263 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5264 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5266 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5267 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5270 SDValue Ops[2] = { Lo, Hi };
5271 return DAG.getMergeValues(Ops, 2, dl);
5274 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5275 EVT SrcVT = Op.getOperand(0).getValueType();
5277 if (SrcVT.isVector()) {
5278 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5284 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5285 "Unknown SINT_TO_FP to lower!");
5287 // These are really Legal; return the operand so the caller accepts it as
5289 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5291 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5292 Subtarget->is64Bit()) {
5296 DebugLoc dl = Op.getDebugLoc();
5297 unsigned Size = SrcVT.getSizeInBits()/8;
5298 MachineFunction &MF = DAG.getMachineFunction();
5299 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5300 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5301 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5303 PseudoSourceValue::getFixedStack(SSFI), 0);
5304 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5307 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5309 SelectionDAG &DAG) {
5311 DebugLoc dl = Op.getDebugLoc();
5313 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5315 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5317 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5318 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5319 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5320 Tys, Ops, array_lengthof(Ops));
5323 Chain = Result.getValue(1);
5324 SDValue InFlag = Result.getValue(2);
5326 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5327 // shouldn't be necessary except that RFP cannot be live across
5328 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5329 MachineFunction &MF = DAG.getMachineFunction();
5330 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5331 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5332 Tys = DAG.getVTList(MVT::Other);
5334 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5336 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5337 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5338 PseudoSourceValue::getFixedStack(SSFI), 0);
5344 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5345 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5346 // This algorithm is not obvious. Here it is in C code, more or less:
5348 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5349 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5350 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5352 // Copy ints to xmm registers.
5353 __m128i xh = _mm_cvtsi32_si128( hi );
5354 __m128i xl = _mm_cvtsi32_si128( lo );
5356 // Combine into low half of a single xmm register.
5357 __m128i x = _mm_unpacklo_epi32( xh, xl );
5361 // Merge in appropriate exponents to give the integer bits the right
5363 x = _mm_unpacklo_epi32( x, exp );
5365 // Subtract away the biases to deal with the IEEE-754 double precision
5367 d = _mm_sub_pd( (__m128d) x, bias );
5369 // All conversions up to here are exact. The correctly rounded result is
5370 // calculated using the current rounding mode using the following
5372 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5373 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5374 // store doesn't really need to be here (except
5375 // maybe to zero the other double)
5380 DebugLoc dl = Op.getDebugLoc();
5381 LLVMContext *Context = DAG.getContext();
5383 // Build some magic constants.
5384 std::vector<Constant*> CV0;
5385 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5386 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5387 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5388 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5389 Constant *C0 = ConstantVector::get(CV0);
5390 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5392 std::vector<Constant*> CV1;
5394 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5396 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5397 Constant *C1 = ConstantVector::get(CV1);
5398 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5400 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5401 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5403 DAG.getIntPtrConstant(1)));
5404 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5405 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5407 DAG.getIntPtrConstant(0)));
5408 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5409 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5410 PseudoSourceValue::getConstantPool(), 0,
5412 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5413 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5414 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5415 PseudoSourceValue::getConstantPool(), 0,
5417 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5419 // Add the halves; easiest way is to swap them into another reg first.
5420 int ShufMask[2] = { 1, -1 };
5421 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5422 DAG.getUNDEF(MVT::v2f64), ShufMask);
5423 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5424 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5425 DAG.getIntPtrConstant(0));
5428 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5429 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5430 DebugLoc dl = Op.getDebugLoc();
5431 // FP constant to bias correct the final result.
5432 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5435 // Load the 32-bit value into an XMM register.
5436 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5437 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5439 DAG.getIntPtrConstant(0)));
5441 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5442 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5443 DAG.getIntPtrConstant(0));
5445 // Or the load with the bias.
5446 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5447 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5448 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5450 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5451 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5452 MVT::v2f64, Bias)));
5453 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5454 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5455 DAG.getIntPtrConstant(0));
5457 // Subtract the bias.
5458 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5460 // Handle final rounding.
5461 EVT DestVT = Op.getValueType();
5463 if (DestVT.bitsLT(MVT::f64)) {
5464 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5465 DAG.getIntPtrConstant(0));
5466 } else if (DestVT.bitsGT(MVT::f64)) {
5467 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5470 // Handle final rounding.
5474 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5475 SDValue N0 = Op.getOperand(0);
5476 DebugLoc dl = Op.getDebugLoc();
5478 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5479 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5480 // the optimization here.
5481 if (DAG.SignBitIsZero(N0))
5482 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5484 EVT SrcVT = N0.getValueType();
5485 if (SrcVT == MVT::i64) {
5486 // We only handle SSE2 f64 target here; caller can expand the rest.
5487 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5490 return LowerUINT_TO_FP_i64(Op, DAG);
5491 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5492 return LowerUINT_TO_FP_i32(Op, DAG);
5495 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5497 // Make a 64-bit buffer, and use it to build an FILD.
5498 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5499 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5500 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5501 getPointerTy(), StackSlot, WordOff);
5502 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5503 StackSlot, NULL, 0);
5504 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5505 OffsetSlot, NULL, 0);
5506 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5509 std::pair<SDValue,SDValue> X86TargetLowering::
5510 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5511 DebugLoc dl = Op.getDebugLoc();
5513 EVT DstTy = Op.getValueType();
5516 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5520 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5521 DstTy.getSimpleVT() >= MVT::i16 &&
5522 "Unknown FP_TO_SINT to lower!");
5524 // These are really Legal.
5525 if (DstTy == MVT::i32 &&
5526 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5527 return std::make_pair(SDValue(), SDValue());
5528 if (Subtarget->is64Bit() &&
5529 DstTy == MVT::i64 &&
5530 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5531 return std::make_pair(SDValue(), SDValue());
5533 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5535 MachineFunction &MF = DAG.getMachineFunction();
5536 unsigned MemSize = DstTy.getSizeInBits()/8;
5537 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5538 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5541 switch (DstTy.getSimpleVT().SimpleTy) {
5542 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5543 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5544 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5545 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5548 SDValue Chain = DAG.getEntryNode();
5549 SDValue Value = Op.getOperand(0);
5550 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5551 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5552 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5553 PseudoSourceValue::getFixedStack(SSFI), 0);
5554 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5556 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5558 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5559 Chain = Value.getValue(1);
5560 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5561 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5564 // Build the FP_TO_INT*_IN_MEM
5565 SDValue Ops[] = { Chain, Value, StackSlot };
5566 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5568 return std::make_pair(FIST, StackSlot);
5571 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5572 if (Op.getValueType().isVector()) {
5573 if (Op.getValueType() == MVT::v2i32 &&
5574 Op.getOperand(0).getValueType() == MVT::v2f64) {
5580 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5581 SDValue FIST = Vals.first, StackSlot = Vals.second;
5582 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5583 if (FIST.getNode() == 0) return Op;
5586 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5587 FIST, StackSlot, NULL, 0);
5590 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5591 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5592 SDValue FIST = Vals.first, StackSlot = Vals.second;
5593 assert(FIST.getNode() && "Unexpected failure");
5596 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5597 FIST, StackSlot, NULL, 0);
5600 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5601 LLVMContext *Context = DAG.getContext();
5602 DebugLoc dl = Op.getDebugLoc();
5603 EVT VT = Op.getValueType();
5606 EltVT = VT.getVectorElementType();
5607 std::vector<Constant*> CV;
5608 if (EltVT == MVT::f64) {
5609 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5613 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5619 Constant *C = ConstantVector::get(CV);
5620 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5621 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5622 PseudoSourceValue::getConstantPool(), 0,
5624 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5627 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5628 LLVMContext *Context = DAG.getContext();
5629 DebugLoc dl = Op.getDebugLoc();
5630 EVT VT = Op.getValueType();
5633 EltVT = VT.getVectorElementType();
5634 std::vector<Constant*> CV;
5635 if (EltVT == MVT::f64) {
5636 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5640 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5646 Constant *C = ConstantVector::get(CV);
5647 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5648 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5649 PseudoSourceValue::getConstantPool(), 0,
5651 if (VT.isVector()) {
5652 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5653 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5654 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5656 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5658 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5662 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5663 LLVMContext *Context = DAG.getContext();
5664 SDValue Op0 = Op.getOperand(0);
5665 SDValue Op1 = Op.getOperand(1);
5666 DebugLoc dl = Op.getDebugLoc();
5667 EVT VT = Op.getValueType();
5668 EVT SrcVT = Op1.getValueType();
5670 // If second operand is smaller, extend it first.
5671 if (SrcVT.bitsLT(VT)) {
5672 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5675 // And if it is bigger, shrink it first.
5676 if (SrcVT.bitsGT(VT)) {
5677 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5681 // At this point the operands and the result should have the same
5682 // type, and that won't be f80 since that is not custom lowered.
5684 // First get the sign bit of second operand.
5685 std::vector<Constant*> CV;
5686 if (SrcVT == MVT::f64) {
5687 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5688 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5690 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5691 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5692 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5693 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5695 Constant *C = ConstantVector::get(CV);
5696 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5697 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5698 PseudoSourceValue::getConstantPool(), 0,
5700 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5702 // Shift sign bit right or left if the two operands have different types.
5703 if (SrcVT.bitsGT(VT)) {
5704 // Op0 is MVT::f32, Op1 is MVT::f64.
5705 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5706 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5707 DAG.getConstant(32, MVT::i32));
5708 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5709 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5710 DAG.getIntPtrConstant(0));
5713 // Clear first operand sign bit.
5715 if (VT == MVT::f64) {
5716 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5717 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5719 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5720 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5721 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5722 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5724 C = ConstantVector::get(CV);
5725 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5726 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5727 PseudoSourceValue::getConstantPool(), 0,
5729 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5731 // Or the value with the sign bit.
5732 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5735 /// Emit nodes that will be selected as "test Op0,Op0", or something
5737 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5738 SelectionDAG &DAG) {
5739 DebugLoc dl = Op.getDebugLoc();
5741 // CF and OF aren't always set the way we want. Determine which
5742 // of these we need.
5743 bool NeedCF = false;
5744 bool NeedOF = false;
5746 case X86::COND_A: case X86::COND_AE:
5747 case X86::COND_B: case X86::COND_BE:
5750 case X86::COND_G: case X86::COND_GE:
5751 case X86::COND_L: case X86::COND_LE:
5752 case X86::COND_O: case X86::COND_NO:
5758 // See if we can use the EFLAGS value from the operand instead of
5759 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5760 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5761 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5762 unsigned Opcode = 0;
5763 unsigned NumOperands = 0;
5764 switch (Op.getNode()->getOpcode()) {
5766 // Due to an isel shortcoming, be conservative if this add is likely to
5767 // be selected as part of a load-modify-store instruction. When the root
5768 // node in a match is a store, isel doesn't know how to remap non-chain
5769 // non-flag uses of other nodes in the match, such as the ADD in this
5770 // case. This leads to the ADD being left around and reselected, with
5771 // the result being two adds in the output.
5772 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5773 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5774 if (UI->getOpcode() == ISD::STORE)
5776 if (ConstantSDNode *C =
5777 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5778 // An add of one will be selected as an INC.
5779 if (C->getAPIntValue() == 1) {
5780 Opcode = X86ISD::INC;
5784 // An add of negative one (subtract of one) will be selected as a DEC.
5785 if (C->getAPIntValue().isAllOnesValue()) {
5786 Opcode = X86ISD::DEC;
5791 // Otherwise use a regular EFLAGS-setting add.
5792 Opcode = X86ISD::ADD;
5796 // If the primary and result isn't used, don't bother using X86ISD::AND,
5797 // because a TEST instruction will be better.
5798 bool NonFlagUse = false;
5799 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5800 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
5802 unsigned UOpNo = UI.getOperandNo();
5803 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
5804 // Look pass truncate.
5805 UOpNo = User->use_begin().getOperandNo();
5806 User = *User->use_begin();
5808 if (User->getOpcode() != ISD::BRCOND &&
5809 User->getOpcode() != ISD::SETCC &&
5810 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
5822 // Due to the ISEL shortcoming noted above, be conservative if this op is
5823 // likely to be selected as part of a load-modify-store instruction.
5824 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5825 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5826 if (UI->getOpcode() == ISD::STORE)
5828 // Otherwise use a regular EFLAGS-setting instruction.
5829 switch (Op.getNode()->getOpcode()) {
5830 case ISD::SUB: Opcode = X86ISD::SUB; break;
5831 case ISD::OR: Opcode = X86ISD::OR; break;
5832 case ISD::XOR: Opcode = X86ISD::XOR; break;
5833 case ISD::AND: Opcode = X86ISD::AND; break;
5834 default: llvm_unreachable("unexpected operator!");
5845 return SDValue(Op.getNode(), 1);
5851 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5852 SmallVector<SDValue, 4> Ops;
5853 for (unsigned i = 0; i != NumOperands; ++i)
5854 Ops.push_back(Op.getOperand(i));
5855 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5856 DAG.ReplaceAllUsesWith(Op, New);
5857 return SDValue(New.getNode(), 1);
5861 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5862 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5863 DAG.getConstant(0, Op.getValueType()));
5866 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5868 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5869 SelectionDAG &DAG) {
5870 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5871 if (C->getAPIntValue() == 0)
5872 return EmitTest(Op0, X86CC, DAG);
5874 DebugLoc dl = Op0.getDebugLoc();
5875 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5878 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
5879 /// if it's possible.
5880 static SDValue LowerToBT(SDValue Op0, ISD::CondCode CC,
5881 DebugLoc dl, SelectionDAG &DAG) {
5883 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5884 if (ConstantSDNode *Op010C =
5885 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5886 if (Op010C->getZExtValue() == 1) {
5887 LHS = Op0.getOperand(0);
5888 RHS = Op0.getOperand(1).getOperand(1);
5890 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5891 if (ConstantSDNode *Op000C =
5892 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5893 if (Op000C->getZExtValue() == 1) {
5894 LHS = Op0.getOperand(1);
5895 RHS = Op0.getOperand(0).getOperand(1);
5897 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5898 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5899 SDValue AndLHS = Op0.getOperand(0);
5900 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5901 LHS = AndLHS.getOperand(0);
5902 RHS = AndLHS.getOperand(1);
5906 if (LHS.getNode()) {
5907 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5908 // instruction. Since the shift amount is in-range-or-undefined, we know
5909 // that doing a bittest on the i16 value is ok. We extend to i32 because
5910 // the encoding for the i16 version is larger than the i32 version.
5911 if (LHS.getValueType() == MVT::i8)
5912 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5914 // If the operand types disagree, extend the shift amount to match. Since
5915 // BT ignores high bits (like shifts) we can use anyextend.
5916 if (LHS.getValueType() != RHS.getValueType())
5917 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5919 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5920 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5921 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5922 DAG.getConstant(Cond, MVT::i8), BT);
5928 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5929 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5930 SDValue Op0 = Op.getOperand(0);
5931 SDValue Op1 = Op.getOperand(1);
5932 DebugLoc dl = Op.getDebugLoc();
5933 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5935 // Optimize to BT if possible.
5936 // Lower (X & (1 << N)) == 0 to BT(X, N).
5937 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5938 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5939 if (Op0.getOpcode() == ISD::AND &&
5941 Op1.getOpcode() == ISD::Constant &&
5942 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5943 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5944 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
5945 if (NewSetCC.getNode())
5949 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5950 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5951 if (X86CC == X86::COND_INVALID)
5954 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5956 // Use sbb x, x to materialize carry bit into a GPR.
5957 if (X86CC == X86::COND_B)
5958 return DAG.getNode(ISD::AND, dl, MVT::i8,
5959 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
5960 DAG.getConstant(X86CC, MVT::i8), Cond),
5961 DAG.getConstant(1, MVT::i8));
5963 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5964 DAG.getConstant(X86CC, MVT::i8), Cond);
5967 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5969 SDValue Op0 = Op.getOperand(0);
5970 SDValue Op1 = Op.getOperand(1);
5971 SDValue CC = Op.getOperand(2);
5972 EVT VT = Op.getValueType();
5973 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5974 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5975 DebugLoc dl = Op.getDebugLoc();
5979 EVT VT0 = Op0.getValueType();
5980 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5981 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5984 switch (SetCCOpcode) {
5987 case ISD::SETEQ: SSECC = 0; break;
5989 case ISD::SETGT: Swap = true; // Fallthrough
5991 case ISD::SETOLT: SSECC = 1; break;
5993 case ISD::SETGE: Swap = true; // Fallthrough
5995 case ISD::SETOLE: SSECC = 2; break;
5996 case ISD::SETUO: SSECC = 3; break;
5998 case ISD::SETNE: SSECC = 4; break;
5999 case ISD::SETULE: Swap = true;
6000 case ISD::SETUGE: SSECC = 5; break;
6001 case ISD::SETULT: Swap = true;
6002 case ISD::SETUGT: SSECC = 6; break;
6003 case ISD::SETO: SSECC = 7; break;
6006 std::swap(Op0, Op1);
6008 // In the two special cases we can't handle, emit two comparisons.
6010 if (SetCCOpcode == ISD::SETUEQ) {
6012 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6013 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6014 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6016 else if (SetCCOpcode == ISD::SETONE) {
6018 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6019 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6020 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6022 llvm_unreachable("Illegal FP comparison");
6024 // Handle all other FP comparisons here.
6025 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6028 // We are handling one of the integer comparisons here. Since SSE only has
6029 // GT and EQ comparisons for integer, swapping operands and multiple
6030 // operations may be required for some comparisons.
6031 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6032 bool Swap = false, Invert = false, FlipSigns = false;
6034 switch (VT.getSimpleVT().SimpleTy) {
6037 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6039 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6041 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6042 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6045 switch (SetCCOpcode) {
6047 case ISD::SETNE: Invert = true;
6048 case ISD::SETEQ: Opc = EQOpc; break;
6049 case ISD::SETLT: Swap = true;
6050 case ISD::SETGT: Opc = GTOpc; break;
6051 case ISD::SETGE: Swap = true;
6052 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6053 case ISD::SETULT: Swap = true;
6054 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6055 case ISD::SETUGE: Swap = true;
6056 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6059 std::swap(Op0, Op1);
6061 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6062 // bits of the inputs before performing those operations.
6064 EVT EltVT = VT.getVectorElementType();
6065 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6067 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6068 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6070 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6071 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6074 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6076 // If the logical-not of the result is required, perform that now.
6078 Result = DAG.getNOT(dl, Result, VT);
6083 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6084 static bool isX86LogicalCmp(SDValue Op) {
6085 unsigned Opc = Op.getNode()->getOpcode();
6086 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6088 if (Op.getResNo() == 1 &&
6089 (Opc == X86ISD::ADD ||
6090 Opc == X86ISD::SUB ||
6091 Opc == X86ISD::SMUL ||
6092 Opc == X86ISD::UMUL ||
6093 Opc == X86ISD::INC ||
6094 Opc == X86ISD::DEC ||
6095 Opc == X86ISD::OR ||
6096 Opc == X86ISD::XOR ||
6097 Opc == X86ISD::AND))
6103 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
6104 bool addTest = true;
6105 SDValue Cond = Op.getOperand(0);
6106 DebugLoc dl = Op.getDebugLoc();
6109 if (Cond.getOpcode() == ISD::SETCC) {
6110 SDValue NewCond = LowerSETCC(Cond, DAG);
6111 if (NewCond.getNode())
6115 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6116 SDValue Op1 = Op.getOperand(1);
6117 SDValue Op2 = Op.getOperand(2);
6118 if (Cond.getOpcode() == X86ISD::SETCC &&
6119 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6120 SDValue Cmp = Cond.getOperand(1);
6121 if (Cmp.getOpcode() == X86ISD::CMP) {
6122 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6123 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6124 ConstantSDNode *RHSC =
6125 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6126 if (N1C && N1C->isAllOnesValue() &&
6127 N2C && N2C->isNullValue() &&
6128 RHSC && RHSC->isNullValue()) {
6129 SDValue CmpOp0 = Cmp.getOperand(0);
6130 Cmp = DAG.getNode(X86ISD::CMP, dl, CmpOp0.getValueType(),
6131 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6132 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6133 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6138 // Look pass (and (setcc_carry (cmp ...)), 1).
6139 if (Cond.getOpcode() == ISD::AND &&
6140 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6141 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6142 if (C && C->getAPIntValue() == 1)
6143 Cond = Cond.getOperand(0);
6146 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6147 // setting operand in place of the X86ISD::SETCC.
6148 if (Cond.getOpcode() == X86ISD::SETCC ||
6149 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6150 CC = Cond.getOperand(0);
6152 SDValue Cmp = Cond.getOperand(1);
6153 unsigned Opc = Cmp.getOpcode();
6154 EVT VT = Op.getValueType();
6156 bool IllegalFPCMov = false;
6157 if (VT.isFloatingPoint() && !VT.isVector() &&
6158 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6159 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6161 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6162 Opc == X86ISD::BT) { // FIXME
6169 // Look pass the truncate.
6170 if (Cond.getOpcode() == ISD::TRUNCATE)
6171 Cond = Cond.getOperand(0);
6173 // We know the result of AND is compared against zero. Try to match
6175 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6176 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6177 if (NewSetCC.getNode()) {
6178 CC = NewSetCC.getOperand(0);
6179 Cond = NewSetCC.getOperand(1);
6186 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6187 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6190 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6191 // condition is true.
6192 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6193 SDValue Ops[] = { Op2, Op1, CC, Cond };
6194 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6197 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6198 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6199 // from the AND / OR.
6200 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6201 Opc = Op.getOpcode();
6202 if (Opc != ISD::OR && Opc != ISD::AND)
6204 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6205 Op.getOperand(0).hasOneUse() &&
6206 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6207 Op.getOperand(1).hasOneUse());
6210 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6211 // 1 and that the SETCC node has a single use.
6212 static bool isXor1OfSetCC(SDValue Op) {
6213 if (Op.getOpcode() != ISD::XOR)
6215 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6216 if (N1C && N1C->getAPIntValue() == 1) {
6217 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6218 Op.getOperand(0).hasOneUse();
6223 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
6224 bool addTest = true;
6225 SDValue Chain = Op.getOperand(0);
6226 SDValue Cond = Op.getOperand(1);
6227 SDValue Dest = Op.getOperand(2);
6228 DebugLoc dl = Op.getDebugLoc();
6231 if (Cond.getOpcode() == ISD::SETCC) {
6232 SDValue NewCond = LowerSETCC(Cond, DAG);
6233 if (NewCond.getNode())
6237 // FIXME: LowerXALUO doesn't handle these!!
6238 else if (Cond.getOpcode() == X86ISD::ADD ||
6239 Cond.getOpcode() == X86ISD::SUB ||
6240 Cond.getOpcode() == X86ISD::SMUL ||
6241 Cond.getOpcode() == X86ISD::UMUL)
6242 Cond = LowerXALUO(Cond, DAG);
6245 // Look pass (and (setcc_carry (cmp ...)), 1).
6246 if (Cond.getOpcode() == ISD::AND &&
6247 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6248 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6249 if (C && C->getAPIntValue() == 1)
6250 Cond = Cond.getOperand(0);
6253 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6254 // setting operand in place of the X86ISD::SETCC.
6255 if (Cond.getOpcode() == X86ISD::SETCC ||
6256 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6257 CC = Cond.getOperand(0);
6259 SDValue Cmp = Cond.getOperand(1);
6260 unsigned Opc = Cmp.getOpcode();
6261 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6262 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6266 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6270 // These can only come from an arithmetic instruction with overflow,
6271 // e.g. SADDO, UADDO.
6272 Cond = Cond.getNode()->getOperand(1);
6279 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6280 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6281 if (CondOpc == ISD::OR) {
6282 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6283 // two branches instead of an explicit OR instruction with a
6285 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6286 isX86LogicalCmp(Cmp)) {
6287 CC = Cond.getOperand(0).getOperand(0);
6288 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6289 Chain, Dest, CC, Cmp);
6290 CC = Cond.getOperand(1).getOperand(0);
6294 } else { // ISD::AND
6295 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6296 // two branches instead of an explicit AND instruction with a
6297 // separate test. However, we only do this if this block doesn't
6298 // have a fall-through edge, because this requires an explicit
6299 // jmp when the condition is false.
6300 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6301 isX86LogicalCmp(Cmp) &&
6302 Op.getNode()->hasOneUse()) {
6303 X86::CondCode CCode =
6304 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6305 CCode = X86::GetOppositeBranchCondition(CCode);
6306 CC = DAG.getConstant(CCode, MVT::i8);
6307 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6308 // Look for an unconditional branch following this conditional branch.
6309 // We need this because we need to reverse the successors in order
6310 // to implement FCMP_OEQ.
6311 if (User.getOpcode() == ISD::BR) {
6312 SDValue FalseBB = User.getOperand(1);
6314 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6315 assert(NewBR == User);
6318 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6319 Chain, Dest, CC, Cmp);
6320 X86::CondCode CCode =
6321 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6322 CCode = X86::GetOppositeBranchCondition(CCode);
6323 CC = DAG.getConstant(CCode, MVT::i8);
6329 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6330 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6331 // It should be transformed during dag combiner except when the condition
6332 // is set by a arithmetics with overflow node.
6333 X86::CondCode CCode =
6334 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6335 CCode = X86::GetOppositeBranchCondition(CCode);
6336 CC = DAG.getConstant(CCode, MVT::i8);
6337 Cond = Cond.getOperand(0).getOperand(1);
6343 // Look pass the truncate.
6344 if (Cond.getOpcode() == ISD::TRUNCATE)
6345 Cond = Cond.getOperand(0);
6347 // We know the result of AND is compared against zero. Try to match
6349 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6350 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6351 if (NewSetCC.getNode()) {
6352 CC = NewSetCC.getOperand(0);
6353 Cond = NewSetCC.getOperand(1);
6360 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6361 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6363 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6364 Chain, Dest, CC, Cond);
6368 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6369 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6370 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6371 // that the guard pages used by the OS virtual memory manager are allocated in
6372 // correct sequence.
6374 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6375 SelectionDAG &DAG) {
6376 assert(Subtarget->isTargetCygMing() &&
6377 "This should be used only on Cygwin/Mingw targets");
6378 DebugLoc dl = Op.getDebugLoc();
6381 SDValue Chain = Op.getOperand(0);
6382 SDValue Size = Op.getOperand(1);
6383 // FIXME: Ensure alignment here
6387 EVT IntPtr = getPointerTy();
6388 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6390 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
6392 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6393 Flag = Chain.getValue(1);
6395 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6396 SDValue Ops[] = { Chain,
6397 DAG.getTargetExternalSymbol("_alloca", IntPtr),
6398 DAG.getRegister(X86::EAX, IntPtr),
6399 DAG.getRegister(X86StackPtr, SPTy),
6401 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
6402 Flag = Chain.getValue(1);
6404 Chain = DAG.getCALLSEQ_END(Chain,
6405 DAG.getIntPtrConstant(0, true),
6406 DAG.getIntPtrConstant(0, true),
6409 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6411 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6412 return DAG.getMergeValues(Ops1, 2, dl);
6416 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6418 SDValue Dst, SDValue Src,
6419 SDValue Size, unsigned Align,
6421 uint64_t DstSVOff) {
6422 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6424 // If not DWORD aligned or size is more than the threshold, call the library.
6425 // The libc version is likely to be faster for these cases. It can use the
6426 // address value and run time information about the CPU.
6427 if ((Align & 3) != 0 ||
6429 ConstantSize->getZExtValue() >
6430 getSubtarget()->getMaxInlineSizeThreshold()) {
6431 SDValue InFlag(0, 0);
6433 // Check to see if there is a specialized entry-point for memory zeroing.
6434 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6436 if (const char *bzeroEntry = V &&
6437 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6438 EVT IntPtr = getPointerTy();
6439 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6440 TargetLowering::ArgListTy Args;
6441 TargetLowering::ArgListEntry Entry;
6443 Entry.Ty = IntPtrTy;
6444 Args.push_back(Entry);
6446 Args.push_back(Entry);
6447 std::pair<SDValue,SDValue> CallResult =
6448 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6449 false, false, false, false,
6450 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6451 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl,
6452 DAG.GetOrdering(Chain.getNode()));
6453 return CallResult.second;
6456 // Otherwise have the target-independent code call memset.
6460 uint64_t SizeVal = ConstantSize->getZExtValue();
6461 SDValue InFlag(0, 0);
6464 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6465 unsigned BytesLeft = 0;
6466 bool TwoRepStos = false;
6469 uint64_t Val = ValC->getZExtValue() & 255;
6471 // If the value is a constant, then we can potentially use larger sets.
6472 switch (Align & 3) {
6473 case 2: // WORD aligned
6476 Val = (Val << 8) | Val;
6478 case 0: // DWORD aligned
6481 Val = (Val << 8) | Val;
6482 Val = (Val << 16) | Val;
6483 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6486 Val = (Val << 32) | Val;
6489 default: // Byte aligned
6492 Count = DAG.getIntPtrConstant(SizeVal);
6496 if (AVT.bitsGT(MVT::i8)) {
6497 unsigned UBytes = AVT.getSizeInBits() / 8;
6498 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6499 BytesLeft = SizeVal % UBytes;
6502 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6504 InFlag = Chain.getValue(1);
6507 Count = DAG.getIntPtrConstant(SizeVal);
6508 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6509 InFlag = Chain.getValue(1);
6512 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6515 InFlag = Chain.getValue(1);
6516 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6519 InFlag = Chain.getValue(1);
6521 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6522 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6523 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6526 InFlag = Chain.getValue(1);
6528 EVT CVT = Count.getValueType();
6529 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6530 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6531 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6534 InFlag = Chain.getValue(1);
6535 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6536 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
6537 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6538 } else if (BytesLeft) {
6539 // Handle the last 1 - 7 bytes.
6540 unsigned Offset = SizeVal - BytesLeft;
6541 EVT AddrVT = Dst.getValueType();
6542 EVT SizeVT = Size.getValueType();
6544 Chain = DAG.getMemset(Chain, dl,
6545 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6546 DAG.getConstant(Offset, AddrVT)),
6548 DAG.getConstant(BytesLeft, SizeVT),
6549 Align, DstSV, DstSVOff + Offset);
6552 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6557 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6558 SDValue Chain, SDValue Dst, SDValue Src,
6559 SDValue Size, unsigned Align,
6561 const Value *DstSV, uint64_t DstSVOff,
6562 const Value *SrcSV, uint64_t SrcSVOff) {
6563 // This requires the copy size to be a constant, preferrably
6564 // within a subtarget-specific limit.
6565 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6568 uint64_t SizeVal = ConstantSize->getZExtValue();
6569 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6572 /// If not DWORD aligned, call the library.
6573 if ((Align & 3) != 0)
6578 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6581 unsigned UBytes = AVT.getSizeInBits() / 8;
6582 unsigned CountVal = SizeVal / UBytes;
6583 SDValue Count = DAG.getIntPtrConstant(CountVal);
6584 unsigned BytesLeft = SizeVal % UBytes;
6586 SDValue InFlag(0, 0);
6587 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6590 InFlag = Chain.getValue(1);
6591 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6594 InFlag = Chain.getValue(1);
6595 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6598 InFlag = Chain.getValue(1);
6600 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6601 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6602 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
6603 array_lengthof(Ops));
6605 SmallVector<SDValue, 4> Results;
6606 Results.push_back(RepMovs);
6608 // Handle the last 1 - 7 bytes.
6609 unsigned Offset = SizeVal - BytesLeft;
6610 EVT DstVT = Dst.getValueType();
6611 EVT SrcVT = Src.getValueType();
6612 EVT SizeVT = Size.getValueType();
6613 Results.push_back(DAG.getMemcpy(Chain, dl,
6614 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6615 DAG.getConstant(Offset, DstVT)),
6616 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6617 DAG.getConstant(Offset, SrcVT)),
6618 DAG.getConstant(BytesLeft, SizeVT),
6619 Align, AlwaysInline,
6620 DstSV, DstSVOff + Offset,
6621 SrcSV, SrcSVOff + Offset));
6624 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6625 &Results[0], Results.size());
6628 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6629 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6630 DebugLoc dl = Op.getDebugLoc();
6632 if (!Subtarget->is64Bit()) {
6633 // vastart just stores the address of the VarArgsFrameIndex slot into the
6634 // memory location argument.
6635 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6636 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6640 // gp_offset (0 - 6 * 8)
6641 // fp_offset (48 - 48 + 8 * 16)
6642 // overflow_arg_area (point to parameters coming in memory).
6644 SmallVector<SDValue, 8> MemOps;
6645 SDValue FIN = Op.getOperand(1);
6647 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6648 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6650 MemOps.push_back(Store);
6653 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6654 FIN, DAG.getIntPtrConstant(4));
6655 Store = DAG.getStore(Op.getOperand(0), dl,
6656 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6658 MemOps.push_back(Store);
6660 // Store ptr to overflow_arg_area
6661 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6662 FIN, DAG.getIntPtrConstant(4));
6663 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6664 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6665 MemOps.push_back(Store);
6667 // Store ptr to reg_save_area.
6668 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6669 FIN, DAG.getIntPtrConstant(8));
6670 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6671 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6672 MemOps.push_back(Store);
6673 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6674 &MemOps[0], MemOps.size());
6677 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6678 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6679 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6680 SDValue Chain = Op.getOperand(0);
6681 SDValue SrcPtr = Op.getOperand(1);
6682 SDValue SrcSV = Op.getOperand(2);
6684 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6688 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6689 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6690 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6691 SDValue Chain = Op.getOperand(0);
6692 SDValue DstPtr = Op.getOperand(1);
6693 SDValue SrcPtr = Op.getOperand(2);
6694 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6695 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6696 DebugLoc dl = Op.getDebugLoc();
6698 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6699 DAG.getIntPtrConstant(24), 8, false,
6700 DstSV, 0, SrcSV, 0);
6704 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6705 DebugLoc dl = Op.getDebugLoc();
6706 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6708 default: return SDValue(); // Don't custom lower most intrinsics.
6709 // Comparison intrinsics.
6710 case Intrinsic::x86_sse_comieq_ss:
6711 case Intrinsic::x86_sse_comilt_ss:
6712 case Intrinsic::x86_sse_comile_ss:
6713 case Intrinsic::x86_sse_comigt_ss:
6714 case Intrinsic::x86_sse_comige_ss:
6715 case Intrinsic::x86_sse_comineq_ss:
6716 case Intrinsic::x86_sse_ucomieq_ss:
6717 case Intrinsic::x86_sse_ucomilt_ss:
6718 case Intrinsic::x86_sse_ucomile_ss:
6719 case Intrinsic::x86_sse_ucomigt_ss:
6720 case Intrinsic::x86_sse_ucomige_ss:
6721 case Intrinsic::x86_sse_ucomineq_ss:
6722 case Intrinsic::x86_sse2_comieq_sd:
6723 case Intrinsic::x86_sse2_comilt_sd:
6724 case Intrinsic::x86_sse2_comile_sd:
6725 case Intrinsic::x86_sse2_comigt_sd:
6726 case Intrinsic::x86_sse2_comige_sd:
6727 case Intrinsic::x86_sse2_comineq_sd:
6728 case Intrinsic::x86_sse2_ucomieq_sd:
6729 case Intrinsic::x86_sse2_ucomilt_sd:
6730 case Intrinsic::x86_sse2_ucomile_sd:
6731 case Intrinsic::x86_sse2_ucomigt_sd:
6732 case Intrinsic::x86_sse2_ucomige_sd:
6733 case Intrinsic::x86_sse2_ucomineq_sd: {
6735 ISD::CondCode CC = ISD::SETCC_INVALID;
6738 case Intrinsic::x86_sse_comieq_ss:
6739 case Intrinsic::x86_sse2_comieq_sd:
6743 case Intrinsic::x86_sse_comilt_ss:
6744 case Intrinsic::x86_sse2_comilt_sd:
6748 case Intrinsic::x86_sse_comile_ss:
6749 case Intrinsic::x86_sse2_comile_sd:
6753 case Intrinsic::x86_sse_comigt_ss:
6754 case Intrinsic::x86_sse2_comigt_sd:
6758 case Intrinsic::x86_sse_comige_ss:
6759 case Intrinsic::x86_sse2_comige_sd:
6763 case Intrinsic::x86_sse_comineq_ss:
6764 case Intrinsic::x86_sse2_comineq_sd:
6768 case Intrinsic::x86_sse_ucomieq_ss:
6769 case Intrinsic::x86_sse2_ucomieq_sd:
6770 Opc = X86ISD::UCOMI;
6773 case Intrinsic::x86_sse_ucomilt_ss:
6774 case Intrinsic::x86_sse2_ucomilt_sd:
6775 Opc = X86ISD::UCOMI;
6778 case Intrinsic::x86_sse_ucomile_ss:
6779 case Intrinsic::x86_sse2_ucomile_sd:
6780 Opc = X86ISD::UCOMI;
6783 case Intrinsic::x86_sse_ucomigt_ss:
6784 case Intrinsic::x86_sse2_ucomigt_sd:
6785 Opc = X86ISD::UCOMI;
6788 case Intrinsic::x86_sse_ucomige_ss:
6789 case Intrinsic::x86_sse2_ucomige_sd:
6790 Opc = X86ISD::UCOMI;
6793 case Intrinsic::x86_sse_ucomineq_ss:
6794 case Intrinsic::x86_sse2_ucomineq_sd:
6795 Opc = X86ISD::UCOMI;
6800 SDValue LHS = Op.getOperand(1);
6801 SDValue RHS = Op.getOperand(2);
6802 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6803 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6804 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6805 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6806 DAG.getConstant(X86CC, MVT::i8), Cond);
6807 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6809 // ptest intrinsics. The intrinsic these come from are designed to return
6810 // an integer value, not just an instruction so lower it to the ptest
6811 // pattern and a setcc for the result.
6812 case Intrinsic::x86_sse41_ptestz:
6813 case Intrinsic::x86_sse41_ptestc:
6814 case Intrinsic::x86_sse41_ptestnzc:{
6817 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6818 case Intrinsic::x86_sse41_ptestz:
6820 X86CC = X86::COND_E;
6822 case Intrinsic::x86_sse41_ptestc:
6824 X86CC = X86::COND_B;
6826 case Intrinsic::x86_sse41_ptestnzc:
6828 X86CC = X86::COND_A;
6832 SDValue LHS = Op.getOperand(1);
6833 SDValue RHS = Op.getOperand(2);
6834 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6835 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6836 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6837 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6840 // Fix vector shift instructions where the last operand is a non-immediate
6842 case Intrinsic::x86_sse2_pslli_w:
6843 case Intrinsic::x86_sse2_pslli_d:
6844 case Intrinsic::x86_sse2_pslli_q:
6845 case Intrinsic::x86_sse2_psrli_w:
6846 case Intrinsic::x86_sse2_psrli_d:
6847 case Intrinsic::x86_sse2_psrli_q:
6848 case Intrinsic::x86_sse2_psrai_w:
6849 case Intrinsic::x86_sse2_psrai_d:
6850 case Intrinsic::x86_mmx_pslli_w:
6851 case Intrinsic::x86_mmx_pslli_d:
6852 case Intrinsic::x86_mmx_pslli_q:
6853 case Intrinsic::x86_mmx_psrli_w:
6854 case Intrinsic::x86_mmx_psrli_d:
6855 case Intrinsic::x86_mmx_psrli_q:
6856 case Intrinsic::x86_mmx_psrai_w:
6857 case Intrinsic::x86_mmx_psrai_d: {
6858 SDValue ShAmt = Op.getOperand(2);
6859 if (isa<ConstantSDNode>(ShAmt))
6862 unsigned NewIntNo = 0;
6863 EVT ShAmtVT = MVT::v4i32;
6865 case Intrinsic::x86_sse2_pslli_w:
6866 NewIntNo = Intrinsic::x86_sse2_psll_w;
6868 case Intrinsic::x86_sse2_pslli_d:
6869 NewIntNo = Intrinsic::x86_sse2_psll_d;
6871 case Intrinsic::x86_sse2_pslli_q:
6872 NewIntNo = Intrinsic::x86_sse2_psll_q;
6874 case Intrinsic::x86_sse2_psrli_w:
6875 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6877 case Intrinsic::x86_sse2_psrli_d:
6878 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6880 case Intrinsic::x86_sse2_psrli_q:
6881 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6883 case Intrinsic::x86_sse2_psrai_w:
6884 NewIntNo = Intrinsic::x86_sse2_psra_w;
6886 case Intrinsic::x86_sse2_psrai_d:
6887 NewIntNo = Intrinsic::x86_sse2_psra_d;
6890 ShAmtVT = MVT::v2i32;
6892 case Intrinsic::x86_mmx_pslli_w:
6893 NewIntNo = Intrinsic::x86_mmx_psll_w;
6895 case Intrinsic::x86_mmx_pslli_d:
6896 NewIntNo = Intrinsic::x86_mmx_psll_d;
6898 case Intrinsic::x86_mmx_pslli_q:
6899 NewIntNo = Intrinsic::x86_mmx_psll_q;
6901 case Intrinsic::x86_mmx_psrli_w:
6902 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6904 case Intrinsic::x86_mmx_psrli_d:
6905 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6907 case Intrinsic::x86_mmx_psrli_q:
6908 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6910 case Intrinsic::x86_mmx_psrai_w:
6911 NewIntNo = Intrinsic::x86_mmx_psra_w;
6913 case Intrinsic::x86_mmx_psrai_d:
6914 NewIntNo = Intrinsic::x86_mmx_psra_d;
6916 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6922 // The vector shift intrinsics with scalars uses 32b shift amounts but
6923 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
6927 ShOps[1] = DAG.getConstant(0, MVT::i32);
6928 if (ShAmtVT == MVT::v4i32) {
6929 ShOps[2] = DAG.getUNDEF(MVT::i32);
6930 ShOps[3] = DAG.getUNDEF(MVT::i32);
6931 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
6933 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
6936 EVT VT = Op.getValueType();
6937 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
6938 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6939 DAG.getConstant(NewIntNo, MVT::i32),
6940 Op.getOperand(1), ShAmt);
6945 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6946 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6947 DebugLoc dl = Op.getDebugLoc();
6950 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6952 DAG.getConstant(TD->getPointerSize(),
6953 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6954 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6955 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6960 // Just load the return address.
6961 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6962 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6963 RetAddrFI, NULL, 0);
6966 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6967 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6968 MFI->setFrameAddressIsTaken(true);
6969 EVT VT = Op.getValueType();
6970 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6971 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6972 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6973 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6975 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6979 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6980 SelectionDAG &DAG) {
6981 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6984 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6986 MachineFunction &MF = DAG.getMachineFunction();
6987 SDValue Chain = Op.getOperand(0);
6988 SDValue Offset = Op.getOperand(1);
6989 SDValue Handler = Op.getOperand(2);
6990 DebugLoc dl = Op.getDebugLoc();
6992 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6994 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6996 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6997 DAG.getIntPtrConstant(-TD->getPointerSize()));
6998 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6999 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
7000 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7001 MF.getRegInfo().addLiveOut(StoreAddrReg);
7003 return DAG.getNode(X86ISD::EH_RETURN, dl,
7005 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7008 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7009 SelectionDAG &DAG) {
7010 SDValue Root = Op.getOperand(0);
7011 SDValue Trmp = Op.getOperand(1); // trampoline
7012 SDValue FPtr = Op.getOperand(2); // nested function
7013 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7014 DebugLoc dl = Op.getDebugLoc();
7016 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7018 if (Subtarget->is64Bit()) {
7019 SDValue OutChains[6];
7021 // Large code-model.
7022 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7023 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7025 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7026 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7028 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7030 // Load the pointer to the nested function into R11.
7031 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7032 SDValue Addr = Trmp;
7033 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7036 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7037 DAG.getConstant(2, MVT::i64));
7038 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
7040 // Load the 'nest' parameter value into R10.
7041 // R10 is specified in X86CallingConv.td
7042 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7043 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7044 DAG.getConstant(10, MVT::i64));
7045 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7046 Addr, TrmpAddr, 10);
7048 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7049 DAG.getConstant(12, MVT::i64));
7050 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
7052 // Jump to the nested function.
7053 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7054 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7055 DAG.getConstant(20, MVT::i64));
7056 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7057 Addr, TrmpAddr, 20);
7059 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7060 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7061 DAG.getConstant(22, MVT::i64));
7062 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7066 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7067 return DAG.getMergeValues(Ops, 2, dl);
7069 const Function *Func =
7070 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7071 CallingConv::ID CC = Func->getCallingConv();
7076 llvm_unreachable("Unsupported calling convention");
7077 case CallingConv::C:
7078 case CallingConv::X86_StdCall: {
7079 // Pass 'nest' parameter in ECX.
7080 // Must be kept in sync with X86CallingConv.td
7083 // Check that ECX wasn't needed by an 'inreg' parameter.
7084 const FunctionType *FTy = Func->getFunctionType();
7085 const AttrListPtr &Attrs = Func->getAttributes();
7087 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7088 unsigned InRegCount = 0;
7091 for (FunctionType::param_iterator I = FTy->param_begin(),
7092 E = FTy->param_end(); I != E; ++I, ++Idx)
7093 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7094 // FIXME: should only count parameters that are lowered to integers.
7095 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7097 if (InRegCount > 2) {
7098 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
7103 case CallingConv::X86_FastCall:
7104 case CallingConv::Fast:
7105 // Pass 'nest' parameter in EAX.
7106 // Must be kept in sync with X86CallingConv.td
7111 SDValue OutChains[4];
7114 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7115 DAG.getConstant(10, MVT::i32));
7116 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7118 // This is storing the opcode for MOV32ri.
7119 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7120 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7121 OutChains[0] = DAG.getStore(Root, dl,
7122 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7125 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7126 DAG.getConstant(1, MVT::i32));
7127 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
7129 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7130 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7131 DAG.getConstant(5, MVT::i32));
7132 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7133 TrmpAddr, 5, false, 1);
7135 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7136 DAG.getConstant(6, MVT::i32));
7137 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
7140 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7141 return DAG.getMergeValues(Ops, 2, dl);
7145 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
7147 The rounding mode is in bits 11:10 of FPSR, and has the following
7154 FLT_ROUNDS, on the other hand, expects the following:
7161 To perform the conversion, we do:
7162 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7165 MachineFunction &MF = DAG.getMachineFunction();
7166 const TargetMachine &TM = MF.getTarget();
7167 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7168 unsigned StackAlignment = TFI.getStackAlignment();
7169 EVT VT = Op.getValueType();
7170 DebugLoc dl = Op.getDebugLoc();
7172 // Save FP Control Word to stack slot
7173 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7174 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7176 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7177 DAG.getEntryNode(), StackSlot);
7179 // Load FP Control Word from stack slot
7180 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
7182 // Transform as necessary
7184 DAG.getNode(ISD::SRL, dl, MVT::i16,
7185 DAG.getNode(ISD::AND, dl, MVT::i16,
7186 CWD, DAG.getConstant(0x800, MVT::i16)),
7187 DAG.getConstant(11, MVT::i8));
7189 DAG.getNode(ISD::SRL, dl, MVT::i16,
7190 DAG.getNode(ISD::AND, dl, MVT::i16,
7191 CWD, DAG.getConstant(0x400, MVT::i16)),
7192 DAG.getConstant(9, MVT::i8));
7195 DAG.getNode(ISD::AND, dl, MVT::i16,
7196 DAG.getNode(ISD::ADD, dl, MVT::i16,
7197 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7198 DAG.getConstant(1, MVT::i16)),
7199 DAG.getConstant(3, MVT::i16));
7202 return DAG.getNode((VT.getSizeInBits() < 16 ?
7203 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7206 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
7207 EVT VT = Op.getValueType();
7209 unsigned NumBits = VT.getSizeInBits();
7210 DebugLoc dl = Op.getDebugLoc();
7212 Op = Op.getOperand(0);
7213 if (VT == MVT::i8) {
7214 // Zero extend to i32 since there is not an i8 bsr.
7216 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7219 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7220 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7221 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7223 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7226 DAG.getConstant(NumBits+NumBits-1, OpVT),
7227 DAG.getConstant(X86::COND_E, MVT::i8),
7230 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7232 // Finally xor with NumBits-1.
7233 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7236 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7240 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
7241 EVT VT = Op.getValueType();
7243 unsigned NumBits = VT.getSizeInBits();
7244 DebugLoc dl = Op.getDebugLoc();
7246 Op = Op.getOperand(0);
7247 if (VT == MVT::i8) {
7249 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7252 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7253 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7254 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7256 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7259 DAG.getConstant(NumBits, OpVT),
7260 DAG.getConstant(X86::COND_E, MVT::i8),
7263 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7266 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7270 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
7271 EVT VT = Op.getValueType();
7272 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7273 DebugLoc dl = Op.getDebugLoc();
7275 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7276 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7277 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7278 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7279 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7281 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7282 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7283 // return AloBlo + AloBhi + AhiBlo;
7285 SDValue A = Op.getOperand(0);
7286 SDValue B = Op.getOperand(1);
7288 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7289 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7290 A, DAG.getConstant(32, MVT::i32));
7291 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7292 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7293 B, DAG.getConstant(32, MVT::i32));
7294 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7295 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7297 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7298 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7300 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7301 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7303 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7304 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7305 AloBhi, DAG.getConstant(32, MVT::i32));
7306 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7307 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7308 AhiBlo, DAG.getConstant(32, MVT::i32));
7309 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7310 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7315 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
7316 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7317 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7318 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7319 // has only one use.
7320 SDNode *N = Op.getNode();
7321 SDValue LHS = N->getOperand(0);
7322 SDValue RHS = N->getOperand(1);
7323 unsigned BaseOp = 0;
7325 DebugLoc dl = Op.getDebugLoc();
7327 switch (Op.getOpcode()) {
7328 default: llvm_unreachable("Unknown ovf instruction!");
7330 // A subtract of one will be selected as a INC. Note that INC doesn't
7331 // set CF, so we can't do this for UADDO.
7332 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7333 if (C->getAPIntValue() == 1) {
7334 BaseOp = X86ISD::INC;
7338 BaseOp = X86ISD::ADD;
7342 BaseOp = X86ISD::ADD;
7346 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7347 // set CF, so we can't do this for USUBO.
7348 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7349 if (C->getAPIntValue() == 1) {
7350 BaseOp = X86ISD::DEC;
7354 BaseOp = X86ISD::SUB;
7358 BaseOp = X86ISD::SUB;
7362 BaseOp = X86ISD::SMUL;
7366 BaseOp = X86ISD::UMUL;
7371 // Also sets EFLAGS.
7372 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7373 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7376 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7377 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7379 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7383 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
7384 EVT T = Op.getValueType();
7385 DebugLoc dl = Op.getDebugLoc();
7388 switch(T.getSimpleVT().SimpleTy) {
7390 assert(false && "Invalid value type!");
7391 case MVT::i8: Reg = X86::AL; size = 1; break;
7392 case MVT::i16: Reg = X86::AX; size = 2; break;
7393 case MVT::i32: Reg = X86::EAX; size = 4; break;
7395 assert(Subtarget->is64Bit() && "Node not type legal!");
7396 Reg = X86::RAX; size = 8;
7399 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7400 Op.getOperand(2), SDValue());
7401 SDValue Ops[] = { cpIn.getValue(0),
7404 DAG.getTargetConstant(size, MVT::i8),
7406 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7407 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7409 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7413 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7414 SelectionDAG &DAG) {
7415 assert(Subtarget->is64Bit() && "Result not type legalized?");
7416 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7417 SDValue TheChain = Op.getOperand(0);
7418 DebugLoc dl = Op.getDebugLoc();
7419 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7420 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7421 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7423 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7424 DAG.getConstant(32, MVT::i8));
7426 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7429 return DAG.getMergeValues(Ops, 2, dl);
7432 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7433 SDNode *Node = Op.getNode();
7434 DebugLoc dl = Node->getDebugLoc();
7435 EVT T = Node->getValueType(0);
7436 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7437 DAG.getConstant(0, T), Node->getOperand(2));
7438 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7439 cast<AtomicSDNode>(Node)->getMemoryVT(),
7440 Node->getOperand(0),
7441 Node->getOperand(1), negOp,
7442 cast<AtomicSDNode>(Node)->getSrcValue(),
7443 cast<AtomicSDNode>(Node)->getAlignment());
7446 /// LowerOperation - Provide custom lowering hooks for some operations.
7448 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7449 switch (Op.getOpcode()) {
7450 default: llvm_unreachable("Should not custom lower this!");
7451 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7452 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7453 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7454 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7455 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7456 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7457 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7458 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7459 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7460 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7461 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7462 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7463 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7464 case ISD::SHL_PARTS:
7465 case ISD::SRA_PARTS:
7466 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7467 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7468 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7469 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7470 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7471 case ISD::FABS: return LowerFABS(Op, DAG);
7472 case ISD::FNEG: return LowerFNEG(Op, DAG);
7473 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7474 case ISD::SETCC: return LowerSETCC(Op, DAG);
7475 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7476 case ISD::SELECT: return LowerSELECT(Op, DAG);
7477 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7478 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7479 case ISD::VASTART: return LowerVASTART(Op, DAG);
7480 case ISD::VAARG: return LowerVAARG(Op, DAG);
7481 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7482 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7483 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7484 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7485 case ISD::FRAME_TO_ARGS_OFFSET:
7486 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7487 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7488 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7489 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7490 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7491 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7492 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7493 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7499 case ISD::UMULO: return LowerXALUO(Op, DAG);
7500 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7504 void X86TargetLowering::
7505 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7506 SelectionDAG &DAG, unsigned NewOp) {
7507 EVT T = Node->getValueType(0);
7508 DebugLoc dl = Node->getDebugLoc();
7509 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7511 SDValue Chain = Node->getOperand(0);
7512 SDValue In1 = Node->getOperand(1);
7513 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7514 Node->getOperand(2), DAG.getIntPtrConstant(0));
7515 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7516 Node->getOperand(2), DAG.getIntPtrConstant(1));
7517 SDValue Ops[] = { Chain, In1, In2L, In2H };
7518 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7520 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7521 cast<MemSDNode>(Node)->getMemOperand());
7522 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7523 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7524 Results.push_back(Result.getValue(2));
7527 /// ReplaceNodeResults - Replace a node with an illegal result type
7528 /// with a new node built out of custom code.
7529 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7530 SmallVectorImpl<SDValue>&Results,
7531 SelectionDAG &DAG) {
7532 DebugLoc dl = N->getDebugLoc();
7533 switch (N->getOpcode()) {
7535 assert(false && "Do not know how to custom type legalize this operation!");
7537 case ISD::FP_TO_SINT: {
7538 std::pair<SDValue,SDValue> Vals =
7539 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7540 SDValue FIST = Vals.first, StackSlot = Vals.second;
7541 if (FIST.getNode() != 0) {
7542 EVT VT = N->getValueType(0);
7543 // Return a load from the stack slot.
7544 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
7548 case ISD::READCYCLECOUNTER: {
7549 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7550 SDValue TheChain = N->getOperand(0);
7551 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7552 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7554 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7556 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7557 SDValue Ops[] = { eax, edx };
7558 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7559 Results.push_back(edx.getValue(1));
7562 case ISD::ATOMIC_CMP_SWAP: {
7563 EVT T = N->getValueType(0);
7564 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7565 SDValue cpInL, cpInH;
7566 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7567 DAG.getConstant(0, MVT::i32));
7568 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7569 DAG.getConstant(1, MVT::i32));
7570 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7571 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7573 SDValue swapInL, swapInH;
7574 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7575 DAG.getConstant(0, MVT::i32));
7576 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7577 DAG.getConstant(1, MVT::i32));
7578 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7580 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7581 swapInL.getValue(1));
7582 SDValue Ops[] = { swapInH.getValue(0),
7584 swapInH.getValue(1) };
7585 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7586 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7587 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7588 MVT::i32, Result.getValue(1));
7589 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7590 MVT::i32, cpOutL.getValue(2));
7591 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7592 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7593 Results.push_back(cpOutH.getValue(1));
7596 case ISD::ATOMIC_LOAD_ADD:
7597 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7599 case ISD::ATOMIC_LOAD_AND:
7600 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7602 case ISD::ATOMIC_LOAD_NAND:
7603 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7605 case ISD::ATOMIC_LOAD_OR:
7606 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7608 case ISD::ATOMIC_LOAD_SUB:
7609 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7611 case ISD::ATOMIC_LOAD_XOR:
7612 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7614 case ISD::ATOMIC_SWAP:
7615 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7620 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7622 default: return NULL;
7623 case X86ISD::BSF: return "X86ISD::BSF";
7624 case X86ISD::BSR: return "X86ISD::BSR";
7625 case X86ISD::SHLD: return "X86ISD::SHLD";
7626 case X86ISD::SHRD: return "X86ISD::SHRD";
7627 case X86ISD::FAND: return "X86ISD::FAND";
7628 case X86ISD::FOR: return "X86ISD::FOR";
7629 case X86ISD::FXOR: return "X86ISD::FXOR";
7630 case X86ISD::FSRL: return "X86ISD::FSRL";
7631 case X86ISD::FILD: return "X86ISD::FILD";
7632 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7633 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7634 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7635 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7636 case X86ISD::FLD: return "X86ISD::FLD";
7637 case X86ISD::FST: return "X86ISD::FST";
7638 case X86ISD::CALL: return "X86ISD::CALL";
7639 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7640 case X86ISD::BT: return "X86ISD::BT";
7641 case X86ISD::CMP: return "X86ISD::CMP";
7642 case X86ISD::COMI: return "X86ISD::COMI";
7643 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7644 case X86ISD::SETCC: return "X86ISD::SETCC";
7645 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7646 case X86ISD::CMOV: return "X86ISD::CMOV";
7647 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7648 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7649 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7650 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7651 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7652 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7653 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7654 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7655 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7656 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7657 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7658 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7659 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7660 case X86ISD::FMAX: return "X86ISD::FMAX";
7661 case X86ISD::FMIN: return "X86ISD::FMIN";
7662 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7663 case X86ISD::FRCP: return "X86ISD::FRCP";
7664 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7665 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7666 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7667 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7668 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7669 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7670 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7671 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7672 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7673 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7674 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7675 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7676 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7677 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7678 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7679 case X86ISD::VSHL: return "X86ISD::VSHL";
7680 case X86ISD::VSRL: return "X86ISD::VSRL";
7681 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7682 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7683 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7684 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7685 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7686 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7687 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7688 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7689 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7690 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7691 case X86ISD::ADD: return "X86ISD::ADD";
7692 case X86ISD::SUB: return "X86ISD::SUB";
7693 case X86ISD::SMUL: return "X86ISD::SMUL";
7694 case X86ISD::UMUL: return "X86ISD::UMUL";
7695 case X86ISD::INC: return "X86ISD::INC";
7696 case X86ISD::DEC: return "X86ISD::DEC";
7697 case X86ISD::OR: return "X86ISD::OR";
7698 case X86ISD::XOR: return "X86ISD::XOR";
7699 case X86ISD::AND: return "X86ISD::AND";
7700 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7701 case X86ISD::PTEST: return "X86ISD::PTEST";
7702 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7706 // isLegalAddressingMode - Return true if the addressing mode represented
7707 // by AM is legal for this target, for a load/store of the specified type.
7708 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7709 const Type *Ty) const {
7710 // X86 supports extremely general addressing modes.
7711 CodeModel::Model M = getTargetMachine().getCodeModel();
7713 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7714 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7719 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7721 // If a reference to this global requires an extra load, we can't fold it.
7722 if (isGlobalStubReference(GVFlags))
7725 // If BaseGV requires a register for the PIC base, we cannot also have a
7726 // BaseReg specified.
7727 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7730 // If lower 4G is not available, then we must use rip-relative addressing.
7731 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7741 // These scales always work.
7746 // These scales are formed with basereg+scalereg. Only accept if there is
7751 default: // Other stuff never works.
7759 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7760 if (!Ty1->isInteger() || !Ty2->isInteger())
7762 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7763 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7764 if (NumBits1 <= NumBits2)
7766 return Subtarget->is64Bit() || NumBits1 < 64;
7769 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7770 if (!VT1.isInteger() || !VT2.isInteger())
7772 unsigned NumBits1 = VT1.getSizeInBits();
7773 unsigned NumBits2 = VT2.getSizeInBits();
7774 if (NumBits1 <= NumBits2)
7776 return Subtarget->is64Bit() || NumBits1 < 64;
7779 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7780 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7781 return Ty1->isInteger(32) && Ty2->isInteger(64) && Subtarget->is64Bit();
7784 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7785 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7786 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7789 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7790 // i16 instructions are longer (0x66 prefix) and potentially slower.
7791 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7794 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7795 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7796 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7797 /// are assumed to be legal.
7799 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7801 // Only do shuffles on 128-bit vector types for now.
7802 if (VT.getSizeInBits() == 64)
7805 // FIXME: pshufb, blends, shifts.
7806 return (VT.getVectorNumElements() == 2 ||
7807 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7808 isMOVLMask(M, VT) ||
7809 isSHUFPMask(M, VT) ||
7810 isPSHUFDMask(M, VT) ||
7811 isPSHUFHWMask(M, VT) ||
7812 isPSHUFLWMask(M, VT) ||
7813 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7814 isUNPCKLMask(M, VT) ||
7815 isUNPCKHMask(M, VT) ||
7816 isUNPCKL_v_undef_Mask(M, VT) ||
7817 isUNPCKH_v_undef_Mask(M, VT));
7821 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7823 unsigned NumElts = VT.getVectorNumElements();
7824 // FIXME: This collection of masks seems suspect.
7827 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7828 return (isMOVLMask(Mask, VT) ||
7829 isCommutedMOVLMask(Mask, VT, true) ||
7830 isSHUFPMask(Mask, VT) ||
7831 isCommutedSHUFPMask(Mask, VT));
7836 //===----------------------------------------------------------------------===//
7837 // X86 Scheduler Hooks
7838 //===----------------------------------------------------------------------===//
7840 // private utility function
7842 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7843 MachineBasicBlock *MBB,
7851 TargetRegisterClass *RC,
7852 bool invSrc) const {
7853 // For the atomic bitwise operator, we generate
7856 // ld t1 = [bitinstr.addr]
7857 // op t2 = t1, [bitinstr.val]
7859 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7861 // fallthrough -->nextMBB
7862 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7863 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7864 MachineFunction::iterator MBBIter = MBB;
7867 /// First build the CFG
7868 MachineFunction *F = MBB->getParent();
7869 MachineBasicBlock *thisMBB = MBB;
7870 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7871 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7872 F->insert(MBBIter, newMBB);
7873 F->insert(MBBIter, nextMBB);
7875 // Move all successors to thisMBB to nextMBB
7876 nextMBB->transferSuccessors(thisMBB);
7878 // Update thisMBB to fall through to newMBB
7879 thisMBB->addSuccessor(newMBB);
7881 // newMBB jumps to itself and fall through to nextMBB
7882 newMBB->addSuccessor(nextMBB);
7883 newMBB->addSuccessor(newMBB);
7885 // Insert instructions into newMBB based on incoming instruction
7886 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7887 "unexpected number of operands");
7888 DebugLoc dl = bInstr->getDebugLoc();
7889 MachineOperand& destOper = bInstr->getOperand(0);
7890 MachineOperand* argOpers[2 + X86AddrNumOperands];
7891 int numArgs = bInstr->getNumOperands() - 1;
7892 for (int i=0; i < numArgs; ++i)
7893 argOpers[i] = &bInstr->getOperand(i+1);
7895 // x86 address has 4 operands: base, index, scale, and displacement
7896 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7897 int valArgIndx = lastAddrIndx + 1;
7899 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7900 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7901 for (int i=0; i <= lastAddrIndx; ++i)
7902 (*MIB).addOperand(*argOpers[i]);
7904 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7906 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7911 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7912 assert((argOpers[valArgIndx]->isReg() ||
7913 argOpers[valArgIndx]->isImm()) &&
7915 if (argOpers[valArgIndx]->isReg())
7916 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7918 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7920 (*MIB).addOperand(*argOpers[valArgIndx]);
7922 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7925 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7926 for (int i=0; i <= lastAddrIndx; ++i)
7927 (*MIB).addOperand(*argOpers[i]);
7929 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7930 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7931 bInstr->memoperands_end());
7933 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7937 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
7939 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7943 // private utility function: 64 bit atomics on 32 bit host.
7945 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7946 MachineBasicBlock *MBB,
7951 bool invSrc) const {
7952 // For the atomic bitwise operator, we generate
7953 // thisMBB (instructions are in pairs, except cmpxchg8b)
7954 // ld t1,t2 = [bitinstr.addr]
7956 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7957 // op t5, t6 <- out1, out2, [bitinstr.val]
7958 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7959 // mov ECX, EBX <- t5, t6
7960 // mov EAX, EDX <- t1, t2
7961 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7962 // mov t3, t4 <- EAX, EDX
7964 // result in out1, out2
7965 // fallthrough -->nextMBB
7967 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7968 const unsigned LoadOpc = X86::MOV32rm;
7969 const unsigned copyOpc = X86::MOV32rr;
7970 const unsigned NotOpc = X86::NOT32r;
7971 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7972 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7973 MachineFunction::iterator MBBIter = MBB;
7976 /// First build the CFG
7977 MachineFunction *F = MBB->getParent();
7978 MachineBasicBlock *thisMBB = MBB;
7979 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7980 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7981 F->insert(MBBIter, newMBB);
7982 F->insert(MBBIter, nextMBB);
7984 // Move all successors to thisMBB to nextMBB
7985 nextMBB->transferSuccessors(thisMBB);
7987 // Update thisMBB to fall through to newMBB
7988 thisMBB->addSuccessor(newMBB);
7990 // newMBB jumps to itself and fall through to nextMBB
7991 newMBB->addSuccessor(nextMBB);
7992 newMBB->addSuccessor(newMBB);
7994 DebugLoc dl = bInstr->getDebugLoc();
7995 // Insert instructions into newMBB based on incoming instruction
7996 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7997 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7998 "unexpected number of operands");
7999 MachineOperand& dest1Oper = bInstr->getOperand(0);
8000 MachineOperand& dest2Oper = bInstr->getOperand(1);
8001 MachineOperand* argOpers[2 + X86AddrNumOperands];
8002 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
8003 argOpers[i] = &bInstr->getOperand(i+2);
8005 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8006 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8008 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8009 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8010 for (int i=0; i <= lastAddrIndx; ++i)
8011 (*MIB).addOperand(*argOpers[i]);
8012 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8013 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8014 // add 4 to displacement.
8015 for (int i=0; i <= lastAddrIndx-2; ++i)
8016 (*MIB).addOperand(*argOpers[i]);
8017 MachineOperand newOp3 = *(argOpers[3]);
8019 newOp3.setImm(newOp3.getImm()+4);
8021 newOp3.setOffset(newOp3.getOffset()+4);
8022 (*MIB).addOperand(newOp3);
8023 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8025 // t3/4 are defined later, at the bottom of the loop
8026 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8027 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8028 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8029 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8030 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8031 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8033 // The subsequent operations should be using the destination registers of
8034 //the PHI instructions.
8036 t1 = F->getRegInfo().createVirtualRegister(RC);
8037 t2 = F->getRegInfo().createVirtualRegister(RC);
8038 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8039 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8041 t1 = dest1Oper.getReg();
8042 t2 = dest2Oper.getReg();
8045 int valArgIndx = lastAddrIndx + 1;
8046 assert((argOpers[valArgIndx]->isReg() ||
8047 argOpers[valArgIndx]->isImm()) &&
8049 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8050 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8051 if (argOpers[valArgIndx]->isReg())
8052 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8054 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8055 if (regOpcL != X86::MOV32rr)
8057 (*MIB).addOperand(*argOpers[valArgIndx]);
8058 assert(argOpers[valArgIndx + 1]->isReg() ==
8059 argOpers[valArgIndx]->isReg());
8060 assert(argOpers[valArgIndx + 1]->isImm() ==
8061 argOpers[valArgIndx]->isImm());
8062 if (argOpers[valArgIndx + 1]->isReg())
8063 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8065 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8066 if (regOpcH != X86::MOV32rr)
8068 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8070 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8072 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8075 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8077 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8080 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8081 for (int i=0; i <= lastAddrIndx; ++i)
8082 (*MIB).addOperand(*argOpers[i]);
8084 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8085 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8086 bInstr->memoperands_end());
8088 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8089 MIB.addReg(X86::EAX);
8090 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8091 MIB.addReg(X86::EDX);
8094 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8096 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8100 // private utility function
8102 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8103 MachineBasicBlock *MBB,
8104 unsigned cmovOpc) const {
8105 // For the atomic min/max operator, we generate
8108 // ld t1 = [min/max.addr]
8109 // mov t2 = [min/max.val]
8111 // cmov[cond] t2 = t1
8113 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8115 // fallthrough -->nextMBB
8117 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8118 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8119 MachineFunction::iterator MBBIter = MBB;
8122 /// First build the CFG
8123 MachineFunction *F = MBB->getParent();
8124 MachineBasicBlock *thisMBB = MBB;
8125 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8126 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8127 F->insert(MBBIter, newMBB);
8128 F->insert(MBBIter, nextMBB);
8130 // Move all successors of thisMBB to nextMBB
8131 nextMBB->transferSuccessors(thisMBB);
8133 // Update thisMBB to fall through to newMBB
8134 thisMBB->addSuccessor(newMBB);
8136 // newMBB jumps to newMBB and fall through to nextMBB
8137 newMBB->addSuccessor(nextMBB);
8138 newMBB->addSuccessor(newMBB);
8140 DebugLoc dl = mInstr->getDebugLoc();
8141 // Insert instructions into newMBB based on incoming instruction
8142 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8143 "unexpected number of operands");
8144 MachineOperand& destOper = mInstr->getOperand(0);
8145 MachineOperand* argOpers[2 + X86AddrNumOperands];
8146 int numArgs = mInstr->getNumOperands() - 1;
8147 for (int i=0; i < numArgs; ++i)
8148 argOpers[i] = &mInstr->getOperand(i+1);
8150 // x86 address has 4 operands: base, index, scale, and displacement
8151 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8152 int valArgIndx = lastAddrIndx + 1;
8154 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8155 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8156 for (int i=0; i <= lastAddrIndx; ++i)
8157 (*MIB).addOperand(*argOpers[i]);
8159 // We only support register and immediate values
8160 assert((argOpers[valArgIndx]->isReg() ||
8161 argOpers[valArgIndx]->isImm()) &&
8164 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8165 if (argOpers[valArgIndx]->isReg())
8166 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8168 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8169 (*MIB).addOperand(*argOpers[valArgIndx]);
8171 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8174 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8179 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8180 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8184 // Cmp and exchange if none has modified the memory location
8185 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8186 for (int i=0; i <= lastAddrIndx; ++i)
8187 (*MIB).addOperand(*argOpers[i]);
8189 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8190 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8191 mInstr->memoperands_end());
8193 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8194 MIB.addReg(X86::EAX);
8197 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8199 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8203 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8204 // all of this code can be replaced with that in the .td file.
8206 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8207 unsigned numArgs, bool memArg) const {
8209 MachineFunction *F = BB->getParent();
8210 DebugLoc dl = MI->getDebugLoc();
8211 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8215 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8217 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8219 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8221 for (unsigned i = 0; i < numArgs; ++i) {
8222 MachineOperand &Op = MI->getOperand(i+1);
8224 if (!(Op.isReg() && Op.isImplicit()))
8228 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8231 F->DeleteMachineInstr(MI);
8237 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8239 MachineBasicBlock *MBB) const {
8240 // Emit code to save XMM registers to the stack. The ABI says that the
8241 // number of registers to save is given in %al, so it's theoretically
8242 // possible to do an indirect jump trick to avoid saving all of them,
8243 // however this code takes a simpler approach and just executes all
8244 // of the stores if %al is non-zero. It's less code, and it's probably
8245 // easier on the hardware branch predictor, and stores aren't all that
8246 // expensive anyway.
8248 // Create the new basic blocks. One block contains all the XMM stores,
8249 // and one block is the final destination regardless of whether any
8250 // stores were performed.
8251 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8252 MachineFunction *F = MBB->getParent();
8253 MachineFunction::iterator MBBIter = MBB;
8255 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8256 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8257 F->insert(MBBIter, XMMSaveMBB);
8258 F->insert(MBBIter, EndMBB);
8261 // Move any original successors of MBB to the end block.
8262 EndMBB->transferSuccessors(MBB);
8263 // The original block will now fall through to the XMM save block.
8264 MBB->addSuccessor(XMMSaveMBB);
8265 // The XMMSaveMBB will fall through to the end block.
8266 XMMSaveMBB->addSuccessor(EndMBB);
8268 // Now add the instructions.
8269 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8270 DebugLoc DL = MI->getDebugLoc();
8272 unsigned CountReg = MI->getOperand(0).getReg();
8273 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8274 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8276 if (!Subtarget->isTargetWin64()) {
8277 // If %al is 0, branch around the XMM save block.
8278 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8279 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8280 MBB->addSuccessor(EndMBB);
8283 // In the XMM save block, save all the XMM argument registers.
8284 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8285 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8286 MachineMemOperand *MMO =
8287 F->getMachineMemOperand(
8288 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8289 MachineMemOperand::MOStore, Offset,
8290 /*Size=*/16, /*Align=*/16);
8291 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8292 .addFrameIndex(RegSaveFrameIndex)
8293 .addImm(/*Scale=*/1)
8294 .addReg(/*IndexReg=*/0)
8295 .addImm(/*Disp=*/Offset)
8296 .addReg(/*Segment=*/0)
8297 .addReg(MI->getOperand(i).getReg())
8298 .addMemOperand(MMO);
8301 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8307 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8308 MachineBasicBlock *BB,
8309 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8310 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8311 DebugLoc DL = MI->getDebugLoc();
8313 // To "insert" a SELECT_CC instruction, we actually have to insert the
8314 // diamond control-flow pattern. The incoming instruction knows the
8315 // destination vreg to set, the condition code register to branch on, the
8316 // true/false values to select between, and a branch opcode to use.
8317 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8318 MachineFunction::iterator It = BB;
8324 // cmpTY ccX, r1, r2
8326 // fallthrough --> copy0MBB
8327 MachineBasicBlock *thisMBB = BB;
8328 MachineFunction *F = BB->getParent();
8329 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8330 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8332 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8333 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8334 F->insert(It, copy0MBB);
8335 F->insert(It, sinkMBB);
8336 // Update machine-CFG edges by first adding all successors of the current
8337 // block to the new block which will contain the Phi node for the select.
8338 // Also inform sdisel of the edge changes.
8339 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8340 E = BB->succ_end(); I != E; ++I) {
8341 EM->insert(std::make_pair(*I, sinkMBB));
8342 sinkMBB->addSuccessor(*I);
8344 // Next, remove all successors of the current block, and add the true
8345 // and fallthrough blocks as its successors.
8346 while (!BB->succ_empty())
8347 BB->removeSuccessor(BB->succ_begin());
8348 // Add the true and fallthrough blocks as its successors.
8349 BB->addSuccessor(copy0MBB);
8350 BB->addSuccessor(sinkMBB);
8353 // %FalseValue = ...
8354 // # fallthrough to sinkMBB
8357 // Update machine-CFG edges
8358 BB->addSuccessor(sinkMBB);
8361 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8364 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8365 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8366 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8368 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8374 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8375 MachineBasicBlock *BB,
8376 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8377 switch (MI->getOpcode()) {
8378 default: assert(false && "Unexpected instr type to insert");
8380 case X86::CMOV_V1I64:
8381 case X86::CMOV_FR32:
8382 case X86::CMOV_FR64:
8383 case X86::CMOV_V4F32:
8384 case X86::CMOV_V2F64:
8385 case X86::CMOV_V2I64:
8386 return EmitLoweredSelect(MI, BB, EM);
8388 case X86::FP32_TO_INT16_IN_MEM:
8389 case X86::FP32_TO_INT32_IN_MEM:
8390 case X86::FP32_TO_INT64_IN_MEM:
8391 case X86::FP64_TO_INT16_IN_MEM:
8392 case X86::FP64_TO_INT32_IN_MEM:
8393 case X86::FP64_TO_INT64_IN_MEM:
8394 case X86::FP80_TO_INT16_IN_MEM:
8395 case X86::FP80_TO_INT32_IN_MEM:
8396 case X86::FP80_TO_INT64_IN_MEM: {
8397 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8398 DebugLoc DL = MI->getDebugLoc();
8400 // Change the floating point control register to use "round towards zero"
8401 // mode when truncating to an integer value.
8402 MachineFunction *F = BB->getParent();
8403 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8404 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8406 // Load the old value of the high byte of the control word...
8408 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8409 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8412 // Set the high part to be round to zero...
8413 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8416 // Reload the modified control word now...
8417 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8419 // Restore the memory image of control word to original value
8420 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8423 // Get the X86 opcode to use.
8425 switch (MI->getOpcode()) {
8426 default: llvm_unreachable("illegal opcode!");
8427 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8428 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8429 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8430 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8431 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8432 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8433 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8434 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8435 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8439 MachineOperand &Op = MI->getOperand(0);
8441 AM.BaseType = X86AddressMode::RegBase;
8442 AM.Base.Reg = Op.getReg();
8444 AM.BaseType = X86AddressMode::FrameIndexBase;
8445 AM.Base.FrameIndex = Op.getIndex();
8447 Op = MI->getOperand(1);
8449 AM.Scale = Op.getImm();
8450 Op = MI->getOperand(2);
8452 AM.IndexReg = Op.getImm();
8453 Op = MI->getOperand(3);
8454 if (Op.isGlobal()) {
8455 AM.GV = Op.getGlobal();
8457 AM.Disp = Op.getImm();
8459 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8460 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8462 // Reload the original control word now.
8463 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8465 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8468 // String/text processing lowering.
8469 case X86::PCMPISTRM128REG:
8470 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8471 case X86::PCMPISTRM128MEM:
8472 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8473 case X86::PCMPESTRM128REG:
8474 return EmitPCMP(MI, BB, 5, false /* in mem */);
8475 case X86::PCMPESTRM128MEM:
8476 return EmitPCMP(MI, BB, 5, true /* in mem */);
8479 case X86::ATOMAND32:
8480 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8481 X86::AND32ri, X86::MOV32rm,
8482 X86::LCMPXCHG32, X86::MOV32rr,
8483 X86::NOT32r, X86::EAX,
8484 X86::GR32RegisterClass);
8486 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8487 X86::OR32ri, X86::MOV32rm,
8488 X86::LCMPXCHG32, X86::MOV32rr,
8489 X86::NOT32r, X86::EAX,
8490 X86::GR32RegisterClass);
8491 case X86::ATOMXOR32:
8492 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8493 X86::XOR32ri, X86::MOV32rm,
8494 X86::LCMPXCHG32, X86::MOV32rr,
8495 X86::NOT32r, X86::EAX,
8496 X86::GR32RegisterClass);
8497 case X86::ATOMNAND32:
8498 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8499 X86::AND32ri, X86::MOV32rm,
8500 X86::LCMPXCHG32, X86::MOV32rr,
8501 X86::NOT32r, X86::EAX,
8502 X86::GR32RegisterClass, true);
8503 case X86::ATOMMIN32:
8504 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8505 case X86::ATOMMAX32:
8506 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8507 case X86::ATOMUMIN32:
8508 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8509 case X86::ATOMUMAX32:
8510 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8512 case X86::ATOMAND16:
8513 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8514 X86::AND16ri, X86::MOV16rm,
8515 X86::LCMPXCHG16, X86::MOV16rr,
8516 X86::NOT16r, X86::AX,
8517 X86::GR16RegisterClass);
8519 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8520 X86::OR16ri, X86::MOV16rm,
8521 X86::LCMPXCHG16, X86::MOV16rr,
8522 X86::NOT16r, X86::AX,
8523 X86::GR16RegisterClass);
8524 case X86::ATOMXOR16:
8525 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8526 X86::XOR16ri, X86::MOV16rm,
8527 X86::LCMPXCHG16, X86::MOV16rr,
8528 X86::NOT16r, X86::AX,
8529 X86::GR16RegisterClass);
8530 case X86::ATOMNAND16:
8531 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8532 X86::AND16ri, X86::MOV16rm,
8533 X86::LCMPXCHG16, X86::MOV16rr,
8534 X86::NOT16r, X86::AX,
8535 X86::GR16RegisterClass, true);
8536 case X86::ATOMMIN16:
8537 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8538 case X86::ATOMMAX16:
8539 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8540 case X86::ATOMUMIN16:
8541 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8542 case X86::ATOMUMAX16:
8543 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8546 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8547 X86::AND8ri, X86::MOV8rm,
8548 X86::LCMPXCHG8, X86::MOV8rr,
8549 X86::NOT8r, X86::AL,
8550 X86::GR8RegisterClass);
8552 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8553 X86::OR8ri, X86::MOV8rm,
8554 X86::LCMPXCHG8, X86::MOV8rr,
8555 X86::NOT8r, X86::AL,
8556 X86::GR8RegisterClass);
8558 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8559 X86::XOR8ri, X86::MOV8rm,
8560 X86::LCMPXCHG8, X86::MOV8rr,
8561 X86::NOT8r, X86::AL,
8562 X86::GR8RegisterClass);
8563 case X86::ATOMNAND8:
8564 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8565 X86::AND8ri, X86::MOV8rm,
8566 X86::LCMPXCHG8, X86::MOV8rr,
8567 X86::NOT8r, X86::AL,
8568 X86::GR8RegisterClass, true);
8569 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8570 // This group is for 64-bit host.
8571 case X86::ATOMAND64:
8572 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8573 X86::AND64ri32, X86::MOV64rm,
8574 X86::LCMPXCHG64, X86::MOV64rr,
8575 X86::NOT64r, X86::RAX,
8576 X86::GR64RegisterClass);
8578 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8579 X86::OR64ri32, X86::MOV64rm,
8580 X86::LCMPXCHG64, X86::MOV64rr,
8581 X86::NOT64r, X86::RAX,
8582 X86::GR64RegisterClass);
8583 case X86::ATOMXOR64:
8584 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8585 X86::XOR64ri32, X86::MOV64rm,
8586 X86::LCMPXCHG64, X86::MOV64rr,
8587 X86::NOT64r, X86::RAX,
8588 X86::GR64RegisterClass);
8589 case X86::ATOMNAND64:
8590 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8591 X86::AND64ri32, X86::MOV64rm,
8592 X86::LCMPXCHG64, X86::MOV64rr,
8593 X86::NOT64r, X86::RAX,
8594 X86::GR64RegisterClass, true);
8595 case X86::ATOMMIN64:
8596 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8597 case X86::ATOMMAX64:
8598 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8599 case X86::ATOMUMIN64:
8600 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8601 case X86::ATOMUMAX64:
8602 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8604 // This group does 64-bit operations on a 32-bit host.
8605 case X86::ATOMAND6432:
8606 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8607 X86::AND32rr, X86::AND32rr,
8608 X86::AND32ri, X86::AND32ri,
8610 case X86::ATOMOR6432:
8611 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8612 X86::OR32rr, X86::OR32rr,
8613 X86::OR32ri, X86::OR32ri,
8615 case X86::ATOMXOR6432:
8616 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8617 X86::XOR32rr, X86::XOR32rr,
8618 X86::XOR32ri, X86::XOR32ri,
8620 case X86::ATOMNAND6432:
8621 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8622 X86::AND32rr, X86::AND32rr,
8623 X86::AND32ri, X86::AND32ri,
8625 case X86::ATOMADD6432:
8626 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8627 X86::ADD32rr, X86::ADC32rr,
8628 X86::ADD32ri, X86::ADC32ri,
8630 case X86::ATOMSUB6432:
8631 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8632 X86::SUB32rr, X86::SBB32rr,
8633 X86::SUB32ri, X86::SBB32ri,
8635 case X86::ATOMSWAP6432:
8636 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8637 X86::MOV32rr, X86::MOV32rr,
8638 X86::MOV32ri, X86::MOV32ri,
8640 case X86::VASTART_SAVE_XMM_REGS:
8641 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8645 //===----------------------------------------------------------------------===//
8646 // X86 Optimization Hooks
8647 //===----------------------------------------------------------------------===//
8649 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8653 const SelectionDAG &DAG,
8654 unsigned Depth) const {
8655 unsigned Opc = Op.getOpcode();
8656 assert((Opc >= ISD::BUILTIN_OP_END ||
8657 Opc == ISD::INTRINSIC_WO_CHAIN ||
8658 Opc == ISD::INTRINSIC_W_CHAIN ||
8659 Opc == ISD::INTRINSIC_VOID) &&
8660 "Should use MaskedValueIsZero if you don't know whether Op"
8661 " is a target node!");
8663 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8675 // These nodes' second result is a boolean.
8676 if (Op.getResNo() == 0)
8680 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8681 Mask.getBitWidth() - 1);
8686 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8687 /// node is a GlobalAddress + offset.
8688 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8689 GlobalValue* &GA, int64_t &Offset) const{
8690 if (N->getOpcode() == X86ISD::Wrapper) {
8691 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8692 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8693 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8697 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8700 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8701 EVT EltVT, LoadSDNode *&LDBase,
8702 unsigned &LastLoadedElt,
8703 SelectionDAG &DAG, MachineFrameInfo *MFI,
8704 const TargetLowering &TLI) {
8706 LastLoadedElt = -1U;
8707 for (unsigned i = 0; i < NumElems; ++i) {
8708 if (N->getMaskElt(i) < 0) {
8714 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8715 if (!Elt.getNode() ||
8716 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8719 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8721 LDBase = cast<LoadSDNode>(Elt.getNode());
8725 if (Elt.getOpcode() == ISD::UNDEF)
8728 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8729 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
8736 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8737 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8738 /// if the load addresses are consecutive, non-overlapping, and in the right
8739 /// order. In the case of v2i64, it will see if it can rewrite the
8740 /// shuffle to be an appropriate build vector so it can take advantage of
8741 // performBuildVectorCombine.
8742 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8743 const TargetLowering &TLI) {
8744 DebugLoc dl = N->getDebugLoc();
8745 EVT VT = N->getValueType(0);
8746 EVT EltVT = VT.getVectorElementType();
8747 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8748 unsigned NumElems = VT.getVectorNumElements();
8750 if (VT.getSizeInBits() != 128)
8753 // Try to combine a vector_shuffle into a 128-bit load.
8754 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8755 LoadSDNode *LD = NULL;
8756 unsigned LastLoadedElt;
8757 if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
8761 if (LastLoadedElt == NumElems - 1) {
8762 if (DAG.InferPtrAlignment(LD->getBasePtr()) >= 16)
8763 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8764 LD->getSrcValue(), LD->getSrcValueOffset(),
8766 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8767 LD->getSrcValue(), LD->getSrcValueOffset(),
8768 LD->isVolatile(), LD->getAlignment());
8769 } else if (NumElems == 4 && LastLoadedElt == 1) {
8770 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8771 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8772 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8773 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8778 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8779 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8780 const X86Subtarget *Subtarget) {
8781 DebugLoc DL = N->getDebugLoc();
8782 SDValue Cond = N->getOperand(0);
8783 // Get the LHS/RHS of the select.
8784 SDValue LHS = N->getOperand(1);
8785 SDValue RHS = N->getOperand(2);
8787 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8788 // instructions have the peculiarity that if either operand is a NaN,
8789 // they chose what we call the RHS operand (and as such are not symmetric).
8790 // It happens that this matches the semantics of the common C idiom
8791 // x<y?x:y and related forms, so we can recognize these cases.
8792 if (Subtarget->hasSSE2() &&
8793 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8794 Cond.getOpcode() == ISD::SETCC) {
8795 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8797 unsigned Opcode = 0;
8798 // Check for x CC y ? x : y.
8799 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8803 // This can be a min if we can prove that at least one of the operands
8805 if (!FiniteOnlyFPMath()) {
8806 if (DAG.isKnownNeverNaN(RHS)) {
8807 // Put the potential NaN in the RHS so that SSE will preserve it.
8808 std::swap(LHS, RHS);
8809 } else if (!DAG.isKnownNeverNaN(LHS))
8812 Opcode = X86ISD::FMIN;
8815 // This can be a min if we can prove that at least one of the operands
8817 if (!FiniteOnlyFPMath()) {
8818 if (DAG.isKnownNeverNaN(LHS)) {
8819 // Put the potential NaN in the RHS so that SSE will preserve it.
8820 std::swap(LHS, RHS);
8821 } else if (!DAG.isKnownNeverNaN(RHS))
8824 Opcode = X86ISD::FMIN;
8827 // This can be a min, but if either operand is a NaN we need it to
8828 // preserve the original LHS.
8829 std::swap(LHS, RHS);
8833 Opcode = X86ISD::FMIN;
8837 // This can be a max if we can prove that at least one of the operands
8839 if (!FiniteOnlyFPMath()) {
8840 if (DAG.isKnownNeverNaN(LHS)) {
8841 // Put the potential NaN in the RHS so that SSE will preserve it.
8842 std::swap(LHS, RHS);
8843 } else if (!DAG.isKnownNeverNaN(RHS))
8846 Opcode = X86ISD::FMAX;
8849 // This can be a max if we can prove that at least one of the operands
8851 if (!FiniteOnlyFPMath()) {
8852 if (DAG.isKnownNeverNaN(RHS)) {
8853 // Put the potential NaN in the RHS so that SSE will preserve it.
8854 std::swap(LHS, RHS);
8855 } else if (!DAG.isKnownNeverNaN(LHS))
8858 Opcode = X86ISD::FMAX;
8861 // This can be a max, but if either operand is a NaN we need it to
8862 // preserve the original LHS.
8863 std::swap(LHS, RHS);
8867 Opcode = X86ISD::FMAX;
8870 // Check for x CC y ? y : x -- a min/max with reversed arms.
8871 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8875 // This can be a min if we can prove that at least one of the operands
8877 if (!FiniteOnlyFPMath()) {
8878 if (DAG.isKnownNeverNaN(RHS)) {
8879 // Put the potential NaN in the RHS so that SSE will preserve it.
8880 std::swap(LHS, RHS);
8881 } else if (!DAG.isKnownNeverNaN(LHS))
8884 Opcode = X86ISD::FMIN;
8887 // This can be a min if we can prove that at least one of the operands
8889 if (!FiniteOnlyFPMath()) {
8890 if (DAG.isKnownNeverNaN(LHS)) {
8891 // Put the potential NaN in the RHS so that SSE will preserve it.
8892 std::swap(LHS, RHS);
8893 } else if (!DAG.isKnownNeverNaN(RHS))
8896 Opcode = X86ISD::FMIN;
8899 // This can be a min, but if either operand is a NaN we need it to
8900 // preserve the original LHS.
8901 std::swap(LHS, RHS);
8905 Opcode = X86ISD::FMIN;
8909 // This can be a max if we can prove that at least one of the operands
8911 if (!FiniteOnlyFPMath()) {
8912 if (DAG.isKnownNeverNaN(LHS)) {
8913 // Put the potential NaN in the RHS so that SSE will preserve it.
8914 std::swap(LHS, RHS);
8915 } else if (!DAG.isKnownNeverNaN(RHS))
8918 Opcode = X86ISD::FMAX;
8921 // This can be a max if we can prove that at least one of the operands
8923 if (!FiniteOnlyFPMath()) {
8924 if (DAG.isKnownNeverNaN(RHS)) {
8925 // Put the potential NaN in the RHS so that SSE will preserve it.
8926 std::swap(LHS, RHS);
8927 } else if (!DAG.isKnownNeverNaN(LHS))
8930 Opcode = X86ISD::FMAX;
8933 // This can be a max, but if either operand is a NaN we need it to
8934 // preserve the original LHS.
8935 std::swap(LHS, RHS);
8939 Opcode = X86ISD::FMAX;
8945 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8948 // If this is a select between two integer constants, try to do some
8950 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8951 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8952 // Don't do this for crazy integer types.
8953 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8954 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8955 // so that TrueC (the true value) is larger than FalseC.
8956 bool NeedsCondInvert = false;
8958 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8959 // Efficiently invertible.
8960 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8961 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8962 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8963 NeedsCondInvert = true;
8964 std::swap(TrueC, FalseC);
8967 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8968 if (FalseC->getAPIntValue() == 0 &&
8969 TrueC->getAPIntValue().isPowerOf2()) {
8970 if (NeedsCondInvert) // Invert the condition if needed.
8971 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8972 DAG.getConstant(1, Cond.getValueType()));
8974 // Zero extend the condition if needed.
8975 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8977 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8978 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8979 DAG.getConstant(ShAmt, MVT::i8));
8982 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8983 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8984 if (NeedsCondInvert) // Invert the condition if needed.
8985 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8986 DAG.getConstant(1, Cond.getValueType()));
8988 // Zero extend the condition if needed.
8989 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8990 FalseC->getValueType(0), Cond);
8991 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8992 SDValue(FalseC, 0));
8995 // Optimize cases that will turn into an LEA instruction. This requires
8996 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8997 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8998 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8999 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9001 bool isFastMultiplier = false;
9003 switch ((unsigned char)Diff) {
9005 case 1: // result = add base, cond
9006 case 2: // result = lea base( , cond*2)
9007 case 3: // result = lea base(cond, cond*2)
9008 case 4: // result = lea base( , cond*4)
9009 case 5: // result = lea base(cond, cond*4)
9010 case 8: // result = lea base( , cond*8)
9011 case 9: // result = lea base(cond, cond*8)
9012 isFastMultiplier = true;
9017 if (isFastMultiplier) {
9018 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9019 if (NeedsCondInvert) // Invert the condition if needed.
9020 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9021 DAG.getConstant(1, Cond.getValueType()));
9023 // Zero extend the condition if needed.
9024 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9026 // Scale the condition by the difference.
9028 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9029 DAG.getConstant(Diff, Cond.getValueType()));
9031 // Add the base if non-zero.
9032 if (FalseC->getAPIntValue() != 0)
9033 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9034 SDValue(FalseC, 0));
9044 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9045 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9046 TargetLowering::DAGCombinerInfo &DCI) {
9047 DebugLoc DL = N->getDebugLoc();
9049 // If the flag operand isn't dead, don't touch this CMOV.
9050 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9053 // If this is a select between two integer constants, try to do some
9054 // optimizations. Note that the operands are ordered the opposite of SELECT
9056 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9057 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9058 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9059 // larger than FalseC (the false value).
9060 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9062 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9063 CC = X86::GetOppositeBranchCondition(CC);
9064 std::swap(TrueC, FalseC);
9067 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9068 // This is efficient for any integer data type (including i8/i16) and
9070 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9071 SDValue Cond = N->getOperand(3);
9072 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9073 DAG.getConstant(CC, MVT::i8), Cond);
9075 // Zero extend the condition if needed.
9076 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9078 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9079 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9080 DAG.getConstant(ShAmt, MVT::i8));
9081 if (N->getNumValues() == 2) // Dead flag value?
9082 return DCI.CombineTo(N, Cond, SDValue());
9086 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9087 // for any integer data type, including i8/i16.
9088 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9089 SDValue Cond = N->getOperand(3);
9090 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9091 DAG.getConstant(CC, MVT::i8), Cond);
9093 // Zero extend the condition if needed.
9094 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9095 FalseC->getValueType(0), Cond);
9096 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9097 SDValue(FalseC, 0));
9099 if (N->getNumValues() == 2) // Dead flag value?
9100 return DCI.CombineTo(N, Cond, SDValue());
9104 // Optimize cases that will turn into an LEA instruction. This requires
9105 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9106 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9107 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9108 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9110 bool isFastMultiplier = false;
9112 switch ((unsigned char)Diff) {
9114 case 1: // result = add base, cond
9115 case 2: // result = lea base( , cond*2)
9116 case 3: // result = lea base(cond, cond*2)
9117 case 4: // result = lea base( , cond*4)
9118 case 5: // result = lea base(cond, cond*4)
9119 case 8: // result = lea base( , cond*8)
9120 case 9: // result = lea base(cond, cond*8)
9121 isFastMultiplier = true;
9126 if (isFastMultiplier) {
9127 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9128 SDValue Cond = N->getOperand(3);
9129 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9130 DAG.getConstant(CC, MVT::i8), Cond);
9131 // Zero extend the condition if needed.
9132 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9134 // Scale the condition by the difference.
9136 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9137 DAG.getConstant(Diff, Cond.getValueType()));
9139 // Add the base if non-zero.
9140 if (FalseC->getAPIntValue() != 0)
9141 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9142 SDValue(FalseC, 0));
9143 if (N->getNumValues() == 2) // Dead flag value?
9144 return DCI.CombineTo(N, Cond, SDValue());
9154 /// PerformMulCombine - Optimize a single multiply with constant into two
9155 /// in order to implement it with two cheaper instructions, e.g.
9156 /// LEA + SHL, LEA + LEA.
9157 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9158 TargetLowering::DAGCombinerInfo &DCI) {
9159 if (DAG.getMachineFunction().
9160 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
9163 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9166 EVT VT = N->getValueType(0);
9170 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9173 uint64_t MulAmt = C->getZExtValue();
9174 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9177 uint64_t MulAmt1 = 0;
9178 uint64_t MulAmt2 = 0;
9179 if ((MulAmt % 9) == 0) {
9181 MulAmt2 = MulAmt / 9;
9182 } else if ((MulAmt % 5) == 0) {
9184 MulAmt2 = MulAmt / 5;
9185 } else if ((MulAmt % 3) == 0) {
9187 MulAmt2 = MulAmt / 3;
9190 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9191 DebugLoc DL = N->getDebugLoc();
9193 if (isPowerOf2_64(MulAmt2) &&
9194 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9195 // If second multiplifer is pow2, issue it first. We want the multiply by
9196 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9198 std::swap(MulAmt1, MulAmt2);
9201 if (isPowerOf2_64(MulAmt1))
9202 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9203 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9205 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9206 DAG.getConstant(MulAmt1, VT));
9208 if (isPowerOf2_64(MulAmt2))
9209 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9210 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9212 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9213 DAG.getConstant(MulAmt2, VT));
9215 // Do not add new nodes to DAG combiner worklist.
9216 DCI.CombineTo(N, NewMul, false);
9221 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9222 SDValue N0 = N->getOperand(0);
9223 SDValue N1 = N->getOperand(1);
9224 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9225 EVT VT = N0.getValueType();
9227 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9228 // since the result of setcc_c is all zero's or all ones.
9229 if (N1C && N0.getOpcode() == ISD::AND &&
9230 N0.getOperand(1).getOpcode() == ISD::Constant) {
9231 SDValue N00 = N0.getOperand(0);
9232 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9233 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9234 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9235 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9236 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9237 APInt ShAmt = N1C->getAPIntValue();
9238 Mask = Mask.shl(ShAmt);
9240 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9241 N00, DAG.getConstant(Mask, VT));
9248 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9250 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9251 const X86Subtarget *Subtarget) {
9252 EVT VT = N->getValueType(0);
9253 if (!VT.isVector() && VT.isInteger() &&
9254 N->getOpcode() == ISD::SHL)
9255 return PerformSHLCombine(N, DAG);
9257 // On X86 with SSE2 support, we can transform this to a vector shift if
9258 // all elements are shifted by the same amount. We can't do this in legalize
9259 // because the a constant vector is typically transformed to a constant pool
9260 // so we have no knowledge of the shift amount.
9261 if (!Subtarget->hasSSE2())
9264 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9267 SDValue ShAmtOp = N->getOperand(1);
9268 EVT EltVT = VT.getVectorElementType();
9269 DebugLoc DL = N->getDebugLoc();
9270 SDValue BaseShAmt = SDValue();
9271 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9272 unsigned NumElts = VT.getVectorNumElements();
9274 for (; i != NumElts; ++i) {
9275 SDValue Arg = ShAmtOp.getOperand(i);
9276 if (Arg.getOpcode() == ISD::UNDEF) continue;
9280 for (; i != NumElts; ++i) {
9281 SDValue Arg = ShAmtOp.getOperand(i);
9282 if (Arg.getOpcode() == ISD::UNDEF) continue;
9283 if (Arg != BaseShAmt) {
9287 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9288 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9289 SDValue InVec = ShAmtOp.getOperand(0);
9290 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9291 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9293 for (; i != NumElts; ++i) {
9294 SDValue Arg = InVec.getOperand(i);
9295 if (Arg.getOpcode() == ISD::UNDEF) continue;
9299 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9300 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9301 unsigned SplatIdx = cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9302 if (C->getZExtValue() == SplatIdx)
9303 BaseShAmt = InVec.getOperand(1);
9306 if (BaseShAmt.getNode() == 0)
9307 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9308 DAG.getIntPtrConstant(0));
9312 // The shift amount is an i32.
9313 if (EltVT.bitsGT(MVT::i32))
9314 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9315 else if (EltVT.bitsLT(MVT::i32))
9316 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9318 // The shift amount is identical so we can do a vector shift.
9319 SDValue ValOp = N->getOperand(0);
9320 switch (N->getOpcode()) {
9322 llvm_unreachable("Unknown shift opcode!");
9325 if (VT == MVT::v2i64)
9326 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9327 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9329 if (VT == MVT::v4i32)
9330 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9331 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9333 if (VT == MVT::v8i16)
9334 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9335 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9339 if (VT == MVT::v4i32)
9340 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9341 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9343 if (VT == MVT::v8i16)
9344 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9345 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9349 if (VT == MVT::v2i64)
9350 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9351 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9353 if (VT == MVT::v4i32)
9354 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9355 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9357 if (VT == MVT::v8i16)
9358 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9359 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9366 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9367 const X86Subtarget *Subtarget) {
9368 EVT VT = N->getValueType(0);
9369 if (VT != MVT::i64 || !Subtarget->is64Bit())
9372 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9373 SDValue N0 = N->getOperand(0);
9374 SDValue N1 = N->getOperand(1);
9375 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9377 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9380 SDValue ShAmt0 = N0.getOperand(1);
9381 if (ShAmt0.getValueType() != MVT::i8)
9383 SDValue ShAmt1 = N1.getOperand(1);
9384 if (ShAmt1.getValueType() != MVT::i8)
9386 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9387 ShAmt0 = ShAmt0.getOperand(0);
9388 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9389 ShAmt1 = ShAmt1.getOperand(0);
9391 DebugLoc DL = N->getDebugLoc();
9392 unsigned Opc = X86ISD::SHLD;
9393 SDValue Op0 = N0.getOperand(0);
9394 SDValue Op1 = N1.getOperand(0);
9395 if (ShAmt0.getOpcode() == ISD::SUB) {
9397 std::swap(Op0, Op1);
9398 std::swap(ShAmt0, ShAmt1);
9401 if (ShAmt1.getOpcode() == ISD::SUB) {
9402 SDValue Sum = ShAmt1.getOperand(0);
9403 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9404 if (SumC->getSExtValue() == 64 &&
9405 ShAmt1.getOperand(1) == ShAmt0)
9406 return DAG.getNode(Opc, DL, VT,
9408 DAG.getNode(ISD::TRUNCATE, DL,
9411 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9412 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9414 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
9415 return DAG.getNode(Opc, DL, VT,
9416 N0.getOperand(0), N1.getOperand(0),
9417 DAG.getNode(ISD::TRUNCATE, DL,
9424 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9425 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9426 const X86Subtarget *Subtarget) {
9427 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9428 // the FP state in cases where an emms may be missing.
9429 // A preferable solution to the general problem is to figure out the right
9430 // places to insert EMMS. This qualifies as a quick hack.
9432 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9433 StoreSDNode *St = cast<StoreSDNode>(N);
9434 EVT VT = St->getValue().getValueType();
9435 if (VT.getSizeInBits() != 64)
9438 const Function *F = DAG.getMachineFunction().getFunction();
9439 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9440 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9441 && Subtarget->hasSSE2();
9442 if ((VT.isVector() ||
9443 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9444 isa<LoadSDNode>(St->getValue()) &&
9445 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9446 St->getChain().hasOneUse() && !St->isVolatile()) {
9447 SDNode* LdVal = St->getValue().getNode();
9449 int TokenFactorIndex = -1;
9450 SmallVector<SDValue, 8> Ops;
9451 SDNode* ChainVal = St->getChain().getNode();
9452 // Must be a store of a load. We currently handle two cases: the load
9453 // is a direct child, and it's under an intervening TokenFactor. It is
9454 // possible to dig deeper under nested TokenFactors.
9455 if (ChainVal == LdVal)
9456 Ld = cast<LoadSDNode>(St->getChain());
9457 else if (St->getValue().hasOneUse() &&
9458 ChainVal->getOpcode() == ISD::TokenFactor) {
9459 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9460 if (ChainVal->getOperand(i).getNode() == LdVal) {
9461 TokenFactorIndex = i;
9462 Ld = cast<LoadSDNode>(St->getValue());
9464 Ops.push_back(ChainVal->getOperand(i));
9468 if (!Ld || !ISD::isNormalLoad(Ld))
9471 // If this is not the MMX case, i.e. we are just turning i64 load/store
9472 // into f64 load/store, avoid the transformation if there are multiple
9473 // uses of the loaded value.
9474 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9477 DebugLoc LdDL = Ld->getDebugLoc();
9478 DebugLoc StDL = N->getDebugLoc();
9479 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9480 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9482 if (Subtarget->is64Bit() || F64IsLegal) {
9483 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9484 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9485 Ld->getBasePtr(), Ld->getSrcValue(),
9486 Ld->getSrcValueOffset(), Ld->isVolatile(),
9487 Ld->getAlignment());
9488 SDValue NewChain = NewLd.getValue(1);
9489 if (TokenFactorIndex != -1) {
9490 Ops.push_back(NewChain);
9491 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9494 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9495 St->getSrcValue(), St->getSrcValueOffset(),
9496 St->isVolatile(), St->getAlignment());
9499 // Otherwise, lower to two pairs of 32-bit loads / stores.
9500 SDValue LoAddr = Ld->getBasePtr();
9501 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9502 DAG.getConstant(4, MVT::i32));
9504 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9505 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9506 Ld->isVolatile(), Ld->getAlignment());
9507 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9508 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9510 MinAlign(Ld->getAlignment(), 4));
9512 SDValue NewChain = LoLd.getValue(1);
9513 if (TokenFactorIndex != -1) {
9514 Ops.push_back(LoLd);
9515 Ops.push_back(HiLd);
9516 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9520 LoAddr = St->getBasePtr();
9521 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9522 DAG.getConstant(4, MVT::i32));
9524 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9525 St->getSrcValue(), St->getSrcValueOffset(),
9526 St->isVolatile(), St->getAlignment());
9527 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9529 St->getSrcValueOffset() + 4,
9531 MinAlign(St->getAlignment(), 4));
9532 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9537 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9538 /// X86ISD::FXOR nodes.
9539 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9540 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9541 // F[X]OR(0.0, x) -> x
9542 // F[X]OR(x, 0.0) -> x
9543 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9544 if (C->getValueAPF().isPosZero())
9545 return N->getOperand(1);
9546 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9547 if (C->getValueAPF().isPosZero())
9548 return N->getOperand(0);
9552 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9553 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9554 // FAND(0.0, x) -> 0.0
9555 // FAND(x, 0.0) -> 0.0
9556 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9557 if (C->getValueAPF().isPosZero())
9558 return N->getOperand(0);
9559 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9560 if (C->getValueAPF().isPosZero())
9561 return N->getOperand(1);
9565 static SDValue PerformBTCombine(SDNode *N,
9567 TargetLowering::DAGCombinerInfo &DCI) {
9568 // BT ignores high bits in the bit index operand.
9569 SDValue Op1 = N->getOperand(1);
9570 if (Op1.hasOneUse()) {
9571 unsigned BitWidth = Op1.getValueSizeInBits();
9572 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9573 APInt KnownZero, KnownOne;
9574 TargetLowering::TargetLoweringOpt TLO(DAG);
9575 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9576 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9577 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9578 DCI.CommitTargetLoweringOpt(TLO);
9583 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9584 SDValue Op = N->getOperand(0);
9585 if (Op.getOpcode() == ISD::BIT_CONVERT)
9586 Op = Op.getOperand(0);
9587 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9588 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9589 VT.getVectorElementType().getSizeInBits() ==
9590 OpVT.getVectorElementType().getSizeInBits()) {
9591 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9596 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9597 // Locked instructions, in turn, have implicit fence semantics (all memory
9598 // operations are flushed before issuing the locked instruction, and the
9599 // are not buffered), so we can fold away the common pattern of
9600 // fence-atomic-fence.
9601 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9602 SDValue atomic = N->getOperand(0);
9603 switch (atomic.getOpcode()) {
9604 case ISD::ATOMIC_CMP_SWAP:
9605 case ISD::ATOMIC_SWAP:
9606 case ISD::ATOMIC_LOAD_ADD:
9607 case ISD::ATOMIC_LOAD_SUB:
9608 case ISD::ATOMIC_LOAD_AND:
9609 case ISD::ATOMIC_LOAD_OR:
9610 case ISD::ATOMIC_LOAD_XOR:
9611 case ISD::ATOMIC_LOAD_NAND:
9612 case ISD::ATOMIC_LOAD_MIN:
9613 case ISD::ATOMIC_LOAD_MAX:
9614 case ISD::ATOMIC_LOAD_UMIN:
9615 case ISD::ATOMIC_LOAD_UMAX:
9621 SDValue fence = atomic.getOperand(0);
9622 if (fence.getOpcode() != ISD::MEMBARRIER)
9625 switch (atomic.getOpcode()) {
9626 case ISD::ATOMIC_CMP_SWAP:
9627 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9628 atomic.getOperand(1), atomic.getOperand(2),
9629 atomic.getOperand(3));
9630 case ISD::ATOMIC_SWAP:
9631 case ISD::ATOMIC_LOAD_ADD:
9632 case ISD::ATOMIC_LOAD_SUB:
9633 case ISD::ATOMIC_LOAD_AND:
9634 case ISD::ATOMIC_LOAD_OR:
9635 case ISD::ATOMIC_LOAD_XOR:
9636 case ISD::ATOMIC_LOAD_NAND:
9637 case ISD::ATOMIC_LOAD_MIN:
9638 case ISD::ATOMIC_LOAD_MAX:
9639 case ISD::ATOMIC_LOAD_UMIN:
9640 case ISD::ATOMIC_LOAD_UMAX:
9641 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9642 atomic.getOperand(1), atomic.getOperand(2));
9648 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9649 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9650 // (and (i32 x86isd::setcc_carry), 1)
9651 // This eliminates the zext. This transformation is necessary because
9652 // ISD::SETCC is always legalized to i8.
9653 DebugLoc dl = N->getDebugLoc();
9654 SDValue N0 = N->getOperand(0);
9655 EVT VT = N->getValueType(0);
9656 if (N0.getOpcode() == ISD::AND &&
9658 N0.getOperand(0).hasOneUse()) {
9659 SDValue N00 = N0.getOperand(0);
9660 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9662 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9663 if (!C || C->getZExtValue() != 1)
9665 return DAG.getNode(ISD::AND, dl, VT,
9666 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9667 N00.getOperand(0), N00.getOperand(1)),
9668 DAG.getConstant(1, VT));
9674 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9675 DAGCombinerInfo &DCI) const {
9676 SelectionDAG &DAG = DCI.DAG;
9677 switch (N->getOpcode()) {
9679 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9680 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9681 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9682 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9685 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9686 case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
9687 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9689 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9690 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9691 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9692 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9693 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9694 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9700 //===----------------------------------------------------------------------===//
9701 // X86 Inline Assembly Support
9702 //===----------------------------------------------------------------------===//
9704 static bool LowerToBSwap(CallInst *CI) {
9705 // FIXME: this should verify that we are targetting a 486 or better. If not,
9706 // we will turn this bswap into something that will be lowered to logical ops
9707 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9708 // so don't worry about this.
9710 // Verify this is a simple bswap.
9711 if (CI->getNumOperands() != 2 ||
9712 CI->getType() != CI->getOperand(1)->getType() ||
9713 !CI->getType()->isInteger())
9716 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9717 if (!Ty || Ty->getBitWidth() % 16 != 0)
9720 // Okay, we can do this xform, do so now.
9721 const Type *Tys[] = { Ty };
9722 Module *M = CI->getParent()->getParent()->getParent();
9723 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9725 Value *Op = CI->getOperand(1);
9726 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9728 CI->replaceAllUsesWith(Op);
9729 CI->eraseFromParent();
9733 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9734 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9735 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9737 std::string AsmStr = IA->getAsmString();
9739 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9740 SmallVector<StringRef, 4> AsmPieces;
9741 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9743 switch (AsmPieces.size()) {
9744 default: return false;
9746 AsmStr = AsmPieces[0];
9748 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9751 if (AsmPieces.size() == 2 &&
9752 (AsmPieces[0] == "bswap" ||
9753 AsmPieces[0] == "bswapq" ||
9754 AsmPieces[0] == "bswapl") &&
9755 (AsmPieces[1] == "$0" ||
9756 AsmPieces[1] == "${0:q}")) {
9757 // No need to check constraints, nothing other than the equivalent of
9758 // "=r,0" would be valid here.
9759 return LowerToBSwap(CI);
9761 // rorw $$8, ${0:w} --> llvm.bswap.i16
9762 if (CI->getType()->isInteger(16) &&
9763 AsmPieces.size() == 3 &&
9764 AsmPieces[0] == "rorw" &&
9765 AsmPieces[1] == "$$8," &&
9766 AsmPieces[2] == "${0:w}" &&
9767 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
9768 return LowerToBSwap(CI);
9772 if (CI->getType()->isInteger(64) &&
9773 Constraints.size() >= 2 &&
9774 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9775 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9776 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9777 SmallVector<StringRef, 4> Words;
9778 SplitString(AsmPieces[0], Words, " \t");
9779 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9781 SplitString(AsmPieces[1], Words, " \t");
9782 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9784 SplitString(AsmPieces[2], Words, " \t,");
9785 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9786 Words[2] == "%edx") {
9787 return LowerToBSwap(CI);
9799 /// getConstraintType - Given a constraint letter, return the type of
9800 /// constraint it is for this target.
9801 X86TargetLowering::ConstraintType
9802 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
9803 if (Constraint.size() == 1) {
9804 switch (Constraint[0]) {
9816 return C_RegisterClass;
9824 return TargetLowering::getConstraintType(Constraint);
9827 /// LowerXConstraint - try to replace an X constraint, which matches anything,
9828 /// with another that has more specific requirements based on the type of the
9829 /// corresponding operand.
9830 const char *X86TargetLowering::
9831 LowerXConstraint(EVT ConstraintVT) const {
9832 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
9833 // 'f' like normal targets.
9834 if (ConstraintVT.isFloatingPoint()) {
9835 if (Subtarget->hasSSE2())
9837 if (Subtarget->hasSSE1())
9841 return TargetLowering::LowerXConstraint(ConstraintVT);
9844 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9845 /// vector. If it is invalid, don't add anything to Ops.
9846 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9849 std::vector<SDValue>&Ops,
9850 SelectionDAG &DAG) const {
9851 SDValue Result(0, 0);
9853 switch (Constraint) {
9856 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9857 if (C->getZExtValue() <= 31) {
9858 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9864 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9865 if (C->getZExtValue() <= 63) {
9866 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9872 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9873 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9874 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9880 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9881 if (C->getZExtValue() <= 255) {
9882 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9888 // 32-bit signed 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->getSExtValue())) {
9893 // Widen to 64 bits here to get it sign extended.
9894 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9897 // FIXME gcc accepts some relocatable values here too, but only in certain
9898 // memory models; it's complicated.
9903 // 32-bit unsigned value
9904 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9905 const ConstantInt *CI = C->getConstantIntValue();
9906 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9907 C->getZExtValue())) {
9908 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9912 // FIXME gcc accepts some relocatable values here too, but only in certain
9913 // memory models; it's complicated.
9917 // Literal immediates are always ok.
9918 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
9919 // Widen to 64 bits here to get it sign extended.
9920 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
9924 // If we are in non-pic codegen mode, we allow the address of a global (with
9925 // an optional displacement) to be used with 'i'.
9926 GlobalAddressSDNode *GA = 0;
9929 // Match either (GA), (GA+C), (GA+C1+C2), etc.
9931 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
9932 Offset += GA->getOffset();
9934 } else if (Op.getOpcode() == ISD::ADD) {
9935 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9936 Offset += C->getZExtValue();
9937 Op = Op.getOperand(0);
9940 } else if (Op.getOpcode() == ISD::SUB) {
9941 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9942 Offset += -C->getZExtValue();
9943 Op = Op.getOperand(0);
9948 // Otherwise, this isn't something we can handle, reject it.
9952 GlobalValue *GV = GA->getGlobal();
9953 // If we require an extra load to get this address, as in PIC mode, we
9955 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
9956 getTargetMachine())))
9960 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
9962 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
9968 if (Result.getNode()) {
9969 Ops.push_back(Result);
9972 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
9976 std::vector<unsigned> X86TargetLowering::
9977 getRegClassForInlineAsmConstraint(const std::string &Constraint,
9979 if (Constraint.size() == 1) {
9980 // FIXME: not handling fp-stack yet!
9981 switch (Constraint[0]) { // GCC X86 Constraint Letters
9982 default: break; // Unknown constraint letter
9983 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
9984 if (Subtarget->is64Bit()) {
9986 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
9987 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
9988 X86::R10D,X86::R11D,X86::R12D,
9989 X86::R13D,X86::R14D,X86::R15D,
9990 X86::EBP, X86::ESP, 0);
9991 else if (VT == MVT::i16)
9992 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
9993 X86::SI, X86::DI, X86::R8W,X86::R9W,
9994 X86::R10W,X86::R11W,X86::R12W,
9995 X86::R13W,X86::R14W,X86::R15W,
9996 X86::BP, X86::SP, 0);
9997 else if (VT == MVT::i8)
9998 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
9999 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10000 X86::R10B,X86::R11B,X86::R12B,
10001 X86::R13B,X86::R14B,X86::R15B,
10002 X86::BPL, X86::SPL, 0);
10004 else if (VT == MVT::i64)
10005 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10006 X86::RSI, X86::RDI, X86::R8, X86::R9,
10007 X86::R10, X86::R11, X86::R12,
10008 X86::R13, X86::R14, X86::R15,
10009 X86::RBP, X86::RSP, 0);
10013 // 32-bit fallthrough
10014 case 'Q': // Q_REGS
10015 if (VT == MVT::i32)
10016 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10017 else if (VT == MVT::i16)
10018 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10019 else if (VT == MVT::i8)
10020 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10021 else if (VT == MVT::i64)
10022 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10027 return std::vector<unsigned>();
10030 std::pair<unsigned, const TargetRegisterClass*>
10031 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10033 // First, see if this is a constraint that directly corresponds to an LLVM
10035 if (Constraint.size() == 1) {
10036 // GCC Constraint Letters
10037 switch (Constraint[0]) {
10039 case 'r': // GENERAL_REGS
10040 case 'l': // INDEX_REGS
10042 return std::make_pair(0U, X86::GR8RegisterClass);
10043 if (VT == MVT::i16)
10044 return std::make_pair(0U, X86::GR16RegisterClass);
10045 if (VT == MVT::i32 || !Subtarget->is64Bit())
10046 return std::make_pair(0U, X86::GR32RegisterClass);
10047 return std::make_pair(0U, X86::GR64RegisterClass);
10048 case 'R': // LEGACY_REGS
10050 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10051 if (VT == MVT::i16)
10052 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10053 if (VT == MVT::i32 || !Subtarget->is64Bit())
10054 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10055 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10056 case 'f': // FP Stack registers.
10057 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10058 // value to the correct fpstack register class.
10059 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10060 return std::make_pair(0U, X86::RFP32RegisterClass);
10061 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10062 return std::make_pair(0U, X86::RFP64RegisterClass);
10063 return std::make_pair(0U, X86::RFP80RegisterClass);
10064 case 'y': // MMX_REGS if MMX allowed.
10065 if (!Subtarget->hasMMX()) break;
10066 return std::make_pair(0U, X86::VR64RegisterClass);
10067 case 'Y': // SSE_REGS if SSE2 allowed
10068 if (!Subtarget->hasSSE2()) break;
10070 case 'x': // SSE_REGS if SSE1 allowed
10071 if (!Subtarget->hasSSE1()) break;
10073 switch (VT.getSimpleVT().SimpleTy) {
10075 // Scalar SSE types.
10078 return std::make_pair(0U, X86::FR32RegisterClass);
10081 return std::make_pair(0U, X86::FR64RegisterClass);
10089 return std::make_pair(0U, X86::VR128RegisterClass);
10095 // Use the default implementation in TargetLowering to convert the register
10096 // constraint into a member of a register class.
10097 std::pair<unsigned, const TargetRegisterClass*> Res;
10098 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10100 // Not found as a standard register?
10101 if (Res.second == 0) {
10102 // Map st(0) -> st(7) -> ST0
10103 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10104 tolower(Constraint[1]) == 's' &&
10105 tolower(Constraint[2]) == 't' &&
10106 Constraint[3] == '(' &&
10107 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10108 Constraint[5] == ')' &&
10109 Constraint[6] == '}') {
10111 Res.first = X86::ST0+Constraint[4]-'0';
10112 Res.second = X86::RFP80RegisterClass;
10116 // GCC allows "st(0)" to be called just plain "st".
10117 if (StringRef("{st}").equals_lower(Constraint)) {
10118 Res.first = X86::ST0;
10119 Res.second = X86::RFP80RegisterClass;
10124 if (StringRef("{flags}").equals_lower(Constraint)) {
10125 Res.first = X86::EFLAGS;
10126 Res.second = X86::CCRRegisterClass;
10130 // 'A' means EAX + EDX.
10131 if (Constraint == "A") {
10132 Res.first = X86::EAX;
10133 Res.second = X86::GR32_ADRegisterClass;
10139 // Otherwise, check to see if this is a register class of the wrong value
10140 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10141 // turn into {ax},{dx}.
10142 if (Res.second->hasType(VT))
10143 return Res; // Correct type already, nothing to do.
10145 // All of the single-register GCC register classes map their values onto
10146 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10147 // really want an 8-bit or 32-bit register, map to the appropriate register
10148 // class and return the appropriate register.
10149 if (Res.second == X86::GR16RegisterClass) {
10150 if (VT == MVT::i8) {
10151 unsigned DestReg = 0;
10152 switch (Res.first) {
10154 case X86::AX: DestReg = X86::AL; break;
10155 case X86::DX: DestReg = X86::DL; break;
10156 case X86::CX: DestReg = X86::CL; break;
10157 case X86::BX: DestReg = X86::BL; break;
10160 Res.first = DestReg;
10161 Res.second = X86::GR8RegisterClass;
10163 } else if (VT == MVT::i32) {
10164 unsigned DestReg = 0;
10165 switch (Res.first) {
10167 case X86::AX: DestReg = X86::EAX; break;
10168 case X86::DX: DestReg = X86::EDX; break;
10169 case X86::CX: DestReg = X86::ECX; break;
10170 case X86::BX: DestReg = X86::EBX; break;
10171 case X86::SI: DestReg = X86::ESI; break;
10172 case X86::DI: DestReg = X86::EDI; break;
10173 case X86::BP: DestReg = X86::EBP; break;
10174 case X86::SP: DestReg = X86::ESP; break;
10177 Res.first = DestReg;
10178 Res.second = X86::GR32RegisterClass;
10180 } else if (VT == MVT::i64) {
10181 unsigned DestReg = 0;
10182 switch (Res.first) {
10184 case X86::AX: DestReg = X86::RAX; break;
10185 case X86::DX: DestReg = X86::RDX; break;
10186 case X86::CX: DestReg = X86::RCX; break;
10187 case X86::BX: DestReg = X86::RBX; break;
10188 case X86::SI: DestReg = X86::RSI; break;
10189 case X86::DI: DestReg = X86::RDI; break;
10190 case X86::BP: DestReg = X86::RBP; break;
10191 case X86::SP: DestReg = X86::RSP; break;
10194 Res.first = DestReg;
10195 Res.second = X86::GR64RegisterClass;
10198 } else if (Res.second == X86::FR32RegisterClass ||
10199 Res.second == X86::FR64RegisterClass ||
10200 Res.second == X86::VR128RegisterClass) {
10201 // Handle references to XMM physical registers that got mapped into the
10202 // wrong class. This can happen with constraints like {xmm0} where the
10203 // target independent register mapper will just pick the first match it can
10204 // find, ignoring the required type.
10205 if (VT == MVT::f32)
10206 Res.second = X86::FR32RegisterClass;
10207 else if (VT == MVT::f64)
10208 Res.second = X86::FR64RegisterClass;
10209 else if (X86::VR128RegisterClass->hasType(VT))
10210 Res.second = X86::VR128RegisterClass;
10216 //===----------------------------------------------------------------------===//
10217 // X86 Widen vector type
10218 //===----------------------------------------------------------------------===//
10220 /// getWidenVectorType: given a vector type, returns the type to widen
10221 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
10222 /// If there is no vector type that we want to widen to, returns MVT::Other
10223 /// When and where to widen is target dependent based on the cost of
10224 /// scalarizing vs using the wider vector type.
10226 EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
10227 assert(VT.isVector());
10228 if (isTypeLegal(VT))
10231 // TODO: In computeRegisterProperty, we can compute the list of legal vector
10232 // type based on element type. This would speed up our search (though
10233 // it may not be worth it since the size of the list is relatively
10235 EVT EltVT = VT.getVectorElementType();
10236 unsigned NElts = VT.getVectorNumElements();
10238 // On X86, it make sense to widen any vector wider than 1
10242 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
10243 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
10244 EVT SVT = (MVT::SimpleValueType)nVT;
10246 if (isTypeLegal(SVT) &&
10247 SVT.getVectorElementType() == EltVT &&
10248 SVT.getVectorNumElements() > NElts)