1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "x86-isel"
17 #include "X86InstrBuilder.h"
18 #include "X86ISelLowering.h"
19 #include "X86TargetMachine.h"
20 #include "X86TargetObjectFile.h"
21 #include "llvm/CallingConv.h"
22 #include "llvm/Constants.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/GlobalAlias.h"
25 #include "llvm/GlobalVariable.h"
26 #include "llvm/Function.h"
27 #include "llvm/Instructions.h"
28 #include "llvm/Intrinsics.h"
29 #include "llvm/LLVMContext.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/PseudoSourceValue.h"
37 #include "llvm/MC/MCAsmInfo.h"
38 #include "llvm/MC/MCContext.h"
39 #include "llvm/MC/MCExpr.h"
40 #include "llvm/MC/MCSymbol.h"
41 #include "llvm/ADT/BitVector.h"
42 #include "llvm/ADT/SmallSet.h"
43 #include "llvm/ADT/Statistic.h"
44 #include "llvm/ADT/StringExtras.h"
45 #include "llvm/ADT/VectorExtras.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/ErrorHandling.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Support/raw_ostream.h"
53 STATISTIC(NumTailCalls, "Number of tail calls");
56 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
58 // Disable16Bit - 16-bit operations typically have a larger encoding than
59 // corresponding 32-bit instructions, and 16-bit code is slow on some
60 // processors. This is an experimental flag to disable 16-bit operations
61 // (which forces them to be Legalized to 32-bit operations).
63 Disable16Bit("disable-16bit", cl::Hidden,
64 cl::desc("Disable use of 16-bit instructions"));
66 // Forward declarations.
67 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
70 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
71 switch (TM.getSubtarget<X86Subtarget>().TargetType) {
72 default: llvm_unreachable("unknown subtarget type");
73 case X86Subtarget::isDarwin:
74 if (TM.getSubtarget<X86Subtarget>().is64Bit())
75 return new X8664_MachoTargetObjectFile();
76 return new X8632_MachoTargetObjectFile();
77 case X86Subtarget::isELF:
78 return new TargetLoweringObjectFileELF();
79 case X86Subtarget::isMingw:
80 case X86Subtarget::isCygwin:
81 case X86Subtarget::isWindows:
82 return new TargetLoweringObjectFileCOFF();
87 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
88 : TargetLowering(TM, createTLOF(TM)) {
89 Subtarget = &TM.getSubtarget<X86Subtarget>();
90 X86ScalarSSEf64 = Subtarget->hasSSE2();
91 X86ScalarSSEf32 = Subtarget->hasSSE1();
92 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
94 RegInfo = TM.getRegisterInfo();
97 // Set up the TargetLowering object.
99 // X86 is weird, it always uses i8 for shift amounts and setcc results.
100 setShiftAmountType(MVT::i8);
101 setBooleanContents(ZeroOrOneBooleanContent);
102 setSchedulingPreference(SchedulingForRegPressure);
103 setStackPointerRegisterToSaveRestore(X86StackPtr);
105 if (Subtarget->isTargetDarwin()) {
106 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
107 setUseUnderscoreSetJmp(false);
108 setUseUnderscoreLongJmp(false);
109 } else if (Subtarget->isTargetMingw()) {
110 // MS runtime is weird: it exports _setjmp, but longjmp!
111 setUseUnderscoreSetJmp(true);
112 setUseUnderscoreLongJmp(false);
114 setUseUnderscoreSetJmp(true);
115 setUseUnderscoreLongJmp(true);
118 // Set up the register classes.
119 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
121 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
122 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
123 if (Subtarget->is64Bit())
124 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
126 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
128 // We don't accept any truncstore of integer registers.
129 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
131 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
132 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
134 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
135 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
136 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
138 // SETOEQ and SETUNE require checking two conditions.
139 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
140 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
141 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
142 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
143 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
144 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
146 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
148 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
149 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
150 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
152 if (Subtarget->is64Bit()) {
153 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
154 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
155 } else if (!UseSoftFloat) {
156 if (X86ScalarSSEf64) {
157 // We have an impenetrably clever algorithm for ui64->double only.
158 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
160 // We have an algorithm for SSE2, and we turn this into a 64-bit
161 // FILD for other targets.
162 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
165 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
167 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
168 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
171 // SSE has no i16 to fp conversion, only i32
172 if (X86ScalarSSEf32) {
173 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
174 // f32 and f64 cases are Legal, f80 case is not
175 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
177 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
178 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
181 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
182 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
185 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
186 // are Legal, f80 is custom lowered.
187 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
188 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
190 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
192 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
193 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
195 if (X86ScalarSSEf32) {
196 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
197 // f32 and f64 cases are Legal, f80 case is not
198 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
200 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
201 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
204 // Handle FP_TO_UINT by promoting the destination to a larger signed
206 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
207 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
208 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
210 if (Subtarget->is64Bit()) {
211 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
212 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
213 } else if (!UseSoftFloat) {
214 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
215 // Expand FP_TO_UINT into a select.
216 // FIXME: We would like to use a Custom expander here eventually to do
217 // the optimal thing for SSE vs. the default expansion in the legalizer.
218 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
220 // With SSE3 we can use fisttpll to convert to a signed i64; without
221 // SSE, we're stuck with a fistpll.
222 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
225 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
226 if (!X86ScalarSSEf64) {
227 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
228 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
231 // Scalar integer divide and remainder are lowered to use operations that
232 // produce two results, to match the available instructions. This exposes
233 // the two-result form to trivial CSE, which is able to combine x/y and x%y
234 // into a single instruction.
236 // Scalar integer multiply-high is also lowered to use two-result
237 // operations, to match the available instructions. However, plain multiply
238 // (low) operations are left as Legal, as there are single-result
239 // instructions for this in x86. Using the two-result multiply instructions
240 // when both high and low results are needed must be arranged by dagcombine.
241 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
242 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
243 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
244 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
245 setOperationAction(ISD::SREM , MVT::i8 , Expand);
246 setOperationAction(ISD::UREM , MVT::i8 , Expand);
247 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
248 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
249 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
250 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
251 setOperationAction(ISD::SREM , MVT::i16 , Expand);
252 setOperationAction(ISD::UREM , MVT::i16 , Expand);
253 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
254 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
255 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
256 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
257 setOperationAction(ISD::SREM , MVT::i32 , Expand);
258 setOperationAction(ISD::UREM , MVT::i32 , Expand);
259 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
260 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
261 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
262 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
263 setOperationAction(ISD::SREM , MVT::i64 , Expand);
264 setOperationAction(ISD::UREM , MVT::i64 , Expand);
266 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
267 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
268 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
269 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
270 if (Subtarget->is64Bit())
271 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
272 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
273 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
274 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
275 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
276 setOperationAction(ISD::FREM , MVT::f32 , Expand);
277 setOperationAction(ISD::FREM , MVT::f64 , Expand);
278 setOperationAction(ISD::FREM , MVT::f80 , Expand);
279 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
281 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
282 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
283 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
284 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
286 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
287 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
289 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
290 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
292 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
293 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
294 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
295 if (Subtarget->is64Bit()) {
296 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
297 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
298 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
301 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
302 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
304 // These should be promoted to a larger select which is supported.
305 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
306 // X86 wants to expand cmov itself.
307 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
309 setOperationAction(ISD::SELECT , MVT::i16 , Expand);
311 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
312 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
313 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
314 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
315 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
316 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
318 setOperationAction(ISD::SETCC , MVT::i16 , Expand);
320 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
321 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
322 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
323 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
324 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
325 if (Subtarget->is64Bit()) {
326 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
327 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
329 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
332 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
333 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
334 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
335 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
336 if (Subtarget->is64Bit())
337 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
338 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
339 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
340 if (Subtarget->is64Bit()) {
341 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
342 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
343 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
344 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
345 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
347 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
348 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
349 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
350 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
351 if (Subtarget->is64Bit()) {
352 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
353 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
354 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
357 if (Subtarget->hasSSE1())
358 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
360 if (!Subtarget->hasSSE2())
361 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
363 // Expand certain atomics
364 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
365 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
366 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
367 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
369 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
370 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
371 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
372 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
374 if (!Subtarget->is64Bit()) {
375 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
376 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
377 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
378 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
379 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
380 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
381 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
384 // FIXME - use subtarget debug flags
385 if (!Subtarget->isTargetDarwin() &&
386 !Subtarget->isTargetELF() &&
387 !Subtarget->isTargetCygMing()) {
388 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
391 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
392 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
393 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
394 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
395 if (Subtarget->is64Bit()) {
396 setExceptionPointerRegister(X86::RAX);
397 setExceptionSelectorRegister(X86::RDX);
399 setExceptionPointerRegister(X86::EAX);
400 setExceptionSelectorRegister(X86::EDX);
402 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
403 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
405 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
407 setOperationAction(ISD::TRAP, MVT::Other, Legal);
409 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
410 setOperationAction(ISD::VASTART , MVT::Other, Custom);
411 setOperationAction(ISD::VAEND , MVT::Other, Expand);
412 if (Subtarget->is64Bit()) {
413 setOperationAction(ISD::VAARG , MVT::Other, Custom);
414 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
416 setOperationAction(ISD::VAARG , MVT::Other, Expand);
417 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
420 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
421 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
422 if (Subtarget->is64Bit())
423 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
424 if (Subtarget->isTargetCygMing())
425 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
427 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
429 if (!UseSoftFloat && X86ScalarSSEf64) {
430 // f32 and f64 use SSE.
431 // Set up the FP register classes.
432 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
433 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
435 // Use ANDPD to simulate FABS.
436 setOperationAction(ISD::FABS , MVT::f64, Custom);
437 setOperationAction(ISD::FABS , MVT::f32, Custom);
439 // Use XORP to simulate FNEG.
440 setOperationAction(ISD::FNEG , MVT::f64, Custom);
441 setOperationAction(ISD::FNEG , MVT::f32, Custom);
443 // Use ANDPD and ORPD to simulate FCOPYSIGN.
444 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
445 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
447 // We don't support sin/cos/fmod
448 setOperationAction(ISD::FSIN , MVT::f64, Expand);
449 setOperationAction(ISD::FCOS , MVT::f64, Expand);
450 setOperationAction(ISD::FSIN , MVT::f32, Expand);
451 setOperationAction(ISD::FCOS , MVT::f32, Expand);
453 // Expand FP immediates into loads from the stack, except for the special
455 addLegalFPImmediate(APFloat(+0.0)); // xorpd
456 addLegalFPImmediate(APFloat(+0.0f)); // xorps
457 } else if (!UseSoftFloat && X86ScalarSSEf32) {
458 // Use SSE for f32, x87 for f64.
459 // Set up the FP register classes.
460 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
461 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
463 // Use ANDPS to simulate FABS.
464 setOperationAction(ISD::FABS , MVT::f32, Custom);
466 // Use XORP to simulate FNEG.
467 setOperationAction(ISD::FNEG , MVT::f32, Custom);
469 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
471 // Use ANDPS and ORPS to simulate FCOPYSIGN.
472 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
473 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
475 // We don't support sin/cos/fmod
476 setOperationAction(ISD::FSIN , MVT::f32, Expand);
477 setOperationAction(ISD::FCOS , MVT::f32, Expand);
479 // Special cases we handle for FP constants.
480 addLegalFPImmediate(APFloat(+0.0f)); // xorps
481 addLegalFPImmediate(APFloat(+0.0)); // FLD0
482 addLegalFPImmediate(APFloat(+1.0)); // FLD1
483 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
484 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
487 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
488 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
490 } else if (!UseSoftFloat) {
491 // f32 and f64 in x87.
492 // Set up the FP register classes.
493 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
494 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
496 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
497 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
498 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
499 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
502 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
503 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
505 addLegalFPImmediate(APFloat(+0.0)); // FLD0
506 addLegalFPImmediate(APFloat(+1.0)); // FLD1
507 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
508 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
509 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
510 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
511 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
512 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
515 // Long double always uses X87.
517 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
518 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
519 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
522 APFloat TmpFlt(+0.0);
523 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
525 addLegalFPImmediate(TmpFlt); // FLD0
527 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
528 APFloat TmpFlt2(+1.0);
529 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
531 addLegalFPImmediate(TmpFlt2); // FLD1
532 TmpFlt2.changeSign();
533 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
537 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
538 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
542 // Always use a library call for pow.
543 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
544 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
545 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
547 setOperationAction(ISD::FLOG, MVT::f80, Expand);
548 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
549 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
550 setOperationAction(ISD::FEXP, MVT::f80, Expand);
551 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
553 // First set operation action for all vector types to either promote
554 // (for widening) or expand (for scalarization). Then we will selectively
555 // turn on ones that can be effectively codegen'd.
556 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
557 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
558 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
573 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
574 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
599 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
601 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
603 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
604 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
605 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
606 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
607 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
608 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
609 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
610 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
611 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
612 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
613 setTruncStoreAction((MVT::SimpleValueType)VT,
614 (MVT::SimpleValueType)InnerVT, Expand);
615 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
616 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
617 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
620 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
621 // with -msoft-float, disable use of MMX as well.
622 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
623 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
624 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
625 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
626 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
627 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
629 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
630 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
631 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
632 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
634 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
635 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
636 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
637 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
639 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
640 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
642 setOperationAction(ISD::AND, MVT::v8i8, Promote);
643 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
644 setOperationAction(ISD::AND, MVT::v4i16, Promote);
645 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
646 setOperationAction(ISD::AND, MVT::v2i32, Promote);
647 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
648 setOperationAction(ISD::AND, MVT::v1i64, Legal);
650 setOperationAction(ISD::OR, MVT::v8i8, Promote);
651 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
652 setOperationAction(ISD::OR, MVT::v4i16, Promote);
653 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
654 setOperationAction(ISD::OR, MVT::v2i32, Promote);
655 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
656 setOperationAction(ISD::OR, MVT::v1i64, Legal);
658 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
659 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
660 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
661 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
662 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
663 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
664 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
666 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
667 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
668 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
669 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
670 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
671 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
672 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
673 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
674 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
676 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
677 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
678 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
679 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
680 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
682 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
683 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
684 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
685 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
687 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
688 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
689 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
690 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
692 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
694 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
695 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
696 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
697 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
698 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
699 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
700 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
703 if (!UseSoftFloat && Subtarget->hasSSE1()) {
704 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
706 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
707 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
708 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
709 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
710 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
711 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
712 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
713 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
714 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
715 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
716 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
717 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
720 if (!UseSoftFloat && Subtarget->hasSSE2()) {
721 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
723 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
724 // registers cannot be used even for integer operations.
725 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
726 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
727 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
728 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
730 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
731 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
732 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
733 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
734 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
735 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
736 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
737 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
738 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
739 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
740 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
741 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
742 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
743 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
744 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
745 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
747 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
748 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
749 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
750 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
752 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
753 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
754 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
755 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
756 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
758 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
759 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
760 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
761 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
762 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
764 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
765 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
766 EVT VT = (MVT::SimpleValueType)i;
767 // Do not attempt to custom lower non-power-of-2 vectors
768 if (!isPowerOf2_32(VT.getVectorNumElements()))
770 // Do not attempt to custom lower non-128-bit vectors
771 if (!VT.is128BitVector())
773 setOperationAction(ISD::BUILD_VECTOR,
774 VT.getSimpleVT().SimpleTy, Custom);
775 setOperationAction(ISD::VECTOR_SHUFFLE,
776 VT.getSimpleVT().SimpleTy, Custom);
777 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
778 VT.getSimpleVT().SimpleTy, Custom);
781 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
782 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
783 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
784 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
785 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
786 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
788 if (Subtarget->is64Bit()) {
789 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
790 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
793 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
794 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
795 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
798 // Do not attempt to promote non-128-bit vectors
799 if (!VT.is128BitVector()) {
802 setOperationAction(ISD::AND, SVT, Promote);
803 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
804 setOperationAction(ISD::OR, SVT, Promote);
805 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
806 setOperationAction(ISD::XOR, SVT, Promote);
807 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
808 setOperationAction(ISD::LOAD, SVT, Promote);
809 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
810 setOperationAction(ISD::SELECT, SVT, Promote);
811 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
814 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
816 // Custom lower v2i64 and v2f64 selects.
817 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
818 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
819 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
820 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
822 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
823 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
824 if (!DisableMMX && Subtarget->hasMMX()) {
825 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
826 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
830 if (Subtarget->hasSSE41()) {
831 // FIXME: Do we need to handle scalar-to-vector here?
832 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
834 // i8 and i16 vectors are custom , because the source register and source
835 // source memory operand types are not the same width. f32 vectors are
836 // custom since the immediate controlling the insert encodes additional
838 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
839 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
840 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
841 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
843 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
844 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
845 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
846 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
848 if (Subtarget->is64Bit()) {
849 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
850 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
854 if (Subtarget->hasSSE42()) {
855 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
858 if (!UseSoftFloat && Subtarget->hasAVX()) {
859 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
860 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
861 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
862 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
864 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
865 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
866 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
867 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
868 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
869 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
870 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
871 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
872 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
873 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
874 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
875 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
876 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
877 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
878 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
880 // Operations to consider commented out -v16i16 v32i8
881 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
882 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
883 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
884 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
885 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
886 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
887 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
888 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
889 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
890 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
891 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
892 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
893 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
894 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
896 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
897 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
898 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
899 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
901 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
902 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
903 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
904 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
905 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
907 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
908 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
909 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
910 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
911 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
912 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
915 // Not sure we want to do this since there are no 256-bit integer
918 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
919 // This includes 256-bit vectors
920 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
921 EVT VT = (MVT::SimpleValueType)i;
923 // Do not attempt to custom lower non-power-of-2 vectors
924 if (!isPowerOf2_32(VT.getVectorNumElements()))
927 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
928 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
929 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
932 if (Subtarget->is64Bit()) {
933 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
934 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
939 // Not sure we want to do this since there are no 256-bit integer
942 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
943 // Including 256-bit vectors
944 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
945 EVT VT = (MVT::SimpleValueType)i;
947 if (!VT.is256BitVector()) {
950 setOperationAction(ISD::AND, VT, Promote);
951 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
952 setOperationAction(ISD::OR, VT, Promote);
953 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
954 setOperationAction(ISD::XOR, VT, Promote);
955 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
956 setOperationAction(ISD::LOAD, VT, Promote);
957 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
958 setOperationAction(ISD::SELECT, VT, Promote);
959 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
962 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
966 // We want to custom lower some of our intrinsics.
967 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
969 // Add/Sub/Mul with overflow operations are custom lowered.
970 setOperationAction(ISD::SADDO, MVT::i32, Custom);
971 setOperationAction(ISD::SADDO, MVT::i64, Custom);
972 setOperationAction(ISD::UADDO, MVT::i32, Custom);
973 setOperationAction(ISD::UADDO, MVT::i64, Custom);
974 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
975 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
976 setOperationAction(ISD::USUBO, MVT::i32, Custom);
977 setOperationAction(ISD::USUBO, MVT::i64, Custom);
978 setOperationAction(ISD::SMULO, MVT::i32, Custom);
979 setOperationAction(ISD::SMULO, MVT::i64, Custom);
981 if (!Subtarget->is64Bit()) {
982 // These libcalls are not available in 32-bit.
983 setLibcallName(RTLIB::SHL_I128, 0);
984 setLibcallName(RTLIB::SRL_I128, 0);
985 setLibcallName(RTLIB::SRA_I128, 0);
988 // We have target-specific dag combine patterns for the following nodes:
989 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
990 setTargetDAGCombine(ISD::BUILD_VECTOR);
991 setTargetDAGCombine(ISD::SELECT);
992 setTargetDAGCombine(ISD::SHL);
993 setTargetDAGCombine(ISD::SRA);
994 setTargetDAGCombine(ISD::SRL);
995 setTargetDAGCombine(ISD::OR);
996 setTargetDAGCombine(ISD::STORE);
997 setTargetDAGCombine(ISD::MEMBARRIER);
998 setTargetDAGCombine(ISD::ZERO_EXTEND);
999 if (Subtarget->is64Bit())
1000 setTargetDAGCombine(ISD::MUL);
1002 computeRegisterProperties();
1004 // Divide and reminder operations have no vector equivalent and can
1005 // trap. Do a custom widening for these operations in which we never
1006 // generate more divides/remainder than the original vector width.
1007 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1008 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
1009 if (!isTypeLegal((MVT::SimpleValueType)VT)) {
1010 setOperationAction(ISD::SDIV, (MVT::SimpleValueType) VT, Custom);
1011 setOperationAction(ISD::UDIV, (MVT::SimpleValueType) VT, Custom);
1012 setOperationAction(ISD::SREM, (MVT::SimpleValueType) VT, Custom);
1013 setOperationAction(ISD::UREM, (MVT::SimpleValueType) VT, Custom);
1017 // FIXME: These should be based on subtarget info. Plus, the values should
1018 // be smaller when we are in optimizing for size mode.
1019 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1020 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
1021 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1022 setPrefLoopAlignment(16);
1023 benefitFromCodePlacementOpt = true;
1027 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1032 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1033 /// the desired ByVal argument alignment.
1034 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1037 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1038 if (VTy->getBitWidth() == 128)
1040 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1041 unsigned EltAlign = 0;
1042 getMaxByValAlign(ATy->getElementType(), EltAlign);
1043 if (EltAlign > MaxAlign)
1044 MaxAlign = EltAlign;
1045 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1046 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1047 unsigned EltAlign = 0;
1048 getMaxByValAlign(STy->getElementType(i), EltAlign);
1049 if (EltAlign > MaxAlign)
1050 MaxAlign = EltAlign;
1058 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1059 /// function arguments in the caller parameter area. For X86, aggregates
1060 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1061 /// are at 4-byte boundaries.
1062 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1063 if (Subtarget->is64Bit()) {
1064 // Max of 8 and alignment of type.
1065 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1072 if (Subtarget->hasSSE1())
1073 getMaxByValAlign(Ty, Align);
1077 /// getOptimalMemOpType - Returns the target specific optimal type for load
1078 /// and store operations as a result of memset, memcpy, and memmove
1079 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
1082 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1083 bool isSrcConst, bool isSrcStr,
1084 SelectionDAG &DAG) const {
1085 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1086 // linux. This is because the stack realignment code can't handle certain
1087 // cases like PR2962. This should be removed when PR2962 is fixed.
1088 const Function *F = DAG.getMachineFunction().getFunction();
1089 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1090 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1091 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1093 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1096 if (Subtarget->is64Bit() && Size >= 8)
1101 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1102 /// current function. The returned value is a member of the
1103 /// MachineJumpTableInfo::JTEntryKind enum.
1104 unsigned X86TargetLowering::getJumpTableEncoding() const {
1105 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1107 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1108 Subtarget->isPICStyleGOT())
1109 return MachineJumpTableInfo::EK_Custom32;
1111 // Otherwise, use the normal jump table encoding heuristics.
1112 return TargetLowering::getJumpTableEncoding();
1115 /// getPICBaseSymbol - Return the X86-32 PIC base.
1117 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1118 MCContext &Ctx) const {
1119 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1120 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1121 Twine(MF->getFunctionNumber())+"$pb");
1126 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1127 const MachineBasicBlock *MBB,
1128 unsigned uid,MCContext &Ctx) const{
1129 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1130 Subtarget->isPICStyleGOT());
1131 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1134 // FIXME: @GOTOFF should be a property of MCSymbolRefExpr not in the MCSymbol.
1135 std::string Name = MBB->getSymbol(Ctx)->getName() + "@GOTOFF";
1136 return MCSymbolRefExpr::Create(Ctx.GetOrCreateSymbol(StringRef(Name)), Ctx);
1139 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1141 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1142 SelectionDAG &DAG) const {
1143 if (!Subtarget->is64Bit())
1144 // This doesn't have DebugLoc associated with it, but is not really the
1145 // same as a Register.
1146 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1151 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1152 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1154 const MCExpr *X86TargetLowering::
1155 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1156 MCContext &Ctx) const {
1157 // X86-64 uses RIP relative addressing based on the jump table label.
1158 if (Subtarget->isPICStyleRIPRel())
1159 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1161 // Otherwise, the reference is relative to the PIC base.
1162 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1165 /// getFunctionAlignment - Return the Log2 alignment of this function.
1166 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1167 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1170 //===----------------------------------------------------------------------===//
1171 // Return Value Calling Convention Implementation
1172 //===----------------------------------------------------------------------===//
1174 #include "X86GenCallingConv.inc"
1177 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1178 const SmallVectorImpl<EVT> &OutTys,
1179 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1180 SelectionDAG &DAG) {
1181 SmallVector<CCValAssign, 16> RVLocs;
1182 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1183 RVLocs, *DAG.getContext());
1184 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1188 X86TargetLowering::LowerReturn(SDValue Chain,
1189 CallingConv::ID CallConv, bool isVarArg,
1190 const SmallVectorImpl<ISD::OutputArg> &Outs,
1191 DebugLoc dl, SelectionDAG &DAG) {
1193 SmallVector<CCValAssign, 16> RVLocs;
1194 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1195 RVLocs, *DAG.getContext());
1196 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1198 // Add the regs to the liveout set for the function.
1199 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1200 for (unsigned i = 0; i != RVLocs.size(); ++i)
1201 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1202 MRI.addLiveOut(RVLocs[i].getLocReg());
1206 SmallVector<SDValue, 6> RetOps;
1207 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1208 // Operand #1 = Bytes To Pop
1209 RetOps.push_back(DAG.getTargetConstant(getBytesToPopOnReturn(), MVT::i16));
1211 // Copy the result values into the output registers.
1212 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1213 CCValAssign &VA = RVLocs[i];
1214 assert(VA.isRegLoc() && "Can only return in registers!");
1215 SDValue ValToCopy = Outs[i].Val;
1217 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1218 // the RET instruction and handled by the FP Stackifier.
1219 if (VA.getLocReg() == X86::ST0 ||
1220 VA.getLocReg() == X86::ST1) {
1221 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1222 // change the value to the FP stack register class.
1223 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1224 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1225 RetOps.push_back(ValToCopy);
1226 // Don't emit a copytoreg.
1230 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1231 // which is returned in RAX / RDX.
1232 if (Subtarget->is64Bit()) {
1233 EVT ValVT = ValToCopy.getValueType();
1234 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1235 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1236 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1237 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1241 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1242 Flag = Chain.getValue(1);
1245 // The x86-64 ABI for returning structs by value requires that we copy
1246 // the sret argument into %rax for the return. We saved the argument into
1247 // a virtual register in the entry block, so now we copy the value out
1249 if (Subtarget->is64Bit() &&
1250 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1251 MachineFunction &MF = DAG.getMachineFunction();
1252 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1253 unsigned Reg = FuncInfo->getSRetReturnReg();
1255 Reg = MRI.createVirtualRegister(getRegClassFor(MVT::i64));
1256 FuncInfo->setSRetReturnReg(Reg);
1258 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1260 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1261 Flag = Chain.getValue(1);
1263 // RAX now acts like a return value.
1264 MRI.addLiveOut(X86::RAX);
1267 RetOps[0] = Chain; // Update chain.
1269 // Add the flag if we have it.
1271 RetOps.push_back(Flag);
1273 return DAG.getNode(X86ISD::RET_FLAG, dl,
1274 MVT::Other, &RetOps[0], RetOps.size());
1277 /// LowerCallResult - Lower the result values of a call into the
1278 /// appropriate copies out of appropriate physical registers.
1281 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1282 CallingConv::ID CallConv, bool isVarArg,
1283 const SmallVectorImpl<ISD::InputArg> &Ins,
1284 DebugLoc dl, SelectionDAG &DAG,
1285 SmallVectorImpl<SDValue> &InVals) {
1287 // Assign locations to each value returned by this call.
1288 SmallVector<CCValAssign, 16> RVLocs;
1289 bool Is64Bit = Subtarget->is64Bit();
1290 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1291 RVLocs, *DAG.getContext());
1292 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1294 // Copy all of the result registers out of their specified physreg.
1295 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1296 CCValAssign &VA = RVLocs[i];
1297 EVT CopyVT = VA.getValVT();
1299 // If this is x86-64, and we disabled SSE, we can't return FP values
1300 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1301 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1302 llvm_report_error("SSE register return with SSE disabled");
1305 // If this is a call to a function that returns an fp value on the floating
1306 // point stack, but where we prefer to use the value in xmm registers, copy
1307 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1308 if ((VA.getLocReg() == X86::ST0 ||
1309 VA.getLocReg() == X86::ST1) &&
1310 isScalarFPTypeInSSEReg(VA.getValVT())) {
1315 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1316 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1317 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1318 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1319 MVT::v2i64, InFlag).getValue(1);
1320 Val = Chain.getValue(0);
1321 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1322 Val, DAG.getConstant(0, MVT::i64));
1324 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1325 MVT::i64, InFlag).getValue(1);
1326 Val = Chain.getValue(0);
1328 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1330 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1331 CopyVT, InFlag).getValue(1);
1332 Val = Chain.getValue(0);
1334 InFlag = Chain.getValue(2);
1336 if (CopyVT != VA.getValVT()) {
1337 // Round the F80 the right size, which also moves to the appropriate xmm
1339 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1340 // This truncation won't change the value.
1341 DAG.getIntPtrConstant(1));
1344 InVals.push_back(Val);
1351 //===----------------------------------------------------------------------===//
1352 // C & StdCall & Fast Calling Convention implementation
1353 //===----------------------------------------------------------------------===//
1354 // StdCall calling convention seems to be standard for many Windows' API
1355 // routines and around. It differs from C calling convention just a little:
1356 // callee should clean up the stack, not caller. Symbols should be also
1357 // decorated in some fancy way :) It doesn't support any vector arguments.
1358 // For info on fast calling convention see Fast Calling Convention (tail call)
1359 // implementation LowerX86_32FastCCCallTo.
1361 /// CallIsStructReturn - Determines whether a call uses struct return
1363 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1367 return Outs[0].Flags.isSRet();
1370 /// ArgsAreStructReturn - Determines whether a function uses struct
1371 /// return semantics.
1373 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1377 return Ins[0].Flags.isSRet();
1380 /// IsCalleePop - Determines whether the callee is required to pop its
1381 /// own arguments. Callee pop is necessary to support tail calls.
1382 bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
1386 switch (CallingConv) {
1389 case CallingConv::X86_StdCall:
1390 return !Subtarget->is64Bit();
1391 case CallingConv::X86_FastCall:
1392 return !Subtarget->is64Bit();
1393 case CallingConv::Fast:
1394 return PerformTailCallOpt;
1398 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1399 /// given CallingConvention value.
1400 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1401 if (Subtarget->is64Bit()) {
1402 if (Subtarget->isTargetWin64())
1403 return CC_X86_Win64_C;
1408 if (CC == CallingConv::X86_FastCall)
1409 return CC_X86_32_FastCall;
1410 else if (CC == CallingConv::Fast)
1411 return CC_X86_32_FastCC;
1416 /// NameDecorationForCallConv - Selects the appropriate decoration to
1417 /// apply to a MachineFunction containing a given calling convention.
1419 X86TargetLowering::NameDecorationForCallConv(CallingConv::ID CallConv) {
1420 if (CallConv == CallingConv::X86_FastCall)
1422 else if (CallConv == CallingConv::X86_StdCall)
1428 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1429 /// by "Src" to address "Dst" with size and alignment information specified by
1430 /// the specific parameter attribute. The copy will be passed as a byval
1431 /// function parameter.
1433 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1434 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1436 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1437 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1438 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1441 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1442 /// a tailcall target by changing its ABI.
1443 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1444 return PerformTailCallOpt && CC == CallingConv::Fast;
1448 X86TargetLowering::LowerMemArgument(SDValue Chain,
1449 CallingConv::ID CallConv,
1450 const SmallVectorImpl<ISD::InputArg> &Ins,
1451 DebugLoc dl, SelectionDAG &DAG,
1452 const CCValAssign &VA,
1453 MachineFrameInfo *MFI,
1455 // Create the nodes corresponding to a load from this parameter slot.
1456 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1457 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1458 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1461 // If value is passed by pointer we have address passed instead of the value
1463 if (VA.getLocInfo() == CCValAssign::Indirect)
1464 ValVT = VA.getLocVT();
1466 ValVT = VA.getValVT();
1468 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1469 // changed with more analysis.
1470 // In case of tail call optimization mark all arguments mutable. Since they
1471 // could be overwritten by lowering of arguments in case of a tail call.
1472 if (Flags.isByVal()) {
1473 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1474 VA.getLocMemOffset(), isImmutable, false);
1475 return DAG.getFrameIndex(FI, getPointerTy());
1477 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1478 VA.getLocMemOffset(), isImmutable, false);
1479 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1480 return DAG.getLoad(ValVT, dl, Chain, FIN,
1481 PseudoSourceValue::getFixedStack(FI), 0);
1486 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1487 CallingConv::ID CallConv,
1489 const SmallVectorImpl<ISD::InputArg> &Ins,
1492 SmallVectorImpl<SDValue> &InVals) {
1494 MachineFunction &MF = DAG.getMachineFunction();
1495 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1497 const Function* Fn = MF.getFunction();
1498 if (Fn->hasExternalLinkage() &&
1499 Subtarget->isTargetCygMing() &&
1500 Fn->getName() == "main")
1501 FuncInfo->setForceFramePointer(true);
1503 // Decorate the function name.
1504 FuncInfo->setDecorationStyle(NameDecorationForCallConv(CallConv));
1506 MachineFrameInfo *MFI = MF.getFrameInfo();
1507 bool Is64Bit = Subtarget->is64Bit();
1508 bool IsWin64 = Subtarget->isTargetWin64();
1510 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1511 "Var args not supported with calling convention fastcc");
1513 // Assign locations to all of the incoming arguments.
1514 SmallVector<CCValAssign, 16> ArgLocs;
1515 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1516 ArgLocs, *DAG.getContext());
1517 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1519 unsigned LastVal = ~0U;
1521 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1522 CCValAssign &VA = ArgLocs[i];
1523 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1525 assert(VA.getValNo() != LastVal &&
1526 "Don't support value assigned to multiple locs yet");
1527 LastVal = VA.getValNo();
1529 if (VA.isRegLoc()) {
1530 EVT RegVT = VA.getLocVT();
1531 TargetRegisterClass *RC = NULL;
1532 if (RegVT == MVT::i32)
1533 RC = X86::GR32RegisterClass;
1534 else if (Is64Bit && RegVT == MVT::i64)
1535 RC = X86::GR64RegisterClass;
1536 else if (RegVT == MVT::f32)
1537 RC = X86::FR32RegisterClass;
1538 else if (RegVT == MVT::f64)
1539 RC = X86::FR64RegisterClass;
1540 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1541 RC = X86::VR128RegisterClass;
1542 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1543 RC = X86::VR64RegisterClass;
1545 llvm_unreachable("Unknown argument type!");
1547 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1548 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1550 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1551 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1553 if (VA.getLocInfo() == CCValAssign::SExt)
1554 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1555 DAG.getValueType(VA.getValVT()));
1556 else if (VA.getLocInfo() == CCValAssign::ZExt)
1557 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1558 DAG.getValueType(VA.getValVT()));
1559 else if (VA.getLocInfo() == CCValAssign::BCvt)
1560 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1562 if (VA.isExtInLoc()) {
1563 // Handle MMX values passed in XMM regs.
1564 if (RegVT.isVector()) {
1565 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1566 ArgValue, DAG.getConstant(0, MVT::i64));
1567 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1569 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1572 assert(VA.isMemLoc());
1573 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1576 // If value is passed via pointer - do a load.
1577 if (VA.getLocInfo() == CCValAssign::Indirect)
1578 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0);
1580 InVals.push_back(ArgValue);
1583 // The x86-64 ABI for returning structs by value requires that we copy
1584 // the sret argument into %rax for the return. Save the argument into
1585 // a virtual register so that we can access it from the return points.
1586 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1587 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1588 unsigned Reg = FuncInfo->getSRetReturnReg();
1590 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1591 FuncInfo->setSRetReturnReg(Reg);
1593 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1594 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1597 unsigned StackSize = CCInfo.getNextStackOffset();
1598 // Align stack specially for tail calls.
1599 if (FuncIsMadeTailCallSafe(CallConv))
1600 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1602 // If the function takes variable number of arguments, make a frame index for
1603 // the start of the first vararg value... for expansion of llvm.va_start.
1605 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1606 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize, true, false);
1609 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1611 // FIXME: We should really autogenerate these arrays
1612 static const unsigned GPR64ArgRegsWin64[] = {
1613 X86::RCX, X86::RDX, X86::R8, X86::R9
1615 static const unsigned XMMArgRegsWin64[] = {
1616 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1618 static const unsigned GPR64ArgRegs64Bit[] = {
1619 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1621 static const unsigned XMMArgRegs64Bit[] = {
1622 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1623 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1625 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1628 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1629 GPR64ArgRegs = GPR64ArgRegsWin64;
1630 XMMArgRegs = XMMArgRegsWin64;
1632 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1633 GPR64ArgRegs = GPR64ArgRegs64Bit;
1634 XMMArgRegs = XMMArgRegs64Bit;
1636 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1638 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1641 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1642 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1643 "SSE register cannot be used when SSE is disabled!");
1644 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1645 "SSE register cannot be used when SSE is disabled!");
1646 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1647 // Kernel mode asks for SSE to be disabled, so don't push them
1649 TotalNumXMMRegs = 0;
1651 // For X86-64, if there are vararg parameters that are passed via
1652 // registers, then we must store them to their spots on the stack so they
1653 // may be loaded by deferencing the result of va_next.
1654 VarArgsGPOffset = NumIntRegs * 8;
1655 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1656 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1657 TotalNumXMMRegs * 16, 16,
1660 // Store the integer parameter registers.
1661 SmallVector<SDValue, 8> MemOps;
1662 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1663 unsigned Offset = VarArgsGPOffset;
1664 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1665 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1666 DAG.getIntPtrConstant(Offset));
1667 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1668 X86::GR64RegisterClass);
1669 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1671 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1672 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1674 MemOps.push_back(Store);
1678 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1679 // Now store the XMM (fp + vector) parameter registers.
1680 SmallVector<SDValue, 11> SaveXMMOps;
1681 SaveXMMOps.push_back(Chain);
1683 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1684 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1685 SaveXMMOps.push_back(ALVal);
1687 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1688 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1690 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1691 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1692 X86::VR128RegisterClass);
1693 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1694 SaveXMMOps.push_back(Val);
1696 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1698 &SaveXMMOps[0], SaveXMMOps.size()));
1701 if (!MemOps.empty())
1702 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1703 &MemOps[0], MemOps.size());
1707 // Some CCs need callee pop.
1708 if (IsCalleePop(isVarArg, CallConv)) {
1709 BytesToPopOnReturn = StackSize; // Callee pops everything.
1711 BytesToPopOnReturn = 0; // Callee pops nothing.
1712 // If this is an sret function, the return should pop the hidden pointer.
1713 if (!Is64Bit && CallConv != CallingConv::Fast && ArgsAreStructReturn(Ins))
1714 BytesToPopOnReturn = 4;
1718 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1719 if (CallConv == CallingConv::X86_FastCall)
1720 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1723 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1729 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1730 SDValue StackPtr, SDValue Arg,
1731 DebugLoc dl, SelectionDAG &DAG,
1732 const CCValAssign &VA,
1733 ISD::ArgFlagsTy Flags) {
1734 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1735 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1736 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1737 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1738 if (Flags.isByVal()) {
1739 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1741 return DAG.getStore(Chain, dl, Arg, PtrOff,
1742 PseudoSourceValue::getStack(), LocMemOffset);
1745 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1746 /// optimization is performed and it is required.
1748 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1749 SDValue &OutRetAddr, SDValue Chain,
1750 bool IsTailCall, bool Is64Bit,
1751 int FPDiff, DebugLoc dl) {
1752 // Adjust the Return address stack slot.
1753 EVT VT = getPointerTy();
1754 OutRetAddr = getReturnAddressFrameIndex(DAG);
1756 // Load the "old" Return address.
1757 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1758 return SDValue(OutRetAddr.getNode(), 1);
1761 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1762 /// optimization is performed and it is required (FPDiff!=0).
1764 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1765 SDValue Chain, SDValue RetAddrFrIdx,
1766 bool Is64Bit, int FPDiff, DebugLoc dl) {
1767 // Store the return address to the appropriate stack slot.
1768 if (!FPDiff) return Chain;
1769 // Calculate the new stack slot for the return address.
1770 int SlotSize = Is64Bit ? 8 : 4;
1771 int NewReturnAddrFI =
1772 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, true,false);
1773 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1774 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1775 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1776 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1781 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1782 CallingConv::ID CallConv, bool isVarArg,
1784 const SmallVectorImpl<ISD::OutputArg> &Outs,
1785 const SmallVectorImpl<ISD::InputArg> &Ins,
1786 DebugLoc dl, SelectionDAG &DAG,
1787 SmallVectorImpl<SDValue> &InVals) {
1788 MachineFunction &MF = DAG.getMachineFunction();
1789 bool Is64Bit = Subtarget->is64Bit();
1790 bool IsStructRet = CallIsStructReturn(Outs);
1791 bool IsSibcall = false;
1794 // Check if it's really possible to do a tail call.
1795 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
1798 // Sibcalls are automatically detected tailcalls which do not require
1800 if (!PerformTailCallOpt && isTailCall)
1807 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1808 "Var args not supported with calling convention fastcc");
1810 // Analyze operands of the call, assigning locations to each operand.
1811 SmallVector<CCValAssign, 16> ArgLocs;
1812 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1813 ArgLocs, *DAG.getContext());
1814 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1816 // Get a count of how many bytes are to be pushed on the stack.
1817 unsigned NumBytes = CCInfo.getNextStackOffset();
1819 // This is a sibcall. The memory operands are available in caller's
1820 // own caller's stack.
1822 else if (PerformTailCallOpt && CallConv == CallingConv::Fast)
1823 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1826 if (isTailCall && !IsSibcall) {
1827 // Lower arguments at fp - stackoffset + fpdiff.
1828 unsigned NumBytesCallerPushed =
1829 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1830 FPDiff = NumBytesCallerPushed - NumBytes;
1832 // Set the delta of movement of the returnaddr stackslot.
1833 // But only set if delta is greater than previous delta.
1834 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1835 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1839 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1841 SDValue RetAddrFrIdx;
1842 // Load return adress for tail calls.
1843 if (isTailCall && FPDiff)
1844 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1845 Is64Bit, FPDiff, dl);
1847 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1848 SmallVector<SDValue, 8> MemOpChains;
1851 // Walk the register/memloc assignments, inserting copies/loads. In the case
1852 // of tail call optimization arguments are handle later.
1853 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1854 CCValAssign &VA = ArgLocs[i];
1855 EVT RegVT = VA.getLocVT();
1856 SDValue Arg = Outs[i].Val;
1857 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1858 bool isByVal = Flags.isByVal();
1860 // Promote the value if needed.
1861 switch (VA.getLocInfo()) {
1862 default: llvm_unreachable("Unknown loc info!");
1863 case CCValAssign::Full: break;
1864 case CCValAssign::SExt:
1865 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1867 case CCValAssign::ZExt:
1868 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1870 case CCValAssign::AExt:
1871 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1872 // Special case: passing MMX values in XMM registers.
1873 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1874 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1875 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1877 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1879 case CCValAssign::BCvt:
1880 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1882 case CCValAssign::Indirect: {
1883 // Store the argument.
1884 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1885 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1886 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1887 PseudoSourceValue::getFixedStack(FI), 0);
1893 if (VA.isRegLoc()) {
1894 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1895 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1896 assert(VA.isMemLoc());
1897 if (StackPtr.getNode() == 0)
1898 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1899 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1900 dl, DAG, VA, Flags));
1904 if (!MemOpChains.empty())
1905 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1906 &MemOpChains[0], MemOpChains.size());
1908 // Build a sequence of copy-to-reg nodes chained together with token chain
1909 // and flag operands which copy the outgoing args into registers.
1911 // Tail call byval lowering might overwrite argument registers so in case of
1912 // tail call optimization the copies to registers are lowered later.
1914 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1915 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1916 RegsToPass[i].second, InFlag);
1917 InFlag = Chain.getValue(1);
1920 if (Subtarget->isPICStyleGOT()) {
1921 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1924 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1925 DAG.getNode(X86ISD::GlobalBaseReg,
1926 DebugLoc::getUnknownLoc(),
1929 InFlag = Chain.getValue(1);
1931 // If we are tail calling and generating PIC/GOT style code load the
1932 // address of the callee into ECX. The value in ecx is used as target of
1933 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1934 // for tail calls on PIC/GOT architectures. Normally we would just put the
1935 // address of GOT into ebx and then call target@PLT. But for tail calls
1936 // ebx would be restored (since ebx is callee saved) before jumping to the
1939 // Note: The actual moving to ECX is done further down.
1940 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1941 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1942 !G->getGlobal()->hasProtectedVisibility())
1943 Callee = LowerGlobalAddress(Callee, DAG);
1944 else if (isa<ExternalSymbolSDNode>(Callee))
1945 Callee = LowerExternalSymbol(Callee, DAG);
1949 if (Is64Bit && isVarArg) {
1950 // From AMD64 ABI document:
1951 // For calls that may call functions that use varargs or stdargs
1952 // (prototype-less calls or calls to functions containing ellipsis (...) in
1953 // the declaration) %al is used as hidden argument to specify the number
1954 // of SSE registers used. The contents of %al do not need to match exactly
1955 // the number of registers, but must be an ubound on the number of SSE
1956 // registers used and is in the range 0 - 8 inclusive.
1958 // FIXME: Verify this on Win64
1959 // Count the number of XMM registers allocated.
1960 static const unsigned XMMArgRegs[] = {
1961 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1962 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1964 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1965 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1966 && "SSE registers cannot be used when SSE is disabled");
1968 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1969 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1970 InFlag = Chain.getValue(1);
1974 // For tail calls lower the arguments to the 'real' stack slot.
1976 // Force all the incoming stack arguments to be loaded from the stack
1977 // before any new outgoing arguments are stored to the stack, because the
1978 // outgoing stack slots may alias the incoming argument stack slots, and
1979 // the alias isn't otherwise explicit. This is slightly more conservative
1980 // than necessary, because it means that each store effectively depends
1981 // on every argument instead of just those arguments it would clobber.
1982 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1984 SmallVector<SDValue, 8> MemOpChains2;
1987 // Do not flag preceeding copytoreg stuff together with the following stuff.
1989 if (PerformTailCallOpt) {
1990 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1991 CCValAssign &VA = ArgLocs[i];
1994 assert(VA.isMemLoc());
1995 SDValue Arg = Outs[i].Val;
1996 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1997 // Create frame index.
1998 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1999 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2000 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
2001 FIN = DAG.getFrameIndex(FI, getPointerTy());
2003 if (Flags.isByVal()) {
2004 // Copy relative to framepointer.
2005 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2006 if (StackPtr.getNode() == 0)
2007 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2009 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2011 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2015 // Store relative to framepointer.
2016 MemOpChains2.push_back(
2017 DAG.getStore(ArgChain, dl, Arg, FIN,
2018 PseudoSourceValue::getFixedStack(FI), 0));
2023 if (!MemOpChains2.empty())
2024 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2025 &MemOpChains2[0], MemOpChains2.size());
2027 // Copy arguments to their registers.
2028 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2029 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2030 RegsToPass[i].second, InFlag);
2031 InFlag = Chain.getValue(1);
2035 // Store the return address to the appropriate stack slot.
2036 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2040 bool WasGlobalOrExternal = false;
2041 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2042 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2043 // In the 64-bit large code model, we have to make all calls
2044 // through a register, since the call instruction's 32-bit
2045 // pc-relative offset may not be large enough to hold the whole
2047 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2048 WasGlobalOrExternal = true;
2049 // If the callee is a GlobalAddress node (quite common, every direct call
2050 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2053 // We should use extra load for direct calls to dllimported functions in
2055 GlobalValue *GV = G->getGlobal();
2056 if (!GV->hasDLLImportLinkage()) {
2057 unsigned char OpFlags = 0;
2059 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2060 // external symbols most go through the PLT in PIC mode. If the symbol
2061 // has hidden or protected visibility, or if it is static or local, then
2062 // we don't need to use the PLT - we can directly call it.
2063 if (Subtarget->isTargetELF() &&
2064 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2065 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2066 OpFlags = X86II::MO_PLT;
2067 } else if (Subtarget->isPICStyleStubAny() &&
2068 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2069 Subtarget->getDarwinVers() < 9) {
2070 // PC-relative references to external symbols should go through $stub,
2071 // unless we're building with the leopard linker or later, which
2072 // automatically synthesizes these stubs.
2073 OpFlags = X86II::MO_DARWIN_STUB;
2076 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2077 G->getOffset(), OpFlags);
2079 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2080 WasGlobalOrExternal = true;
2081 unsigned char OpFlags = 0;
2083 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2084 // symbols should go through the PLT.
2085 if (Subtarget->isTargetELF() &&
2086 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2087 OpFlags = X86II::MO_PLT;
2088 } else if (Subtarget->isPICStyleStubAny() &&
2089 Subtarget->getDarwinVers() < 9) {
2090 // PC-relative references to external symbols should go through $stub,
2091 // unless we're building with the leopard linker or later, which
2092 // automatically synthesizes these stubs.
2093 OpFlags = X86II::MO_DARWIN_STUB;
2096 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2100 if (isTailCall && !WasGlobalOrExternal) {
2101 // Force the address into a (call preserved) caller-saved register since
2102 // tailcall must happen after callee-saved registers are poped.
2103 // FIXME: Give it a special register class that contains caller-saved
2104 // register instead?
2105 unsigned TCReg = Is64Bit ? X86::R11 : X86::EAX;
2106 Chain = DAG.getCopyToReg(Chain, dl,
2107 DAG.getRegister(TCReg, getPointerTy()),
2109 Callee = DAG.getRegister(TCReg, getPointerTy());
2112 // Returns a chain & a flag for retval copy to use.
2113 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2114 SmallVector<SDValue, 8> Ops;
2116 if (!IsSibcall && isTailCall) {
2117 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2118 DAG.getIntPtrConstant(0, true), InFlag);
2119 InFlag = Chain.getValue(1);
2122 Ops.push_back(Chain);
2123 Ops.push_back(Callee);
2126 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2128 // Add argument registers to the end of the list so that they are known live
2130 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2131 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2132 RegsToPass[i].second.getValueType()));
2134 // Add an implicit use GOT pointer in EBX.
2135 if (!isTailCall && Subtarget->isPICStyleGOT())
2136 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2138 // Add an implicit use of AL for x86 vararg functions.
2139 if (Is64Bit && isVarArg)
2140 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2142 if (InFlag.getNode())
2143 Ops.push_back(InFlag);
2146 // If this is the first return lowered for this function, add the regs
2147 // to the liveout set for the function.
2148 if (MF.getRegInfo().liveout_empty()) {
2149 SmallVector<CCValAssign, 16> RVLocs;
2150 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2152 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2153 for (unsigned i = 0; i != RVLocs.size(); ++i)
2154 if (RVLocs[i].isRegLoc())
2155 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2158 assert(((Callee.getOpcode() == ISD::Register &&
2159 (cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
2160 cast<RegisterSDNode>(Callee)->getReg() == X86::R11)) ||
2161 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2162 Callee.getOpcode() == ISD::TargetGlobalAddress) &&
2163 "Expecting a global address, external symbol, or scratch register");
2165 return DAG.getNode(X86ISD::TC_RETURN, dl,
2166 NodeTys, &Ops[0], Ops.size());
2169 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2170 InFlag = Chain.getValue(1);
2172 // Create the CALLSEQ_END node.
2173 unsigned NumBytesForCalleeToPush;
2174 if (IsCalleePop(isVarArg, CallConv))
2175 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2176 else if (!Is64Bit && CallConv != CallingConv::Fast && IsStructRet)
2177 // If this is is a call to a struct-return function, the callee
2178 // pops the hidden struct pointer, so we have to push it back.
2179 // This is common for Darwin/X86, Linux & Mingw32 targets.
2180 NumBytesForCalleeToPush = 4;
2182 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2184 // Returns a flag for retval copy to use.
2186 Chain = DAG.getCALLSEQ_END(Chain,
2187 DAG.getIntPtrConstant(NumBytes, true),
2188 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2191 InFlag = Chain.getValue(1);
2194 // Handle result values, copying them out of physregs into vregs that we
2196 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2197 Ins, dl, DAG, InVals);
2201 //===----------------------------------------------------------------------===//
2202 // Fast Calling Convention (tail call) implementation
2203 //===----------------------------------------------------------------------===//
2205 // Like std call, callee cleans arguments, convention except that ECX is
2206 // reserved for storing the tail called function address. Only 2 registers are
2207 // free for argument passing (inreg). Tail call optimization is performed
2209 // * tailcallopt is enabled
2210 // * caller/callee are fastcc
2211 // On X86_64 architecture with GOT-style position independent code only local
2212 // (within module) calls are supported at the moment.
2213 // To keep the stack aligned according to platform abi the function
2214 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2215 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2216 // If a tail called function callee has more arguments than the caller the
2217 // caller needs to make sure that there is room to move the RETADDR to. This is
2218 // achieved by reserving an area the size of the argument delta right after the
2219 // original REtADDR, but before the saved framepointer or the spilled registers
2220 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2232 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2233 /// for a 16 byte align requirement.
2234 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2235 SelectionDAG& DAG) {
2236 MachineFunction &MF = DAG.getMachineFunction();
2237 const TargetMachine &TM = MF.getTarget();
2238 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2239 unsigned StackAlignment = TFI.getStackAlignment();
2240 uint64_t AlignMask = StackAlignment - 1;
2241 int64_t Offset = StackSize;
2242 uint64_t SlotSize = TD->getPointerSize();
2243 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2244 // Number smaller than 12 so just add the difference.
2245 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2247 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2248 Offset = ((~AlignMask) & Offset) + StackAlignment +
2249 (StackAlignment-SlotSize);
2254 /// MatchingStackOffset - Return true if the given stack call argument is
2255 /// already available in the same position (relatively) of the caller's
2256 /// incoming argument stack.
2258 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2259 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2260 const X86InstrInfo *TII) {
2262 if (Arg.getOpcode() == ISD::CopyFromReg) {
2263 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2264 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2266 MachineInstr *Def = MRI->getVRegDef(VR);
2269 if (!Flags.isByVal()) {
2270 if (!TII->isLoadFromStackSlot(Def, FI))
2273 unsigned Opcode = Def->getOpcode();
2274 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2275 Def->getOperand(1).isFI()) {
2276 FI = Def->getOperand(1).getIndex();
2277 if (MFI->getObjectSize(FI) != Flags.getByValSize())
2283 LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg);
2286 SDValue Ptr = Ld->getBasePtr();
2287 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2290 FI = FINode->getIndex();
2293 if (!MFI->isFixedObjectIndex(FI))
2295 return Offset == MFI->getObjectOffset(FI);
2298 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2299 /// for tail call optimization. Targets which want to do tail call
2300 /// optimization should implement this function.
2302 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2303 CallingConv::ID CalleeCC,
2305 const SmallVectorImpl<ISD::OutputArg> &Outs,
2306 const SmallVectorImpl<ISD::InputArg> &Ins,
2307 SelectionDAG& DAG) const {
2308 if (CalleeCC != CallingConv::Fast &&
2309 CalleeCC != CallingConv::C)
2312 // If -tailcallopt is specified, make fastcc functions tail-callable.
2313 const Function *CallerF = DAG.getMachineFunction().getFunction();
2314 if (PerformTailCallOpt) {
2315 if (CalleeCC == CallingConv::Fast &&
2316 CallerF->getCallingConv() == CalleeCC)
2321 // Look for obvious safe cases to perform tail call optimization that does not
2322 // requite ABI changes. This is what gcc calls sibcall.
2324 // Do not tail call optimize vararg calls for now.
2328 // If the callee takes no arguments then go on to check the results of the
2330 if (!Outs.empty()) {
2331 // Check if stack adjustment is needed. For now, do not do this if any
2332 // argument is passed on the stack.
2333 SmallVector<CCValAssign, 16> ArgLocs;
2334 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2335 ArgLocs, *DAG.getContext());
2336 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2337 if (CCInfo.getNextStackOffset()) {
2338 MachineFunction &MF = DAG.getMachineFunction();
2339 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2341 if (Subtarget->isTargetWin64())
2342 // Win64 ABI has additional complications.
2345 // Check if the arguments are already laid out in the right way as
2346 // the caller's fixed stack objects.
2347 MachineFrameInfo *MFI = MF.getFrameInfo();
2348 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2349 const X86InstrInfo *TII =
2350 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2351 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2352 CCValAssign &VA = ArgLocs[i];
2353 EVT RegVT = VA.getLocVT();
2354 SDValue Arg = Outs[i].Val;
2355 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2356 if (VA.getLocInfo() == CCValAssign::Indirect)
2358 if (!VA.isRegLoc()) {
2359 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2371 X86TargetLowering::createFastISel(MachineFunction &mf, MachineModuleInfo *mmo,
2373 DenseMap<const Value *, unsigned> &vm,
2374 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2375 DenseMap<const AllocaInst *, int> &am
2377 , SmallSet<Instruction*, 8> &cil
2380 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2388 //===----------------------------------------------------------------------===//
2389 // Other Lowering Hooks
2390 //===----------------------------------------------------------------------===//
2393 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2394 MachineFunction &MF = DAG.getMachineFunction();
2395 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2396 int ReturnAddrIndex = FuncInfo->getRAIndex();
2398 if (ReturnAddrIndex == 0) {
2399 // Set up a frame object for the return address.
2400 uint64_t SlotSize = TD->getPointerSize();
2401 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2403 FuncInfo->setRAIndex(ReturnAddrIndex);
2406 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2410 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2411 bool hasSymbolicDisplacement) {
2412 // Offset should fit into 32 bit immediate field.
2413 if (!isInt32(Offset))
2416 // If we don't have a symbolic displacement - we don't have any extra
2418 if (!hasSymbolicDisplacement)
2421 // FIXME: Some tweaks might be needed for medium code model.
2422 if (M != CodeModel::Small && M != CodeModel::Kernel)
2425 // For small code model we assume that latest object is 16MB before end of 31
2426 // bits boundary. We may also accept pretty large negative constants knowing
2427 // that all objects are in the positive half of address space.
2428 if (M == CodeModel::Small && Offset < 16*1024*1024)
2431 // For kernel code model we know that all object resist in the negative half
2432 // of 32bits address space. We may not accept negative offsets, since they may
2433 // be just off and we may accept pretty large positive ones.
2434 if (M == CodeModel::Kernel && Offset > 0)
2440 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2441 /// specific condition code, returning the condition code and the LHS/RHS of the
2442 /// comparison to make.
2443 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2444 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2446 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2447 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2448 // X > -1 -> X == 0, jump !sign.
2449 RHS = DAG.getConstant(0, RHS.getValueType());
2450 return X86::COND_NS;
2451 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2452 // X < 0 -> X == 0, jump on sign.
2454 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2456 RHS = DAG.getConstant(0, RHS.getValueType());
2457 return X86::COND_LE;
2461 switch (SetCCOpcode) {
2462 default: llvm_unreachable("Invalid integer condition!");
2463 case ISD::SETEQ: return X86::COND_E;
2464 case ISD::SETGT: return X86::COND_G;
2465 case ISD::SETGE: return X86::COND_GE;
2466 case ISD::SETLT: return X86::COND_L;
2467 case ISD::SETLE: return X86::COND_LE;
2468 case ISD::SETNE: return X86::COND_NE;
2469 case ISD::SETULT: return X86::COND_B;
2470 case ISD::SETUGT: return X86::COND_A;
2471 case ISD::SETULE: return X86::COND_BE;
2472 case ISD::SETUGE: return X86::COND_AE;
2476 // First determine if it is required or is profitable to flip the operands.
2478 // If LHS is a foldable load, but RHS is not, flip the condition.
2479 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2480 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2481 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2482 std::swap(LHS, RHS);
2485 switch (SetCCOpcode) {
2491 std::swap(LHS, RHS);
2495 // On a floating point condition, the flags are set as follows:
2497 // 0 | 0 | 0 | X > Y
2498 // 0 | 0 | 1 | X < Y
2499 // 1 | 0 | 0 | X == Y
2500 // 1 | 1 | 1 | unordered
2501 switch (SetCCOpcode) {
2502 default: llvm_unreachable("Condcode should be pre-legalized away");
2504 case ISD::SETEQ: return X86::COND_E;
2505 case ISD::SETOLT: // flipped
2507 case ISD::SETGT: return X86::COND_A;
2508 case ISD::SETOLE: // flipped
2510 case ISD::SETGE: return X86::COND_AE;
2511 case ISD::SETUGT: // flipped
2513 case ISD::SETLT: return X86::COND_B;
2514 case ISD::SETUGE: // flipped
2516 case ISD::SETLE: return X86::COND_BE;
2518 case ISD::SETNE: return X86::COND_NE;
2519 case ISD::SETUO: return X86::COND_P;
2520 case ISD::SETO: return X86::COND_NP;
2522 case ISD::SETUNE: return X86::COND_INVALID;
2526 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2527 /// code. Current x86 isa includes the following FP cmov instructions:
2528 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2529 static bool hasFPCMov(unsigned X86CC) {
2545 /// isFPImmLegal - Returns true if the target can instruction select the
2546 /// specified FP immediate natively. If false, the legalizer will
2547 /// materialize the FP immediate as a load from a constant pool.
2548 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2549 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2550 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2556 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2557 /// the specified range (L, H].
2558 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2559 return (Val < 0) || (Val >= Low && Val < Hi);
2562 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2563 /// specified value.
2564 static bool isUndefOrEqual(int Val, int CmpVal) {
2565 if (Val < 0 || Val == CmpVal)
2570 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2571 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2572 /// the second operand.
2573 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2574 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2575 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2576 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2577 return (Mask[0] < 2 && Mask[1] < 2);
2581 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2582 SmallVector<int, 8> M;
2584 return ::isPSHUFDMask(M, N->getValueType(0));
2587 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2588 /// is suitable for input to PSHUFHW.
2589 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2590 if (VT != MVT::v8i16)
2593 // Lower quadword copied in order or undef.
2594 for (int i = 0; i != 4; ++i)
2595 if (Mask[i] >= 0 && Mask[i] != i)
2598 // Upper quadword shuffled.
2599 for (int i = 4; i != 8; ++i)
2600 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2606 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2607 SmallVector<int, 8> M;
2609 return ::isPSHUFHWMask(M, N->getValueType(0));
2612 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2613 /// is suitable for input to PSHUFLW.
2614 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2615 if (VT != MVT::v8i16)
2618 // Upper quadword copied in order.
2619 for (int i = 4; i != 8; ++i)
2620 if (Mask[i] >= 0 && Mask[i] != i)
2623 // Lower quadword shuffled.
2624 for (int i = 0; i != 4; ++i)
2631 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2632 SmallVector<int, 8> M;
2634 return ::isPSHUFLWMask(M, N->getValueType(0));
2637 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2638 /// is suitable for input to PALIGNR.
2639 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2641 int i, e = VT.getVectorNumElements();
2643 // Do not handle v2i64 / v2f64 shuffles with palignr.
2644 if (e < 4 || !hasSSSE3)
2647 for (i = 0; i != e; ++i)
2651 // All undef, not a palignr.
2655 // Determine if it's ok to perform a palignr with only the LHS, since we
2656 // don't have access to the actual shuffle elements to see if RHS is undef.
2657 bool Unary = Mask[i] < (int)e;
2658 bool NeedsUnary = false;
2660 int s = Mask[i] - i;
2662 // Check the rest of the elements to see if they are consecutive.
2663 for (++i; i != e; ++i) {
2668 Unary = Unary && (m < (int)e);
2669 NeedsUnary = NeedsUnary || (m < s);
2671 if (NeedsUnary && !Unary)
2673 if (Unary && m != ((s+i) & (e-1)))
2675 if (!Unary && m != (s+i))
2681 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2682 SmallVector<int, 8> M;
2684 return ::isPALIGNRMask(M, N->getValueType(0), true);
2687 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2688 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2689 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2690 int NumElems = VT.getVectorNumElements();
2691 if (NumElems != 2 && NumElems != 4)
2694 int Half = NumElems / 2;
2695 for (int i = 0; i < Half; ++i)
2696 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2698 for (int i = Half; i < NumElems; ++i)
2699 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2705 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2706 SmallVector<int, 8> M;
2708 return ::isSHUFPMask(M, N->getValueType(0));
2711 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2712 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2713 /// half elements to come from vector 1 (which would equal the dest.) and
2714 /// the upper half to come from vector 2.
2715 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2716 int NumElems = VT.getVectorNumElements();
2718 if (NumElems != 2 && NumElems != 4)
2721 int Half = NumElems / 2;
2722 for (int i = 0; i < Half; ++i)
2723 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2725 for (int i = Half; i < NumElems; ++i)
2726 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2731 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2732 SmallVector<int, 8> M;
2734 return isCommutedSHUFPMask(M, N->getValueType(0));
2737 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2738 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2739 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2740 if (N->getValueType(0).getVectorNumElements() != 4)
2743 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2744 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2745 isUndefOrEqual(N->getMaskElt(1), 7) &&
2746 isUndefOrEqual(N->getMaskElt(2), 2) &&
2747 isUndefOrEqual(N->getMaskElt(3), 3);
2750 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2751 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2753 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2754 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2759 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2760 isUndefOrEqual(N->getMaskElt(1), 3) &&
2761 isUndefOrEqual(N->getMaskElt(2), 2) &&
2762 isUndefOrEqual(N->getMaskElt(3), 3);
2765 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2766 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2767 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2768 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2770 if (NumElems != 2 && NumElems != 4)
2773 for (unsigned i = 0; i < NumElems/2; ++i)
2774 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2777 for (unsigned i = NumElems/2; i < NumElems; ++i)
2778 if (!isUndefOrEqual(N->getMaskElt(i), i))
2784 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2785 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2786 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2787 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2789 if (NumElems != 2 && NumElems != 4)
2792 for (unsigned i = 0; i < NumElems/2; ++i)
2793 if (!isUndefOrEqual(N->getMaskElt(i), i))
2796 for (unsigned i = 0; i < NumElems/2; ++i)
2797 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2803 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2804 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2805 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2806 bool V2IsSplat = false) {
2807 int NumElts = VT.getVectorNumElements();
2808 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2811 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2813 int BitI1 = Mask[i+1];
2814 if (!isUndefOrEqual(BitI, j))
2817 if (!isUndefOrEqual(BitI1, NumElts))
2820 if (!isUndefOrEqual(BitI1, j + NumElts))
2827 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2828 SmallVector<int, 8> M;
2830 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2833 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2834 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2835 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2836 bool V2IsSplat = false) {
2837 int NumElts = VT.getVectorNumElements();
2838 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2841 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2843 int BitI1 = Mask[i+1];
2844 if (!isUndefOrEqual(BitI, j + NumElts/2))
2847 if (isUndefOrEqual(BitI1, NumElts))
2850 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2857 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2858 SmallVector<int, 8> M;
2860 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2863 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2864 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2866 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2867 int NumElems = VT.getVectorNumElements();
2868 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2871 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2873 int BitI1 = Mask[i+1];
2874 if (!isUndefOrEqual(BitI, j))
2876 if (!isUndefOrEqual(BitI1, j))
2882 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2883 SmallVector<int, 8> M;
2885 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2888 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2889 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2891 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2892 int NumElems = VT.getVectorNumElements();
2893 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2896 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2898 int BitI1 = Mask[i+1];
2899 if (!isUndefOrEqual(BitI, j))
2901 if (!isUndefOrEqual(BitI1, j))
2907 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2908 SmallVector<int, 8> M;
2910 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2913 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2914 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2915 /// MOVSD, and MOVD, i.e. setting the lowest element.
2916 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2917 if (VT.getVectorElementType().getSizeInBits() < 32)
2920 int NumElts = VT.getVectorNumElements();
2922 if (!isUndefOrEqual(Mask[0], NumElts))
2925 for (int i = 1; i < NumElts; ++i)
2926 if (!isUndefOrEqual(Mask[i], i))
2932 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2933 SmallVector<int, 8> M;
2935 return ::isMOVLMask(M, N->getValueType(0));
2938 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2939 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2940 /// element of vector 2 and the other elements to come from vector 1 in order.
2941 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2942 bool V2IsSplat = false, bool V2IsUndef = false) {
2943 int NumOps = VT.getVectorNumElements();
2944 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2947 if (!isUndefOrEqual(Mask[0], 0))
2950 for (int i = 1; i < NumOps; ++i)
2951 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2952 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2953 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2959 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2960 bool V2IsUndef = false) {
2961 SmallVector<int, 8> M;
2963 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2966 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2967 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2968 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2969 if (N->getValueType(0).getVectorNumElements() != 4)
2972 // Expect 1, 1, 3, 3
2973 for (unsigned i = 0; i < 2; ++i) {
2974 int Elt = N->getMaskElt(i);
2975 if (Elt >= 0 && Elt != 1)
2980 for (unsigned i = 2; i < 4; ++i) {
2981 int Elt = N->getMaskElt(i);
2982 if (Elt >= 0 && Elt != 3)
2987 // Don't use movshdup if it can be done with a shufps.
2988 // FIXME: verify that matching u, u, 3, 3 is what we want.
2992 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2993 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2994 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2995 if (N->getValueType(0).getVectorNumElements() != 4)
2998 // Expect 0, 0, 2, 2
2999 for (unsigned i = 0; i < 2; ++i)
3000 if (N->getMaskElt(i) > 0)
3004 for (unsigned i = 2; i < 4; ++i) {
3005 int Elt = N->getMaskElt(i);
3006 if (Elt >= 0 && Elt != 2)
3011 // Don't use movsldup if it can be done with a shufps.
3015 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3016 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3017 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3018 int e = N->getValueType(0).getVectorNumElements() / 2;
3020 for (int i = 0; i < e; ++i)
3021 if (!isUndefOrEqual(N->getMaskElt(i), i))
3023 for (int i = 0; i < e; ++i)
3024 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3029 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3030 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3031 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3032 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3033 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3035 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3037 for (int i = 0; i < NumOperands; ++i) {
3038 int Val = SVOp->getMaskElt(NumOperands-i-1);
3039 if (Val < 0) Val = 0;
3040 if (Val >= NumOperands) Val -= NumOperands;
3042 if (i != NumOperands - 1)
3048 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3049 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3050 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3051 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3053 // 8 nodes, but we only care about the last 4.
3054 for (unsigned i = 7; i >= 4; --i) {
3055 int Val = SVOp->getMaskElt(i);
3064 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3065 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3066 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3067 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3069 // 8 nodes, but we only care about the first 4.
3070 for (int i = 3; i >= 0; --i) {
3071 int Val = SVOp->getMaskElt(i);
3080 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3081 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3082 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3083 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3084 EVT VVT = N->getValueType(0);
3085 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3089 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3090 Val = SVOp->getMaskElt(i);
3094 return (Val - i) * EltSize;
3097 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3099 bool X86::isZeroNode(SDValue Elt) {
3100 return ((isa<ConstantSDNode>(Elt) &&
3101 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
3102 (isa<ConstantFPSDNode>(Elt) &&
3103 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3106 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3107 /// their permute mask.
3108 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3109 SelectionDAG &DAG) {
3110 EVT VT = SVOp->getValueType(0);
3111 unsigned NumElems = VT.getVectorNumElements();
3112 SmallVector<int, 8> MaskVec;
3114 for (unsigned i = 0; i != NumElems; ++i) {
3115 int idx = SVOp->getMaskElt(i);
3117 MaskVec.push_back(idx);
3118 else if (idx < (int)NumElems)
3119 MaskVec.push_back(idx + NumElems);
3121 MaskVec.push_back(idx - NumElems);
3123 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3124 SVOp->getOperand(0), &MaskVec[0]);
3127 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3128 /// the two vector operands have swapped position.
3129 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3130 unsigned NumElems = VT.getVectorNumElements();
3131 for (unsigned i = 0; i != NumElems; ++i) {
3135 else if (idx < (int)NumElems)
3136 Mask[i] = idx + NumElems;
3138 Mask[i] = idx - NumElems;
3142 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3143 /// match movhlps. The lower half elements should come from upper half of
3144 /// V1 (and in order), and the upper half elements should come from the upper
3145 /// half of V2 (and in order).
3146 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3147 if (Op->getValueType(0).getVectorNumElements() != 4)
3149 for (unsigned i = 0, e = 2; i != e; ++i)
3150 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3152 for (unsigned i = 2; i != 4; ++i)
3153 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3158 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3159 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3161 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3162 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3164 N = N->getOperand(0).getNode();
3165 if (!ISD::isNON_EXTLoad(N))
3168 *LD = cast<LoadSDNode>(N);
3172 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3173 /// match movlp{s|d}. The lower half elements should come from lower half of
3174 /// V1 (and in order), and the upper half elements should come from the upper
3175 /// half of V2 (and in order). And since V1 will become the source of the
3176 /// MOVLP, it must be either a vector load or a scalar load to vector.
3177 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3178 ShuffleVectorSDNode *Op) {
3179 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3181 // Is V2 is a vector load, don't do this transformation. We will try to use
3182 // load folding shufps op.
3183 if (ISD::isNON_EXTLoad(V2))
3186 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3188 if (NumElems != 2 && NumElems != 4)
3190 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3191 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3193 for (unsigned i = NumElems/2; i != NumElems; ++i)
3194 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3199 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3201 static bool isSplatVector(SDNode *N) {
3202 if (N->getOpcode() != ISD::BUILD_VECTOR)
3205 SDValue SplatValue = N->getOperand(0);
3206 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3207 if (N->getOperand(i) != SplatValue)
3212 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3213 /// to an zero vector.
3214 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3215 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3216 SDValue V1 = N->getOperand(0);
3217 SDValue V2 = N->getOperand(1);
3218 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3219 for (unsigned i = 0; i != NumElems; ++i) {
3220 int Idx = N->getMaskElt(i);
3221 if (Idx >= (int)NumElems) {
3222 unsigned Opc = V2.getOpcode();
3223 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3225 if (Opc != ISD::BUILD_VECTOR ||
3226 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3228 } else if (Idx >= 0) {
3229 unsigned Opc = V1.getOpcode();
3230 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3232 if (Opc != ISD::BUILD_VECTOR ||
3233 !X86::isZeroNode(V1.getOperand(Idx)))
3240 /// getZeroVector - Returns a vector of specified type with all zero elements.
3242 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3244 assert(VT.isVector() && "Expected a vector type");
3246 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3247 // type. This ensures they get CSE'd.
3249 if (VT.getSizeInBits() == 64) { // MMX
3250 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3251 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3252 } else if (HasSSE2) { // SSE2
3253 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3254 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3256 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3257 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3259 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3262 /// getOnesVector - Returns a vector of specified type with all bits set.
3264 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3265 assert(VT.isVector() && "Expected a vector type");
3267 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3268 // type. This ensures they get CSE'd.
3269 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3271 if (VT.getSizeInBits() == 64) // MMX
3272 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3274 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3275 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3279 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3280 /// that point to V2 points to its first element.
3281 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3282 EVT VT = SVOp->getValueType(0);
3283 unsigned NumElems = VT.getVectorNumElements();
3285 bool Changed = false;
3286 SmallVector<int, 8> MaskVec;
3287 SVOp->getMask(MaskVec);
3289 for (unsigned i = 0; i != NumElems; ++i) {
3290 if (MaskVec[i] > (int)NumElems) {
3291 MaskVec[i] = NumElems;
3296 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3297 SVOp->getOperand(1), &MaskVec[0]);
3298 return SDValue(SVOp, 0);
3301 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3302 /// operation of specified width.
3303 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3305 unsigned NumElems = VT.getVectorNumElements();
3306 SmallVector<int, 8> Mask;
3307 Mask.push_back(NumElems);
3308 for (unsigned i = 1; i != NumElems; ++i)
3310 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3313 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3314 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3316 unsigned NumElems = VT.getVectorNumElements();
3317 SmallVector<int, 8> Mask;
3318 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3320 Mask.push_back(i + NumElems);
3322 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3325 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3326 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3328 unsigned NumElems = VT.getVectorNumElements();
3329 unsigned Half = NumElems/2;
3330 SmallVector<int, 8> Mask;
3331 for (unsigned i = 0; i != Half; ++i) {
3332 Mask.push_back(i + Half);
3333 Mask.push_back(i + NumElems + Half);
3335 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3338 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3339 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3341 if (SV->getValueType(0).getVectorNumElements() <= 4)
3342 return SDValue(SV, 0);
3344 EVT PVT = MVT::v4f32;
3345 EVT VT = SV->getValueType(0);
3346 DebugLoc dl = SV->getDebugLoc();
3347 SDValue V1 = SV->getOperand(0);
3348 int NumElems = VT.getVectorNumElements();
3349 int EltNo = SV->getSplatIndex();
3351 // unpack elements to the correct location
3352 while (NumElems > 4) {
3353 if (EltNo < NumElems/2) {
3354 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3356 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3357 EltNo -= NumElems/2;
3362 // Perform the splat.
3363 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3364 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3365 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3366 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3369 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3370 /// vector of zero or undef vector. This produces a shuffle where the low
3371 /// element of V2 is swizzled into the zero/undef vector, landing at element
3372 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3373 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3374 bool isZero, bool HasSSE2,
3375 SelectionDAG &DAG) {
3376 EVT VT = V2.getValueType();
3378 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3379 unsigned NumElems = VT.getVectorNumElements();
3380 SmallVector<int, 16> MaskVec;
3381 for (unsigned i = 0; i != NumElems; ++i)
3382 // If this is the insertion idx, put the low elt of V2 here.
3383 MaskVec.push_back(i == Idx ? NumElems : i);
3384 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3387 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3388 /// a shuffle that is zero.
3390 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3391 bool Low, SelectionDAG &DAG) {
3392 unsigned NumZeros = 0;
3393 for (int i = 0; i < NumElems; ++i) {
3394 unsigned Index = Low ? i : NumElems-i-1;
3395 int Idx = SVOp->getMaskElt(Index);
3400 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3401 if (Elt.getNode() && X86::isZeroNode(Elt))
3409 /// isVectorShift - Returns true if the shuffle can be implemented as a
3410 /// logical left or right shift of a vector.
3411 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3412 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3413 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3414 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3417 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3420 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3424 bool SeenV1 = false;
3425 bool SeenV2 = false;
3426 for (int i = NumZeros; i < NumElems; ++i) {
3427 int Val = isLeft ? (i - NumZeros) : i;
3428 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3440 if (SeenV1 && SeenV2)
3443 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3449 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3451 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3452 unsigned NumNonZero, unsigned NumZero,
3453 SelectionDAG &DAG, TargetLowering &TLI) {
3457 DebugLoc dl = Op.getDebugLoc();
3460 for (unsigned i = 0; i < 16; ++i) {
3461 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3462 if (ThisIsNonZero && First) {
3464 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3466 V = DAG.getUNDEF(MVT::v8i16);
3471 SDValue ThisElt(0, 0), LastElt(0, 0);
3472 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3473 if (LastIsNonZero) {
3474 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3475 MVT::i16, Op.getOperand(i-1));
3477 if (ThisIsNonZero) {
3478 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3479 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3480 ThisElt, DAG.getConstant(8, MVT::i8));
3482 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3486 if (ThisElt.getNode())
3487 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3488 DAG.getIntPtrConstant(i/2));
3492 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3495 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3497 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3498 unsigned NumNonZero, unsigned NumZero,
3499 SelectionDAG &DAG, TargetLowering &TLI) {
3503 DebugLoc dl = Op.getDebugLoc();
3506 for (unsigned i = 0; i < 8; ++i) {
3507 bool isNonZero = (NonZeros & (1 << i)) != 0;
3511 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3513 V = DAG.getUNDEF(MVT::v8i16);
3516 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3517 MVT::v8i16, V, Op.getOperand(i),
3518 DAG.getIntPtrConstant(i));
3525 /// getVShift - Return a vector logical shift node.
3527 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3528 unsigned NumBits, SelectionDAG &DAG,
3529 const TargetLowering &TLI, DebugLoc dl) {
3530 bool isMMX = VT.getSizeInBits() == 64;
3531 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3532 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3533 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3534 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3535 DAG.getNode(Opc, dl, ShVT, SrcOp,
3536 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3540 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3541 SelectionDAG &DAG) {
3543 // Check if the scalar load can be widened into a vector load. And if
3544 // the address is "base + cst" see if the cst can be "absorbed" into
3545 // the shuffle mask.
3546 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3547 SDValue Ptr = LD->getBasePtr();
3548 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3550 EVT PVT = LD->getValueType(0);
3551 if (PVT != MVT::i32 && PVT != MVT::f32)
3556 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3557 FI = FINode->getIndex();
3559 } else if (Ptr.getOpcode() == ISD::ADD &&
3560 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3561 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3562 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3563 Offset = Ptr.getConstantOperandVal(1);
3564 Ptr = Ptr.getOperand(0);
3569 SDValue Chain = LD->getChain();
3570 // Make sure the stack object alignment is at least 16.
3571 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3572 if (DAG.InferPtrAlignment(Ptr) < 16) {
3573 if (MFI->isFixedObjectIndex(FI)) {
3574 // Can't change the alignment. FIXME: It's possible to compute
3575 // the exact stack offset and reference FI + adjust offset instead.
3576 // If someone *really* cares about this. That's the way to implement it.
3579 MFI->setObjectAlignment(FI, 16);
3583 // (Offset % 16) must be multiple of 4. Then address is then
3584 // Ptr + (Offset & ~15).
3587 if ((Offset % 16) & 3)
3589 int64_t StartOffset = Offset & ~15;
3591 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3592 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3594 int EltNo = (Offset - StartOffset) >> 2;
3595 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3596 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3597 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0);
3598 // Canonicalize it to a v4i32 shuffle.
3599 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3600 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3601 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3602 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3609 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3610 DebugLoc dl = Op.getDebugLoc();
3611 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3612 if (ISD::isBuildVectorAllZeros(Op.getNode())
3613 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3614 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3615 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3616 // eliminated on x86-32 hosts.
3617 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3620 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3621 return getOnesVector(Op.getValueType(), DAG, dl);
3622 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3625 EVT VT = Op.getValueType();
3626 EVT ExtVT = VT.getVectorElementType();
3627 unsigned EVTBits = ExtVT.getSizeInBits();
3629 unsigned NumElems = Op.getNumOperands();
3630 unsigned NumZero = 0;
3631 unsigned NumNonZero = 0;
3632 unsigned NonZeros = 0;
3633 bool IsAllConstants = true;
3634 SmallSet<SDValue, 8> Values;
3635 for (unsigned i = 0; i < NumElems; ++i) {
3636 SDValue Elt = Op.getOperand(i);
3637 if (Elt.getOpcode() == ISD::UNDEF)
3640 if (Elt.getOpcode() != ISD::Constant &&
3641 Elt.getOpcode() != ISD::ConstantFP)
3642 IsAllConstants = false;
3643 if (X86::isZeroNode(Elt))
3646 NonZeros |= (1 << i);
3651 if (NumNonZero == 0) {
3652 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3653 return DAG.getUNDEF(VT);
3656 // Special case for single non-zero, non-undef, element.
3657 if (NumNonZero == 1) {
3658 unsigned Idx = CountTrailingZeros_32(NonZeros);
3659 SDValue Item = Op.getOperand(Idx);
3661 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3662 // the value are obviously zero, truncate the value to i32 and do the
3663 // insertion that way. Only do this if the value is non-constant or if the
3664 // value is a constant being inserted into element 0. It is cheaper to do
3665 // a constant pool load than it is to do a movd + shuffle.
3666 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3667 (!IsAllConstants || Idx == 0)) {
3668 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3669 // Handle MMX and SSE both.
3670 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3671 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3673 // Truncate the value (which may itself be a constant) to i32, and
3674 // convert it to a vector with movd (S2V+shuffle to zero extend).
3675 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3676 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3677 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3678 Subtarget->hasSSE2(), DAG);
3680 // Now we have our 32-bit value zero extended in the low element of
3681 // a vector. If Idx != 0, swizzle it into place.
3683 SmallVector<int, 4> Mask;
3684 Mask.push_back(Idx);
3685 for (unsigned i = 1; i != VecElts; ++i)
3687 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3688 DAG.getUNDEF(Item.getValueType()),
3691 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3695 // If we have a constant or non-constant insertion into the low element of
3696 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3697 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3698 // depending on what the source datatype is.
3701 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3702 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3703 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3704 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3705 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3706 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3708 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3709 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3710 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3711 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3712 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3713 Subtarget->hasSSE2(), DAG);
3714 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3718 // Is it a vector logical left shift?
3719 if (NumElems == 2 && Idx == 1 &&
3720 X86::isZeroNode(Op.getOperand(0)) &&
3721 !X86::isZeroNode(Op.getOperand(1))) {
3722 unsigned NumBits = VT.getSizeInBits();
3723 return getVShift(true, VT,
3724 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3725 VT, Op.getOperand(1)),
3726 NumBits/2, DAG, *this, dl);
3729 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3732 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3733 // is a non-constant being inserted into an element other than the low one,
3734 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3735 // movd/movss) to move this into the low element, then shuffle it into
3737 if (EVTBits == 32) {
3738 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3740 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3741 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3742 Subtarget->hasSSE2(), DAG);
3743 SmallVector<int, 8> MaskVec;
3744 for (unsigned i = 0; i < NumElems; i++)
3745 MaskVec.push_back(i == Idx ? 0 : 1);
3746 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3750 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3751 if (Values.size() == 1) {
3752 if (EVTBits == 32) {
3753 // Instead of a shuffle like this:
3754 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3755 // Check if it's possible to issue this instead.
3756 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3757 unsigned Idx = CountTrailingZeros_32(NonZeros);
3758 SDValue Item = Op.getOperand(Idx);
3759 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3760 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3765 // A vector full of immediates; various special cases are already
3766 // handled, so this is best done with a single constant-pool load.
3770 // Let legalizer expand 2-wide build_vectors.
3771 if (EVTBits == 64) {
3772 if (NumNonZero == 1) {
3773 // One half is zero or undef.
3774 unsigned Idx = CountTrailingZeros_32(NonZeros);
3775 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3776 Op.getOperand(Idx));
3777 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3778 Subtarget->hasSSE2(), DAG);
3783 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3784 if (EVTBits == 8 && NumElems == 16) {
3785 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3787 if (V.getNode()) return V;
3790 if (EVTBits == 16 && NumElems == 8) {
3791 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3793 if (V.getNode()) return V;
3796 // If element VT is == 32 bits, turn it into a number of shuffles.
3797 SmallVector<SDValue, 8> V;
3799 if (NumElems == 4 && NumZero > 0) {
3800 for (unsigned i = 0; i < 4; ++i) {
3801 bool isZero = !(NonZeros & (1 << i));
3803 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3805 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3808 for (unsigned i = 0; i < 2; ++i) {
3809 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3812 V[i] = V[i*2]; // Must be a zero vector.
3815 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3818 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3821 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3826 SmallVector<int, 8> MaskVec;
3827 bool Reverse = (NonZeros & 0x3) == 2;
3828 for (unsigned i = 0; i < 2; ++i)
3829 MaskVec.push_back(Reverse ? 1-i : i);
3830 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3831 for (unsigned i = 0; i < 2; ++i)
3832 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3833 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3836 if (Values.size() > 2) {
3837 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3838 // values to be inserted is equal to the number of elements, in which case
3839 // use the unpack code below in the hopes of matching the consecutive elts
3840 // load merge pattern for shuffles.
3841 // FIXME: We could probably just check that here directly.
3842 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3843 getSubtarget()->hasSSE41()) {
3844 V[0] = DAG.getUNDEF(VT);
3845 for (unsigned i = 0; i < NumElems; ++i)
3846 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3847 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3848 Op.getOperand(i), DAG.getIntPtrConstant(i));
3851 // Expand into a number of unpckl*.
3853 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3854 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3855 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3856 for (unsigned i = 0; i < NumElems; ++i)
3857 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3859 while (NumElems != 0) {
3860 for (unsigned i = 0; i < NumElems; ++i)
3861 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3871 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3872 // We support concatenate two MMX registers and place them in a MMX
3873 // register. This is better than doing a stack convert.
3874 DebugLoc dl = Op.getDebugLoc();
3875 EVT ResVT = Op.getValueType();
3876 assert(Op.getNumOperands() == 2);
3877 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
3878 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
3880 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
3881 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3882 InVec = Op.getOperand(1);
3883 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3884 unsigned NumElts = ResVT.getVectorNumElements();
3885 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3886 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
3887 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
3889 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
3890 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3891 Mask[0] = 0; Mask[1] = 2;
3892 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
3894 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3897 // v8i16 shuffles - Prefer shuffles in the following order:
3898 // 1. [all] pshuflw, pshufhw, optional move
3899 // 2. [ssse3] 1 x pshufb
3900 // 3. [ssse3] 2 x pshufb + 1 x por
3901 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3903 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3904 SelectionDAG &DAG, X86TargetLowering &TLI) {
3905 SDValue V1 = SVOp->getOperand(0);
3906 SDValue V2 = SVOp->getOperand(1);
3907 DebugLoc dl = SVOp->getDebugLoc();
3908 SmallVector<int, 8> MaskVals;
3910 // Determine if more than 1 of the words in each of the low and high quadwords
3911 // of the result come from the same quadword of one of the two inputs. Undef
3912 // mask values count as coming from any quadword, for better codegen.
3913 SmallVector<unsigned, 4> LoQuad(4);
3914 SmallVector<unsigned, 4> HiQuad(4);
3915 BitVector InputQuads(4);
3916 for (unsigned i = 0; i < 8; ++i) {
3917 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3918 int EltIdx = SVOp->getMaskElt(i);
3919 MaskVals.push_back(EltIdx);
3928 InputQuads.set(EltIdx / 4);
3931 int BestLoQuad = -1;
3932 unsigned MaxQuad = 1;
3933 for (unsigned i = 0; i < 4; ++i) {
3934 if (LoQuad[i] > MaxQuad) {
3936 MaxQuad = LoQuad[i];
3940 int BestHiQuad = -1;
3942 for (unsigned i = 0; i < 4; ++i) {
3943 if (HiQuad[i] > MaxQuad) {
3945 MaxQuad = HiQuad[i];
3949 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3950 // of the two input vectors, shuffle them into one input vector so only a
3951 // single pshufb instruction is necessary. If There are more than 2 input
3952 // quads, disable the next transformation since it does not help SSSE3.
3953 bool V1Used = InputQuads[0] || InputQuads[1];
3954 bool V2Used = InputQuads[2] || InputQuads[3];
3955 if (TLI.getSubtarget()->hasSSSE3()) {
3956 if (InputQuads.count() == 2 && V1Used && V2Used) {
3957 BestLoQuad = InputQuads.find_first();
3958 BestHiQuad = InputQuads.find_next(BestLoQuad);
3960 if (InputQuads.count() > 2) {
3966 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3967 // the shuffle mask. If a quad is scored as -1, that means that it contains
3968 // words from all 4 input quadwords.
3970 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3971 SmallVector<int, 8> MaskV;
3972 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3973 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3974 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3975 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3976 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3977 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3979 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3980 // source words for the shuffle, to aid later transformations.
3981 bool AllWordsInNewV = true;
3982 bool InOrder[2] = { true, true };
3983 for (unsigned i = 0; i != 8; ++i) {
3984 int idx = MaskVals[i];
3986 InOrder[i/4] = false;
3987 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3989 AllWordsInNewV = false;
3993 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3994 if (AllWordsInNewV) {
3995 for (int i = 0; i != 8; ++i) {
3996 int idx = MaskVals[i];
3999 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4000 if ((idx != i) && idx < 4)
4002 if ((idx != i) && idx > 3)
4011 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4012 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4013 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4014 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4015 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4019 // If we have SSSE3, and all words of the result are from 1 input vector,
4020 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4021 // is present, fall back to case 4.
4022 if (TLI.getSubtarget()->hasSSSE3()) {
4023 SmallVector<SDValue,16> pshufbMask;
4025 // If we have elements from both input vectors, set the high bit of the
4026 // shuffle mask element to zero out elements that come from V2 in the V1
4027 // mask, and elements that come from V1 in the V2 mask, so that the two
4028 // results can be OR'd together.
4029 bool TwoInputs = V1Used && V2Used;
4030 for (unsigned i = 0; i != 8; ++i) {
4031 int EltIdx = MaskVals[i] * 2;
4032 if (TwoInputs && (EltIdx >= 16)) {
4033 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4034 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4037 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4038 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4040 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4041 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4042 DAG.getNode(ISD::BUILD_VECTOR, dl,
4043 MVT::v16i8, &pshufbMask[0], 16));
4045 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4047 // Calculate the shuffle mask for the second input, shuffle it, and
4048 // OR it with the first shuffled input.
4050 for (unsigned i = 0; i != 8; ++i) {
4051 int EltIdx = MaskVals[i] * 2;
4053 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4054 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4057 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4058 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4060 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4061 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4062 DAG.getNode(ISD::BUILD_VECTOR, dl,
4063 MVT::v16i8, &pshufbMask[0], 16));
4064 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4065 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4068 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4069 // and update MaskVals with new element order.
4070 BitVector InOrder(8);
4071 if (BestLoQuad >= 0) {
4072 SmallVector<int, 8> MaskV;
4073 for (int i = 0; i != 4; ++i) {
4074 int idx = MaskVals[i];
4076 MaskV.push_back(-1);
4078 } else if ((idx / 4) == BestLoQuad) {
4079 MaskV.push_back(idx & 3);
4082 MaskV.push_back(-1);
4085 for (unsigned i = 4; i != 8; ++i)
4087 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4091 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4092 // and update MaskVals with the new element order.
4093 if (BestHiQuad >= 0) {
4094 SmallVector<int, 8> MaskV;
4095 for (unsigned i = 0; i != 4; ++i)
4097 for (unsigned i = 4; i != 8; ++i) {
4098 int idx = MaskVals[i];
4100 MaskV.push_back(-1);
4102 } else if ((idx / 4) == BestHiQuad) {
4103 MaskV.push_back((idx & 3) + 4);
4106 MaskV.push_back(-1);
4109 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4113 // In case BestHi & BestLo were both -1, which means each quadword has a word
4114 // from each of the four input quadwords, calculate the InOrder bitvector now
4115 // before falling through to the insert/extract cleanup.
4116 if (BestLoQuad == -1 && BestHiQuad == -1) {
4118 for (int i = 0; i != 8; ++i)
4119 if (MaskVals[i] < 0 || MaskVals[i] == i)
4123 // The other elements are put in the right place using pextrw and pinsrw.
4124 for (unsigned i = 0; i != 8; ++i) {
4127 int EltIdx = MaskVals[i];
4130 SDValue ExtOp = (EltIdx < 8)
4131 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4132 DAG.getIntPtrConstant(EltIdx))
4133 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4134 DAG.getIntPtrConstant(EltIdx - 8));
4135 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4136 DAG.getIntPtrConstant(i));
4141 // v16i8 shuffles - Prefer shuffles in the following order:
4142 // 1. [ssse3] 1 x pshufb
4143 // 2. [ssse3] 2 x pshufb + 1 x por
4144 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4146 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4147 SelectionDAG &DAG, X86TargetLowering &TLI) {
4148 SDValue V1 = SVOp->getOperand(0);
4149 SDValue V2 = SVOp->getOperand(1);
4150 DebugLoc dl = SVOp->getDebugLoc();
4151 SmallVector<int, 16> MaskVals;
4152 SVOp->getMask(MaskVals);
4154 // If we have SSSE3, case 1 is generated when all result bytes come from
4155 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4156 // present, fall back to case 3.
4157 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4160 for (unsigned i = 0; i < 16; ++i) {
4161 int EltIdx = MaskVals[i];
4170 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4171 if (TLI.getSubtarget()->hasSSSE3()) {
4172 SmallVector<SDValue,16> pshufbMask;
4174 // If all result elements are from one input vector, then only translate
4175 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4177 // Otherwise, we have elements from both input vectors, and must zero out
4178 // elements that come from V2 in the first mask, and V1 in the second mask
4179 // so that we can OR them together.
4180 bool TwoInputs = !(V1Only || V2Only);
4181 for (unsigned i = 0; i != 16; ++i) {
4182 int EltIdx = MaskVals[i];
4183 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4184 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4187 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4189 // If all the elements are from V2, assign it to V1 and return after
4190 // building the first pshufb.
4193 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4194 DAG.getNode(ISD::BUILD_VECTOR, dl,
4195 MVT::v16i8, &pshufbMask[0], 16));
4199 // Calculate the shuffle mask for the second input, shuffle it, and
4200 // OR it with the first shuffled input.
4202 for (unsigned i = 0; i != 16; ++i) {
4203 int EltIdx = MaskVals[i];
4205 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4208 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4210 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4211 DAG.getNode(ISD::BUILD_VECTOR, dl,
4212 MVT::v16i8, &pshufbMask[0], 16));
4213 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4216 // No SSSE3 - Calculate in place words and then fix all out of place words
4217 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4218 // the 16 different words that comprise the two doublequadword input vectors.
4219 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4220 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4221 SDValue NewV = V2Only ? V2 : V1;
4222 for (int i = 0; i != 8; ++i) {
4223 int Elt0 = MaskVals[i*2];
4224 int Elt1 = MaskVals[i*2+1];
4226 // This word of the result is all undef, skip it.
4227 if (Elt0 < 0 && Elt1 < 0)
4230 // This word of the result is already in the correct place, skip it.
4231 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4233 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4236 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4237 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4240 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4241 // using a single extract together, load it and store it.
4242 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4243 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4244 DAG.getIntPtrConstant(Elt1 / 2));
4245 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4246 DAG.getIntPtrConstant(i));
4250 // If Elt1 is defined, extract it from the appropriate source. If the
4251 // source byte is not also odd, shift the extracted word left 8 bits
4252 // otherwise clear the bottom 8 bits if we need to do an or.
4254 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4255 DAG.getIntPtrConstant(Elt1 / 2));
4256 if ((Elt1 & 1) == 0)
4257 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4258 DAG.getConstant(8, TLI.getShiftAmountTy()));
4260 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4261 DAG.getConstant(0xFF00, MVT::i16));
4263 // If Elt0 is defined, extract it from the appropriate source. If the
4264 // source byte is not also even, shift the extracted word right 8 bits. If
4265 // Elt1 was also defined, OR the extracted values together before
4266 // inserting them in the result.
4268 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4269 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4270 if ((Elt0 & 1) != 0)
4271 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4272 DAG.getConstant(8, TLI.getShiftAmountTy()));
4274 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4275 DAG.getConstant(0x00FF, MVT::i16));
4276 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4279 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4280 DAG.getIntPtrConstant(i));
4282 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4285 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4286 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4287 /// done when every pair / quad of shuffle mask elements point to elements in
4288 /// the right sequence. e.g.
4289 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4291 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4293 TargetLowering &TLI, DebugLoc dl) {
4294 EVT VT = SVOp->getValueType(0);
4295 SDValue V1 = SVOp->getOperand(0);
4296 SDValue V2 = SVOp->getOperand(1);
4297 unsigned NumElems = VT.getVectorNumElements();
4298 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4299 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4300 EVT MaskEltVT = MaskVT.getVectorElementType();
4302 switch (VT.getSimpleVT().SimpleTy) {
4303 default: assert(false && "Unexpected!");
4304 case MVT::v4f32: NewVT = MVT::v2f64; break;
4305 case MVT::v4i32: NewVT = MVT::v2i64; break;
4306 case MVT::v8i16: NewVT = MVT::v4i32; break;
4307 case MVT::v16i8: NewVT = MVT::v4i32; break;
4310 if (NewWidth == 2) {
4316 int Scale = NumElems / NewWidth;
4317 SmallVector<int, 8> MaskVec;
4318 for (unsigned i = 0; i < NumElems; i += Scale) {
4320 for (int j = 0; j < Scale; ++j) {
4321 int EltIdx = SVOp->getMaskElt(i+j);
4325 StartIdx = EltIdx - (EltIdx % Scale);
4326 if (EltIdx != StartIdx + j)
4330 MaskVec.push_back(-1);
4332 MaskVec.push_back(StartIdx / Scale);
4335 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4336 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4337 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4340 /// getVZextMovL - Return a zero-extending vector move low node.
4342 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4343 SDValue SrcOp, SelectionDAG &DAG,
4344 const X86Subtarget *Subtarget, DebugLoc dl) {
4345 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4346 LoadSDNode *LD = NULL;
4347 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4348 LD = dyn_cast<LoadSDNode>(SrcOp);
4350 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4352 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4353 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4354 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4355 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4356 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4358 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4359 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4360 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4361 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4369 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4370 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4371 DAG.getNode(ISD::BIT_CONVERT, dl,
4375 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4378 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4379 SDValue V1 = SVOp->getOperand(0);
4380 SDValue V2 = SVOp->getOperand(1);
4381 DebugLoc dl = SVOp->getDebugLoc();
4382 EVT VT = SVOp->getValueType(0);
4384 SmallVector<std::pair<int, int>, 8> Locs;
4386 SmallVector<int, 8> Mask1(4U, -1);
4387 SmallVector<int, 8> PermMask;
4388 SVOp->getMask(PermMask);
4392 for (unsigned i = 0; i != 4; ++i) {
4393 int Idx = PermMask[i];
4395 Locs[i] = std::make_pair(-1, -1);
4397 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4399 Locs[i] = std::make_pair(0, NumLo);
4403 Locs[i] = std::make_pair(1, NumHi);
4405 Mask1[2+NumHi] = Idx;
4411 if (NumLo <= 2 && NumHi <= 2) {
4412 // If no more than two elements come from either vector. This can be
4413 // implemented with two shuffles. First shuffle gather the elements.
4414 // The second shuffle, which takes the first shuffle as both of its
4415 // vector operands, put the elements into the right order.
4416 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4418 SmallVector<int, 8> Mask2(4U, -1);
4420 for (unsigned i = 0; i != 4; ++i) {
4421 if (Locs[i].first == -1)
4424 unsigned Idx = (i < 2) ? 0 : 4;
4425 Idx += Locs[i].first * 2 + Locs[i].second;
4430 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4431 } else if (NumLo == 3 || NumHi == 3) {
4432 // Otherwise, we must have three elements from one vector, call it X, and
4433 // one element from the other, call it Y. First, use a shufps to build an
4434 // intermediate vector with the one element from Y and the element from X
4435 // that will be in the same half in the final destination (the indexes don't
4436 // matter). Then, use a shufps to build the final vector, taking the half
4437 // containing the element from Y from the intermediate, and the other half
4440 // Normalize it so the 3 elements come from V1.
4441 CommuteVectorShuffleMask(PermMask, VT);
4445 // Find the element from V2.
4447 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4448 int Val = PermMask[HiIndex];
4455 Mask1[0] = PermMask[HiIndex];
4457 Mask1[2] = PermMask[HiIndex^1];
4459 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4462 Mask1[0] = PermMask[0];
4463 Mask1[1] = PermMask[1];
4464 Mask1[2] = HiIndex & 1 ? 6 : 4;
4465 Mask1[3] = HiIndex & 1 ? 4 : 6;
4466 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4468 Mask1[0] = HiIndex & 1 ? 2 : 0;
4469 Mask1[1] = HiIndex & 1 ? 0 : 2;
4470 Mask1[2] = PermMask[2];
4471 Mask1[3] = PermMask[3];
4476 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4480 // Break it into (shuffle shuffle_hi, shuffle_lo).
4482 SmallVector<int,8> LoMask(4U, -1);
4483 SmallVector<int,8> HiMask(4U, -1);
4485 SmallVector<int,8> *MaskPtr = &LoMask;
4486 unsigned MaskIdx = 0;
4489 for (unsigned i = 0; i != 4; ++i) {
4496 int Idx = PermMask[i];
4498 Locs[i] = std::make_pair(-1, -1);
4499 } else if (Idx < 4) {
4500 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4501 (*MaskPtr)[LoIdx] = Idx;
4504 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4505 (*MaskPtr)[HiIdx] = Idx;
4510 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4511 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4512 SmallVector<int, 8> MaskOps;
4513 for (unsigned i = 0; i != 4; ++i) {
4514 if (Locs[i].first == -1) {
4515 MaskOps.push_back(-1);
4517 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4518 MaskOps.push_back(Idx);
4521 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4525 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4526 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4527 SDValue V1 = Op.getOperand(0);
4528 SDValue V2 = Op.getOperand(1);
4529 EVT VT = Op.getValueType();
4530 DebugLoc dl = Op.getDebugLoc();
4531 unsigned NumElems = VT.getVectorNumElements();
4532 bool isMMX = VT.getSizeInBits() == 64;
4533 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4534 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4535 bool V1IsSplat = false;
4536 bool V2IsSplat = false;
4538 if (isZeroShuffle(SVOp))
4539 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4541 // Promote splats to v4f32.
4542 if (SVOp->isSplat()) {
4543 if (isMMX || NumElems < 4)
4545 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4548 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4550 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4551 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4552 if (NewOp.getNode())
4553 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4554 LowerVECTOR_SHUFFLE(NewOp, DAG));
4555 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4556 // FIXME: Figure out a cleaner way to do this.
4557 // Try to make use of movq to zero out the top part.
4558 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4559 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4560 if (NewOp.getNode()) {
4561 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4562 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4563 DAG, Subtarget, dl);
4565 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4566 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4567 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4568 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4569 DAG, Subtarget, dl);
4573 if (X86::isPSHUFDMask(SVOp))
4576 // Check if this can be converted into a logical shift.
4577 bool isLeft = false;
4580 bool isShift = getSubtarget()->hasSSE2() &&
4581 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4582 if (isShift && ShVal.hasOneUse()) {
4583 // If the shifted value has multiple uses, it may be cheaper to use
4584 // v_set0 + movlhps or movhlps, etc.
4585 EVT EltVT = VT.getVectorElementType();
4586 ShAmt *= EltVT.getSizeInBits();
4587 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4590 if (X86::isMOVLMask(SVOp)) {
4593 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4594 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4599 // FIXME: fold these into legal mask.
4600 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4601 X86::isMOVSLDUPMask(SVOp) ||
4602 X86::isMOVHLPSMask(SVOp) ||
4603 X86::isMOVLHPSMask(SVOp) ||
4604 X86::isMOVLPMask(SVOp)))
4607 if (ShouldXformToMOVHLPS(SVOp) ||
4608 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4609 return CommuteVectorShuffle(SVOp, DAG);
4612 // No better options. Use a vshl / vsrl.
4613 EVT EltVT = VT.getVectorElementType();
4614 ShAmt *= EltVT.getSizeInBits();
4615 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4618 bool Commuted = false;
4619 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4620 // 1,1,1,1 -> v8i16 though.
4621 V1IsSplat = isSplatVector(V1.getNode());
4622 V2IsSplat = isSplatVector(V2.getNode());
4624 // Canonicalize the splat or undef, if present, to be on the RHS.
4625 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4626 Op = CommuteVectorShuffle(SVOp, DAG);
4627 SVOp = cast<ShuffleVectorSDNode>(Op);
4628 V1 = SVOp->getOperand(0);
4629 V2 = SVOp->getOperand(1);
4630 std::swap(V1IsSplat, V2IsSplat);
4631 std::swap(V1IsUndef, V2IsUndef);
4635 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4636 // Shuffling low element of v1 into undef, just return v1.
4639 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4640 // the instruction selector will not match, so get a canonical MOVL with
4641 // swapped operands to undo the commute.
4642 return getMOVL(DAG, dl, VT, V2, V1);
4645 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4646 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4647 X86::isUNPCKLMask(SVOp) ||
4648 X86::isUNPCKHMask(SVOp))
4652 // Normalize mask so all entries that point to V2 points to its first
4653 // element then try to match unpck{h|l} again. If match, return a
4654 // new vector_shuffle with the corrected mask.
4655 SDValue NewMask = NormalizeMask(SVOp, DAG);
4656 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4657 if (NSVOp != SVOp) {
4658 if (X86::isUNPCKLMask(NSVOp, true)) {
4660 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4667 // Commute is back and try unpck* again.
4668 // FIXME: this seems wrong.
4669 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4670 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4671 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4672 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4673 X86::isUNPCKLMask(NewSVOp) ||
4674 X86::isUNPCKHMask(NewSVOp))
4678 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4680 // Normalize the node to match x86 shuffle ops if needed
4681 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4682 return CommuteVectorShuffle(SVOp, DAG);
4684 // Check for legal shuffle and return?
4685 SmallVector<int, 16> PermMask;
4686 SVOp->getMask(PermMask);
4687 if (isShuffleMaskLegal(PermMask, VT))
4690 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4691 if (VT == MVT::v8i16) {
4692 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4693 if (NewOp.getNode())
4697 if (VT == MVT::v16i8) {
4698 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4699 if (NewOp.getNode())
4703 // Handle all 4 wide cases with a number of shuffles except for MMX.
4704 if (NumElems == 4 && !isMMX)
4705 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4711 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4712 SelectionDAG &DAG) {
4713 EVT VT = Op.getValueType();
4714 DebugLoc dl = Op.getDebugLoc();
4715 if (VT.getSizeInBits() == 8) {
4716 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, 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.getSizeInBits() == 16) {
4722 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4723 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4725 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4726 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4727 DAG.getNode(ISD::BIT_CONVERT, dl,
4731 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4732 Op.getOperand(0), Op.getOperand(1));
4733 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4734 DAG.getValueType(VT));
4735 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4736 } else if (VT == MVT::f32) {
4737 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4738 // the result back to FR32 register. It's only worth matching if the
4739 // result has a single use which is a store or a bitcast to i32. And in
4740 // the case of a store, it's not worth it if the index is a constant 0,
4741 // because a MOVSSmr can be used instead, which is smaller and faster.
4742 if (!Op.hasOneUse())
4744 SDNode *User = *Op.getNode()->use_begin();
4745 if ((User->getOpcode() != ISD::STORE ||
4746 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4747 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4748 (User->getOpcode() != ISD::BIT_CONVERT ||
4749 User->getValueType(0) != MVT::i32))
4751 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4752 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4755 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4756 } else if (VT == MVT::i32) {
4757 // ExtractPS works with constant index.
4758 if (isa<ConstantSDNode>(Op.getOperand(1)))
4766 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4767 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4770 if (Subtarget->hasSSE41()) {
4771 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4776 EVT VT = Op.getValueType();
4777 DebugLoc dl = Op.getDebugLoc();
4778 // TODO: handle v16i8.
4779 if (VT.getSizeInBits() == 16) {
4780 SDValue Vec = Op.getOperand(0);
4781 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4783 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4784 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4785 DAG.getNode(ISD::BIT_CONVERT, dl,
4788 // Transform it so it match pextrw which produces a 32-bit result.
4789 EVT EltVT = MVT::i32;
4790 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4791 Op.getOperand(0), Op.getOperand(1));
4792 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4793 DAG.getValueType(VT));
4794 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4795 } else if (VT.getSizeInBits() == 32) {
4796 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4800 // SHUFPS the element to the lowest double word, then movss.
4801 int Mask[4] = { Idx, -1, -1, -1 };
4802 EVT VVT = Op.getOperand(0).getValueType();
4803 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4804 DAG.getUNDEF(VVT), Mask);
4805 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4806 DAG.getIntPtrConstant(0));
4807 } else if (VT.getSizeInBits() == 64) {
4808 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4809 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4810 // to match extract_elt for f64.
4811 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4815 // UNPCKHPD the element to the lowest double word, then movsd.
4816 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4817 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4818 int Mask[2] = { 1, -1 };
4819 EVT VVT = Op.getOperand(0).getValueType();
4820 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4821 DAG.getUNDEF(VVT), Mask);
4822 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4823 DAG.getIntPtrConstant(0));
4830 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4831 EVT VT = Op.getValueType();
4832 EVT EltVT = VT.getVectorElementType();
4833 DebugLoc dl = Op.getDebugLoc();
4835 SDValue N0 = Op.getOperand(0);
4836 SDValue N1 = Op.getOperand(1);
4837 SDValue N2 = Op.getOperand(2);
4839 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4840 isa<ConstantSDNode>(N2)) {
4841 unsigned Opc = (EltVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4843 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4845 if (N1.getValueType() != MVT::i32)
4846 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4847 if (N2.getValueType() != MVT::i32)
4848 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4849 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4850 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4851 // Bits [7:6] of the constant are the source select. This will always be
4852 // zero here. The DAG Combiner may combine an extract_elt index into these
4853 // bits. For example (insert (extract, 3), 2) could be matched by putting
4854 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4855 // Bits [5:4] of the constant are the destination select. This is the
4856 // value of the incoming immediate.
4857 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4858 // combine either bitwise AND or insert of float 0.0 to set these bits.
4859 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4860 // Create this as a scalar to vector..
4861 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4862 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4863 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4864 // PINSR* works with constant index.
4871 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4872 EVT VT = Op.getValueType();
4873 EVT EltVT = VT.getVectorElementType();
4875 if (Subtarget->hasSSE41())
4876 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4878 if (EltVT == MVT::i8)
4881 DebugLoc dl = Op.getDebugLoc();
4882 SDValue N0 = Op.getOperand(0);
4883 SDValue N1 = Op.getOperand(1);
4884 SDValue N2 = Op.getOperand(2);
4886 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4887 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4888 // as its second argument.
4889 if (N1.getValueType() != MVT::i32)
4890 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4891 if (N2.getValueType() != MVT::i32)
4892 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4893 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4899 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4900 DebugLoc dl = Op.getDebugLoc();
4901 if (Op.getValueType() == MVT::v2f32)
4902 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4903 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4904 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4905 Op.getOperand(0))));
4907 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4908 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4910 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4911 EVT VT = MVT::v2i32;
4912 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4919 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4920 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4923 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4924 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4925 // one of the above mentioned nodes. It has to be wrapped because otherwise
4926 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4927 // be used to form addressing mode. These wrapped nodes will be selected
4930 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4931 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4933 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4935 unsigned char OpFlag = 0;
4936 unsigned WrapperKind = X86ISD::Wrapper;
4937 CodeModel::Model M = getTargetMachine().getCodeModel();
4939 if (Subtarget->isPICStyleRIPRel() &&
4940 (M == CodeModel::Small || M == CodeModel::Kernel))
4941 WrapperKind = X86ISD::WrapperRIP;
4942 else if (Subtarget->isPICStyleGOT())
4943 OpFlag = X86II::MO_GOTOFF;
4944 else if (Subtarget->isPICStyleStubPIC())
4945 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4947 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4949 CP->getOffset(), OpFlag);
4950 DebugLoc DL = CP->getDebugLoc();
4951 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4952 // With PIC, the address is actually $g + Offset.
4954 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4955 DAG.getNode(X86ISD::GlobalBaseReg,
4956 DebugLoc::getUnknownLoc(), getPointerTy()),
4963 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4964 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4966 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4968 unsigned char OpFlag = 0;
4969 unsigned WrapperKind = X86ISD::Wrapper;
4970 CodeModel::Model M = getTargetMachine().getCodeModel();
4972 if (Subtarget->isPICStyleRIPRel() &&
4973 (M == CodeModel::Small || M == CodeModel::Kernel))
4974 WrapperKind = X86ISD::WrapperRIP;
4975 else if (Subtarget->isPICStyleGOT())
4976 OpFlag = X86II::MO_GOTOFF;
4977 else if (Subtarget->isPICStyleStubPIC())
4978 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4980 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4982 DebugLoc DL = JT->getDebugLoc();
4983 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4985 // With PIC, the address is actually $g + Offset.
4987 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4988 DAG.getNode(X86ISD::GlobalBaseReg,
4989 DebugLoc::getUnknownLoc(), getPointerTy()),
4997 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4998 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5000 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5002 unsigned char OpFlag = 0;
5003 unsigned WrapperKind = X86ISD::Wrapper;
5004 CodeModel::Model M = getTargetMachine().getCodeModel();
5006 if (Subtarget->isPICStyleRIPRel() &&
5007 (M == CodeModel::Small || M == CodeModel::Kernel))
5008 WrapperKind = X86ISD::WrapperRIP;
5009 else if (Subtarget->isPICStyleGOT())
5010 OpFlag = X86II::MO_GOTOFF;
5011 else if (Subtarget->isPICStyleStubPIC())
5012 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5014 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5016 DebugLoc DL = Op.getDebugLoc();
5017 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5020 // With PIC, the address is actually $g + Offset.
5021 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5022 !Subtarget->is64Bit()) {
5023 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5024 DAG.getNode(X86ISD::GlobalBaseReg,
5025 DebugLoc::getUnknownLoc(),
5034 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
5035 // Create the TargetBlockAddressAddress node.
5036 unsigned char OpFlags =
5037 Subtarget->ClassifyBlockAddressReference();
5038 CodeModel::Model M = getTargetMachine().getCodeModel();
5039 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5040 DebugLoc dl = Op.getDebugLoc();
5041 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5042 /*isTarget=*/true, OpFlags);
5044 if (Subtarget->isPICStyleRIPRel() &&
5045 (M == CodeModel::Small || M == CodeModel::Kernel))
5046 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5048 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5050 // With PIC, the address is actually $g + Offset.
5051 if (isGlobalRelativeToPICBase(OpFlags)) {
5052 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5053 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5061 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5063 SelectionDAG &DAG) const {
5064 // Create the TargetGlobalAddress node, folding in the constant
5065 // offset if it is legal.
5066 unsigned char OpFlags =
5067 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5068 CodeModel::Model M = getTargetMachine().getCodeModel();
5070 if (OpFlags == X86II::MO_NO_FLAG &&
5071 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5072 // A direct static reference to a global.
5073 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5076 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5079 if (Subtarget->isPICStyleRIPRel() &&
5080 (M == CodeModel::Small || M == CodeModel::Kernel))
5081 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5083 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5085 // With PIC, the address is actually $g + Offset.
5086 if (isGlobalRelativeToPICBase(OpFlags)) {
5087 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5088 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5092 // For globals that require a load from a stub to get the address, emit the
5094 if (isGlobalStubReference(OpFlags))
5095 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5096 PseudoSourceValue::getGOT(), 0);
5098 // If there was a non-zero offset that we didn't fold, create an explicit
5101 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5102 DAG.getConstant(Offset, getPointerTy()));
5108 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
5109 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5110 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5111 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5115 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5116 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5117 unsigned char OperandFlags) {
5118 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5119 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5120 DebugLoc dl = GA->getDebugLoc();
5121 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5122 GA->getValueType(0),
5126 SDValue Ops[] = { Chain, TGA, *InFlag };
5127 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5129 SDValue Ops[] = { Chain, TGA };
5130 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5133 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5134 MFI->setHasCalls(true);
5136 SDValue Flag = Chain.getValue(1);
5137 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5140 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5142 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5145 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5146 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5147 DAG.getNode(X86ISD::GlobalBaseReg,
5148 DebugLoc::getUnknownLoc(),
5150 InFlag = Chain.getValue(1);
5152 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5155 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5157 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5159 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5160 X86::RAX, X86II::MO_TLSGD);
5163 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5164 // "local exec" model.
5165 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5166 const EVT PtrVT, TLSModel::Model model,
5168 DebugLoc dl = GA->getDebugLoc();
5169 // Get the Thread Pointer
5170 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5171 DebugLoc::getUnknownLoc(), PtrVT,
5172 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5175 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5178 unsigned char OperandFlags = 0;
5179 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5181 unsigned WrapperKind = X86ISD::Wrapper;
5182 if (model == TLSModel::LocalExec) {
5183 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5184 } else if (is64Bit) {
5185 assert(model == TLSModel::InitialExec);
5186 OperandFlags = X86II::MO_GOTTPOFF;
5187 WrapperKind = X86ISD::WrapperRIP;
5189 assert(model == TLSModel::InitialExec);
5190 OperandFlags = X86II::MO_INDNTPOFF;
5193 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5195 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5196 GA->getOffset(), OperandFlags);
5197 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5199 if (model == TLSModel::InitialExec)
5200 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5201 PseudoSourceValue::getGOT(), 0);
5203 // The address of the thread local variable is the add of the thread
5204 // pointer with the offset of the variable.
5205 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5209 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
5210 // TODO: implement the "local dynamic" model
5211 // TODO: implement the "initial exec"model for pic executables
5212 assert(Subtarget->isTargetELF() &&
5213 "TLS not implemented for non-ELF targets");
5214 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5215 const GlobalValue *GV = GA->getGlobal();
5217 // If GV is an alias then use the aliasee for determining
5218 // thread-localness.
5219 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5220 GV = GA->resolveAliasedGlobal(false);
5222 TLSModel::Model model = getTLSModel(GV,
5223 getTargetMachine().getRelocationModel());
5226 case TLSModel::GeneralDynamic:
5227 case TLSModel::LocalDynamic: // not implemented
5228 if (Subtarget->is64Bit())
5229 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5230 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5232 case TLSModel::InitialExec:
5233 case TLSModel::LocalExec:
5234 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5235 Subtarget->is64Bit());
5238 llvm_unreachable("Unreachable");
5243 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5244 /// take a 2 x i32 value to shift plus a shift amount.
5245 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
5246 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5247 EVT VT = Op.getValueType();
5248 unsigned VTBits = VT.getSizeInBits();
5249 DebugLoc dl = Op.getDebugLoc();
5250 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5251 SDValue ShOpLo = Op.getOperand(0);
5252 SDValue ShOpHi = Op.getOperand(1);
5253 SDValue ShAmt = Op.getOperand(2);
5254 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5255 DAG.getConstant(VTBits - 1, MVT::i8))
5256 : DAG.getConstant(0, VT);
5259 if (Op.getOpcode() == ISD::SHL_PARTS) {
5260 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5261 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5263 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5264 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5267 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5268 DAG.getConstant(VTBits, MVT::i8));
5269 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
5270 AndNode, DAG.getConstant(0, MVT::i8));
5273 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5274 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5275 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5277 if (Op.getOpcode() == ISD::SHL_PARTS) {
5278 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5279 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5281 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5282 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5285 SDValue Ops[2] = { Lo, Hi };
5286 return DAG.getMergeValues(Ops, 2, dl);
5289 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5290 EVT SrcVT = Op.getOperand(0).getValueType();
5292 if (SrcVT.isVector()) {
5293 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5299 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5300 "Unknown SINT_TO_FP to lower!");
5302 // These are really Legal; return the operand so the caller accepts it as
5304 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5306 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5307 Subtarget->is64Bit()) {
5311 DebugLoc dl = Op.getDebugLoc();
5312 unsigned Size = SrcVT.getSizeInBits()/8;
5313 MachineFunction &MF = DAG.getMachineFunction();
5314 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5315 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5316 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5318 PseudoSourceValue::getFixedStack(SSFI), 0);
5319 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5322 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5324 SelectionDAG &DAG) {
5326 DebugLoc dl = Op.getDebugLoc();
5328 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5330 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5332 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5333 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5334 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5335 Tys, Ops, array_lengthof(Ops));
5338 Chain = Result.getValue(1);
5339 SDValue InFlag = Result.getValue(2);
5341 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5342 // shouldn't be necessary except that RFP cannot be live across
5343 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5344 MachineFunction &MF = DAG.getMachineFunction();
5345 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5346 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5347 Tys = DAG.getVTList(MVT::Other);
5349 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5351 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5352 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5353 PseudoSourceValue::getFixedStack(SSFI), 0);
5359 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5360 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5361 // This algorithm is not obvious. Here it is in C code, more or less:
5363 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5364 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5365 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5367 // Copy ints to xmm registers.
5368 __m128i xh = _mm_cvtsi32_si128( hi );
5369 __m128i xl = _mm_cvtsi32_si128( lo );
5371 // Combine into low half of a single xmm register.
5372 __m128i x = _mm_unpacklo_epi32( xh, xl );
5376 // Merge in appropriate exponents to give the integer bits the right
5378 x = _mm_unpacklo_epi32( x, exp );
5380 // Subtract away the biases to deal with the IEEE-754 double precision
5382 d = _mm_sub_pd( (__m128d) x, bias );
5384 // All conversions up to here are exact. The correctly rounded result is
5385 // calculated using the current rounding mode using the following
5387 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5388 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5389 // store doesn't really need to be here (except
5390 // maybe to zero the other double)
5395 DebugLoc dl = Op.getDebugLoc();
5396 LLVMContext *Context = DAG.getContext();
5398 // Build some magic constants.
5399 std::vector<Constant*> CV0;
5400 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5401 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5402 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5403 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5404 Constant *C0 = ConstantVector::get(CV0);
5405 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5407 std::vector<Constant*> CV1;
5409 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5411 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5412 Constant *C1 = ConstantVector::get(CV1);
5413 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5415 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5416 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5418 DAG.getIntPtrConstant(1)));
5419 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5420 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5422 DAG.getIntPtrConstant(0)));
5423 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5424 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5425 PseudoSourceValue::getConstantPool(), 0,
5427 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5428 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5429 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5430 PseudoSourceValue::getConstantPool(), 0,
5432 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5434 // Add the halves; easiest way is to swap them into another reg first.
5435 int ShufMask[2] = { 1, -1 };
5436 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5437 DAG.getUNDEF(MVT::v2f64), ShufMask);
5438 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5439 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5440 DAG.getIntPtrConstant(0));
5443 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5444 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5445 DebugLoc dl = Op.getDebugLoc();
5446 // FP constant to bias correct the final result.
5447 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5450 // Load the 32-bit value into an XMM register.
5451 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5452 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5454 DAG.getIntPtrConstant(0)));
5456 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5457 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5458 DAG.getIntPtrConstant(0));
5460 // Or the load with the bias.
5461 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5462 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5463 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5465 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5466 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5467 MVT::v2f64, Bias)));
5468 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5469 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5470 DAG.getIntPtrConstant(0));
5472 // Subtract the bias.
5473 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5475 // Handle final rounding.
5476 EVT DestVT = Op.getValueType();
5478 if (DestVT.bitsLT(MVT::f64)) {
5479 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5480 DAG.getIntPtrConstant(0));
5481 } else if (DestVT.bitsGT(MVT::f64)) {
5482 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5485 // Handle final rounding.
5489 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5490 SDValue N0 = Op.getOperand(0);
5491 DebugLoc dl = Op.getDebugLoc();
5493 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5494 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5495 // the optimization here.
5496 if (DAG.SignBitIsZero(N0))
5497 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5499 EVT SrcVT = N0.getValueType();
5500 if (SrcVT == MVT::i64) {
5501 // We only handle SSE2 f64 target here; caller can expand the rest.
5502 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5505 return LowerUINT_TO_FP_i64(Op, DAG);
5506 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5507 return LowerUINT_TO_FP_i32(Op, DAG);
5510 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5512 // Make a 64-bit buffer, and use it to build an FILD.
5513 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5514 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5515 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5516 getPointerTy(), StackSlot, WordOff);
5517 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5518 StackSlot, NULL, 0);
5519 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5520 OffsetSlot, NULL, 0);
5521 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5524 std::pair<SDValue,SDValue> X86TargetLowering::
5525 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5526 DebugLoc dl = Op.getDebugLoc();
5528 EVT DstTy = Op.getValueType();
5531 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5535 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5536 DstTy.getSimpleVT() >= MVT::i16 &&
5537 "Unknown FP_TO_SINT to lower!");
5539 // These are really Legal.
5540 if (DstTy == MVT::i32 &&
5541 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5542 return std::make_pair(SDValue(), SDValue());
5543 if (Subtarget->is64Bit() &&
5544 DstTy == MVT::i64 &&
5545 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5546 return std::make_pair(SDValue(), SDValue());
5548 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5550 MachineFunction &MF = DAG.getMachineFunction();
5551 unsigned MemSize = DstTy.getSizeInBits()/8;
5552 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5553 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5556 switch (DstTy.getSimpleVT().SimpleTy) {
5557 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5558 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5559 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5560 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5563 SDValue Chain = DAG.getEntryNode();
5564 SDValue Value = Op.getOperand(0);
5565 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5566 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5567 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5568 PseudoSourceValue::getFixedStack(SSFI), 0);
5569 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5571 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5573 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5574 Chain = Value.getValue(1);
5575 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5576 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5579 // Build the FP_TO_INT*_IN_MEM
5580 SDValue Ops[] = { Chain, Value, StackSlot };
5581 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5583 return std::make_pair(FIST, StackSlot);
5586 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5587 if (Op.getValueType().isVector()) {
5588 if (Op.getValueType() == MVT::v2i32 &&
5589 Op.getOperand(0).getValueType() == MVT::v2f64) {
5595 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5596 SDValue FIST = Vals.first, StackSlot = Vals.second;
5597 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5598 if (FIST.getNode() == 0) return Op;
5601 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5602 FIST, StackSlot, NULL, 0);
5605 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5606 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5607 SDValue FIST = Vals.first, StackSlot = Vals.second;
5608 assert(FIST.getNode() && "Unexpected failure");
5611 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5612 FIST, StackSlot, NULL, 0);
5615 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5616 LLVMContext *Context = DAG.getContext();
5617 DebugLoc dl = Op.getDebugLoc();
5618 EVT VT = Op.getValueType();
5621 EltVT = VT.getVectorElementType();
5622 std::vector<Constant*> CV;
5623 if (EltVT == MVT::f64) {
5624 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5628 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5634 Constant *C = ConstantVector::get(CV);
5635 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5636 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5637 PseudoSourceValue::getConstantPool(), 0,
5639 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5642 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5643 LLVMContext *Context = DAG.getContext();
5644 DebugLoc dl = Op.getDebugLoc();
5645 EVT VT = Op.getValueType();
5648 EltVT = VT.getVectorElementType();
5649 std::vector<Constant*> CV;
5650 if (EltVT == MVT::f64) {
5651 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5655 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5661 Constant *C = ConstantVector::get(CV);
5662 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5663 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5664 PseudoSourceValue::getConstantPool(), 0,
5666 if (VT.isVector()) {
5667 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5668 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5669 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5671 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5673 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5677 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5678 LLVMContext *Context = DAG.getContext();
5679 SDValue Op0 = Op.getOperand(0);
5680 SDValue Op1 = Op.getOperand(1);
5681 DebugLoc dl = Op.getDebugLoc();
5682 EVT VT = Op.getValueType();
5683 EVT SrcVT = Op1.getValueType();
5685 // If second operand is smaller, extend it first.
5686 if (SrcVT.bitsLT(VT)) {
5687 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5690 // And if it is bigger, shrink it first.
5691 if (SrcVT.bitsGT(VT)) {
5692 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5696 // At this point the operands and the result should have the same
5697 // type, and that won't be f80 since that is not custom lowered.
5699 // First get the sign bit of second operand.
5700 std::vector<Constant*> CV;
5701 if (SrcVT == MVT::f64) {
5702 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5703 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5705 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5706 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5707 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5708 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5710 Constant *C = ConstantVector::get(CV);
5711 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5712 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5713 PseudoSourceValue::getConstantPool(), 0,
5715 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5717 // Shift sign bit right or left if the two operands have different types.
5718 if (SrcVT.bitsGT(VT)) {
5719 // Op0 is MVT::f32, Op1 is MVT::f64.
5720 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5721 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5722 DAG.getConstant(32, MVT::i32));
5723 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5724 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5725 DAG.getIntPtrConstant(0));
5728 // Clear first operand sign bit.
5730 if (VT == MVT::f64) {
5731 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5732 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5734 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5735 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5736 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5737 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5739 C = ConstantVector::get(CV);
5740 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5741 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5742 PseudoSourceValue::getConstantPool(), 0,
5744 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5746 // Or the value with the sign bit.
5747 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5750 /// Emit nodes that will be selected as "test Op0,Op0", or something
5752 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5753 SelectionDAG &DAG) {
5754 DebugLoc dl = Op.getDebugLoc();
5756 // CF and OF aren't always set the way we want. Determine which
5757 // of these we need.
5758 bool NeedCF = false;
5759 bool NeedOF = false;
5761 case X86::COND_A: case X86::COND_AE:
5762 case X86::COND_B: case X86::COND_BE:
5765 case X86::COND_G: case X86::COND_GE:
5766 case X86::COND_L: case X86::COND_LE:
5767 case X86::COND_O: case X86::COND_NO:
5773 // See if we can use the EFLAGS value from the operand instead of
5774 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5775 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5776 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5777 unsigned Opcode = 0;
5778 unsigned NumOperands = 0;
5779 switch (Op.getNode()->getOpcode()) {
5781 // Due to an isel shortcoming, be conservative if this add is likely to
5782 // be selected as part of a load-modify-store instruction. When the root
5783 // node in a match is a store, isel doesn't know how to remap non-chain
5784 // non-flag uses of other nodes in the match, such as the ADD in this
5785 // case. This leads to the ADD being left around and reselected, with
5786 // the result being two adds in the output.
5787 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5788 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5789 if (UI->getOpcode() == ISD::STORE)
5791 if (ConstantSDNode *C =
5792 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5793 // An add of one will be selected as an INC.
5794 if (C->getAPIntValue() == 1) {
5795 Opcode = X86ISD::INC;
5799 // An add of negative one (subtract of one) will be selected as a DEC.
5800 if (C->getAPIntValue().isAllOnesValue()) {
5801 Opcode = X86ISD::DEC;
5806 // Otherwise use a regular EFLAGS-setting add.
5807 Opcode = X86ISD::ADD;
5811 // If the primary and result isn't used, don't bother using X86ISD::AND,
5812 // because a TEST instruction will be better.
5813 bool NonFlagUse = false;
5814 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5815 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
5817 unsigned UOpNo = UI.getOperandNo();
5818 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
5819 // Look pass truncate.
5820 UOpNo = User->use_begin().getOperandNo();
5821 User = *User->use_begin();
5823 if (User->getOpcode() != ISD::BRCOND &&
5824 User->getOpcode() != ISD::SETCC &&
5825 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
5837 // Due to the ISEL shortcoming noted above, be conservative if this op is
5838 // likely to be selected as part of a load-modify-store instruction.
5839 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5840 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5841 if (UI->getOpcode() == ISD::STORE)
5843 // Otherwise use a regular EFLAGS-setting instruction.
5844 switch (Op.getNode()->getOpcode()) {
5845 case ISD::SUB: Opcode = X86ISD::SUB; break;
5846 case ISD::OR: Opcode = X86ISD::OR; break;
5847 case ISD::XOR: Opcode = X86ISD::XOR; break;
5848 case ISD::AND: Opcode = X86ISD::AND; break;
5849 default: llvm_unreachable("unexpected operator!");
5860 return SDValue(Op.getNode(), 1);
5866 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5867 SmallVector<SDValue, 4> Ops;
5868 for (unsigned i = 0; i != NumOperands; ++i)
5869 Ops.push_back(Op.getOperand(i));
5870 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5871 DAG.ReplaceAllUsesWith(Op, New);
5872 return SDValue(New.getNode(), 1);
5876 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5877 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5878 DAG.getConstant(0, Op.getValueType()));
5881 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5883 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5884 SelectionDAG &DAG) {
5885 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5886 if (C->getAPIntValue() == 0)
5887 return EmitTest(Op0, X86CC, DAG);
5889 DebugLoc dl = Op0.getDebugLoc();
5890 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5893 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
5894 /// if it's possible.
5895 static SDValue LowerToBT(SDValue Op0, ISD::CondCode CC,
5896 DebugLoc dl, SelectionDAG &DAG) {
5898 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5899 if (ConstantSDNode *Op010C =
5900 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5901 if (Op010C->getZExtValue() == 1) {
5902 LHS = Op0.getOperand(0);
5903 RHS = Op0.getOperand(1).getOperand(1);
5905 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5906 if (ConstantSDNode *Op000C =
5907 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5908 if (Op000C->getZExtValue() == 1) {
5909 LHS = Op0.getOperand(1);
5910 RHS = Op0.getOperand(0).getOperand(1);
5912 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5913 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5914 SDValue AndLHS = Op0.getOperand(0);
5915 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5916 LHS = AndLHS.getOperand(0);
5917 RHS = AndLHS.getOperand(1);
5921 if (LHS.getNode()) {
5922 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5923 // instruction. Since the shift amount is in-range-or-undefined, we know
5924 // that doing a bittest on the i16 value is ok. We extend to i32 because
5925 // the encoding for the i16 version is larger than the i32 version.
5926 if (LHS.getValueType() == MVT::i8)
5927 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5929 // If the operand types disagree, extend the shift amount to match. Since
5930 // BT ignores high bits (like shifts) we can use anyextend.
5931 if (LHS.getValueType() != RHS.getValueType())
5932 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5934 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5935 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5936 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5937 DAG.getConstant(Cond, MVT::i8), BT);
5943 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5944 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5945 SDValue Op0 = Op.getOperand(0);
5946 SDValue Op1 = Op.getOperand(1);
5947 DebugLoc dl = Op.getDebugLoc();
5948 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5950 // Optimize to BT if possible.
5951 // Lower (X & (1 << N)) == 0 to BT(X, N).
5952 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5953 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5954 if (Op0.getOpcode() == ISD::AND &&
5956 Op1.getOpcode() == ISD::Constant &&
5957 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5958 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5959 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
5960 if (NewSetCC.getNode())
5964 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5965 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5966 if (X86CC == X86::COND_INVALID)
5969 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5971 // Use sbb x, x to materialize carry bit into a GPR.
5972 if (X86CC == X86::COND_B)
5973 return DAG.getNode(ISD::AND, dl, MVT::i8,
5974 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
5975 DAG.getConstant(X86CC, MVT::i8), Cond),
5976 DAG.getConstant(1, MVT::i8));
5978 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5979 DAG.getConstant(X86CC, MVT::i8), Cond);
5982 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5984 SDValue Op0 = Op.getOperand(0);
5985 SDValue Op1 = Op.getOperand(1);
5986 SDValue CC = Op.getOperand(2);
5987 EVT VT = Op.getValueType();
5988 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5989 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5990 DebugLoc dl = Op.getDebugLoc();
5994 EVT VT0 = Op0.getValueType();
5995 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5996 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5999 switch (SetCCOpcode) {
6002 case ISD::SETEQ: SSECC = 0; break;
6004 case ISD::SETGT: Swap = true; // Fallthrough
6006 case ISD::SETOLT: SSECC = 1; break;
6008 case ISD::SETGE: Swap = true; // Fallthrough
6010 case ISD::SETOLE: SSECC = 2; break;
6011 case ISD::SETUO: SSECC = 3; break;
6013 case ISD::SETNE: SSECC = 4; break;
6014 case ISD::SETULE: Swap = true;
6015 case ISD::SETUGE: SSECC = 5; break;
6016 case ISD::SETULT: Swap = true;
6017 case ISD::SETUGT: SSECC = 6; break;
6018 case ISD::SETO: SSECC = 7; break;
6021 std::swap(Op0, Op1);
6023 // In the two special cases we can't handle, emit two comparisons.
6025 if (SetCCOpcode == ISD::SETUEQ) {
6027 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6028 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6029 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6031 else if (SetCCOpcode == ISD::SETONE) {
6033 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6034 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6035 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6037 llvm_unreachable("Illegal FP comparison");
6039 // Handle all other FP comparisons here.
6040 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6043 // We are handling one of the integer comparisons here. Since SSE only has
6044 // GT and EQ comparisons for integer, swapping operands and multiple
6045 // operations may be required for some comparisons.
6046 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6047 bool Swap = false, Invert = false, FlipSigns = false;
6049 switch (VT.getSimpleVT().SimpleTy) {
6052 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6054 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6056 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6057 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6060 switch (SetCCOpcode) {
6062 case ISD::SETNE: Invert = true;
6063 case ISD::SETEQ: Opc = EQOpc; break;
6064 case ISD::SETLT: Swap = true;
6065 case ISD::SETGT: Opc = GTOpc; break;
6066 case ISD::SETGE: Swap = true;
6067 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6068 case ISD::SETULT: Swap = true;
6069 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6070 case ISD::SETUGE: Swap = true;
6071 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6074 std::swap(Op0, Op1);
6076 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6077 // bits of the inputs before performing those operations.
6079 EVT EltVT = VT.getVectorElementType();
6080 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6082 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6083 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6085 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6086 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6089 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6091 // If the logical-not of the result is required, perform that now.
6093 Result = DAG.getNOT(dl, Result, VT);
6098 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6099 static bool isX86LogicalCmp(SDValue Op) {
6100 unsigned Opc = Op.getNode()->getOpcode();
6101 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6103 if (Op.getResNo() == 1 &&
6104 (Opc == X86ISD::ADD ||
6105 Opc == X86ISD::SUB ||
6106 Opc == X86ISD::SMUL ||
6107 Opc == X86ISD::UMUL ||
6108 Opc == X86ISD::INC ||
6109 Opc == X86ISD::DEC ||
6110 Opc == X86ISD::OR ||
6111 Opc == X86ISD::XOR ||
6112 Opc == X86ISD::AND))
6118 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
6119 bool addTest = true;
6120 SDValue Cond = Op.getOperand(0);
6121 DebugLoc dl = Op.getDebugLoc();
6124 if (Cond.getOpcode() == ISD::SETCC) {
6125 SDValue NewCond = LowerSETCC(Cond, DAG);
6126 if (NewCond.getNode())
6130 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6131 SDValue Op1 = Op.getOperand(1);
6132 SDValue Op2 = Op.getOperand(2);
6133 if (Cond.getOpcode() == X86ISD::SETCC &&
6134 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6135 SDValue Cmp = Cond.getOperand(1);
6136 if (Cmp.getOpcode() == X86ISD::CMP) {
6137 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6138 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6139 ConstantSDNode *RHSC =
6140 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6141 if (N1C && N1C->isAllOnesValue() &&
6142 N2C && N2C->isNullValue() &&
6143 RHSC && RHSC->isNullValue()) {
6144 SDValue CmpOp0 = Cmp.getOperand(0);
6145 Cmp = DAG.getNode(X86ISD::CMP, dl, CmpOp0.getValueType(),
6146 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6147 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6148 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6153 // Look pass (and (setcc_carry (cmp ...)), 1).
6154 if (Cond.getOpcode() == ISD::AND &&
6155 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6156 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6157 if (C && C->getAPIntValue() == 1)
6158 Cond = Cond.getOperand(0);
6161 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6162 // setting operand in place of the X86ISD::SETCC.
6163 if (Cond.getOpcode() == X86ISD::SETCC ||
6164 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6165 CC = Cond.getOperand(0);
6167 SDValue Cmp = Cond.getOperand(1);
6168 unsigned Opc = Cmp.getOpcode();
6169 EVT VT = Op.getValueType();
6171 bool IllegalFPCMov = false;
6172 if (VT.isFloatingPoint() && !VT.isVector() &&
6173 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6174 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6176 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6177 Opc == X86ISD::BT) { // FIXME
6184 // Look pass the truncate.
6185 if (Cond.getOpcode() == ISD::TRUNCATE)
6186 Cond = Cond.getOperand(0);
6188 // We know the result of AND is compared against zero. Try to match
6190 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6191 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6192 if (NewSetCC.getNode()) {
6193 CC = NewSetCC.getOperand(0);
6194 Cond = NewSetCC.getOperand(1);
6201 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6202 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6205 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6206 // condition is true.
6207 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6208 SDValue Ops[] = { Op2, Op1, CC, Cond };
6209 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6212 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6213 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6214 // from the AND / OR.
6215 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6216 Opc = Op.getOpcode();
6217 if (Opc != ISD::OR && Opc != ISD::AND)
6219 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6220 Op.getOperand(0).hasOneUse() &&
6221 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6222 Op.getOperand(1).hasOneUse());
6225 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6226 // 1 and that the SETCC node has a single use.
6227 static bool isXor1OfSetCC(SDValue Op) {
6228 if (Op.getOpcode() != ISD::XOR)
6230 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6231 if (N1C && N1C->getAPIntValue() == 1) {
6232 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6233 Op.getOperand(0).hasOneUse();
6238 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
6239 bool addTest = true;
6240 SDValue Chain = Op.getOperand(0);
6241 SDValue Cond = Op.getOperand(1);
6242 SDValue Dest = Op.getOperand(2);
6243 DebugLoc dl = Op.getDebugLoc();
6246 if (Cond.getOpcode() == ISD::SETCC) {
6247 SDValue NewCond = LowerSETCC(Cond, DAG);
6248 if (NewCond.getNode())
6252 // FIXME: LowerXALUO doesn't handle these!!
6253 else if (Cond.getOpcode() == X86ISD::ADD ||
6254 Cond.getOpcode() == X86ISD::SUB ||
6255 Cond.getOpcode() == X86ISD::SMUL ||
6256 Cond.getOpcode() == X86ISD::UMUL)
6257 Cond = LowerXALUO(Cond, DAG);
6260 // Look pass (and (setcc_carry (cmp ...)), 1).
6261 if (Cond.getOpcode() == ISD::AND &&
6262 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6263 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6264 if (C && C->getAPIntValue() == 1)
6265 Cond = Cond.getOperand(0);
6268 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6269 // setting operand in place of the X86ISD::SETCC.
6270 if (Cond.getOpcode() == X86ISD::SETCC ||
6271 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6272 CC = Cond.getOperand(0);
6274 SDValue Cmp = Cond.getOperand(1);
6275 unsigned Opc = Cmp.getOpcode();
6276 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6277 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6281 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6285 // These can only come from an arithmetic instruction with overflow,
6286 // e.g. SADDO, UADDO.
6287 Cond = Cond.getNode()->getOperand(1);
6294 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6295 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6296 if (CondOpc == ISD::OR) {
6297 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6298 // two branches instead of an explicit OR instruction with a
6300 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6301 isX86LogicalCmp(Cmp)) {
6302 CC = Cond.getOperand(0).getOperand(0);
6303 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6304 Chain, Dest, CC, Cmp);
6305 CC = Cond.getOperand(1).getOperand(0);
6309 } else { // ISD::AND
6310 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6311 // two branches instead of an explicit AND instruction with a
6312 // separate test. However, we only do this if this block doesn't
6313 // have a fall-through edge, because this requires an explicit
6314 // jmp when the condition is false.
6315 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6316 isX86LogicalCmp(Cmp) &&
6317 Op.getNode()->hasOneUse()) {
6318 X86::CondCode CCode =
6319 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6320 CCode = X86::GetOppositeBranchCondition(CCode);
6321 CC = DAG.getConstant(CCode, MVT::i8);
6322 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6323 // Look for an unconditional branch following this conditional branch.
6324 // We need this because we need to reverse the successors in order
6325 // to implement FCMP_OEQ.
6326 if (User.getOpcode() == ISD::BR) {
6327 SDValue FalseBB = User.getOperand(1);
6329 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6330 assert(NewBR == User);
6333 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6334 Chain, Dest, CC, Cmp);
6335 X86::CondCode CCode =
6336 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6337 CCode = X86::GetOppositeBranchCondition(CCode);
6338 CC = DAG.getConstant(CCode, MVT::i8);
6344 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6345 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6346 // It should be transformed during dag combiner except when the condition
6347 // is set by a arithmetics with overflow node.
6348 X86::CondCode CCode =
6349 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6350 CCode = X86::GetOppositeBranchCondition(CCode);
6351 CC = DAG.getConstant(CCode, MVT::i8);
6352 Cond = Cond.getOperand(0).getOperand(1);
6358 // Look pass the truncate.
6359 if (Cond.getOpcode() == ISD::TRUNCATE)
6360 Cond = Cond.getOperand(0);
6362 // We know the result of AND is compared against zero. Try to match
6364 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6365 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6366 if (NewSetCC.getNode()) {
6367 CC = NewSetCC.getOperand(0);
6368 Cond = NewSetCC.getOperand(1);
6375 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6376 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6378 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6379 Chain, Dest, CC, Cond);
6383 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6384 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6385 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6386 // that the guard pages used by the OS virtual memory manager are allocated in
6387 // correct sequence.
6389 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6390 SelectionDAG &DAG) {
6391 assert(Subtarget->isTargetCygMing() &&
6392 "This should be used only on Cygwin/Mingw targets");
6393 DebugLoc dl = Op.getDebugLoc();
6396 SDValue Chain = Op.getOperand(0);
6397 SDValue Size = Op.getOperand(1);
6398 // FIXME: Ensure alignment here
6402 EVT IntPtr = getPointerTy();
6403 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6405 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
6407 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6408 Flag = Chain.getValue(1);
6410 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6411 SDValue Ops[] = { Chain,
6412 DAG.getTargetExternalSymbol("_alloca", IntPtr),
6413 DAG.getRegister(X86::EAX, IntPtr),
6414 DAG.getRegister(X86StackPtr, SPTy),
6416 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
6417 Flag = Chain.getValue(1);
6419 Chain = DAG.getCALLSEQ_END(Chain,
6420 DAG.getIntPtrConstant(0, true),
6421 DAG.getIntPtrConstant(0, true),
6424 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6426 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6427 return DAG.getMergeValues(Ops1, 2, dl);
6431 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6433 SDValue Dst, SDValue Src,
6434 SDValue Size, unsigned Align,
6436 uint64_t DstSVOff) {
6437 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6439 // If not DWORD aligned or size is more than the threshold, call the library.
6440 // The libc version is likely to be faster for these cases. It can use the
6441 // address value and run time information about the CPU.
6442 if ((Align & 3) != 0 ||
6444 ConstantSize->getZExtValue() >
6445 getSubtarget()->getMaxInlineSizeThreshold()) {
6446 SDValue InFlag(0, 0);
6448 // Check to see if there is a specialized entry-point for memory zeroing.
6449 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6451 if (const char *bzeroEntry = V &&
6452 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6453 EVT IntPtr = getPointerTy();
6454 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6455 TargetLowering::ArgListTy Args;
6456 TargetLowering::ArgListEntry Entry;
6458 Entry.Ty = IntPtrTy;
6459 Args.push_back(Entry);
6461 Args.push_back(Entry);
6462 std::pair<SDValue,SDValue> CallResult =
6463 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6464 false, false, false, false,
6465 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6466 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl,
6467 DAG.GetOrdering(Chain.getNode()));
6468 return CallResult.second;
6471 // Otherwise have the target-independent code call memset.
6475 uint64_t SizeVal = ConstantSize->getZExtValue();
6476 SDValue InFlag(0, 0);
6479 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6480 unsigned BytesLeft = 0;
6481 bool TwoRepStos = false;
6484 uint64_t Val = ValC->getZExtValue() & 255;
6486 // If the value is a constant, then we can potentially use larger sets.
6487 switch (Align & 3) {
6488 case 2: // WORD aligned
6491 Val = (Val << 8) | Val;
6493 case 0: // DWORD aligned
6496 Val = (Val << 8) | Val;
6497 Val = (Val << 16) | Val;
6498 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6501 Val = (Val << 32) | Val;
6504 default: // Byte aligned
6507 Count = DAG.getIntPtrConstant(SizeVal);
6511 if (AVT.bitsGT(MVT::i8)) {
6512 unsigned UBytes = AVT.getSizeInBits() / 8;
6513 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6514 BytesLeft = SizeVal % UBytes;
6517 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6519 InFlag = Chain.getValue(1);
6522 Count = DAG.getIntPtrConstant(SizeVal);
6523 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6524 InFlag = Chain.getValue(1);
6527 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6530 InFlag = Chain.getValue(1);
6531 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6534 InFlag = Chain.getValue(1);
6536 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6537 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6538 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6541 InFlag = Chain.getValue(1);
6543 EVT CVT = Count.getValueType();
6544 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6545 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6546 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6549 InFlag = Chain.getValue(1);
6550 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6551 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
6552 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6553 } else if (BytesLeft) {
6554 // Handle the last 1 - 7 bytes.
6555 unsigned Offset = SizeVal - BytesLeft;
6556 EVT AddrVT = Dst.getValueType();
6557 EVT SizeVT = Size.getValueType();
6559 Chain = DAG.getMemset(Chain, dl,
6560 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6561 DAG.getConstant(Offset, AddrVT)),
6563 DAG.getConstant(BytesLeft, SizeVT),
6564 Align, DstSV, DstSVOff + Offset);
6567 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6572 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6573 SDValue Chain, SDValue Dst, SDValue Src,
6574 SDValue Size, unsigned Align,
6576 const Value *DstSV, uint64_t DstSVOff,
6577 const Value *SrcSV, uint64_t SrcSVOff) {
6578 // This requires the copy size to be a constant, preferrably
6579 // within a subtarget-specific limit.
6580 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6583 uint64_t SizeVal = ConstantSize->getZExtValue();
6584 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6587 /// If not DWORD aligned, call the library.
6588 if ((Align & 3) != 0)
6593 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6596 unsigned UBytes = AVT.getSizeInBits() / 8;
6597 unsigned CountVal = SizeVal / UBytes;
6598 SDValue Count = DAG.getIntPtrConstant(CountVal);
6599 unsigned BytesLeft = SizeVal % UBytes;
6601 SDValue InFlag(0, 0);
6602 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6605 InFlag = Chain.getValue(1);
6606 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6609 InFlag = Chain.getValue(1);
6610 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6613 InFlag = Chain.getValue(1);
6615 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6616 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6617 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
6618 array_lengthof(Ops));
6620 SmallVector<SDValue, 4> Results;
6621 Results.push_back(RepMovs);
6623 // Handle the last 1 - 7 bytes.
6624 unsigned Offset = SizeVal - BytesLeft;
6625 EVT DstVT = Dst.getValueType();
6626 EVT SrcVT = Src.getValueType();
6627 EVT SizeVT = Size.getValueType();
6628 Results.push_back(DAG.getMemcpy(Chain, dl,
6629 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6630 DAG.getConstant(Offset, DstVT)),
6631 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6632 DAG.getConstant(Offset, SrcVT)),
6633 DAG.getConstant(BytesLeft, SizeVT),
6634 Align, AlwaysInline,
6635 DstSV, DstSVOff + Offset,
6636 SrcSV, SrcSVOff + Offset));
6639 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6640 &Results[0], Results.size());
6643 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6644 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6645 DebugLoc dl = Op.getDebugLoc();
6647 if (!Subtarget->is64Bit()) {
6648 // vastart just stores the address of the VarArgsFrameIndex slot into the
6649 // memory location argument.
6650 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6651 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6655 // gp_offset (0 - 6 * 8)
6656 // fp_offset (48 - 48 + 8 * 16)
6657 // overflow_arg_area (point to parameters coming in memory).
6659 SmallVector<SDValue, 8> MemOps;
6660 SDValue FIN = Op.getOperand(1);
6662 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6663 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6665 MemOps.push_back(Store);
6668 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6669 FIN, DAG.getIntPtrConstant(4));
6670 Store = DAG.getStore(Op.getOperand(0), dl,
6671 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6673 MemOps.push_back(Store);
6675 // Store ptr to overflow_arg_area
6676 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6677 FIN, DAG.getIntPtrConstant(4));
6678 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6679 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6680 MemOps.push_back(Store);
6682 // Store ptr to reg_save_area.
6683 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6684 FIN, DAG.getIntPtrConstant(8));
6685 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6686 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6687 MemOps.push_back(Store);
6688 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6689 &MemOps[0], MemOps.size());
6692 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6693 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6694 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6695 SDValue Chain = Op.getOperand(0);
6696 SDValue SrcPtr = Op.getOperand(1);
6697 SDValue SrcSV = Op.getOperand(2);
6699 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6703 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6704 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6705 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6706 SDValue Chain = Op.getOperand(0);
6707 SDValue DstPtr = Op.getOperand(1);
6708 SDValue SrcPtr = Op.getOperand(2);
6709 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6710 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6711 DebugLoc dl = Op.getDebugLoc();
6713 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6714 DAG.getIntPtrConstant(24), 8, false,
6715 DstSV, 0, SrcSV, 0);
6719 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6720 DebugLoc dl = Op.getDebugLoc();
6721 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6723 default: return SDValue(); // Don't custom lower most intrinsics.
6724 // Comparison intrinsics.
6725 case Intrinsic::x86_sse_comieq_ss:
6726 case Intrinsic::x86_sse_comilt_ss:
6727 case Intrinsic::x86_sse_comile_ss:
6728 case Intrinsic::x86_sse_comigt_ss:
6729 case Intrinsic::x86_sse_comige_ss:
6730 case Intrinsic::x86_sse_comineq_ss:
6731 case Intrinsic::x86_sse_ucomieq_ss:
6732 case Intrinsic::x86_sse_ucomilt_ss:
6733 case Intrinsic::x86_sse_ucomile_ss:
6734 case Intrinsic::x86_sse_ucomigt_ss:
6735 case Intrinsic::x86_sse_ucomige_ss:
6736 case Intrinsic::x86_sse_ucomineq_ss:
6737 case Intrinsic::x86_sse2_comieq_sd:
6738 case Intrinsic::x86_sse2_comilt_sd:
6739 case Intrinsic::x86_sse2_comile_sd:
6740 case Intrinsic::x86_sse2_comigt_sd:
6741 case Intrinsic::x86_sse2_comige_sd:
6742 case Intrinsic::x86_sse2_comineq_sd:
6743 case Intrinsic::x86_sse2_ucomieq_sd:
6744 case Intrinsic::x86_sse2_ucomilt_sd:
6745 case Intrinsic::x86_sse2_ucomile_sd:
6746 case Intrinsic::x86_sse2_ucomigt_sd:
6747 case Intrinsic::x86_sse2_ucomige_sd:
6748 case Intrinsic::x86_sse2_ucomineq_sd: {
6750 ISD::CondCode CC = ISD::SETCC_INVALID;
6753 case Intrinsic::x86_sse_comieq_ss:
6754 case Intrinsic::x86_sse2_comieq_sd:
6758 case Intrinsic::x86_sse_comilt_ss:
6759 case Intrinsic::x86_sse2_comilt_sd:
6763 case Intrinsic::x86_sse_comile_ss:
6764 case Intrinsic::x86_sse2_comile_sd:
6768 case Intrinsic::x86_sse_comigt_ss:
6769 case Intrinsic::x86_sse2_comigt_sd:
6773 case Intrinsic::x86_sse_comige_ss:
6774 case Intrinsic::x86_sse2_comige_sd:
6778 case Intrinsic::x86_sse_comineq_ss:
6779 case Intrinsic::x86_sse2_comineq_sd:
6783 case Intrinsic::x86_sse_ucomieq_ss:
6784 case Intrinsic::x86_sse2_ucomieq_sd:
6785 Opc = X86ISD::UCOMI;
6788 case Intrinsic::x86_sse_ucomilt_ss:
6789 case Intrinsic::x86_sse2_ucomilt_sd:
6790 Opc = X86ISD::UCOMI;
6793 case Intrinsic::x86_sse_ucomile_ss:
6794 case Intrinsic::x86_sse2_ucomile_sd:
6795 Opc = X86ISD::UCOMI;
6798 case Intrinsic::x86_sse_ucomigt_ss:
6799 case Intrinsic::x86_sse2_ucomigt_sd:
6800 Opc = X86ISD::UCOMI;
6803 case Intrinsic::x86_sse_ucomige_ss:
6804 case Intrinsic::x86_sse2_ucomige_sd:
6805 Opc = X86ISD::UCOMI;
6808 case Intrinsic::x86_sse_ucomineq_ss:
6809 case Intrinsic::x86_sse2_ucomineq_sd:
6810 Opc = X86ISD::UCOMI;
6815 SDValue LHS = Op.getOperand(1);
6816 SDValue RHS = Op.getOperand(2);
6817 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6818 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6819 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6820 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6821 DAG.getConstant(X86CC, MVT::i8), Cond);
6822 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6824 // ptest intrinsics. The intrinsic these come from are designed to return
6825 // an integer value, not just an instruction so lower it to the ptest
6826 // pattern and a setcc for the result.
6827 case Intrinsic::x86_sse41_ptestz:
6828 case Intrinsic::x86_sse41_ptestc:
6829 case Intrinsic::x86_sse41_ptestnzc:{
6832 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6833 case Intrinsic::x86_sse41_ptestz:
6835 X86CC = X86::COND_E;
6837 case Intrinsic::x86_sse41_ptestc:
6839 X86CC = X86::COND_B;
6841 case Intrinsic::x86_sse41_ptestnzc:
6843 X86CC = X86::COND_A;
6847 SDValue LHS = Op.getOperand(1);
6848 SDValue RHS = Op.getOperand(2);
6849 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6850 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6851 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6852 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6855 // Fix vector shift instructions where the last operand is a non-immediate
6857 case Intrinsic::x86_sse2_pslli_w:
6858 case Intrinsic::x86_sse2_pslli_d:
6859 case Intrinsic::x86_sse2_pslli_q:
6860 case Intrinsic::x86_sse2_psrli_w:
6861 case Intrinsic::x86_sse2_psrli_d:
6862 case Intrinsic::x86_sse2_psrli_q:
6863 case Intrinsic::x86_sse2_psrai_w:
6864 case Intrinsic::x86_sse2_psrai_d:
6865 case Intrinsic::x86_mmx_pslli_w:
6866 case Intrinsic::x86_mmx_pslli_d:
6867 case Intrinsic::x86_mmx_pslli_q:
6868 case Intrinsic::x86_mmx_psrli_w:
6869 case Intrinsic::x86_mmx_psrli_d:
6870 case Intrinsic::x86_mmx_psrli_q:
6871 case Intrinsic::x86_mmx_psrai_w:
6872 case Intrinsic::x86_mmx_psrai_d: {
6873 SDValue ShAmt = Op.getOperand(2);
6874 if (isa<ConstantSDNode>(ShAmt))
6877 unsigned NewIntNo = 0;
6878 EVT ShAmtVT = MVT::v4i32;
6880 case Intrinsic::x86_sse2_pslli_w:
6881 NewIntNo = Intrinsic::x86_sse2_psll_w;
6883 case Intrinsic::x86_sse2_pslli_d:
6884 NewIntNo = Intrinsic::x86_sse2_psll_d;
6886 case Intrinsic::x86_sse2_pslli_q:
6887 NewIntNo = Intrinsic::x86_sse2_psll_q;
6889 case Intrinsic::x86_sse2_psrli_w:
6890 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6892 case Intrinsic::x86_sse2_psrli_d:
6893 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6895 case Intrinsic::x86_sse2_psrli_q:
6896 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6898 case Intrinsic::x86_sse2_psrai_w:
6899 NewIntNo = Intrinsic::x86_sse2_psra_w;
6901 case Intrinsic::x86_sse2_psrai_d:
6902 NewIntNo = Intrinsic::x86_sse2_psra_d;
6905 ShAmtVT = MVT::v2i32;
6907 case Intrinsic::x86_mmx_pslli_w:
6908 NewIntNo = Intrinsic::x86_mmx_psll_w;
6910 case Intrinsic::x86_mmx_pslli_d:
6911 NewIntNo = Intrinsic::x86_mmx_psll_d;
6913 case Intrinsic::x86_mmx_pslli_q:
6914 NewIntNo = Intrinsic::x86_mmx_psll_q;
6916 case Intrinsic::x86_mmx_psrli_w:
6917 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6919 case Intrinsic::x86_mmx_psrli_d:
6920 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6922 case Intrinsic::x86_mmx_psrli_q:
6923 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6925 case Intrinsic::x86_mmx_psrai_w:
6926 NewIntNo = Intrinsic::x86_mmx_psra_w;
6928 case Intrinsic::x86_mmx_psrai_d:
6929 NewIntNo = Intrinsic::x86_mmx_psra_d;
6931 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6937 // The vector shift intrinsics with scalars uses 32b shift amounts but
6938 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
6942 ShOps[1] = DAG.getConstant(0, MVT::i32);
6943 if (ShAmtVT == MVT::v4i32) {
6944 ShOps[2] = DAG.getUNDEF(MVT::i32);
6945 ShOps[3] = DAG.getUNDEF(MVT::i32);
6946 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
6948 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
6951 EVT VT = Op.getValueType();
6952 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
6953 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6954 DAG.getConstant(NewIntNo, MVT::i32),
6955 Op.getOperand(1), ShAmt);
6960 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6961 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6962 DebugLoc dl = Op.getDebugLoc();
6965 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6967 DAG.getConstant(TD->getPointerSize(),
6968 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6969 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6970 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6975 // Just load the return address.
6976 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6977 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6978 RetAddrFI, NULL, 0);
6981 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6982 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6983 MFI->setFrameAddressIsTaken(true);
6984 EVT VT = Op.getValueType();
6985 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6986 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6987 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6988 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6990 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6994 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6995 SelectionDAG &DAG) {
6996 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6999 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
7001 MachineFunction &MF = DAG.getMachineFunction();
7002 SDValue Chain = Op.getOperand(0);
7003 SDValue Offset = Op.getOperand(1);
7004 SDValue Handler = Op.getOperand(2);
7005 DebugLoc dl = Op.getDebugLoc();
7007 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7009 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7011 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7012 DAG.getIntPtrConstant(-TD->getPointerSize()));
7013 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7014 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
7015 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7016 MF.getRegInfo().addLiveOut(StoreAddrReg);
7018 return DAG.getNode(X86ISD::EH_RETURN, dl,
7020 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7023 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7024 SelectionDAG &DAG) {
7025 SDValue Root = Op.getOperand(0);
7026 SDValue Trmp = Op.getOperand(1); // trampoline
7027 SDValue FPtr = Op.getOperand(2); // nested function
7028 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7029 DebugLoc dl = Op.getDebugLoc();
7031 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7033 if (Subtarget->is64Bit()) {
7034 SDValue OutChains[6];
7036 // Large code-model.
7037 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7038 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7040 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7041 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7043 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7045 // Load the pointer to the nested function into R11.
7046 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7047 SDValue Addr = Trmp;
7048 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7051 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7052 DAG.getConstant(2, MVT::i64));
7053 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
7055 // Load the 'nest' parameter value into R10.
7056 // R10 is specified in X86CallingConv.td
7057 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7058 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7059 DAG.getConstant(10, MVT::i64));
7060 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7061 Addr, TrmpAddr, 10);
7063 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7064 DAG.getConstant(12, MVT::i64));
7065 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
7067 // Jump to the nested function.
7068 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7069 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7070 DAG.getConstant(20, MVT::i64));
7071 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7072 Addr, TrmpAddr, 20);
7074 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7075 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7076 DAG.getConstant(22, MVT::i64));
7077 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7081 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7082 return DAG.getMergeValues(Ops, 2, dl);
7084 const Function *Func =
7085 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7086 CallingConv::ID CC = Func->getCallingConv();
7091 llvm_unreachable("Unsupported calling convention");
7092 case CallingConv::C:
7093 case CallingConv::X86_StdCall: {
7094 // Pass 'nest' parameter in ECX.
7095 // Must be kept in sync with X86CallingConv.td
7098 // Check that ECX wasn't needed by an 'inreg' parameter.
7099 const FunctionType *FTy = Func->getFunctionType();
7100 const AttrListPtr &Attrs = Func->getAttributes();
7102 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7103 unsigned InRegCount = 0;
7106 for (FunctionType::param_iterator I = FTy->param_begin(),
7107 E = FTy->param_end(); I != E; ++I, ++Idx)
7108 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7109 // FIXME: should only count parameters that are lowered to integers.
7110 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7112 if (InRegCount > 2) {
7113 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
7118 case CallingConv::X86_FastCall:
7119 case CallingConv::Fast:
7120 // Pass 'nest' parameter in EAX.
7121 // Must be kept in sync with X86CallingConv.td
7126 SDValue OutChains[4];
7129 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7130 DAG.getConstant(10, MVT::i32));
7131 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7133 // This is storing the opcode for MOV32ri.
7134 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7135 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7136 OutChains[0] = DAG.getStore(Root, dl,
7137 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7140 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7141 DAG.getConstant(1, MVT::i32));
7142 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
7144 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7145 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7146 DAG.getConstant(5, MVT::i32));
7147 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7148 TrmpAddr, 5, false, 1);
7150 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7151 DAG.getConstant(6, MVT::i32));
7152 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
7155 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7156 return DAG.getMergeValues(Ops, 2, dl);
7160 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
7162 The rounding mode is in bits 11:10 of FPSR, and has the following
7169 FLT_ROUNDS, on the other hand, expects the following:
7176 To perform the conversion, we do:
7177 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7180 MachineFunction &MF = DAG.getMachineFunction();
7181 const TargetMachine &TM = MF.getTarget();
7182 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7183 unsigned StackAlignment = TFI.getStackAlignment();
7184 EVT VT = Op.getValueType();
7185 DebugLoc dl = Op.getDebugLoc();
7187 // Save FP Control Word to stack slot
7188 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7189 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7191 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7192 DAG.getEntryNode(), StackSlot);
7194 // Load FP Control Word from stack slot
7195 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
7197 // Transform as necessary
7199 DAG.getNode(ISD::SRL, dl, MVT::i16,
7200 DAG.getNode(ISD::AND, dl, MVT::i16,
7201 CWD, DAG.getConstant(0x800, MVT::i16)),
7202 DAG.getConstant(11, MVT::i8));
7204 DAG.getNode(ISD::SRL, dl, MVT::i16,
7205 DAG.getNode(ISD::AND, dl, MVT::i16,
7206 CWD, DAG.getConstant(0x400, MVT::i16)),
7207 DAG.getConstant(9, MVT::i8));
7210 DAG.getNode(ISD::AND, dl, MVT::i16,
7211 DAG.getNode(ISD::ADD, dl, MVT::i16,
7212 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7213 DAG.getConstant(1, MVT::i16)),
7214 DAG.getConstant(3, MVT::i16));
7217 return DAG.getNode((VT.getSizeInBits() < 16 ?
7218 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7221 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
7222 EVT VT = Op.getValueType();
7224 unsigned NumBits = VT.getSizeInBits();
7225 DebugLoc dl = Op.getDebugLoc();
7227 Op = Op.getOperand(0);
7228 if (VT == MVT::i8) {
7229 // Zero extend to i32 since there is not an i8 bsr.
7231 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7234 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7235 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7236 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7238 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7241 DAG.getConstant(NumBits+NumBits-1, OpVT),
7242 DAG.getConstant(X86::COND_E, MVT::i8),
7245 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7247 // Finally xor with NumBits-1.
7248 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7251 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7255 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
7256 EVT VT = Op.getValueType();
7258 unsigned NumBits = VT.getSizeInBits();
7259 DebugLoc dl = Op.getDebugLoc();
7261 Op = Op.getOperand(0);
7262 if (VT == MVT::i8) {
7264 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7267 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7268 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7269 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7271 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7274 DAG.getConstant(NumBits, OpVT),
7275 DAG.getConstant(X86::COND_E, MVT::i8),
7278 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7281 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7285 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
7286 EVT VT = Op.getValueType();
7287 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7288 DebugLoc dl = Op.getDebugLoc();
7290 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7291 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7292 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7293 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7294 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7296 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7297 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7298 // return AloBlo + AloBhi + AhiBlo;
7300 SDValue A = Op.getOperand(0);
7301 SDValue B = Op.getOperand(1);
7303 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7304 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7305 A, DAG.getConstant(32, MVT::i32));
7306 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7307 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7308 B, DAG.getConstant(32, MVT::i32));
7309 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7310 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7312 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7313 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7315 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7316 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7318 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7319 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7320 AloBhi, DAG.getConstant(32, MVT::i32));
7321 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7322 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7323 AhiBlo, DAG.getConstant(32, MVT::i32));
7324 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7325 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7330 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
7331 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7332 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7333 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7334 // has only one use.
7335 SDNode *N = Op.getNode();
7336 SDValue LHS = N->getOperand(0);
7337 SDValue RHS = N->getOperand(1);
7338 unsigned BaseOp = 0;
7340 DebugLoc dl = Op.getDebugLoc();
7342 switch (Op.getOpcode()) {
7343 default: llvm_unreachable("Unknown ovf instruction!");
7345 // A subtract of one will be selected as a INC. Note that INC doesn't
7346 // set CF, so we can't do this for UADDO.
7347 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7348 if (C->getAPIntValue() == 1) {
7349 BaseOp = X86ISD::INC;
7353 BaseOp = X86ISD::ADD;
7357 BaseOp = X86ISD::ADD;
7361 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7362 // set CF, so we can't do this for USUBO.
7363 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7364 if (C->getAPIntValue() == 1) {
7365 BaseOp = X86ISD::DEC;
7369 BaseOp = X86ISD::SUB;
7373 BaseOp = X86ISD::SUB;
7377 BaseOp = X86ISD::SMUL;
7381 BaseOp = X86ISD::UMUL;
7386 // Also sets EFLAGS.
7387 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7388 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7391 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7392 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7394 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7398 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
7399 EVT T = Op.getValueType();
7400 DebugLoc dl = Op.getDebugLoc();
7403 switch(T.getSimpleVT().SimpleTy) {
7405 assert(false && "Invalid value type!");
7406 case MVT::i8: Reg = X86::AL; size = 1; break;
7407 case MVT::i16: Reg = X86::AX; size = 2; break;
7408 case MVT::i32: Reg = X86::EAX; size = 4; break;
7410 assert(Subtarget->is64Bit() && "Node not type legal!");
7411 Reg = X86::RAX; size = 8;
7414 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7415 Op.getOperand(2), SDValue());
7416 SDValue Ops[] = { cpIn.getValue(0),
7419 DAG.getTargetConstant(size, MVT::i8),
7421 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7422 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7424 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7428 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7429 SelectionDAG &DAG) {
7430 assert(Subtarget->is64Bit() && "Result not type legalized?");
7431 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7432 SDValue TheChain = Op.getOperand(0);
7433 DebugLoc dl = Op.getDebugLoc();
7434 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7435 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7436 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7438 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7439 DAG.getConstant(32, MVT::i8));
7441 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7444 return DAG.getMergeValues(Ops, 2, dl);
7447 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7448 SDNode *Node = Op.getNode();
7449 DebugLoc dl = Node->getDebugLoc();
7450 EVT T = Node->getValueType(0);
7451 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7452 DAG.getConstant(0, T), Node->getOperand(2));
7453 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7454 cast<AtomicSDNode>(Node)->getMemoryVT(),
7455 Node->getOperand(0),
7456 Node->getOperand(1), negOp,
7457 cast<AtomicSDNode>(Node)->getSrcValue(),
7458 cast<AtomicSDNode>(Node)->getAlignment());
7461 /// LowerOperation - Provide custom lowering hooks for some operations.
7463 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7464 switch (Op.getOpcode()) {
7465 default: llvm_unreachable("Should not custom lower this!");
7466 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7467 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7468 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7469 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7470 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7471 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7472 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7473 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7474 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7475 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7476 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7477 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7478 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7479 case ISD::SHL_PARTS:
7480 case ISD::SRA_PARTS:
7481 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7482 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7483 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7484 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7485 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7486 case ISD::FABS: return LowerFABS(Op, DAG);
7487 case ISD::FNEG: return LowerFNEG(Op, DAG);
7488 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7489 case ISD::SETCC: return LowerSETCC(Op, DAG);
7490 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7491 case ISD::SELECT: return LowerSELECT(Op, DAG);
7492 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7493 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7494 case ISD::VASTART: return LowerVASTART(Op, DAG);
7495 case ISD::VAARG: return LowerVAARG(Op, DAG);
7496 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7497 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7498 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7499 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7500 case ISD::FRAME_TO_ARGS_OFFSET:
7501 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7502 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7503 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7504 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7505 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7506 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7507 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7508 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7514 case ISD::UMULO: return LowerXALUO(Op, DAG);
7515 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7519 void X86TargetLowering::
7520 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7521 SelectionDAG &DAG, unsigned NewOp) {
7522 EVT T = Node->getValueType(0);
7523 DebugLoc dl = Node->getDebugLoc();
7524 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7526 SDValue Chain = Node->getOperand(0);
7527 SDValue In1 = Node->getOperand(1);
7528 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7529 Node->getOperand(2), DAG.getIntPtrConstant(0));
7530 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7531 Node->getOperand(2), DAG.getIntPtrConstant(1));
7532 SDValue Ops[] = { Chain, In1, In2L, In2H };
7533 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7535 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7536 cast<MemSDNode>(Node)->getMemOperand());
7537 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7538 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7539 Results.push_back(Result.getValue(2));
7542 /// ReplaceNodeResults - Replace a node with an illegal result type
7543 /// with a new node built out of custom code.
7544 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7545 SmallVectorImpl<SDValue>&Results,
7546 SelectionDAG &DAG) {
7547 DebugLoc dl = N->getDebugLoc();
7548 switch (N->getOpcode()) {
7550 assert(false && "Do not know how to custom type legalize this operation!");
7552 case ISD::FP_TO_SINT: {
7553 std::pair<SDValue,SDValue> Vals =
7554 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7555 SDValue FIST = Vals.first, StackSlot = Vals.second;
7556 if (FIST.getNode() != 0) {
7557 EVT VT = N->getValueType(0);
7558 // Return a load from the stack slot.
7559 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
7563 case ISD::READCYCLECOUNTER: {
7564 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7565 SDValue TheChain = N->getOperand(0);
7566 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7567 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7569 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7571 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7572 SDValue Ops[] = { eax, edx };
7573 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7574 Results.push_back(edx.getValue(1));
7581 EVT WidenVT = getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
7582 Results.push_back(DAG.UnrollVectorOp(N, WidenVT.getVectorNumElements()));
7585 case ISD::ATOMIC_CMP_SWAP: {
7586 EVT T = N->getValueType(0);
7587 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7588 SDValue cpInL, cpInH;
7589 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7590 DAG.getConstant(0, MVT::i32));
7591 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7592 DAG.getConstant(1, MVT::i32));
7593 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7594 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7596 SDValue swapInL, swapInH;
7597 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7598 DAG.getConstant(0, MVT::i32));
7599 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7600 DAG.getConstant(1, MVT::i32));
7601 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7603 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7604 swapInL.getValue(1));
7605 SDValue Ops[] = { swapInH.getValue(0),
7607 swapInH.getValue(1) };
7608 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7609 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7610 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7611 MVT::i32, Result.getValue(1));
7612 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7613 MVT::i32, cpOutL.getValue(2));
7614 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7615 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7616 Results.push_back(cpOutH.getValue(1));
7619 case ISD::ATOMIC_LOAD_ADD:
7620 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7622 case ISD::ATOMIC_LOAD_AND:
7623 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7625 case ISD::ATOMIC_LOAD_NAND:
7626 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7628 case ISD::ATOMIC_LOAD_OR:
7629 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7631 case ISD::ATOMIC_LOAD_SUB:
7632 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7634 case ISD::ATOMIC_LOAD_XOR:
7635 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7637 case ISD::ATOMIC_SWAP:
7638 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7643 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7645 default: return NULL;
7646 case X86ISD::BSF: return "X86ISD::BSF";
7647 case X86ISD::BSR: return "X86ISD::BSR";
7648 case X86ISD::SHLD: return "X86ISD::SHLD";
7649 case X86ISD::SHRD: return "X86ISD::SHRD";
7650 case X86ISD::FAND: return "X86ISD::FAND";
7651 case X86ISD::FOR: return "X86ISD::FOR";
7652 case X86ISD::FXOR: return "X86ISD::FXOR";
7653 case X86ISD::FSRL: return "X86ISD::FSRL";
7654 case X86ISD::FILD: return "X86ISD::FILD";
7655 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7656 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7657 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7658 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7659 case X86ISD::FLD: return "X86ISD::FLD";
7660 case X86ISD::FST: return "X86ISD::FST";
7661 case X86ISD::CALL: return "X86ISD::CALL";
7662 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7663 case X86ISD::BT: return "X86ISD::BT";
7664 case X86ISD::CMP: return "X86ISD::CMP";
7665 case X86ISD::COMI: return "X86ISD::COMI";
7666 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7667 case X86ISD::SETCC: return "X86ISD::SETCC";
7668 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7669 case X86ISD::CMOV: return "X86ISD::CMOV";
7670 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7671 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7672 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7673 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7674 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7675 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7676 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7677 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7678 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7679 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7680 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7681 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7682 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7683 case X86ISD::FMAX: return "X86ISD::FMAX";
7684 case X86ISD::FMIN: return "X86ISD::FMIN";
7685 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7686 case X86ISD::FRCP: return "X86ISD::FRCP";
7687 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7688 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7689 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7690 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7691 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7692 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7693 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7694 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7695 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7696 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7697 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7698 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7699 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7700 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7701 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7702 case X86ISD::VSHL: return "X86ISD::VSHL";
7703 case X86ISD::VSRL: return "X86ISD::VSRL";
7704 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7705 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7706 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7707 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7708 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7709 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7710 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7711 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7712 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7713 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7714 case X86ISD::ADD: return "X86ISD::ADD";
7715 case X86ISD::SUB: return "X86ISD::SUB";
7716 case X86ISD::SMUL: return "X86ISD::SMUL";
7717 case X86ISD::UMUL: return "X86ISD::UMUL";
7718 case X86ISD::INC: return "X86ISD::INC";
7719 case X86ISD::DEC: return "X86ISD::DEC";
7720 case X86ISD::OR: return "X86ISD::OR";
7721 case X86ISD::XOR: return "X86ISD::XOR";
7722 case X86ISD::AND: return "X86ISD::AND";
7723 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7724 case X86ISD::PTEST: return "X86ISD::PTEST";
7725 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7729 // isLegalAddressingMode - Return true if the addressing mode represented
7730 // by AM is legal for this target, for a load/store of the specified type.
7731 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7732 const Type *Ty) const {
7733 // X86 supports extremely general addressing modes.
7734 CodeModel::Model M = getTargetMachine().getCodeModel();
7736 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7737 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7742 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7744 // If a reference to this global requires an extra load, we can't fold it.
7745 if (isGlobalStubReference(GVFlags))
7748 // If BaseGV requires a register for the PIC base, we cannot also have a
7749 // BaseReg specified.
7750 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7753 // If lower 4G is not available, then we must use rip-relative addressing.
7754 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7764 // These scales always work.
7769 // These scales are formed with basereg+scalereg. Only accept if there is
7774 default: // Other stuff never works.
7782 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7783 if (!Ty1->isInteger() || !Ty2->isInteger())
7785 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7786 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7787 if (NumBits1 <= NumBits2)
7789 return Subtarget->is64Bit() || NumBits1 < 64;
7792 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7793 if (!VT1.isInteger() || !VT2.isInteger())
7795 unsigned NumBits1 = VT1.getSizeInBits();
7796 unsigned NumBits2 = VT2.getSizeInBits();
7797 if (NumBits1 <= NumBits2)
7799 return Subtarget->is64Bit() || NumBits1 < 64;
7802 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7803 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7804 return Ty1->isInteger(32) && Ty2->isInteger(64) && Subtarget->is64Bit();
7807 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7808 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7809 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7812 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7813 // i16 instructions are longer (0x66 prefix) and potentially slower.
7814 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7817 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7818 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7819 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7820 /// are assumed to be legal.
7822 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7824 // Only do shuffles on 128-bit vector types for now.
7825 if (VT.getSizeInBits() == 64)
7828 // FIXME: pshufb, blends, shifts.
7829 return (VT.getVectorNumElements() == 2 ||
7830 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7831 isMOVLMask(M, VT) ||
7832 isSHUFPMask(M, VT) ||
7833 isPSHUFDMask(M, VT) ||
7834 isPSHUFHWMask(M, VT) ||
7835 isPSHUFLWMask(M, VT) ||
7836 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7837 isUNPCKLMask(M, VT) ||
7838 isUNPCKHMask(M, VT) ||
7839 isUNPCKL_v_undef_Mask(M, VT) ||
7840 isUNPCKH_v_undef_Mask(M, VT));
7844 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7846 unsigned NumElts = VT.getVectorNumElements();
7847 // FIXME: This collection of masks seems suspect.
7850 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7851 return (isMOVLMask(Mask, VT) ||
7852 isCommutedMOVLMask(Mask, VT, true) ||
7853 isSHUFPMask(Mask, VT) ||
7854 isCommutedSHUFPMask(Mask, VT));
7859 //===----------------------------------------------------------------------===//
7860 // X86 Scheduler Hooks
7861 //===----------------------------------------------------------------------===//
7863 // private utility function
7865 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7866 MachineBasicBlock *MBB,
7874 TargetRegisterClass *RC,
7875 bool invSrc) const {
7876 // For the atomic bitwise operator, we generate
7879 // ld t1 = [bitinstr.addr]
7880 // op t2 = t1, [bitinstr.val]
7882 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7884 // fallthrough -->nextMBB
7885 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7886 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7887 MachineFunction::iterator MBBIter = MBB;
7890 /// First build the CFG
7891 MachineFunction *F = MBB->getParent();
7892 MachineBasicBlock *thisMBB = MBB;
7893 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7894 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7895 F->insert(MBBIter, newMBB);
7896 F->insert(MBBIter, nextMBB);
7898 // Move all successors to thisMBB to nextMBB
7899 nextMBB->transferSuccessors(thisMBB);
7901 // Update thisMBB to fall through to newMBB
7902 thisMBB->addSuccessor(newMBB);
7904 // newMBB jumps to itself and fall through to nextMBB
7905 newMBB->addSuccessor(nextMBB);
7906 newMBB->addSuccessor(newMBB);
7908 // Insert instructions into newMBB based on incoming instruction
7909 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7910 "unexpected number of operands");
7911 DebugLoc dl = bInstr->getDebugLoc();
7912 MachineOperand& destOper = bInstr->getOperand(0);
7913 MachineOperand* argOpers[2 + X86AddrNumOperands];
7914 int numArgs = bInstr->getNumOperands() - 1;
7915 for (int i=0; i < numArgs; ++i)
7916 argOpers[i] = &bInstr->getOperand(i+1);
7918 // x86 address has 4 operands: base, index, scale, and displacement
7919 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7920 int valArgIndx = lastAddrIndx + 1;
7922 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7923 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7924 for (int i=0; i <= lastAddrIndx; ++i)
7925 (*MIB).addOperand(*argOpers[i]);
7927 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7929 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7934 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7935 assert((argOpers[valArgIndx]->isReg() ||
7936 argOpers[valArgIndx]->isImm()) &&
7938 if (argOpers[valArgIndx]->isReg())
7939 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7941 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7943 (*MIB).addOperand(*argOpers[valArgIndx]);
7945 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7948 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7949 for (int i=0; i <= lastAddrIndx; ++i)
7950 (*MIB).addOperand(*argOpers[i]);
7952 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7953 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7954 bInstr->memoperands_end());
7956 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7960 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7962 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7966 // private utility function: 64 bit atomics on 32 bit host.
7968 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7969 MachineBasicBlock *MBB,
7974 bool invSrc) const {
7975 // For the atomic bitwise operator, we generate
7976 // thisMBB (instructions are in pairs, except cmpxchg8b)
7977 // ld t1,t2 = [bitinstr.addr]
7979 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7980 // op t5, t6 <- out1, out2, [bitinstr.val]
7981 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7982 // mov ECX, EBX <- t5, t6
7983 // mov EAX, EDX <- t1, t2
7984 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7985 // mov t3, t4 <- EAX, EDX
7987 // result in out1, out2
7988 // fallthrough -->nextMBB
7990 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7991 const unsigned LoadOpc = X86::MOV32rm;
7992 const unsigned copyOpc = X86::MOV32rr;
7993 const unsigned NotOpc = X86::NOT32r;
7994 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7995 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7996 MachineFunction::iterator MBBIter = MBB;
7999 /// First build the CFG
8000 MachineFunction *F = MBB->getParent();
8001 MachineBasicBlock *thisMBB = MBB;
8002 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8003 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8004 F->insert(MBBIter, newMBB);
8005 F->insert(MBBIter, nextMBB);
8007 // Move all successors to thisMBB to nextMBB
8008 nextMBB->transferSuccessors(thisMBB);
8010 // Update thisMBB to fall through to newMBB
8011 thisMBB->addSuccessor(newMBB);
8013 // newMBB jumps to itself and fall through to nextMBB
8014 newMBB->addSuccessor(nextMBB);
8015 newMBB->addSuccessor(newMBB);
8017 DebugLoc dl = bInstr->getDebugLoc();
8018 // Insert instructions into newMBB based on incoming instruction
8019 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8020 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
8021 "unexpected number of operands");
8022 MachineOperand& dest1Oper = bInstr->getOperand(0);
8023 MachineOperand& dest2Oper = bInstr->getOperand(1);
8024 MachineOperand* argOpers[2 + X86AddrNumOperands];
8025 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
8026 argOpers[i] = &bInstr->getOperand(i+2);
8028 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8029 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8031 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8032 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8033 for (int i=0; i <= lastAddrIndx; ++i)
8034 (*MIB).addOperand(*argOpers[i]);
8035 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8036 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8037 // add 4 to displacement.
8038 for (int i=0; i <= lastAddrIndx-2; ++i)
8039 (*MIB).addOperand(*argOpers[i]);
8040 MachineOperand newOp3 = *(argOpers[3]);
8042 newOp3.setImm(newOp3.getImm()+4);
8044 newOp3.setOffset(newOp3.getOffset()+4);
8045 (*MIB).addOperand(newOp3);
8046 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8048 // t3/4 are defined later, at the bottom of the loop
8049 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8050 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8051 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8052 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8053 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8054 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8056 // The subsequent operations should be using the destination registers of
8057 //the PHI instructions.
8059 t1 = F->getRegInfo().createVirtualRegister(RC);
8060 t2 = F->getRegInfo().createVirtualRegister(RC);
8061 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8062 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8064 t1 = dest1Oper.getReg();
8065 t2 = dest2Oper.getReg();
8068 int valArgIndx = lastAddrIndx + 1;
8069 assert((argOpers[valArgIndx]->isReg() ||
8070 argOpers[valArgIndx]->isImm()) &&
8072 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8073 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8074 if (argOpers[valArgIndx]->isReg())
8075 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8077 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8078 if (regOpcL != X86::MOV32rr)
8080 (*MIB).addOperand(*argOpers[valArgIndx]);
8081 assert(argOpers[valArgIndx + 1]->isReg() ==
8082 argOpers[valArgIndx]->isReg());
8083 assert(argOpers[valArgIndx + 1]->isImm() ==
8084 argOpers[valArgIndx]->isImm());
8085 if (argOpers[valArgIndx + 1]->isReg())
8086 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8088 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8089 if (regOpcH != X86::MOV32rr)
8091 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8093 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8095 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8098 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8100 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8103 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8104 for (int i=0; i <= lastAddrIndx; ++i)
8105 (*MIB).addOperand(*argOpers[i]);
8107 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8108 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8109 bInstr->memoperands_end());
8111 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8112 MIB.addReg(X86::EAX);
8113 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8114 MIB.addReg(X86::EDX);
8117 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
8119 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8123 // private utility function
8125 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8126 MachineBasicBlock *MBB,
8127 unsigned cmovOpc) const {
8128 // For the atomic min/max operator, we generate
8131 // ld t1 = [min/max.addr]
8132 // mov t2 = [min/max.val]
8134 // cmov[cond] t2 = t1
8136 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8138 // fallthrough -->nextMBB
8140 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8141 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8142 MachineFunction::iterator MBBIter = MBB;
8145 /// First build the CFG
8146 MachineFunction *F = MBB->getParent();
8147 MachineBasicBlock *thisMBB = MBB;
8148 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8149 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8150 F->insert(MBBIter, newMBB);
8151 F->insert(MBBIter, nextMBB);
8153 // Move all successors of thisMBB to nextMBB
8154 nextMBB->transferSuccessors(thisMBB);
8156 // Update thisMBB to fall through to newMBB
8157 thisMBB->addSuccessor(newMBB);
8159 // newMBB jumps to newMBB and fall through to nextMBB
8160 newMBB->addSuccessor(nextMBB);
8161 newMBB->addSuccessor(newMBB);
8163 DebugLoc dl = mInstr->getDebugLoc();
8164 // Insert instructions into newMBB based on incoming instruction
8165 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8166 "unexpected number of operands");
8167 MachineOperand& destOper = mInstr->getOperand(0);
8168 MachineOperand* argOpers[2 + X86AddrNumOperands];
8169 int numArgs = mInstr->getNumOperands() - 1;
8170 for (int i=0; i < numArgs; ++i)
8171 argOpers[i] = &mInstr->getOperand(i+1);
8173 // x86 address has 4 operands: base, index, scale, and displacement
8174 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8175 int valArgIndx = lastAddrIndx + 1;
8177 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8178 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8179 for (int i=0; i <= lastAddrIndx; ++i)
8180 (*MIB).addOperand(*argOpers[i]);
8182 // We only support register and immediate values
8183 assert((argOpers[valArgIndx]->isReg() ||
8184 argOpers[valArgIndx]->isImm()) &&
8187 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8188 if (argOpers[valArgIndx]->isReg())
8189 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8191 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8192 (*MIB).addOperand(*argOpers[valArgIndx]);
8194 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8197 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8202 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8203 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8207 // Cmp and exchange if none has modified the memory location
8208 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8209 for (int i=0; i <= lastAddrIndx; ++i)
8210 (*MIB).addOperand(*argOpers[i]);
8212 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8213 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8214 mInstr->memoperands_end());
8216 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8217 MIB.addReg(X86::EAX);
8220 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
8222 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8226 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8227 // all of this code can be replaced with that in the .td file.
8229 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8230 unsigned numArgs, bool memArg) const {
8232 MachineFunction *F = BB->getParent();
8233 DebugLoc dl = MI->getDebugLoc();
8234 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8238 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8240 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8242 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8244 for (unsigned i = 0; i < numArgs; ++i) {
8245 MachineOperand &Op = MI->getOperand(i+1);
8247 if (!(Op.isReg() && Op.isImplicit()))
8251 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8254 F->DeleteMachineInstr(MI);
8260 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8262 MachineBasicBlock *MBB) const {
8263 // Emit code to save XMM registers to the stack. The ABI says that the
8264 // number of registers to save is given in %al, so it's theoretically
8265 // possible to do an indirect jump trick to avoid saving all of them,
8266 // however this code takes a simpler approach and just executes all
8267 // of the stores if %al is non-zero. It's less code, and it's probably
8268 // easier on the hardware branch predictor, and stores aren't all that
8269 // expensive anyway.
8271 // Create the new basic blocks. One block contains all the XMM stores,
8272 // and one block is the final destination regardless of whether any
8273 // stores were performed.
8274 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8275 MachineFunction *F = MBB->getParent();
8276 MachineFunction::iterator MBBIter = MBB;
8278 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8279 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8280 F->insert(MBBIter, XMMSaveMBB);
8281 F->insert(MBBIter, EndMBB);
8284 // Move any original successors of MBB to the end block.
8285 EndMBB->transferSuccessors(MBB);
8286 // The original block will now fall through to the XMM save block.
8287 MBB->addSuccessor(XMMSaveMBB);
8288 // The XMMSaveMBB will fall through to the end block.
8289 XMMSaveMBB->addSuccessor(EndMBB);
8291 // Now add the instructions.
8292 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8293 DebugLoc DL = MI->getDebugLoc();
8295 unsigned CountReg = MI->getOperand(0).getReg();
8296 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8297 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8299 if (!Subtarget->isTargetWin64()) {
8300 // If %al is 0, branch around the XMM save block.
8301 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8302 BuildMI(MBB, DL, TII->get(X86::JE)).addMBB(EndMBB);
8303 MBB->addSuccessor(EndMBB);
8306 // In the XMM save block, save all the XMM argument registers.
8307 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8308 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8309 MachineMemOperand *MMO =
8310 F->getMachineMemOperand(
8311 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8312 MachineMemOperand::MOStore, Offset,
8313 /*Size=*/16, /*Align=*/16);
8314 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8315 .addFrameIndex(RegSaveFrameIndex)
8316 .addImm(/*Scale=*/1)
8317 .addReg(/*IndexReg=*/0)
8318 .addImm(/*Disp=*/Offset)
8319 .addReg(/*Segment=*/0)
8320 .addReg(MI->getOperand(i).getReg())
8321 .addMemOperand(MMO);
8324 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8330 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8331 MachineBasicBlock *BB,
8332 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8333 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8334 DebugLoc DL = MI->getDebugLoc();
8336 // To "insert" a SELECT_CC instruction, we actually have to insert the
8337 // diamond control-flow pattern. The incoming instruction knows the
8338 // destination vreg to set, the condition code register to branch on, the
8339 // true/false values to select between, and a branch opcode to use.
8340 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8341 MachineFunction::iterator It = BB;
8347 // cmpTY ccX, r1, r2
8349 // fallthrough --> copy0MBB
8350 MachineBasicBlock *thisMBB = BB;
8351 MachineFunction *F = BB->getParent();
8352 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8353 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8355 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8356 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8357 F->insert(It, copy0MBB);
8358 F->insert(It, sinkMBB);
8359 // Update machine-CFG edges by first adding all successors of the current
8360 // block to the new block which will contain the Phi node for the select.
8361 // Also inform sdisel of the edge changes.
8362 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8363 E = BB->succ_end(); I != E; ++I) {
8364 EM->insert(std::make_pair(*I, sinkMBB));
8365 sinkMBB->addSuccessor(*I);
8367 // Next, remove all successors of the current block, and add the true
8368 // and fallthrough blocks as its successors.
8369 while (!BB->succ_empty())
8370 BB->removeSuccessor(BB->succ_begin());
8371 // Add the true and fallthrough blocks as its successors.
8372 BB->addSuccessor(copy0MBB);
8373 BB->addSuccessor(sinkMBB);
8376 // %FalseValue = ...
8377 // # fallthrough to sinkMBB
8380 // Update machine-CFG edges
8381 BB->addSuccessor(sinkMBB);
8384 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8387 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8388 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8389 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8391 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8397 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8398 MachineBasicBlock *BB,
8399 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8400 switch (MI->getOpcode()) {
8401 default: assert(false && "Unexpected instr type to insert");
8403 case X86::CMOV_V1I64:
8404 case X86::CMOV_FR32:
8405 case X86::CMOV_FR64:
8406 case X86::CMOV_V4F32:
8407 case X86::CMOV_V2F64:
8408 case X86::CMOV_V2I64:
8409 return EmitLoweredSelect(MI, BB, EM);
8411 case X86::FP32_TO_INT16_IN_MEM:
8412 case X86::FP32_TO_INT32_IN_MEM:
8413 case X86::FP32_TO_INT64_IN_MEM:
8414 case X86::FP64_TO_INT16_IN_MEM:
8415 case X86::FP64_TO_INT32_IN_MEM:
8416 case X86::FP64_TO_INT64_IN_MEM:
8417 case X86::FP80_TO_INT16_IN_MEM:
8418 case X86::FP80_TO_INT32_IN_MEM:
8419 case X86::FP80_TO_INT64_IN_MEM: {
8420 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8421 DebugLoc DL = MI->getDebugLoc();
8423 // Change the floating point control register to use "round towards zero"
8424 // mode when truncating to an integer value.
8425 MachineFunction *F = BB->getParent();
8426 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8427 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8429 // Load the old value of the high byte of the control word...
8431 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8432 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8435 // Set the high part to be round to zero...
8436 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8439 // Reload the modified control word now...
8440 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8442 // Restore the memory image of control word to original value
8443 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8446 // Get the X86 opcode to use.
8448 switch (MI->getOpcode()) {
8449 default: llvm_unreachable("illegal opcode!");
8450 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8451 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8452 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8453 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8454 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8455 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8456 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8457 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8458 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8462 MachineOperand &Op = MI->getOperand(0);
8464 AM.BaseType = X86AddressMode::RegBase;
8465 AM.Base.Reg = Op.getReg();
8467 AM.BaseType = X86AddressMode::FrameIndexBase;
8468 AM.Base.FrameIndex = Op.getIndex();
8470 Op = MI->getOperand(1);
8472 AM.Scale = Op.getImm();
8473 Op = MI->getOperand(2);
8475 AM.IndexReg = Op.getImm();
8476 Op = MI->getOperand(3);
8477 if (Op.isGlobal()) {
8478 AM.GV = Op.getGlobal();
8480 AM.Disp = Op.getImm();
8482 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8483 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8485 // Reload the original control word now.
8486 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8488 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8491 // String/text processing lowering.
8492 case X86::PCMPISTRM128REG:
8493 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8494 case X86::PCMPISTRM128MEM:
8495 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8496 case X86::PCMPESTRM128REG:
8497 return EmitPCMP(MI, BB, 5, false /* in mem */);
8498 case X86::PCMPESTRM128MEM:
8499 return EmitPCMP(MI, BB, 5, true /* in mem */);
8502 case X86::ATOMAND32:
8503 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8504 X86::AND32ri, X86::MOV32rm,
8505 X86::LCMPXCHG32, X86::MOV32rr,
8506 X86::NOT32r, X86::EAX,
8507 X86::GR32RegisterClass);
8509 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8510 X86::OR32ri, X86::MOV32rm,
8511 X86::LCMPXCHG32, X86::MOV32rr,
8512 X86::NOT32r, X86::EAX,
8513 X86::GR32RegisterClass);
8514 case X86::ATOMXOR32:
8515 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8516 X86::XOR32ri, X86::MOV32rm,
8517 X86::LCMPXCHG32, X86::MOV32rr,
8518 X86::NOT32r, X86::EAX,
8519 X86::GR32RegisterClass);
8520 case X86::ATOMNAND32:
8521 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8522 X86::AND32ri, X86::MOV32rm,
8523 X86::LCMPXCHG32, X86::MOV32rr,
8524 X86::NOT32r, X86::EAX,
8525 X86::GR32RegisterClass, true);
8526 case X86::ATOMMIN32:
8527 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8528 case X86::ATOMMAX32:
8529 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8530 case X86::ATOMUMIN32:
8531 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8532 case X86::ATOMUMAX32:
8533 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8535 case X86::ATOMAND16:
8536 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8537 X86::AND16ri, X86::MOV16rm,
8538 X86::LCMPXCHG16, X86::MOV16rr,
8539 X86::NOT16r, X86::AX,
8540 X86::GR16RegisterClass);
8542 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8543 X86::OR16ri, X86::MOV16rm,
8544 X86::LCMPXCHG16, X86::MOV16rr,
8545 X86::NOT16r, X86::AX,
8546 X86::GR16RegisterClass);
8547 case X86::ATOMXOR16:
8548 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8549 X86::XOR16ri, X86::MOV16rm,
8550 X86::LCMPXCHG16, X86::MOV16rr,
8551 X86::NOT16r, X86::AX,
8552 X86::GR16RegisterClass);
8553 case X86::ATOMNAND16:
8554 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8555 X86::AND16ri, X86::MOV16rm,
8556 X86::LCMPXCHG16, X86::MOV16rr,
8557 X86::NOT16r, X86::AX,
8558 X86::GR16RegisterClass, true);
8559 case X86::ATOMMIN16:
8560 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8561 case X86::ATOMMAX16:
8562 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8563 case X86::ATOMUMIN16:
8564 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8565 case X86::ATOMUMAX16:
8566 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8569 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8570 X86::AND8ri, X86::MOV8rm,
8571 X86::LCMPXCHG8, X86::MOV8rr,
8572 X86::NOT8r, X86::AL,
8573 X86::GR8RegisterClass);
8575 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8576 X86::OR8ri, X86::MOV8rm,
8577 X86::LCMPXCHG8, X86::MOV8rr,
8578 X86::NOT8r, X86::AL,
8579 X86::GR8RegisterClass);
8581 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8582 X86::XOR8ri, X86::MOV8rm,
8583 X86::LCMPXCHG8, X86::MOV8rr,
8584 X86::NOT8r, X86::AL,
8585 X86::GR8RegisterClass);
8586 case X86::ATOMNAND8:
8587 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8588 X86::AND8ri, X86::MOV8rm,
8589 X86::LCMPXCHG8, X86::MOV8rr,
8590 X86::NOT8r, X86::AL,
8591 X86::GR8RegisterClass, true);
8592 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8593 // This group is for 64-bit host.
8594 case X86::ATOMAND64:
8595 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8596 X86::AND64ri32, X86::MOV64rm,
8597 X86::LCMPXCHG64, X86::MOV64rr,
8598 X86::NOT64r, X86::RAX,
8599 X86::GR64RegisterClass);
8601 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8602 X86::OR64ri32, X86::MOV64rm,
8603 X86::LCMPXCHG64, X86::MOV64rr,
8604 X86::NOT64r, X86::RAX,
8605 X86::GR64RegisterClass);
8606 case X86::ATOMXOR64:
8607 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8608 X86::XOR64ri32, X86::MOV64rm,
8609 X86::LCMPXCHG64, X86::MOV64rr,
8610 X86::NOT64r, X86::RAX,
8611 X86::GR64RegisterClass);
8612 case X86::ATOMNAND64:
8613 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8614 X86::AND64ri32, X86::MOV64rm,
8615 X86::LCMPXCHG64, X86::MOV64rr,
8616 X86::NOT64r, X86::RAX,
8617 X86::GR64RegisterClass, true);
8618 case X86::ATOMMIN64:
8619 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8620 case X86::ATOMMAX64:
8621 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8622 case X86::ATOMUMIN64:
8623 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8624 case X86::ATOMUMAX64:
8625 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8627 // This group does 64-bit operations on a 32-bit host.
8628 case X86::ATOMAND6432:
8629 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8630 X86::AND32rr, X86::AND32rr,
8631 X86::AND32ri, X86::AND32ri,
8633 case X86::ATOMOR6432:
8634 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8635 X86::OR32rr, X86::OR32rr,
8636 X86::OR32ri, X86::OR32ri,
8638 case X86::ATOMXOR6432:
8639 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8640 X86::XOR32rr, X86::XOR32rr,
8641 X86::XOR32ri, X86::XOR32ri,
8643 case X86::ATOMNAND6432:
8644 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8645 X86::AND32rr, X86::AND32rr,
8646 X86::AND32ri, X86::AND32ri,
8648 case X86::ATOMADD6432:
8649 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8650 X86::ADD32rr, X86::ADC32rr,
8651 X86::ADD32ri, X86::ADC32ri,
8653 case X86::ATOMSUB6432:
8654 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8655 X86::SUB32rr, X86::SBB32rr,
8656 X86::SUB32ri, X86::SBB32ri,
8658 case X86::ATOMSWAP6432:
8659 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8660 X86::MOV32rr, X86::MOV32rr,
8661 X86::MOV32ri, X86::MOV32ri,
8663 case X86::VASTART_SAVE_XMM_REGS:
8664 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8668 //===----------------------------------------------------------------------===//
8669 // X86 Optimization Hooks
8670 //===----------------------------------------------------------------------===//
8672 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8676 const SelectionDAG &DAG,
8677 unsigned Depth) const {
8678 unsigned Opc = Op.getOpcode();
8679 assert((Opc >= ISD::BUILTIN_OP_END ||
8680 Opc == ISD::INTRINSIC_WO_CHAIN ||
8681 Opc == ISD::INTRINSIC_W_CHAIN ||
8682 Opc == ISD::INTRINSIC_VOID) &&
8683 "Should use MaskedValueIsZero if you don't know whether Op"
8684 " is a target node!");
8686 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8698 // These nodes' second result is a boolean.
8699 if (Op.getResNo() == 0)
8703 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8704 Mask.getBitWidth() - 1);
8709 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8710 /// node is a GlobalAddress + offset.
8711 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8712 GlobalValue* &GA, int64_t &Offset) const{
8713 if (N->getOpcode() == X86ISD::Wrapper) {
8714 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8715 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8716 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8720 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8723 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8724 EVT EltVT, LoadSDNode *&LDBase,
8725 unsigned &LastLoadedElt,
8726 SelectionDAG &DAG, MachineFrameInfo *MFI,
8727 const TargetLowering &TLI) {
8729 LastLoadedElt = -1U;
8730 for (unsigned i = 0; i < NumElems; ++i) {
8731 if (N->getMaskElt(i) < 0) {
8737 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8738 if (!Elt.getNode() ||
8739 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8742 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8744 LDBase = cast<LoadSDNode>(Elt.getNode());
8748 if (Elt.getOpcode() == ISD::UNDEF)
8751 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8752 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
8759 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8760 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8761 /// if the load addresses are consecutive, non-overlapping, and in the right
8762 /// order. In the case of v2i64, it will see if it can rewrite the
8763 /// shuffle to be an appropriate build vector so it can take advantage of
8764 // performBuildVectorCombine.
8765 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8766 const TargetLowering &TLI) {
8767 DebugLoc dl = N->getDebugLoc();
8768 EVT VT = N->getValueType(0);
8769 EVT EltVT = VT.getVectorElementType();
8770 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8771 unsigned NumElems = VT.getVectorNumElements();
8773 if (VT.getSizeInBits() != 128)
8776 // Try to combine a vector_shuffle into a 128-bit load.
8777 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8778 LoadSDNode *LD = NULL;
8779 unsigned LastLoadedElt;
8780 if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
8784 if (LastLoadedElt == NumElems - 1) {
8785 if (DAG.InferPtrAlignment(LD->getBasePtr()) >= 16)
8786 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8787 LD->getSrcValue(), LD->getSrcValueOffset(),
8789 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8790 LD->getSrcValue(), LD->getSrcValueOffset(),
8791 LD->isVolatile(), LD->getAlignment());
8792 } else if (NumElems == 4 && LastLoadedElt == 1) {
8793 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8794 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8795 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8796 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8801 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8802 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8803 const X86Subtarget *Subtarget) {
8804 DebugLoc DL = N->getDebugLoc();
8805 SDValue Cond = N->getOperand(0);
8806 // Get the LHS/RHS of the select.
8807 SDValue LHS = N->getOperand(1);
8808 SDValue RHS = N->getOperand(2);
8810 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8811 // instructions have the peculiarity that if either operand is a NaN,
8812 // they chose what we call the RHS operand (and as such are not symmetric).
8813 // It happens that this matches the semantics of the common C idiom
8814 // x<y?x:y and related forms, so we can recognize these cases.
8815 if (Subtarget->hasSSE2() &&
8816 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8817 Cond.getOpcode() == ISD::SETCC) {
8818 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8820 unsigned Opcode = 0;
8821 // Check for x CC y ? x : y.
8822 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8826 // This can be a min if we can prove that at least one of the operands
8828 if (!FiniteOnlyFPMath()) {
8829 if (DAG.isKnownNeverNaN(RHS)) {
8830 // Put the potential NaN in the RHS so that SSE will preserve it.
8831 std::swap(LHS, RHS);
8832 } else if (!DAG.isKnownNeverNaN(LHS))
8835 Opcode = X86ISD::FMIN;
8838 // This can be a min if we can prove that at least one of the operands
8840 if (!FiniteOnlyFPMath()) {
8841 if (DAG.isKnownNeverNaN(LHS)) {
8842 // Put the potential NaN in the RHS so that SSE will preserve it.
8843 std::swap(LHS, RHS);
8844 } else if (!DAG.isKnownNeverNaN(RHS))
8847 Opcode = X86ISD::FMIN;
8850 // This can be a min, but if either operand is a NaN we need it to
8851 // preserve the original LHS.
8852 std::swap(LHS, RHS);
8856 Opcode = X86ISD::FMIN;
8860 // This can be a max if we can prove that at least one of the operands
8862 if (!FiniteOnlyFPMath()) {
8863 if (DAG.isKnownNeverNaN(LHS)) {
8864 // Put the potential NaN in the RHS so that SSE will preserve it.
8865 std::swap(LHS, RHS);
8866 } else if (!DAG.isKnownNeverNaN(RHS))
8869 Opcode = X86ISD::FMAX;
8872 // This can be a max if we can prove that at least one of the operands
8874 if (!FiniteOnlyFPMath()) {
8875 if (DAG.isKnownNeverNaN(RHS)) {
8876 // Put the potential NaN in the RHS so that SSE will preserve it.
8877 std::swap(LHS, RHS);
8878 } else if (!DAG.isKnownNeverNaN(LHS))
8881 Opcode = X86ISD::FMAX;
8884 // This can be a max, but if either operand is a NaN we need it to
8885 // preserve the original LHS.
8886 std::swap(LHS, RHS);
8890 Opcode = X86ISD::FMAX;
8893 // Check for x CC y ? y : x -- a min/max with reversed arms.
8894 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8898 // This can be a min if we can prove that at least one of the operands
8900 if (!FiniteOnlyFPMath()) {
8901 if (DAG.isKnownNeverNaN(RHS)) {
8902 // Put the potential NaN in the RHS so that SSE will preserve it.
8903 std::swap(LHS, RHS);
8904 } else if (!DAG.isKnownNeverNaN(LHS))
8907 Opcode = X86ISD::FMIN;
8910 // This can be a min if we can prove that at least one of the operands
8912 if (!FiniteOnlyFPMath()) {
8913 if (DAG.isKnownNeverNaN(LHS)) {
8914 // Put the potential NaN in the RHS so that SSE will preserve it.
8915 std::swap(LHS, RHS);
8916 } else if (!DAG.isKnownNeverNaN(RHS))
8919 Opcode = X86ISD::FMIN;
8922 // This can be a min, but if either operand is a NaN we need it to
8923 // preserve the original LHS.
8924 std::swap(LHS, RHS);
8928 Opcode = X86ISD::FMIN;
8932 // This can be a max if we can prove that at least one of the operands
8934 if (!FiniteOnlyFPMath()) {
8935 if (DAG.isKnownNeverNaN(LHS)) {
8936 // Put the potential NaN in the RHS so that SSE will preserve it.
8937 std::swap(LHS, RHS);
8938 } else if (!DAG.isKnownNeverNaN(RHS))
8941 Opcode = X86ISD::FMAX;
8944 // This can be a max if we can prove that at least one of the operands
8946 if (!FiniteOnlyFPMath()) {
8947 if (DAG.isKnownNeverNaN(RHS)) {
8948 // Put the potential NaN in the RHS so that SSE will preserve it.
8949 std::swap(LHS, RHS);
8950 } else if (!DAG.isKnownNeverNaN(LHS))
8953 Opcode = X86ISD::FMAX;
8956 // This can be a max, but if either operand is a NaN we need it to
8957 // preserve the original LHS.
8958 std::swap(LHS, RHS);
8962 Opcode = X86ISD::FMAX;
8968 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8971 // If this is a select between two integer constants, try to do some
8973 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8974 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8975 // Don't do this for crazy integer types.
8976 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8977 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8978 // so that TrueC (the true value) is larger than FalseC.
8979 bool NeedsCondInvert = false;
8981 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8982 // Efficiently invertible.
8983 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8984 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8985 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8986 NeedsCondInvert = true;
8987 std::swap(TrueC, FalseC);
8990 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8991 if (FalseC->getAPIntValue() == 0 &&
8992 TrueC->getAPIntValue().isPowerOf2()) {
8993 if (NeedsCondInvert) // Invert the condition if needed.
8994 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8995 DAG.getConstant(1, Cond.getValueType()));
8997 // Zero extend the condition if needed.
8998 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9000 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9001 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9002 DAG.getConstant(ShAmt, MVT::i8));
9005 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9006 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9007 if (NeedsCondInvert) // Invert the condition if needed.
9008 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9009 DAG.getConstant(1, Cond.getValueType()));
9011 // Zero extend the condition if needed.
9012 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9013 FalseC->getValueType(0), Cond);
9014 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9015 SDValue(FalseC, 0));
9018 // Optimize cases that will turn into an LEA instruction. This requires
9019 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9020 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9021 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9022 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9024 bool isFastMultiplier = false;
9026 switch ((unsigned char)Diff) {
9028 case 1: // result = add base, cond
9029 case 2: // result = lea base( , cond*2)
9030 case 3: // result = lea base(cond, cond*2)
9031 case 4: // result = lea base( , cond*4)
9032 case 5: // result = lea base(cond, cond*4)
9033 case 8: // result = lea base( , cond*8)
9034 case 9: // result = lea base(cond, cond*8)
9035 isFastMultiplier = true;
9040 if (isFastMultiplier) {
9041 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9042 if (NeedsCondInvert) // Invert the condition if needed.
9043 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9044 DAG.getConstant(1, Cond.getValueType()));
9046 // Zero extend the condition if needed.
9047 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9049 // Scale the condition by the difference.
9051 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9052 DAG.getConstant(Diff, Cond.getValueType()));
9054 // Add the base if non-zero.
9055 if (FalseC->getAPIntValue() != 0)
9056 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9057 SDValue(FalseC, 0));
9067 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9068 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9069 TargetLowering::DAGCombinerInfo &DCI) {
9070 DebugLoc DL = N->getDebugLoc();
9072 // If the flag operand isn't dead, don't touch this CMOV.
9073 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9076 // If this is a select between two integer constants, try to do some
9077 // optimizations. Note that the operands are ordered the opposite of SELECT
9079 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9080 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9081 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9082 // larger than FalseC (the false value).
9083 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9085 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9086 CC = X86::GetOppositeBranchCondition(CC);
9087 std::swap(TrueC, FalseC);
9090 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9091 // This is efficient for any integer data type (including i8/i16) and
9093 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9094 SDValue Cond = N->getOperand(3);
9095 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9096 DAG.getConstant(CC, MVT::i8), Cond);
9098 // Zero extend the condition if needed.
9099 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9101 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9102 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9103 DAG.getConstant(ShAmt, MVT::i8));
9104 if (N->getNumValues() == 2) // Dead flag value?
9105 return DCI.CombineTo(N, Cond, SDValue());
9109 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9110 // for any integer data type, including i8/i16.
9111 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9112 SDValue Cond = N->getOperand(3);
9113 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9114 DAG.getConstant(CC, MVT::i8), Cond);
9116 // Zero extend the condition if needed.
9117 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9118 FalseC->getValueType(0), Cond);
9119 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9120 SDValue(FalseC, 0));
9122 if (N->getNumValues() == 2) // Dead flag value?
9123 return DCI.CombineTo(N, Cond, SDValue());
9127 // Optimize cases that will turn into an LEA instruction. This requires
9128 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9129 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9130 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9131 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9133 bool isFastMultiplier = false;
9135 switch ((unsigned char)Diff) {
9137 case 1: // result = add base, cond
9138 case 2: // result = lea base( , cond*2)
9139 case 3: // result = lea base(cond, cond*2)
9140 case 4: // result = lea base( , cond*4)
9141 case 5: // result = lea base(cond, cond*4)
9142 case 8: // result = lea base( , cond*8)
9143 case 9: // result = lea base(cond, cond*8)
9144 isFastMultiplier = true;
9149 if (isFastMultiplier) {
9150 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9151 SDValue Cond = N->getOperand(3);
9152 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9153 DAG.getConstant(CC, MVT::i8), Cond);
9154 // Zero extend the condition if needed.
9155 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9157 // Scale the condition by the difference.
9159 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9160 DAG.getConstant(Diff, Cond.getValueType()));
9162 // Add the base if non-zero.
9163 if (FalseC->getAPIntValue() != 0)
9164 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9165 SDValue(FalseC, 0));
9166 if (N->getNumValues() == 2) // Dead flag value?
9167 return DCI.CombineTo(N, Cond, SDValue());
9177 /// PerformMulCombine - Optimize a single multiply with constant into two
9178 /// in order to implement it with two cheaper instructions, e.g.
9179 /// LEA + SHL, LEA + LEA.
9180 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9181 TargetLowering::DAGCombinerInfo &DCI) {
9182 if (DAG.getMachineFunction().
9183 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
9186 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9189 EVT VT = N->getValueType(0);
9193 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9196 uint64_t MulAmt = C->getZExtValue();
9197 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9200 uint64_t MulAmt1 = 0;
9201 uint64_t MulAmt2 = 0;
9202 if ((MulAmt % 9) == 0) {
9204 MulAmt2 = MulAmt / 9;
9205 } else if ((MulAmt % 5) == 0) {
9207 MulAmt2 = MulAmt / 5;
9208 } else if ((MulAmt % 3) == 0) {
9210 MulAmt2 = MulAmt / 3;
9213 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9214 DebugLoc DL = N->getDebugLoc();
9216 if (isPowerOf2_64(MulAmt2) &&
9217 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9218 // If second multiplifer is pow2, issue it first. We want the multiply by
9219 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9221 std::swap(MulAmt1, MulAmt2);
9224 if (isPowerOf2_64(MulAmt1))
9225 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9226 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9228 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9229 DAG.getConstant(MulAmt1, VT));
9231 if (isPowerOf2_64(MulAmt2))
9232 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9233 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9235 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9236 DAG.getConstant(MulAmt2, VT));
9238 // Do not add new nodes to DAG combiner worklist.
9239 DCI.CombineTo(N, NewMul, false);
9244 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9245 SDValue N0 = N->getOperand(0);
9246 SDValue N1 = N->getOperand(1);
9247 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9248 EVT VT = N0.getValueType();
9250 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9251 // since the result of setcc_c is all zero's or all ones.
9252 if (N1C && N0.getOpcode() == ISD::AND &&
9253 N0.getOperand(1).getOpcode() == ISD::Constant) {
9254 SDValue N00 = N0.getOperand(0);
9255 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9256 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9257 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9258 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9259 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9260 APInt ShAmt = N1C->getAPIntValue();
9261 Mask = Mask.shl(ShAmt);
9263 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9264 N00, DAG.getConstant(Mask, VT));
9271 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9273 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9274 const X86Subtarget *Subtarget) {
9275 EVT VT = N->getValueType(0);
9276 if (!VT.isVector() && VT.isInteger() &&
9277 N->getOpcode() == ISD::SHL)
9278 return PerformSHLCombine(N, DAG);
9280 // On X86 with SSE2 support, we can transform this to a vector shift if
9281 // all elements are shifted by the same amount. We can't do this in legalize
9282 // because the a constant vector is typically transformed to a constant pool
9283 // so we have no knowledge of the shift amount.
9284 if (!Subtarget->hasSSE2())
9287 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9290 SDValue ShAmtOp = N->getOperand(1);
9291 EVT EltVT = VT.getVectorElementType();
9292 DebugLoc DL = N->getDebugLoc();
9293 SDValue BaseShAmt = SDValue();
9294 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9295 unsigned NumElts = VT.getVectorNumElements();
9297 for (; i != NumElts; ++i) {
9298 SDValue Arg = ShAmtOp.getOperand(i);
9299 if (Arg.getOpcode() == ISD::UNDEF) continue;
9303 for (; i != NumElts; ++i) {
9304 SDValue Arg = ShAmtOp.getOperand(i);
9305 if (Arg.getOpcode() == ISD::UNDEF) continue;
9306 if (Arg != BaseShAmt) {
9310 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9311 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9312 SDValue InVec = ShAmtOp.getOperand(0);
9313 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9314 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9316 for (; i != NumElts; ++i) {
9317 SDValue Arg = InVec.getOperand(i);
9318 if (Arg.getOpcode() == ISD::UNDEF) continue;
9322 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9323 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9324 unsigned SplatIdx = cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9325 if (C->getZExtValue() == SplatIdx)
9326 BaseShAmt = InVec.getOperand(1);
9329 if (BaseShAmt.getNode() == 0)
9330 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9331 DAG.getIntPtrConstant(0));
9335 // The shift amount is an i32.
9336 if (EltVT.bitsGT(MVT::i32))
9337 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9338 else if (EltVT.bitsLT(MVT::i32))
9339 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9341 // The shift amount is identical so we can do a vector shift.
9342 SDValue ValOp = N->getOperand(0);
9343 switch (N->getOpcode()) {
9345 llvm_unreachable("Unknown shift opcode!");
9348 if (VT == MVT::v2i64)
9349 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9350 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9352 if (VT == MVT::v4i32)
9353 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9354 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9356 if (VT == MVT::v8i16)
9357 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9358 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9362 if (VT == MVT::v4i32)
9363 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9364 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9366 if (VT == MVT::v8i16)
9367 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9368 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9372 if (VT == MVT::v2i64)
9373 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9374 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9376 if (VT == MVT::v4i32)
9377 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9378 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9380 if (VT == MVT::v8i16)
9381 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9382 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9389 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9390 const X86Subtarget *Subtarget) {
9391 EVT VT = N->getValueType(0);
9392 if (VT != MVT::i64 || !Subtarget->is64Bit())
9395 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9396 SDValue N0 = N->getOperand(0);
9397 SDValue N1 = N->getOperand(1);
9398 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9400 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9403 SDValue ShAmt0 = N0.getOperand(1);
9404 if (ShAmt0.getValueType() != MVT::i8)
9406 SDValue ShAmt1 = N1.getOperand(1);
9407 if (ShAmt1.getValueType() != MVT::i8)
9409 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9410 ShAmt0 = ShAmt0.getOperand(0);
9411 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9412 ShAmt1 = ShAmt1.getOperand(0);
9414 DebugLoc DL = N->getDebugLoc();
9415 unsigned Opc = X86ISD::SHLD;
9416 SDValue Op0 = N0.getOperand(0);
9417 SDValue Op1 = N1.getOperand(0);
9418 if (ShAmt0.getOpcode() == ISD::SUB) {
9420 std::swap(Op0, Op1);
9421 std::swap(ShAmt0, ShAmt1);
9424 if (ShAmt1.getOpcode() == ISD::SUB) {
9425 SDValue Sum = ShAmt1.getOperand(0);
9426 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9427 if (SumC->getSExtValue() == 64 &&
9428 ShAmt1.getOperand(1) == ShAmt0)
9429 return DAG.getNode(Opc, DL, VT,
9431 DAG.getNode(ISD::TRUNCATE, DL,
9434 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9435 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9437 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
9438 return DAG.getNode(Opc, DL, VT,
9439 N0.getOperand(0), N1.getOperand(0),
9440 DAG.getNode(ISD::TRUNCATE, DL,
9447 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9448 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9449 const X86Subtarget *Subtarget) {
9450 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9451 // the FP state in cases where an emms may be missing.
9452 // A preferable solution to the general problem is to figure out the right
9453 // places to insert EMMS. This qualifies as a quick hack.
9455 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9456 StoreSDNode *St = cast<StoreSDNode>(N);
9457 EVT VT = St->getValue().getValueType();
9458 if (VT.getSizeInBits() != 64)
9461 const Function *F = DAG.getMachineFunction().getFunction();
9462 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9463 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9464 && Subtarget->hasSSE2();
9465 if ((VT.isVector() ||
9466 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9467 isa<LoadSDNode>(St->getValue()) &&
9468 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9469 St->getChain().hasOneUse() && !St->isVolatile()) {
9470 SDNode* LdVal = St->getValue().getNode();
9472 int TokenFactorIndex = -1;
9473 SmallVector<SDValue, 8> Ops;
9474 SDNode* ChainVal = St->getChain().getNode();
9475 // Must be a store of a load. We currently handle two cases: the load
9476 // is a direct child, and it's under an intervening TokenFactor. It is
9477 // possible to dig deeper under nested TokenFactors.
9478 if (ChainVal == LdVal)
9479 Ld = cast<LoadSDNode>(St->getChain());
9480 else if (St->getValue().hasOneUse() &&
9481 ChainVal->getOpcode() == ISD::TokenFactor) {
9482 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9483 if (ChainVal->getOperand(i).getNode() == LdVal) {
9484 TokenFactorIndex = i;
9485 Ld = cast<LoadSDNode>(St->getValue());
9487 Ops.push_back(ChainVal->getOperand(i));
9491 if (!Ld || !ISD::isNormalLoad(Ld))
9494 // If this is not the MMX case, i.e. we are just turning i64 load/store
9495 // into f64 load/store, avoid the transformation if there are multiple
9496 // uses of the loaded value.
9497 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9500 DebugLoc LdDL = Ld->getDebugLoc();
9501 DebugLoc StDL = N->getDebugLoc();
9502 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9503 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9505 if (Subtarget->is64Bit() || F64IsLegal) {
9506 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9507 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9508 Ld->getBasePtr(), Ld->getSrcValue(),
9509 Ld->getSrcValueOffset(), Ld->isVolatile(),
9510 Ld->getAlignment());
9511 SDValue NewChain = NewLd.getValue(1);
9512 if (TokenFactorIndex != -1) {
9513 Ops.push_back(NewChain);
9514 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9517 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9518 St->getSrcValue(), St->getSrcValueOffset(),
9519 St->isVolatile(), St->getAlignment());
9522 // Otherwise, lower to two pairs of 32-bit loads / stores.
9523 SDValue LoAddr = Ld->getBasePtr();
9524 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9525 DAG.getConstant(4, MVT::i32));
9527 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9528 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9529 Ld->isVolatile(), Ld->getAlignment());
9530 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9531 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9533 MinAlign(Ld->getAlignment(), 4));
9535 SDValue NewChain = LoLd.getValue(1);
9536 if (TokenFactorIndex != -1) {
9537 Ops.push_back(LoLd);
9538 Ops.push_back(HiLd);
9539 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9543 LoAddr = St->getBasePtr();
9544 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9545 DAG.getConstant(4, MVT::i32));
9547 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9548 St->getSrcValue(), St->getSrcValueOffset(),
9549 St->isVolatile(), St->getAlignment());
9550 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9552 St->getSrcValueOffset() + 4,
9554 MinAlign(St->getAlignment(), 4));
9555 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9560 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9561 /// X86ISD::FXOR nodes.
9562 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9563 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9564 // F[X]OR(0.0, x) -> x
9565 // F[X]OR(x, 0.0) -> x
9566 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9567 if (C->getValueAPF().isPosZero())
9568 return N->getOperand(1);
9569 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9570 if (C->getValueAPF().isPosZero())
9571 return N->getOperand(0);
9575 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9576 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9577 // FAND(0.0, x) -> 0.0
9578 // FAND(x, 0.0) -> 0.0
9579 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9580 if (C->getValueAPF().isPosZero())
9581 return N->getOperand(0);
9582 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9583 if (C->getValueAPF().isPosZero())
9584 return N->getOperand(1);
9588 static SDValue PerformBTCombine(SDNode *N,
9590 TargetLowering::DAGCombinerInfo &DCI) {
9591 // BT ignores high bits in the bit index operand.
9592 SDValue Op1 = N->getOperand(1);
9593 if (Op1.hasOneUse()) {
9594 unsigned BitWidth = Op1.getValueSizeInBits();
9595 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9596 APInt KnownZero, KnownOne;
9597 TargetLowering::TargetLoweringOpt TLO(DAG);
9598 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9599 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9600 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9601 DCI.CommitTargetLoweringOpt(TLO);
9606 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9607 SDValue Op = N->getOperand(0);
9608 if (Op.getOpcode() == ISD::BIT_CONVERT)
9609 Op = Op.getOperand(0);
9610 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9611 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9612 VT.getVectorElementType().getSizeInBits() ==
9613 OpVT.getVectorElementType().getSizeInBits()) {
9614 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9619 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9620 // Locked instructions, in turn, have implicit fence semantics (all memory
9621 // operations are flushed before issuing the locked instruction, and the
9622 // are not buffered), so we can fold away the common pattern of
9623 // fence-atomic-fence.
9624 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9625 SDValue atomic = N->getOperand(0);
9626 switch (atomic.getOpcode()) {
9627 case ISD::ATOMIC_CMP_SWAP:
9628 case ISD::ATOMIC_SWAP:
9629 case ISD::ATOMIC_LOAD_ADD:
9630 case ISD::ATOMIC_LOAD_SUB:
9631 case ISD::ATOMIC_LOAD_AND:
9632 case ISD::ATOMIC_LOAD_OR:
9633 case ISD::ATOMIC_LOAD_XOR:
9634 case ISD::ATOMIC_LOAD_NAND:
9635 case ISD::ATOMIC_LOAD_MIN:
9636 case ISD::ATOMIC_LOAD_MAX:
9637 case ISD::ATOMIC_LOAD_UMIN:
9638 case ISD::ATOMIC_LOAD_UMAX:
9644 SDValue fence = atomic.getOperand(0);
9645 if (fence.getOpcode() != ISD::MEMBARRIER)
9648 switch (atomic.getOpcode()) {
9649 case ISD::ATOMIC_CMP_SWAP:
9650 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9651 atomic.getOperand(1), atomic.getOperand(2),
9652 atomic.getOperand(3));
9653 case ISD::ATOMIC_SWAP:
9654 case ISD::ATOMIC_LOAD_ADD:
9655 case ISD::ATOMIC_LOAD_SUB:
9656 case ISD::ATOMIC_LOAD_AND:
9657 case ISD::ATOMIC_LOAD_OR:
9658 case ISD::ATOMIC_LOAD_XOR:
9659 case ISD::ATOMIC_LOAD_NAND:
9660 case ISD::ATOMIC_LOAD_MIN:
9661 case ISD::ATOMIC_LOAD_MAX:
9662 case ISD::ATOMIC_LOAD_UMIN:
9663 case ISD::ATOMIC_LOAD_UMAX:
9664 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9665 atomic.getOperand(1), atomic.getOperand(2));
9671 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9672 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9673 // (and (i32 x86isd::setcc_carry), 1)
9674 // This eliminates the zext. This transformation is necessary because
9675 // ISD::SETCC is always legalized to i8.
9676 DebugLoc dl = N->getDebugLoc();
9677 SDValue N0 = N->getOperand(0);
9678 EVT VT = N->getValueType(0);
9679 if (N0.getOpcode() == ISD::AND &&
9681 N0.getOperand(0).hasOneUse()) {
9682 SDValue N00 = N0.getOperand(0);
9683 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9685 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9686 if (!C || C->getZExtValue() != 1)
9688 return DAG.getNode(ISD::AND, dl, VT,
9689 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9690 N00.getOperand(0), N00.getOperand(1)),
9691 DAG.getConstant(1, VT));
9697 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9698 DAGCombinerInfo &DCI) const {
9699 SelectionDAG &DAG = DCI.DAG;
9700 switch (N->getOpcode()) {
9702 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9703 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9704 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9705 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9708 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9709 case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
9710 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9712 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9713 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9714 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9715 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9716 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9717 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9723 //===----------------------------------------------------------------------===//
9724 // X86 Inline Assembly Support
9725 //===----------------------------------------------------------------------===//
9727 static bool LowerToBSwap(CallInst *CI) {
9728 // FIXME: this should verify that we are targetting a 486 or better. If not,
9729 // we will turn this bswap into something that will be lowered to logical ops
9730 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9731 // so don't worry about this.
9733 // Verify this is a simple bswap.
9734 if (CI->getNumOperands() != 2 ||
9735 CI->getType() != CI->getOperand(1)->getType() ||
9736 !CI->getType()->isInteger())
9739 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9740 if (!Ty || Ty->getBitWidth() % 16 != 0)
9743 // Okay, we can do this xform, do so now.
9744 const Type *Tys[] = { Ty };
9745 Module *M = CI->getParent()->getParent()->getParent();
9746 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9748 Value *Op = CI->getOperand(1);
9749 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9751 CI->replaceAllUsesWith(Op);
9752 CI->eraseFromParent();
9756 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9757 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9758 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9760 std::string AsmStr = IA->getAsmString();
9762 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9763 SmallVector<StringRef, 4> AsmPieces;
9764 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9766 switch (AsmPieces.size()) {
9767 default: return false;
9769 AsmStr = AsmPieces[0];
9771 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9774 if (AsmPieces.size() == 2 &&
9775 (AsmPieces[0] == "bswap" ||
9776 AsmPieces[0] == "bswapq" ||
9777 AsmPieces[0] == "bswapl") &&
9778 (AsmPieces[1] == "$0" ||
9779 AsmPieces[1] == "${0:q}")) {
9780 // No need to check constraints, nothing other than the equivalent of
9781 // "=r,0" would be valid here.
9782 return LowerToBSwap(CI);
9784 // rorw $$8, ${0:w} --> llvm.bswap.i16
9785 if (CI->getType()->isInteger(16) &&
9786 AsmPieces.size() == 3 &&
9787 AsmPieces[0] == "rorw" &&
9788 AsmPieces[1] == "$$8," &&
9789 AsmPieces[2] == "${0:w}" &&
9790 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
9791 return LowerToBSwap(CI);
9795 if (CI->getType()->isInteger(64) &&
9796 Constraints.size() >= 2 &&
9797 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9798 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9799 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9800 SmallVector<StringRef, 4> Words;
9801 SplitString(AsmPieces[0], Words, " \t");
9802 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9804 SplitString(AsmPieces[1], Words, " \t");
9805 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9807 SplitString(AsmPieces[2], Words, " \t,");
9808 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9809 Words[2] == "%edx") {
9810 return LowerToBSwap(CI);
9822 /// getConstraintType - Given a constraint letter, return the type of
9823 /// constraint it is for this target.
9824 X86TargetLowering::ConstraintType
9825 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
9826 if (Constraint.size() == 1) {
9827 switch (Constraint[0]) {
9839 return C_RegisterClass;
9847 return TargetLowering::getConstraintType(Constraint);
9850 /// LowerXConstraint - try to replace an X constraint, which matches anything,
9851 /// with another that has more specific requirements based on the type of the
9852 /// corresponding operand.
9853 const char *X86TargetLowering::
9854 LowerXConstraint(EVT ConstraintVT) const {
9855 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
9856 // 'f' like normal targets.
9857 if (ConstraintVT.isFloatingPoint()) {
9858 if (Subtarget->hasSSE2())
9860 if (Subtarget->hasSSE1())
9864 return TargetLowering::LowerXConstraint(ConstraintVT);
9867 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9868 /// vector. If it is invalid, don't add anything to Ops.
9869 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9872 std::vector<SDValue>&Ops,
9873 SelectionDAG &DAG) const {
9874 SDValue Result(0, 0);
9876 switch (Constraint) {
9879 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9880 if (C->getZExtValue() <= 31) {
9881 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9887 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9888 if (C->getZExtValue() <= 63) {
9889 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9895 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9896 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9897 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9903 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9904 if (C->getZExtValue() <= 255) {
9905 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9911 // 32-bit signed value
9912 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9913 const ConstantInt *CI = C->getConstantIntValue();
9914 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9915 C->getSExtValue())) {
9916 // Widen to 64 bits here to get it sign extended.
9917 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9920 // FIXME gcc accepts some relocatable values here too, but only in certain
9921 // memory models; it's complicated.
9926 // 32-bit unsigned value
9927 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9928 const ConstantInt *CI = C->getConstantIntValue();
9929 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9930 C->getZExtValue())) {
9931 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9935 // FIXME gcc accepts some relocatable values here too, but only in certain
9936 // memory models; it's complicated.
9940 // Literal immediates are always ok.
9941 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
9942 // Widen to 64 bits here to get it sign extended.
9943 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
9947 // If we are in non-pic codegen mode, we allow the address of a global (with
9948 // an optional displacement) to be used with 'i'.
9949 GlobalAddressSDNode *GA = 0;
9952 // Match either (GA), (GA+C), (GA+C1+C2), etc.
9954 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
9955 Offset += GA->getOffset();
9957 } else if (Op.getOpcode() == ISD::ADD) {
9958 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9959 Offset += C->getZExtValue();
9960 Op = Op.getOperand(0);
9963 } else if (Op.getOpcode() == ISD::SUB) {
9964 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9965 Offset += -C->getZExtValue();
9966 Op = Op.getOperand(0);
9971 // Otherwise, this isn't something we can handle, reject it.
9975 GlobalValue *GV = GA->getGlobal();
9976 // If we require an extra load to get this address, as in PIC mode, we
9978 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
9979 getTargetMachine())))
9983 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
9985 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
9991 if (Result.getNode()) {
9992 Ops.push_back(Result);
9995 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
9999 std::vector<unsigned> X86TargetLowering::
10000 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10002 if (Constraint.size() == 1) {
10003 // FIXME: not handling fp-stack yet!
10004 switch (Constraint[0]) { // GCC X86 Constraint Letters
10005 default: break; // Unknown constraint letter
10006 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10007 if (Subtarget->is64Bit()) {
10008 if (VT == MVT::i32)
10009 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10010 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10011 X86::R10D,X86::R11D,X86::R12D,
10012 X86::R13D,X86::R14D,X86::R15D,
10013 X86::EBP, X86::ESP, 0);
10014 else if (VT == MVT::i16)
10015 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10016 X86::SI, X86::DI, X86::R8W,X86::R9W,
10017 X86::R10W,X86::R11W,X86::R12W,
10018 X86::R13W,X86::R14W,X86::R15W,
10019 X86::BP, X86::SP, 0);
10020 else if (VT == MVT::i8)
10021 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10022 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10023 X86::R10B,X86::R11B,X86::R12B,
10024 X86::R13B,X86::R14B,X86::R15B,
10025 X86::BPL, X86::SPL, 0);
10027 else if (VT == MVT::i64)
10028 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10029 X86::RSI, X86::RDI, X86::R8, X86::R9,
10030 X86::R10, X86::R11, X86::R12,
10031 X86::R13, X86::R14, X86::R15,
10032 X86::RBP, X86::RSP, 0);
10036 // 32-bit fallthrough
10037 case 'Q': // Q_REGS
10038 if (VT == MVT::i32)
10039 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10040 else if (VT == MVT::i16)
10041 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10042 else if (VT == MVT::i8)
10043 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10044 else if (VT == MVT::i64)
10045 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10050 return std::vector<unsigned>();
10053 std::pair<unsigned, const TargetRegisterClass*>
10054 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10056 // First, see if this is a constraint that directly corresponds to an LLVM
10058 if (Constraint.size() == 1) {
10059 // GCC Constraint Letters
10060 switch (Constraint[0]) {
10062 case 'r': // GENERAL_REGS
10063 case 'l': // INDEX_REGS
10065 return std::make_pair(0U, X86::GR8RegisterClass);
10066 if (VT == MVT::i16)
10067 return std::make_pair(0U, X86::GR16RegisterClass);
10068 if (VT == MVT::i32 || !Subtarget->is64Bit())
10069 return std::make_pair(0U, X86::GR32RegisterClass);
10070 return std::make_pair(0U, X86::GR64RegisterClass);
10071 case 'R': // LEGACY_REGS
10073 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10074 if (VT == MVT::i16)
10075 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10076 if (VT == MVT::i32 || !Subtarget->is64Bit())
10077 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10078 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10079 case 'f': // FP Stack registers.
10080 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10081 // value to the correct fpstack register class.
10082 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10083 return std::make_pair(0U, X86::RFP32RegisterClass);
10084 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10085 return std::make_pair(0U, X86::RFP64RegisterClass);
10086 return std::make_pair(0U, X86::RFP80RegisterClass);
10087 case 'y': // MMX_REGS if MMX allowed.
10088 if (!Subtarget->hasMMX()) break;
10089 return std::make_pair(0U, X86::VR64RegisterClass);
10090 case 'Y': // SSE_REGS if SSE2 allowed
10091 if (!Subtarget->hasSSE2()) break;
10093 case 'x': // SSE_REGS if SSE1 allowed
10094 if (!Subtarget->hasSSE1()) break;
10096 switch (VT.getSimpleVT().SimpleTy) {
10098 // Scalar SSE types.
10101 return std::make_pair(0U, X86::FR32RegisterClass);
10104 return std::make_pair(0U, X86::FR64RegisterClass);
10112 return std::make_pair(0U, X86::VR128RegisterClass);
10118 // Use the default implementation in TargetLowering to convert the register
10119 // constraint into a member of a register class.
10120 std::pair<unsigned, const TargetRegisterClass*> Res;
10121 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10123 // Not found as a standard register?
10124 if (Res.second == 0) {
10125 // Map st(0) -> st(7) -> ST0
10126 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10127 tolower(Constraint[1]) == 's' &&
10128 tolower(Constraint[2]) == 't' &&
10129 Constraint[3] == '(' &&
10130 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10131 Constraint[5] == ')' &&
10132 Constraint[6] == '}') {
10134 Res.first = X86::ST0+Constraint[4]-'0';
10135 Res.second = X86::RFP80RegisterClass;
10139 // GCC allows "st(0)" to be called just plain "st".
10140 if (StringRef("{st}").equals_lower(Constraint)) {
10141 Res.first = X86::ST0;
10142 Res.second = X86::RFP80RegisterClass;
10147 if (StringRef("{flags}").equals_lower(Constraint)) {
10148 Res.first = X86::EFLAGS;
10149 Res.second = X86::CCRRegisterClass;
10153 // 'A' means EAX + EDX.
10154 if (Constraint == "A") {
10155 Res.first = X86::EAX;
10156 Res.second = X86::GR32_ADRegisterClass;
10162 // Otherwise, check to see if this is a register class of the wrong value
10163 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10164 // turn into {ax},{dx}.
10165 if (Res.second->hasType(VT))
10166 return Res; // Correct type already, nothing to do.
10168 // All of the single-register GCC register classes map their values onto
10169 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10170 // really want an 8-bit or 32-bit register, map to the appropriate register
10171 // class and return the appropriate register.
10172 if (Res.second == X86::GR16RegisterClass) {
10173 if (VT == MVT::i8) {
10174 unsigned DestReg = 0;
10175 switch (Res.first) {
10177 case X86::AX: DestReg = X86::AL; break;
10178 case X86::DX: DestReg = X86::DL; break;
10179 case X86::CX: DestReg = X86::CL; break;
10180 case X86::BX: DestReg = X86::BL; break;
10183 Res.first = DestReg;
10184 Res.second = X86::GR8RegisterClass;
10186 } else if (VT == MVT::i32) {
10187 unsigned DestReg = 0;
10188 switch (Res.first) {
10190 case X86::AX: DestReg = X86::EAX; break;
10191 case X86::DX: DestReg = X86::EDX; break;
10192 case X86::CX: DestReg = X86::ECX; break;
10193 case X86::BX: DestReg = X86::EBX; break;
10194 case X86::SI: DestReg = X86::ESI; break;
10195 case X86::DI: DestReg = X86::EDI; break;
10196 case X86::BP: DestReg = X86::EBP; break;
10197 case X86::SP: DestReg = X86::ESP; break;
10200 Res.first = DestReg;
10201 Res.second = X86::GR32RegisterClass;
10203 } else if (VT == MVT::i64) {
10204 unsigned DestReg = 0;
10205 switch (Res.first) {
10207 case X86::AX: DestReg = X86::RAX; break;
10208 case X86::DX: DestReg = X86::RDX; break;
10209 case X86::CX: DestReg = X86::RCX; break;
10210 case X86::BX: DestReg = X86::RBX; break;
10211 case X86::SI: DestReg = X86::RSI; break;
10212 case X86::DI: DestReg = X86::RDI; break;
10213 case X86::BP: DestReg = X86::RBP; break;
10214 case X86::SP: DestReg = X86::RSP; break;
10217 Res.first = DestReg;
10218 Res.second = X86::GR64RegisterClass;
10221 } else if (Res.second == X86::FR32RegisterClass ||
10222 Res.second == X86::FR64RegisterClass ||
10223 Res.second == X86::VR128RegisterClass) {
10224 // Handle references to XMM physical registers that got mapped into the
10225 // wrong class. This can happen with constraints like {xmm0} where the
10226 // target independent register mapper will just pick the first match it can
10227 // find, ignoring the required type.
10228 if (VT == MVT::f32)
10229 Res.second = X86::FR32RegisterClass;
10230 else if (VT == MVT::f64)
10231 Res.second = X86::FR64RegisterClass;
10232 else if (X86::VR128RegisterClass->hasType(VT))
10233 Res.second = X86::VR128RegisterClass;
10239 //===----------------------------------------------------------------------===//
10240 // X86 Widen vector type
10241 //===----------------------------------------------------------------------===//
10243 /// getWidenVectorType: given a vector type, returns the type to widen
10244 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
10245 /// If there is no vector type that we want to widen to, returns MVT::Other
10246 /// When and where to widen is target dependent based on the cost of
10247 /// scalarizing vs using the wider vector type.
10249 EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
10250 assert(VT.isVector());
10251 if (isTypeLegal(VT))
10254 // TODO: In computeRegisterProperty, we can compute the list of legal vector
10255 // type based on element type. This would speed up our search (though
10256 // it may not be worth it since the size of the list is relatively
10258 EVT EltVT = VT.getVectorElementType();
10259 unsigned NElts = VT.getVectorNumElements();
10261 // On X86, it make sense to widen any vector wider than 1
10265 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
10266 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
10267 EVT SVT = (MVT::SimpleValueType)nVT;
10269 if (isTypeLegal(SVT) &&
10270 SVT.getVectorElementType() == EltVT &&
10271 SVT.getVectorNumElements() > NElts)