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 //===----------------------------------------------------------------------===//
16 #include "X86InstrBuilder.h"
17 #include "X86ISelLowering.h"
18 #include "X86TargetMachine.h"
19 #include "X86TargetObjectFile.h"
20 #include "llvm/CallingConv.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/GlobalAlias.h"
24 #include "llvm/GlobalVariable.h"
25 #include "llvm/Function.h"
26 #include "llvm/Instructions.h"
27 #include "llvm/Intrinsics.h"
28 #include "llvm/LLVMContext.h"
29 #include "llvm/ADT/BitVector.h"
30 #include "llvm/ADT/VectorExtras.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/PseudoSourceValue.h"
37 #include "llvm/Support/MathExtras.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/ErrorHandling.h"
40 #include "llvm/Target/TargetOptions.h"
41 #include "llvm/ADT/SmallSet.h"
42 #include "llvm/ADT/StringExtras.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/Support/raw_ostream.h"
48 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
50 // Disable16Bit - 16-bit operations typically have a larger encoding than
51 // corresponding 32-bit instructions, and 16-bit code is slow on some
52 // processors. This is an experimental flag to disable 16-bit operations
53 // (which forces them to be Legalized to 32-bit operations).
55 Disable16Bit("disable-16bit", cl::Hidden,
56 cl::desc("Disable use of 16-bit instructions"));
58 // Forward declarations.
59 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
62 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
63 switch (TM.getSubtarget<X86Subtarget>().TargetType) {
64 default: llvm_unreachable("unknown subtarget type");
65 case X86Subtarget::isDarwin:
66 if (TM.getSubtarget<X86Subtarget>().is64Bit())
67 return new X8664_MachoTargetObjectFile();
68 return new X8632_MachoTargetObjectFile();
69 case X86Subtarget::isELF:
70 return new TargetLoweringObjectFileELF();
71 case X86Subtarget::isMingw:
72 case X86Subtarget::isCygwin:
73 case X86Subtarget::isWindows:
74 return new TargetLoweringObjectFileCOFF();
79 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
80 : TargetLowering(TM, createTLOF(TM)) {
81 Subtarget = &TM.getSubtarget<X86Subtarget>();
82 X86ScalarSSEf64 = Subtarget->hasSSE2();
83 X86ScalarSSEf32 = Subtarget->hasSSE1();
84 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
86 RegInfo = TM.getRegisterInfo();
89 // Set up the TargetLowering object.
91 // X86 is weird, it always uses i8 for shift amounts and setcc results.
92 setShiftAmountType(MVT::i8);
93 setBooleanContents(ZeroOrOneBooleanContent);
94 setSchedulingPreference(SchedulingForRegPressure);
95 setStackPointerRegisterToSaveRestore(X86StackPtr);
97 if (Subtarget->isTargetDarwin()) {
98 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
99 setUseUnderscoreSetJmp(false);
100 setUseUnderscoreLongJmp(false);
101 } else if (Subtarget->isTargetMingw()) {
102 // MS runtime is weird: it exports _setjmp, but longjmp!
103 setUseUnderscoreSetJmp(true);
104 setUseUnderscoreLongJmp(false);
106 setUseUnderscoreSetJmp(true);
107 setUseUnderscoreLongJmp(true);
110 // Set up the register classes.
111 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
113 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
114 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
115 if (Subtarget->is64Bit())
116 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
118 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
120 // We don't accept any truncstore of integer registers.
121 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
123 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
124 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
126 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
127 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
128 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
130 // SETOEQ and SETUNE require checking two conditions.
131 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
132 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
133 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
134 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
135 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
136 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
138 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
140 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
141 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
142 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
144 if (Subtarget->is64Bit()) {
145 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
146 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
147 } else if (!UseSoftFloat) {
148 if (X86ScalarSSEf64) {
149 // We have an impenetrably clever algorithm for ui64->double only.
150 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
152 // We have an algorithm for SSE2, and we turn this into a 64-bit
153 // FILD for other targets.
154 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
157 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
159 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
160 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
163 // SSE has no i16 to fp conversion, only i32
164 if (X86ScalarSSEf32) {
165 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
166 // f32 and f64 cases are Legal, f80 case is not
167 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
169 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
170 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
173 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
174 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
177 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
178 // are Legal, f80 is custom lowered.
179 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
180 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
182 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
184 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
185 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
187 if (X86ScalarSSEf32) {
188 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
189 // f32 and f64 cases are Legal, f80 case is not
190 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
192 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
193 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
196 // Handle FP_TO_UINT by promoting the destination to a larger signed
198 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
199 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
200 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
202 if (Subtarget->is64Bit()) {
203 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
204 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
205 } else if (!UseSoftFloat) {
206 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
207 // Expand FP_TO_UINT into a select.
208 // FIXME: We would like to use a Custom expander here eventually to do
209 // the optimal thing for SSE vs. the default expansion in the legalizer.
210 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
212 // With SSE3 we can use fisttpll to convert to a signed i64; without
213 // SSE, we're stuck with a fistpll.
214 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
217 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
218 if (!X86ScalarSSEf64) {
219 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
220 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
223 // Scalar integer divide and remainder are lowered to use operations that
224 // produce two results, to match the available instructions. This exposes
225 // the two-result form to trivial CSE, which is able to combine x/y and x%y
226 // into a single instruction.
228 // Scalar integer multiply-high is also lowered to use two-result
229 // operations, to match the available instructions. However, plain multiply
230 // (low) operations are left as Legal, as there are single-result
231 // instructions for this in x86. Using the two-result multiply instructions
232 // when both high and low results are needed must be arranged by dagcombine.
233 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
234 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
235 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
236 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
237 setOperationAction(ISD::SREM , MVT::i8 , Expand);
238 setOperationAction(ISD::UREM , MVT::i8 , Expand);
239 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
240 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
241 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
242 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
243 setOperationAction(ISD::SREM , MVT::i16 , Expand);
244 setOperationAction(ISD::UREM , MVT::i16 , Expand);
245 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
246 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
247 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
248 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
249 setOperationAction(ISD::SREM , MVT::i32 , Expand);
250 setOperationAction(ISD::UREM , MVT::i32 , Expand);
251 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
252 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
253 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
254 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
255 setOperationAction(ISD::SREM , MVT::i64 , Expand);
256 setOperationAction(ISD::UREM , MVT::i64 , Expand);
258 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
259 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
260 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
261 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
262 if (Subtarget->is64Bit())
263 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
264 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
265 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
266 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
267 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
268 setOperationAction(ISD::FREM , MVT::f32 , Expand);
269 setOperationAction(ISD::FREM , MVT::f64 , Expand);
270 setOperationAction(ISD::FREM , MVT::f80 , Expand);
271 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
273 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
274 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
275 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
276 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
278 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
279 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
281 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
282 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
284 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
285 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
286 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
287 if (Subtarget->is64Bit()) {
288 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
289 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
290 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
293 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
294 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
296 // These should be promoted to a larger select which is supported.
297 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
298 // X86 wants to expand cmov itself.
299 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
301 setOperationAction(ISD::SELECT , MVT::i16 , Expand);
303 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
304 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
305 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
306 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
307 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
308 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
310 setOperationAction(ISD::SETCC , MVT::i16 , Expand);
312 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
313 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
314 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
315 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
316 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
317 if (Subtarget->is64Bit()) {
318 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
319 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
321 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
324 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
325 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
326 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
327 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
328 if (Subtarget->is64Bit())
329 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
330 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
331 if (Subtarget->is64Bit()) {
332 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
333 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
334 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
335 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
337 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
338 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
339 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
340 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
341 if (Subtarget->is64Bit()) {
342 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
343 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
344 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
347 if (Subtarget->hasSSE1())
348 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
350 if (!Subtarget->hasSSE2())
351 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
353 // Expand certain atomics
354 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
355 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
356 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
357 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
359 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
360 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
361 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
362 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
364 if (!Subtarget->is64Bit()) {
365 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
366 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
367 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
368 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
369 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
370 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
371 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
374 // Use the default ISD::DBG_STOPPOINT.
375 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
376 // FIXME - use subtarget debug flags
377 if (!Subtarget->isTargetDarwin() &&
378 !Subtarget->isTargetELF() &&
379 !Subtarget->isTargetCygMing()) {
380 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
381 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
384 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
385 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
386 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
387 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
388 if (Subtarget->is64Bit()) {
389 setExceptionPointerRegister(X86::RAX);
390 setExceptionSelectorRegister(X86::RDX);
392 setExceptionPointerRegister(X86::EAX);
393 setExceptionSelectorRegister(X86::EDX);
395 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
396 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
398 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
400 setOperationAction(ISD::TRAP, MVT::Other, Legal);
402 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
403 setOperationAction(ISD::VASTART , MVT::Other, Custom);
404 setOperationAction(ISD::VAEND , MVT::Other, Expand);
405 if (Subtarget->is64Bit()) {
406 setOperationAction(ISD::VAARG , MVT::Other, Custom);
407 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
409 setOperationAction(ISD::VAARG , MVT::Other, Expand);
410 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
413 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
414 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
415 if (Subtarget->is64Bit())
416 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
417 if (Subtarget->isTargetCygMing())
418 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
420 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
422 if (!UseSoftFloat && X86ScalarSSEf64) {
423 // f32 and f64 use SSE.
424 // Set up the FP register classes.
425 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
426 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
428 // Use ANDPD to simulate FABS.
429 setOperationAction(ISD::FABS , MVT::f64, Custom);
430 setOperationAction(ISD::FABS , MVT::f32, Custom);
432 // Use XORP to simulate FNEG.
433 setOperationAction(ISD::FNEG , MVT::f64, Custom);
434 setOperationAction(ISD::FNEG , MVT::f32, Custom);
436 // Use ANDPD and ORPD to simulate FCOPYSIGN.
437 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
438 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
440 // We don't support sin/cos/fmod
441 setOperationAction(ISD::FSIN , MVT::f64, Expand);
442 setOperationAction(ISD::FCOS , MVT::f64, Expand);
443 setOperationAction(ISD::FSIN , MVT::f32, Expand);
444 setOperationAction(ISD::FCOS , MVT::f32, Expand);
446 // Expand FP immediates into loads from the stack, except for the special
448 addLegalFPImmediate(APFloat(+0.0)); // xorpd
449 addLegalFPImmediate(APFloat(+0.0f)); // xorps
450 } else if (!UseSoftFloat && X86ScalarSSEf32) {
451 // Use SSE for f32, x87 for f64.
452 // Set up the FP register classes.
453 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
454 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
456 // Use ANDPS to simulate FABS.
457 setOperationAction(ISD::FABS , MVT::f32, Custom);
459 // Use XORP to simulate FNEG.
460 setOperationAction(ISD::FNEG , MVT::f32, Custom);
462 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
464 // Use ANDPS and ORPS to simulate FCOPYSIGN.
465 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
466 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
468 // We don't support sin/cos/fmod
469 setOperationAction(ISD::FSIN , MVT::f32, Expand);
470 setOperationAction(ISD::FCOS , MVT::f32, Expand);
472 // Special cases we handle for FP constants.
473 addLegalFPImmediate(APFloat(+0.0f)); // xorps
474 addLegalFPImmediate(APFloat(+0.0)); // FLD0
475 addLegalFPImmediate(APFloat(+1.0)); // FLD1
476 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
477 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
480 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
481 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
483 } else if (!UseSoftFloat) {
484 // f32 and f64 in x87.
485 // Set up the FP register classes.
486 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
487 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
489 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
490 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
491 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
492 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
495 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
496 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
498 addLegalFPImmediate(APFloat(+0.0)); // FLD0
499 addLegalFPImmediate(APFloat(+1.0)); // FLD1
500 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
501 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
502 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
503 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
504 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
505 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
508 // Long double always uses X87.
510 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
511 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
512 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
515 APFloat TmpFlt(+0.0);
516 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
518 addLegalFPImmediate(TmpFlt); // FLD0
520 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
521 APFloat TmpFlt2(+1.0);
522 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
524 addLegalFPImmediate(TmpFlt2); // FLD1
525 TmpFlt2.changeSign();
526 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
530 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
531 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
535 // Always use a library call for pow.
536 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
537 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
538 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
540 setOperationAction(ISD::FLOG, MVT::f80, Expand);
541 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
542 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
543 setOperationAction(ISD::FEXP, MVT::f80, Expand);
544 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
546 // First set operation action for all vector types to either promote
547 // (for widening) or expand (for scalarization). Then we will selectively
548 // turn on ones that can be effectively codegen'd.
549 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
550 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
551 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
566 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
567 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
601 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
602 // with -msoft-float, disable use of MMX as well.
603 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
604 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
605 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
606 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
607 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
608 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
610 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
611 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
612 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
613 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
615 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
616 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
617 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
618 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
620 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
621 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
623 setOperationAction(ISD::AND, MVT::v8i8, Promote);
624 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
625 setOperationAction(ISD::AND, MVT::v4i16, Promote);
626 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
627 setOperationAction(ISD::AND, MVT::v2i32, Promote);
628 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
629 setOperationAction(ISD::AND, MVT::v1i64, Legal);
631 setOperationAction(ISD::OR, MVT::v8i8, Promote);
632 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
633 setOperationAction(ISD::OR, MVT::v4i16, Promote);
634 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
635 setOperationAction(ISD::OR, MVT::v2i32, Promote);
636 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
637 setOperationAction(ISD::OR, MVT::v1i64, Legal);
639 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
640 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
641 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
642 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
643 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
644 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
645 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
647 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
648 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
649 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
650 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
651 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
652 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
653 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
654 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
655 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
657 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
658 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
659 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
660 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
661 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
663 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
664 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
665 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
666 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
668 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
669 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
670 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
671 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
673 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
675 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Expand);
676 setOperationAction(ISD::TRUNCATE, MVT::v8i8, Expand);
677 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
678 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
679 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
680 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
681 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
682 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
683 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
686 if (!UseSoftFloat && Subtarget->hasSSE1()) {
687 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
689 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
690 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
691 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
692 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
693 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
694 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
695 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
696 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
697 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
698 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
699 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
700 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
703 if (!UseSoftFloat && Subtarget->hasSSE2()) {
704 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
706 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
707 // registers cannot be used even for integer operations.
708 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
709 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
710 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
711 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
713 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
714 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
715 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
716 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
717 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
718 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
719 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
720 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
721 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
722 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
723 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
724 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
725 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
726 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
727 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
728 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
730 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
731 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
732 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
733 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
735 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
736 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
737 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
738 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
739 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
741 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
742 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
743 EVT VT = (MVT::SimpleValueType)i;
744 // Do not attempt to custom lower non-power-of-2 vectors
745 if (!isPowerOf2_32(VT.getVectorNumElements()))
747 // Do not attempt to custom lower non-128-bit vectors
748 if (!VT.is128BitVector())
750 setOperationAction(ISD::BUILD_VECTOR,
751 VT.getSimpleVT().SimpleTy, Custom);
752 setOperationAction(ISD::VECTOR_SHUFFLE,
753 VT.getSimpleVT().SimpleTy, Custom);
754 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
755 VT.getSimpleVT().SimpleTy, Custom);
758 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
759 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
760 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
761 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
762 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
763 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
765 if (Subtarget->is64Bit()) {
766 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
767 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
770 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
771 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
772 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
775 // Do not attempt to promote non-128-bit vectors
776 if (!VT.is128BitVector()) {
779 setOperationAction(ISD::AND, SVT, Promote);
780 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
781 setOperationAction(ISD::OR, SVT, Promote);
782 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
783 setOperationAction(ISD::XOR, SVT, Promote);
784 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
785 setOperationAction(ISD::LOAD, SVT, Promote);
786 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
787 setOperationAction(ISD::SELECT, SVT, Promote);
788 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
791 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
793 // Custom lower v2i64 and v2f64 selects.
794 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
795 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
796 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
797 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
799 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
800 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
801 if (!DisableMMX && Subtarget->hasMMX()) {
802 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
803 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
807 if (Subtarget->hasSSE41()) {
808 // FIXME: Do we need to handle scalar-to-vector here?
809 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
811 // i8 and i16 vectors are custom , because the source register and source
812 // source memory operand types are not the same width. f32 vectors are
813 // custom since the immediate controlling the insert encodes additional
815 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
816 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
817 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
818 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
820 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
821 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
822 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
823 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
825 if (Subtarget->is64Bit()) {
826 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
827 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
831 if (Subtarget->hasSSE42()) {
832 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
835 if (!UseSoftFloat && Subtarget->hasAVX()) {
836 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
837 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
838 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
839 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
841 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
842 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
843 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
844 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
845 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
846 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
847 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
848 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
849 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
850 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
851 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
852 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
853 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
854 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
855 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
857 // Operations to consider commented out -v16i16 v32i8
858 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
859 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
860 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
861 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
862 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
863 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
864 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
865 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
866 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
867 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
868 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
869 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
870 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
871 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
873 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
874 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
875 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
876 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
878 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
879 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
880 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
881 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
882 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
884 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
885 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
886 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
887 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
888 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
889 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
892 // Not sure we want to do this since there are no 256-bit integer
895 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
896 // This includes 256-bit vectors
897 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
898 EVT VT = (MVT::SimpleValueType)i;
900 // Do not attempt to custom lower non-power-of-2 vectors
901 if (!isPowerOf2_32(VT.getVectorNumElements()))
904 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
905 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
906 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
909 if (Subtarget->is64Bit()) {
910 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
911 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
916 // Not sure we want to do this since there are no 256-bit integer
919 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
920 // Including 256-bit vectors
921 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
922 EVT VT = (MVT::SimpleValueType)i;
924 if (!VT.is256BitVector()) {
927 setOperationAction(ISD::AND, VT, Promote);
928 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
929 setOperationAction(ISD::OR, VT, Promote);
930 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
931 setOperationAction(ISD::XOR, VT, Promote);
932 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
933 setOperationAction(ISD::LOAD, VT, Promote);
934 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
935 setOperationAction(ISD::SELECT, VT, Promote);
936 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
939 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
943 // We want to custom lower some of our intrinsics.
944 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
946 // Add/Sub/Mul with overflow operations are custom lowered.
947 setOperationAction(ISD::SADDO, MVT::i32, Custom);
948 setOperationAction(ISD::SADDO, MVT::i64, Custom);
949 setOperationAction(ISD::UADDO, MVT::i32, Custom);
950 setOperationAction(ISD::UADDO, MVT::i64, Custom);
951 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
952 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
953 setOperationAction(ISD::USUBO, MVT::i32, Custom);
954 setOperationAction(ISD::USUBO, MVT::i64, Custom);
955 setOperationAction(ISD::SMULO, MVT::i32, Custom);
956 setOperationAction(ISD::SMULO, MVT::i64, Custom);
958 if (!Subtarget->is64Bit()) {
959 // These libcalls are not available in 32-bit.
960 setLibcallName(RTLIB::SHL_I128, 0);
961 setLibcallName(RTLIB::SRL_I128, 0);
962 setLibcallName(RTLIB::SRA_I128, 0);
965 // We have target-specific dag combine patterns for the following nodes:
966 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
967 setTargetDAGCombine(ISD::BUILD_VECTOR);
968 setTargetDAGCombine(ISD::SELECT);
969 setTargetDAGCombine(ISD::SHL);
970 setTargetDAGCombine(ISD::SRA);
971 setTargetDAGCombine(ISD::SRL);
972 setTargetDAGCombine(ISD::STORE);
973 setTargetDAGCombine(ISD::MEMBARRIER);
974 if (Subtarget->is64Bit())
975 setTargetDAGCombine(ISD::MUL);
977 computeRegisterProperties();
979 // FIXME: These should be based on subtarget info. Plus, the values should
980 // be smaller when we are in optimizing for size mode.
981 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
982 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
983 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
984 setPrefLoopAlignment(16);
985 benefitFromCodePlacementOpt = true;
989 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
994 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
995 /// the desired ByVal argument alignment.
996 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
999 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1000 if (VTy->getBitWidth() == 128)
1002 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1003 unsigned EltAlign = 0;
1004 getMaxByValAlign(ATy->getElementType(), EltAlign);
1005 if (EltAlign > MaxAlign)
1006 MaxAlign = EltAlign;
1007 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1008 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1009 unsigned EltAlign = 0;
1010 getMaxByValAlign(STy->getElementType(i), EltAlign);
1011 if (EltAlign > MaxAlign)
1012 MaxAlign = EltAlign;
1020 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1021 /// function arguments in the caller parameter area. For X86, aggregates
1022 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1023 /// are at 4-byte boundaries.
1024 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1025 if (Subtarget->is64Bit()) {
1026 // Max of 8 and alignment of type.
1027 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1034 if (Subtarget->hasSSE1())
1035 getMaxByValAlign(Ty, Align);
1039 /// getOptimalMemOpType - Returns the target specific optimal type for load
1040 /// and store operations as a result of memset, memcpy, and memmove
1041 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
1044 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1045 bool isSrcConst, bool isSrcStr,
1046 SelectionDAG &DAG) const {
1047 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1048 // linux. This is because the stack realignment code can't handle certain
1049 // cases like PR2962. This should be removed when PR2962 is fixed.
1050 const Function *F = DAG.getMachineFunction().getFunction();
1051 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1052 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1053 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1055 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1058 if (Subtarget->is64Bit() && Size >= 8)
1063 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1065 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1066 SelectionDAG &DAG) const {
1067 if (usesGlobalOffsetTable())
1068 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
1069 if (!Subtarget->is64Bit())
1070 // This doesn't have DebugLoc associated with it, but is not really the
1071 // same as a Register.
1072 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1077 /// getFunctionAlignment - Return the Log2 alignment of this function.
1078 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1079 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1082 //===----------------------------------------------------------------------===//
1083 // Return Value Calling Convention Implementation
1084 //===----------------------------------------------------------------------===//
1086 #include "X86GenCallingConv.inc"
1089 X86TargetLowering::LowerReturn(SDValue Chain,
1090 CallingConv::ID CallConv, bool isVarArg,
1091 const SmallVectorImpl<ISD::OutputArg> &Outs,
1092 DebugLoc dl, SelectionDAG &DAG) {
1094 SmallVector<CCValAssign, 16> RVLocs;
1095 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1096 RVLocs, *DAG.getContext());
1097 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1099 // If this is the first return lowered for this function, add the regs to the
1100 // liveout set for the function.
1101 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1102 for (unsigned i = 0; i != RVLocs.size(); ++i)
1103 if (RVLocs[i].isRegLoc())
1104 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1109 SmallVector<SDValue, 6> RetOps;
1110 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1111 // Operand #1 = Bytes To Pop
1112 RetOps.push_back(DAG.getTargetConstant(getBytesToPopOnReturn(), MVT::i16));
1114 // Copy the result values into the output registers.
1115 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1116 CCValAssign &VA = RVLocs[i];
1117 assert(VA.isRegLoc() && "Can only return in registers!");
1118 SDValue ValToCopy = Outs[i].Val;
1120 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1121 // the RET instruction and handled by the FP Stackifier.
1122 if (VA.getLocReg() == X86::ST0 ||
1123 VA.getLocReg() == X86::ST1) {
1124 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1125 // change the value to the FP stack register class.
1126 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1127 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1128 RetOps.push_back(ValToCopy);
1129 // Don't emit a copytoreg.
1133 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1134 // which is returned in RAX / RDX.
1135 if (Subtarget->is64Bit()) {
1136 EVT ValVT = ValToCopy.getValueType();
1137 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1138 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1139 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1140 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1144 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1145 Flag = Chain.getValue(1);
1148 // The x86-64 ABI for returning structs by value requires that we copy
1149 // the sret argument into %rax for the return. We saved the argument into
1150 // a virtual register in the entry block, so now we copy the value out
1152 if (Subtarget->is64Bit() &&
1153 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1154 MachineFunction &MF = DAG.getMachineFunction();
1155 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1156 unsigned Reg = FuncInfo->getSRetReturnReg();
1158 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1159 FuncInfo->setSRetReturnReg(Reg);
1161 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1163 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1164 Flag = Chain.getValue(1);
1166 // RAX now acts like a return value.
1167 MF.getRegInfo().addLiveOut(X86::RAX);
1170 RetOps[0] = Chain; // Update chain.
1172 // Add the flag if we have it.
1174 RetOps.push_back(Flag);
1176 return DAG.getNode(X86ISD::RET_FLAG, dl,
1177 MVT::Other, &RetOps[0], RetOps.size());
1180 /// LowerCallResult - Lower the result values of a call into the
1181 /// appropriate copies out of appropriate physical registers.
1184 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1185 CallingConv::ID CallConv, bool isVarArg,
1186 const SmallVectorImpl<ISD::InputArg> &Ins,
1187 DebugLoc dl, SelectionDAG &DAG,
1188 SmallVectorImpl<SDValue> &InVals) {
1190 // Assign locations to each value returned by this call.
1191 SmallVector<CCValAssign, 16> RVLocs;
1192 bool Is64Bit = Subtarget->is64Bit();
1193 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1194 RVLocs, *DAG.getContext());
1195 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1197 // Copy all of the result registers out of their specified physreg.
1198 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1199 CCValAssign &VA = RVLocs[i];
1200 EVT CopyVT = VA.getValVT();
1202 // If this is x86-64, and we disabled SSE, we can't return FP values
1203 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1204 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1205 llvm_report_error("SSE register return with SSE disabled");
1208 // If this is a call to a function that returns an fp value on the floating
1209 // point stack, but where we prefer to use the value in xmm registers, copy
1210 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1211 if ((VA.getLocReg() == X86::ST0 ||
1212 VA.getLocReg() == X86::ST1) &&
1213 isScalarFPTypeInSSEReg(VA.getValVT())) {
1218 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1219 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1220 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1221 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1222 MVT::v2i64, InFlag).getValue(1);
1223 Val = Chain.getValue(0);
1224 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1225 Val, DAG.getConstant(0, MVT::i64));
1227 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1228 MVT::i64, InFlag).getValue(1);
1229 Val = Chain.getValue(0);
1231 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1233 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1234 CopyVT, InFlag).getValue(1);
1235 Val = Chain.getValue(0);
1237 InFlag = Chain.getValue(2);
1239 if (CopyVT != VA.getValVT()) {
1240 // Round the F80 the right size, which also moves to the appropriate xmm
1242 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1243 // This truncation won't change the value.
1244 DAG.getIntPtrConstant(1));
1247 InVals.push_back(Val);
1254 //===----------------------------------------------------------------------===//
1255 // C & StdCall & Fast Calling Convention implementation
1256 //===----------------------------------------------------------------------===//
1257 // StdCall calling convention seems to be standard for many Windows' API
1258 // routines and around. It differs from C calling convention just a little:
1259 // callee should clean up the stack, not caller. Symbols should be also
1260 // decorated in some fancy way :) It doesn't support any vector arguments.
1261 // For info on fast calling convention see Fast Calling Convention (tail call)
1262 // implementation LowerX86_32FastCCCallTo.
1264 /// CallIsStructReturn - Determines whether a call uses struct return
1266 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1270 return Outs[0].Flags.isSRet();
1273 /// ArgsAreStructReturn - Determines whether a function uses struct
1274 /// return semantics.
1276 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1280 return Ins[0].Flags.isSRet();
1283 /// IsCalleePop - Determines whether the callee is required to pop its
1284 /// own arguments. Callee pop is necessary to support tail calls.
1285 bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
1289 switch (CallingConv) {
1292 case CallingConv::X86_StdCall:
1293 return !Subtarget->is64Bit();
1294 case CallingConv::X86_FastCall:
1295 return !Subtarget->is64Bit();
1296 case CallingConv::Fast:
1297 return PerformTailCallOpt;
1301 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1302 /// given CallingConvention value.
1303 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1304 if (Subtarget->is64Bit()) {
1305 if (Subtarget->isTargetWin64())
1306 return CC_X86_Win64_C;
1311 if (CC == CallingConv::X86_FastCall)
1312 return CC_X86_32_FastCall;
1313 else if (CC == CallingConv::Fast)
1314 return CC_X86_32_FastCC;
1319 /// NameDecorationForCallConv - Selects the appropriate decoration to
1320 /// apply to a MachineFunction containing a given calling convention.
1322 X86TargetLowering::NameDecorationForCallConv(CallingConv::ID CallConv) {
1323 if (CallConv == CallingConv::X86_FastCall)
1325 else if (CallConv == CallingConv::X86_StdCall)
1331 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1332 /// by "Src" to address "Dst" with size and alignment information specified by
1333 /// the specific parameter attribute. The copy will be passed as a byval
1334 /// function parameter.
1336 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1337 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1339 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1340 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1341 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1345 X86TargetLowering::LowerMemArgument(SDValue Chain,
1346 CallingConv::ID CallConv,
1347 const SmallVectorImpl<ISD::InputArg> &Ins,
1348 DebugLoc dl, SelectionDAG &DAG,
1349 const CCValAssign &VA,
1350 MachineFrameInfo *MFI,
1353 // Create the nodes corresponding to a load from this parameter slot.
1354 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1355 bool AlwaysUseMutable = (CallConv==CallingConv::Fast) && PerformTailCallOpt;
1356 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1359 // If value is passed by pointer we have address passed instead of the value
1361 if (VA.getLocInfo() == CCValAssign::Indirect)
1362 ValVT = VA.getLocVT();
1364 ValVT = VA.getValVT();
1366 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1367 // changed with more analysis.
1368 // In case of tail call optimization mark all arguments mutable. Since they
1369 // could be overwritten by lowering of arguments in case of a tail call.
1370 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1371 VA.getLocMemOffset(), isImmutable);
1372 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1373 if (Flags.isByVal())
1375 return DAG.getLoad(ValVT, dl, Chain, FIN,
1376 PseudoSourceValue::getFixedStack(FI), 0);
1380 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1381 CallingConv::ID CallConv,
1383 const SmallVectorImpl<ISD::InputArg> &Ins,
1386 SmallVectorImpl<SDValue> &InVals) {
1388 MachineFunction &MF = DAG.getMachineFunction();
1389 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1391 const Function* Fn = MF.getFunction();
1392 if (Fn->hasExternalLinkage() &&
1393 Subtarget->isTargetCygMing() &&
1394 Fn->getName() == "main")
1395 FuncInfo->setForceFramePointer(true);
1397 // Decorate the function name.
1398 FuncInfo->setDecorationStyle(NameDecorationForCallConv(CallConv));
1400 MachineFrameInfo *MFI = MF.getFrameInfo();
1401 bool Is64Bit = Subtarget->is64Bit();
1402 bool IsWin64 = Subtarget->isTargetWin64();
1404 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1405 "Var args not supported with calling convention fastcc");
1407 // Assign locations to all of the incoming arguments.
1408 SmallVector<CCValAssign, 16> ArgLocs;
1409 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1410 ArgLocs, *DAG.getContext());
1411 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1413 unsigned LastVal = ~0U;
1415 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1416 CCValAssign &VA = ArgLocs[i];
1417 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1419 assert(VA.getValNo() != LastVal &&
1420 "Don't support value assigned to multiple locs yet");
1421 LastVal = VA.getValNo();
1423 if (VA.isRegLoc()) {
1424 EVT RegVT = VA.getLocVT();
1425 TargetRegisterClass *RC = NULL;
1426 if (RegVT == MVT::i32)
1427 RC = X86::GR32RegisterClass;
1428 else if (Is64Bit && RegVT == MVT::i64)
1429 RC = X86::GR64RegisterClass;
1430 else if (RegVT == MVT::f32)
1431 RC = X86::FR32RegisterClass;
1432 else if (RegVT == MVT::f64)
1433 RC = X86::FR64RegisterClass;
1434 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1435 RC = X86::VR128RegisterClass;
1436 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1437 RC = X86::VR64RegisterClass;
1439 llvm_unreachable("Unknown argument type!");
1441 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1442 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1444 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1445 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1447 if (VA.getLocInfo() == CCValAssign::SExt)
1448 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1449 DAG.getValueType(VA.getValVT()));
1450 else if (VA.getLocInfo() == CCValAssign::ZExt)
1451 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1452 DAG.getValueType(VA.getValVT()));
1453 else if (VA.getLocInfo() == CCValAssign::BCvt)
1454 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1456 if (VA.isExtInLoc()) {
1457 // Handle MMX values passed in XMM regs.
1458 if (RegVT.isVector()) {
1459 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1460 ArgValue, DAG.getConstant(0, MVT::i64));
1461 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1463 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1466 assert(VA.isMemLoc());
1467 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1470 // If value is passed via pointer - do a load.
1471 if (VA.getLocInfo() == CCValAssign::Indirect)
1472 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0);
1474 InVals.push_back(ArgValue);
1477 // The x86-64 ABI for returning structs by value requires that we copy
1478 // the sret argument into %rax for the return. Save the argument into
1479 // a virtual register so that we can access it from the return points.
1480 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1481 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1482 unsigned Reg = FuncInfo->getSRetReturnReg();
1484 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1485 FuncInfo->setSRetReturnReg(Reg);
1487 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1488 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1491 unsigned StackSize = CCInfo.getNextStackOffset();
1492 // align stack specially for tail calls
1493 if (PerformTailCallOpt && CallConv == CallingConv::Fast)
1494 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1496 // If the function takes variable number of arguments, make a frame index for
1497 // the start of the first vararg value... for expansion of llvm.va_start.
1499 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1500 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1503 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1505 // FIXME: We should really autogenerate these arrays
1506 static const unsigned GPR64ArgRegsWin64[] = {
1507 X86::RCX, X86::RDX, X86::R8, X86::R9
1509 static const unsigned XMMArgRegsWin64[] = {
1510 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1512 static const unsigned GPR64ArgRegs64Bit[] = {
1513 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1515 static const unsigned XMMArgRegs64Bit[] = {
1516 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1517 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1519 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1522 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1523 GPR64ArgRegs = GPR64ArgRegsWin64;
1524 XMMArgRegs = XMMArgRegsWin64;
1526 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1527 GPR64ArgRegs = GPR64ArgRegs64Bit;
1528 XMMArgRegs = XMMArgRegs64Bit;
1530 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1532 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1535 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1536 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1537 "SSE register cannot be used when SSE is disabled!");
1538 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1539 "SSE register cannot be used when SSE is disabled!");
1540 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1541 // Kernel mode asks for SSE to be disabled, so don't push them
1543 TotalNumXMMRegs = 0;
1545 // For X86-64, if there are vararg parameters that are passed via
1546 // registers, then we must store them to their spots on the stack so they
1547 // may be loaded by deferencing the result of va_next.
1548 VarArgsGPOffset = NumIntRegs * 8;
1549 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1550 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1551 TotalNumXMMRegs * 16, 16);
1553 // Store the integer parameter registers.
1554 SmallVector<SDValue, 8> MemOps;
1555 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1556 unsigned Offset = VarArgsGPOffset;
1557 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1558 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1559 DAG.getIntPtrConstant(Offset));
1560 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1561 X86::GR64RegisterClass);
1562 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1564 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1565 PseudoSourceValue::getStack(),
1567 MemOps.push_back(Store);
1571 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1572 // Now store the XMM (fp + vector) parameter registers.
1573 SmallVector<SDValue, 11> SaveXMMOps;
1574 SaveXMMOps.push_back(Chain);
1576 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1577 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1578 SaveXMMOps.push_back(ALVal);
1580 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1581 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1583 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1584 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1585 X86::VR128RegisterClass);
1586 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1587 SaveXMMOps.push_back(Val);
1589 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1591 &SaveXMMOps[0], SaveXMMOps.size()));
1594 if (!MemOps.empty())
1595 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1596 &MemOps[0], MemOps.size());
1600 // Some CCs need callee pop.
1601 if (IsCalleePop(isVarArg, CallConv)) {
1602 BytesToPopOnReturn = StackSize; // Callee pops everything.
1603 BytesCallerReserves = 0;
1605 BytesToPopOnReturn = 0; // Callee pops nothing.
1606 // If this is an sret function, the return should pop the hidden pointer.
1607 if (!Is64Bit && CallConv != CallingConv::Fast && ArgsAreStructReturn(Ins))
1608 BytesToPopOnReturn = 4;
1609 BytesCallerReserves = StackSize;
1613 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1614 if (CallConv == CallingConv::X86_FastCall)
1615 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1618 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1624 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1625 SDValue StackPtr, SDValue Arg,
1626 DebugLoc dl, SelectionDAG &DAG,
1627 const CCValAssign &VA,
1628 ISD::ArgFlagsTy Flags) {
1629 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1630 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1631 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1632 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1633 if (Flags.isByVal()) {
1634 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1636 return DAG.getStore(Chain, dl, Arg, PtrOff,
1637 PseudoSourceValue::getStack(), LocMemOffset);
1640 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1641 /// optimization is performed and it is required.
1643 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1644 SDValue &OutRetAddr,
1650 if (!IsTailCall || FPDiff==0) return Chain;
1652 // Adjust the Return address stack slot.
1653 EVT VT = getPointerTy();
1654 OutRetAddr = getReturnAddressFrameIndex(DAG);
1656 // Load the "old" Return address.
1657 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1658 return SDValue(OutRetAddr.getNode(), 1);
1661 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1662 /// optimization is performed and it is required (FPDiff!=0).
1664 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1665 SDValue Chain, SDValue RetAddrFrIdx,
1666 bool Is64Bit, int FPDiff, DebugLoc dl) {
1667 // Store the return address to the appropriate stack slot.
1668 if (!FPDiff) return Chain;
1669 // Calculate the new stack slot for the return address.
1670 int SlotSize = Is64Bit ? 8 : 4;
1671 int NewReturnAddrFI =
1672 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1673 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1674 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1675 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1676 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1681 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1682 CallingConv::ID CallConv, bool isVarArg,
1684 const SmallVectorImpl<ISD::OutputArg> &Outs,
1685 const SmallVectorImpl<ISD::InputArg> &Ins,
1686 DebugLoc dl, SelectionDAG &DAG,
1687 SmallVectorImpl<SDValue> &InVals) {
1689 MachineFunction &MF = DAG.getMachineFunction();
1690 bool Is64Bit = Subtarget->is64Bit();
1691 bool IsStructRet = CallIsStructReturn(Outs);
1693 assert((!isTailCall ||
1694 (CallConv == CallingConv::Fast && PerformTailCallOpt)) &&
1695 "IsEligibleForTailCallOptimization missed a case!");
1696 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1697 "Var args not supported with calling convention fastcc");
1699 // Analyze operands of the call, assigning locations to each operand.
1700 SmallVector<CCValAssign, 16> ArgLocs;
1701 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1702 ArgLocs, *DAG.getContext());
1703 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1705 // Get a count of how many bytes are to be pushed on the stack.
1706 unsigned NumBytes = CCInfo.getNextStackOffset();
1707 if (PerformTailCallOpt && CallConv == CallingConv::Fast)
1708 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1712 // Lower arguments at fp - stackoffset + fpdiff.
1713 unsigned NumBytesCallerPushed =
1714 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1715 FPDiff = NumBytesCallerPushed - NumBytes;
1717 // Set the delta of movement of the returnaddr stackslot.
1718 // But only set if delta is greater than previous delta.
1719 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1720 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1723 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1725 SDValue RetAddrFrIdx;
1726 // Load return adress for tail calls.
1727 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall, Is64Bit,
1730 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1731 SmallVector<SDValue, 8> MemOpChains;
1734 // Walk the register/memloc assignments, inserting copies/loads. In the case
1735 // of tail call optimization arguments are handle later.
1736 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1737 CCValAssign &VA = ArgLocs[i];
1738 EVT RegVT = VA.getLocVT();
1739 SDValue Arg = Outs[i].Val;
1740 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1741 bool isByVal = Flags.isByVal();
1743 // Promote the value if needed.
1744 switch (VA.getLocInfo()) {
1745 default: llvm_unreachable("Unknown loc info!");
1746 case CCValAssign::Full: break;
1747 case CCValAssign::SExt:
1748 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1750 case CCValAssign::ZExt:
1751 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1753 case CCValAssign::AExt:
1754 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1755 // Special case: passing MMX values in XMM registers.
1756 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1757 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1758 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1760 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1762 case CCValAssign::BCvt:
1763 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1765 case CCValAssign::Indirect: {
1766 // Store the argument.
1767 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1768 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1769 PseudoSourceValue::getStack(), 0);
1775 if (VA.isRegLoc()) {
1776 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1778 if (!isTailCall || (isTailCall && isByVal)) {
1779 assert(VA.isMemLoc());
1780 if (StackPtr.getNode() == 0)
1781 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1783 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1784 dl, DAG, VA, Flags));
1789 if (!MemOpChains.empty())
1790 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1791 &MemOpChains[0], MemOpChains.size());
1793 // Build a sequence of copy-to-reg nodes chained together with token chain
1794 // and flag operands which copy the outgoing args into registers.
1796 // Tail call byval lowering might overwrite argument registers so in case of
1797 // tail call optimization the copies to registers are lowered later.
1799 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1800 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1801 RegsToPass[i].second, InFlag);
1802 InFlag = Chain.getValue(1);
1806 if (Subtarget->isPICStyleGOT()) {
1807 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1810 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1811 DAG.getNode(X86ISD::GlobalBaseReg,
1812 DebugLoc::getUnknownLoc(),
1815 InFlag = Chain.getValue(1);
1817 // If we are tail calling and generating PIC/GOT style code load the
1818 // address of the callee into ECX. The value in ecx is used as target of
1819 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1820 // for tail calls on PIC/GOT architectures. Normally we would just put the
1821 // address of GOT into ebx and then call target@PLT. But for tail calls
1822 // ebx would be restored (since ebx is callee saved) before jumping to the
1825 // Note: The actual moving to ECX is done further down.
1826 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1827 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1828 !G->getGlobal()->hasProtectedVisibility())
1829 Callee = LowerGlobalAddress(Callee, DAG);
1830 else if (isa<ExternalSymbolSDNode>(Callee))
1831 Callee = LowerExternalSymbol(Callee, DAG);
1835 if (Is64Bit && isVarArg) {
1836 // From AMD64 ABI document:
1837 // For calls that may call functions that use varargs or stdargs
1838 // (prototype-less calls or calls to functions containing ellipsis (...) in
1839 // the declaration) %al is used as hidden argument to specify the number
1840 // of SSE registers used. The contents of %al do not need to match exactly
1841 // the number of registers, but must be an ubound on the number of SSE
1842 // registers used and is in the range 0 - 8 inclusive.
1844 // FIXME: Verify this on Win64
1845 // Count the number of XMM registers allocated.
1846 static const unsigned XMMArgRegs[] = {
1847 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1848 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1850 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1851 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1852 && "SSE registers cannot be used when SSE is disabled");
1854 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1855 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1856 InFlag = Chain.getValue(1);
1860 // For tail calls lower the arguments to the 'real' stack slot.
1862 // Force all the incoming stack arguments to be loaded from the stack
1863 // before any new outgoing arguments are stored to the stack, because the
1864 // outgoing stack slots may alias the incoming argument stack slots, and
1865 // the alias isn't otherwise explicit. This is slightly more conservative
1866 // than necessary, because it means that each store effectively depends
1867 // on every argument instead of just those arguments it would clobber.
1868 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1870 SmallVector<SDValue, 8> MemOpChains2;
1873 // Do not flag preceeding copytoreg stuff together with the following stuff.
1875 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1876 CCValAssign &VA = ArgLocs[i];
1877 if (!VA.isRegLoc()) {
1878 assert(VA.isMemLoc());
1879 SDValue Arg = Outs[i].Val;
1880 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1881 // Create frame index.
1882 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1883 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1884 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1885 FIN = DAG.getFrameIndex(FI, getPointerTy());
1887 if (Flags.isByVal()) {
1888 // Copy relative to framepointer.
1889 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1890 if (StackPtr.getNode() == 0)
1891 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1893 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1895 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
1899 // Store relative to framepointer.
1900 MemOpChains2.push_back(
1901 DAG.getStore(ArgChain, dl, Arg, FIN,
1902 PseudoSourceValue::getFixedStack(FI), 0));
1907 if (!MemOpChains2.empty())
1908 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1909 &MemOpChains2[0], MemOpChains2.size());
1911 // Copy arguments to their registers.
1912 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1913 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1914 RegsToPass[i].second, InFlag);
1915 InFlag = Chain.getValue(1);
1919 // Store the return address to the appropriate stack slot.
1920 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1924 // If the callee is a GlobalAddress node (quite common, every direct call is)
1925 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1926 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1927 // We should use extra load for direct calls to dllimported functions in
1929 GlobalValue *GV = G->getGlobal();
1930 if (!GV->hasDLLImportLinkage()) {
1931 unsigned char OpFlags = 0;
1933 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
1934 // external symbols most go through the PLT in PIC mode. If the symbol
1935 // has hidden or protected visibility, or if it is static or local, then
1936 // we don't need to use the PLT - we can directly call it.
1937 if (Subtarget->isTargetELF() &&
1938 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1939 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
1940 OpFlags = X86II::MO_PLT;
1941 } else if (Subtarget->isPICStyleStubAny() &&
1942 (GV->isDeclaration() || GV->isWeakForLinker()) &&
1943 Subtarget->getDarwinVers() < 9) {
1944 // PC-relative references to external symbols should go through $stub,
1945 // unless we're building with the leopard linker or later, which
1946 // automatically synthesizes these stubs.
1947 OpFlags = X86II::MO_DARWIN_STUB;
1950 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
1951 G->getOffset(), OpFlags);
1953 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1954 unsigned char OpFlags = 0;
1956 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
1957 // symbols should go through the PLT.
1958 if (Subtarget->isTargetELF() &&
1959 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1960 OpFlags = X86II::MO_PLT;
1961 } else if (Subtarget->isPICStyleStubAny() &&
1962 Subtarget->getDarwinVers() < 9) {
1963 // PC-relative references to external symbols should go through $stub,
1964 // unless we're building with the leopard linker or later, which
1965 // automatically synthesizes these stubs.
1966 OpFlags = X86II::MO_DARWIN_STUB;
1969 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
1971 } else if (isTailCall) {
1972 unsigned Opc = Is64Bit ? X86::R11 : X86::EAX;
1974 Chain = DAG.getCopyToReg(Chain, dl,
1975 DAG.getRegister(Opc, getPointerTy()),
1977 Callee = DAG.getRegister(Opc, getPointerTy());
1978 // Add register as live out.
1979 MF.getRegInfo().addLiveOut(Opc);
1982 // Returns a chain & a flag for retval copy to use.
1983 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1984 SmallVector<SDValue, 8> Ops;
1987 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1988 DAG.getIntPtrConstant(0, true), InFlag);
1989 InFlag = Chain.getValue(1);
1992 Ops.push_back(Chain);
1993 Ops.push_back(Callee);
1996 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1998 // Add argument registers to the end of the list so that they are known live
2000 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2001 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2002 RegsToPass[i].second.getValueType()));
2004 // Add an implicit use GOT pointer in EBX.
2005 if (!isTailCall && Subtarget->isPICStyleGOT())
2006 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2008 // Add an implicit use of AL for x86 vararg functions.
2009 if (Is64Bit && isVarArg)
2010 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2012 if (InFlag.getNode())
2013 Ops.push_back(InFlag);
2016 // If this is the first return lowered for this function, add the regs
2017 // to the liveout set for the function.
2018 if (MF.getRegInfo().liveout_empty()) {
2019 SmallVector<CCValAssign, 16> RVLocs;
2020 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2022 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2023 for (unsigned i = 0; i != RVLocs.size(); ++i)
2024 if (RVLocs[i].isRegLoc())
2025 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2028 assert(((Callee.getOpcode() == ISD::Register &&
2029 (cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
2030 cast<RegisterSDNode>(Callee)->getReg() == X86::R9)) ||
2031 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2032 Callee.getOpcode() == ISD::TargetGlobalAddress) &&
2033 "Expecting an global address, external symbol, or register");
2035 return DAG.getNode(X86ISD::TC_RETURN, dl,
2036 NodeTys, &Ops[0], Ops.size());
2039 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2040 InFlag = Chain.getValue(1);
2042 // Create the CALLSEQ_END node.
2043 unsigned NumBytesForCalleeToPush;
2044 if (IsCalleePop(isVarArg, CallConv))
2045 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2046 else if (!Is64Bit && CallConv != CallingConv::Fast && IsStructRet)
2047 // If this is is a call to a struct-return function, the callee
2048 // pops the hidden struct pointer, so we have to push it back.
2049 // This is common for Darwin/X86, Linux & Mingw32 targets.
2050 NumBytesForCalleeToPush = 4;
2052 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2054 // Returns a flag for retval copy to use.
2055 Chain = DAG.getCALLSEQ_END(Chain,
2056 DAG.getIntPtrConstant(NumBytes, true),
2057 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2060 InFlag = Chain.getValue(1);
2062 // Handle result values, copying them out of physregs into vregs that we
2064 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2065 Ins, dl, DAG, InVals);
2069 //===----------------------------------------------------------------------===//
2070 // Fast Calling Convention (tail call) implementation
2071 //===----------------------------------------------------------------------===//
2073 // Like std call, callee cleans arguments, convention except that ECX is
2074 // reserved for storing the tail called function address. Only 2 registers are
2075 // free for argument passing (inreg). Tail call optimization is performed
2077 // * tailcallopt is enabled
2078 // * caller/callee are fastcc
2079 // On X86_64 architecture with GOT-style position independent code only local
2080 // (within module) calls are supported at the moment.
2081 // To keep the stack aligned according to platform abi the function
2082 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2083 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2084 // If a tail called function callee has more arguments than the caller the
2085 // caller needs to make sure that there is room to move the RETADDR to. This is
2086 // achieved by reserving an area the size of the argument delta right after the
2087 // original REtADDR, but before the saved framepointer or the spilled registers
2088 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2100 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2101 /// for a 16 byte align requirement.
2102 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2103 SelectionDAG& DAG) {
2104 MachineFunction &MF = DAG.getMachineFunction();
2105 const TargetMachine &TM = MF.getTarget();
2106 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2107 unsigned StackAlignment = TFI.getStackAlignment();
2108 uint64_t AlignMask = StackAlignment - 1;
2109 int64_t Offset = StackSize;
2110 uint64_t SlotSize = TD->getPointerSize();
2111 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2112 // Number smaller than 12 so just add the difference.
2113 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2115 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2116 Offset = ((~AlignMask) & Offset) + StackAlignment +
2117 (StackAlignment-SlotSize);
2122 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2123 /// for tail call optimization. Targets which want to do tail call
2124 /// optimization should implement this function.
2126 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2127 CallingConv::ID CalleeCC,
2129 const SmallVectorImpl<ISD::InputArg> &Ins,
2130 SelectionDAG& DAG) const {
2131 MachineFunction &MF = DAG.getMachineFunction();
2132 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2133 return CalleeCC == CallingConv::Fast && CallerCC == CalleeCC;
2137 X86TargetLowering::createFastISel(MachineFunction &mf,
2138 MachineModuleInfo *mmo,
2140 DenseMap<const Value *, unsigned> &vm,
2141 DenseMap<const BasicBlock *,
2142 MachineBasicBlock *> &bm,
2143 DenseMap<const AllocaInst *, int> &am
2145 , SmallSet<Instruction*, 8> &cil
2148 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2156 //===----------------------------------------------------------------------===//
2157 // Other Lowering Hooks
2158 //===----------------------------------------------------------------------===//
2161 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2162 MachineFunction &MF = DAG.getMachineFunction();
2163 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2164 int ReturnAddrIndex = FuncInfo->getRAIndex();
2166 if (ReturnAddrIndex == 0) {
2167 // Set up a frame object for the return address.
2168 uint64_t SlotSize = TD->getPointerSize();
2169 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
2170 FuncInfo->setRAIndex(ReturnAddrIndex);
2173 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2177 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2178 bool hasSymbolicDisplacement) {
2179 // Offset should fit into 32 bit immediate field.
2180 if (!isInt32(Offset))
2183 // If we don't have a symbolic displacement - we don't have any extra
2185 if (!hasSymbolicDisplacement)
2188 // FIXME: Some tweaks might be needed for medium code model.
2189 if (M != CodeModel::Small && M != CodeModel::Kernel)
2192 // For small code model we assume that latest object is 16MB before end of 31
2193 // bits boundary. We may also accept pretty large negative constants knowing
2194 // that all objects are in the positive half of address space.
2195 if (M == CodeModel::Small && Offset < 16*1024*1024)
2198 // For kernel code model we know that all object resist in the negative half
2199 // of 32bits address space. We may not accept negative offsets, since they may
2200 // be just off and we may accept pretty large positive ones.
2201 if (M == CodeModel::Kernel && Offset > 0)
2207 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2208 /// specific condition code, returning the condition code and the LHS/RHS of the
2209 /// comparison to make.
2210 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2211 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2213 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2214 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2215 // X > -1 -> X == 0, jump !sign.
2216 RHS = DAG.getConstant(0, RHS.getValueType());
2217 return X86::COND_NS;
2218 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2219 // X < 0 -> X == 0, jump on sign.
2221 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2223 RHS = DAG.getConstant(0, RHS.getValueType());
2224 return X86::COND_LE;
2228 switch (SetCCOpcode) {
2229 default: llvm_unreachable("Invalid integer condition!");
2230 case ISD::SETEQ: return X86::COND_E;
2231 case ISD::SETGT: return X86::COND_G;
2232 case ISD::SETGE: return X86::COND_GE;
2233 case ISD::SETLT: return X86::COND_L;
2234 case ISD::SETLE: return X86::COND_LE;
2235 case ISD::SETNE: return X86::COND_NE;
2236 case ISD::SETULT: return X86::COND_B;
2237 case ISD::SETUGT: return X86::COND_A;
2238 case ISD::SETULE: return X86::COND_BE;
2239 case ISD::SETUGE: return X86::COND_AE;
2243 // First determine if it is required or is profitable to flip the operands.
2245 // If LHS is a foldable load, but RHS is not, flip the condition.
2246 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2247 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2248 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2249 std::swap(LHS, RHS);
2252 switch (SetCCOpcode) {
2258 std::swap(LHS, RHS);
2262 // On a floating point condition, the flags are set as follows:
2264 // 0 | 0 | 0 | X > Y
2265 // 0 | 0 | 1 | X < Y
2266 // 1 | 0 | 0 | X == Y
2267 // 1 | 1 | 1 | unordered
2268 switch (SetCCOpcode) {
2269 default: llvm_unreachable("Condcode should be pre-legalized away");
2271 case ISD::SETEQ: return X86::COND_E;
2272 case ISD::SETOLT: // flipped
2274 case ISD::SETGT: return X86::COND_A;
2275 case ISD::SETOLE: // flipped
2277 case ISD::SETGE: return X86::COND_AE;
2278 case ISD::SETUGT: // flipped
2280 case ISD::SETLT: return X86::COND_B;
2281 case ISD::SETUGE: // flipped
2283 case ISD::SETLE: return X86::COND_BE;
2285 case ISD::SETNE: return X86::COND_NE;
2286 case ISD::SETUO: return X86::COND_P;
2287 case ISD::SETO: return X86::COND_NP;
2291 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2292 /// code. Current x86 isa includes the following FP cmov instructions:
2293 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2294 static bool hasFPCMov(unsigned X86CC) {
2310 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2311 /// the specified range (L, H].
2312 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2313 return (Val < 0) || (Val >= Low && Val < Hi);
2316 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2317 /// specified value.
2318 static bool isUndefOrEqual(int Val, int CmpVal) {
2319 if (Val < 0 || Val == CmpVal)
2324 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2325 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2326 /// the second operand.
2327 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2328 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2329 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2330 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2331 return (Mask[0] < 2 && Mask[1] < 2);
2335 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2336 SmallVector<int, 8> M;
2338 return ::isPSHUFDMask(M, N->getValueType(0));
2341 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2342 /// is suitable for input to PSHUFHW.
2343 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2344 if (VT != MVT::v8i16)
2347 // Lower quadword copied in order or undef.
2348 for (int i = 0; i != 4; ++i)
2349 if (Mask[i] >= 0 && Mask[i] != i)
2352 // Upper quadword shuffled.
2353 for (int i = 4; i != 8; ++i)
2354 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2360 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2361 SmallVector<int, 8> M;
2363 return ::isPSHUFHWMask(M, N->getValueType(0));
2366 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2367 /// is suitable for input to PSHUFLW.
2368 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2369 if (VT != MVT::v8i16)
2372 // Upper quadword copied in order.
2373 for (int i = 4; i != 8; ++i)
2374 if (Mask[i] >= 0 && Mask[i] != i)
2377 // Lower quadword shuffled.
2378 for (int i = 0; i != 4; ++i)
2385 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2386 SmallVector<int, 8> M;
2388 return ::isPSHUFLWMask(M, N->getValueType(0));
2391 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2392 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2393 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2394 int NumElems = VT.getVectorNumElements();
2395 if (NumElems != 2 && NumElems != 4)
2398 int Half = NumElems / 2;
2399 for (int i = 0; i < Half; ++i)
2400 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2402 for (int i = Half; i < NumElems; ++i)
2403 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2409 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2410 SmallVector<int, 8> M;
2412 return ::isSHUFPMask(M, N->getValueType(0));
2415 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2416 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2417 /// half elements to come from vector 1 (which would equal the dest.) and
2418 /// the upper half to come from vector 2.
2419 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2420 int NumElems = VT.getVectorNumElements();
2422 if (NumElems != 2 && NumElems != 4)
2425 int Half = NumElems / 2;
2426 for (int i = 0; i < Half; ++i)
2427 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2429 for (int i = Half; i < NumElems; ++i)
2430 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2435 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2436 SmallVector<int, 8> M;
2438 return isCommutedSHUFPMask(M, N->getValueType(0));
2441 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2442 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2443 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2444 if (N->getValueType(0).getVectorNumElements() != 4)
2447 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2448 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2449 isUndefOrEqual(N->getMaskElt(1), 7) &&
2450 isUndefOrEqual(N->getMaskElt(2), 2) &&
2451 isUndefOrEqual(N->getMaskElt(3), 3);
2454 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2455 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2456 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2457 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2459 if (NumElems != 2 && NumElems != 4)
2462 for (unsigned i = 0; i < NumElems/2; ++i)
2463 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2466 for (unsigned i = NumElems/2; i < NumElems; ++i)
2467 if (!isUndefOrEqual(N->getMaskElt(i), i))
2473 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2474 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2476 bool X86::isMOVHPMask(ShuffleVectorSDNode *N) {
2477 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2479 if (NumElems != 2 && NumElems != 4)
2482 for (unsigned i = 0; i < NumElems/2; ++i)
2483 if (!isUndefOrEqual(N->getMaskElt(i), i))
2486 for (unsigned i = 0; i < NumElems/2; ++i)
2487 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2493 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2494 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2496 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2497 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2502 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2503 isUndefOrEqual(N->getMaskElt(1), 3) &&
2504 isUndefOrEqual(N->getMaskElt(2), 2) &&
2505 isUndefOrEqual(N->getMaskElt(3), 3);
2508 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2509 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2510 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2511 bool V2IsSplat = false) {
2512 int NumElts = VT.getVectorNumElements();
2513 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2516 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2518 int BitI1 = Mask[i+1];
2519 if (!isUndefOrEqual(BitI, j))
2522 if (!isUndefOrEqual(BitI1, NumElts))
2525 if (!isUndefOrEqual(BitI1, j + NumElts))
2532 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2533 SmallVector<int, 8> M;
2535 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2538 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2539 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2540 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2541 bool V2IsSplat = false) {
2542 int NumElts = VT.getVectorNumElements();
2543 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2546 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2548 int BitI1 = Mask[i+1];
2549 if (!isUndefOrEqual(BitI, j + NumElts/2))
2552 if (isUndefOrEqual(BitI1, NumElts))
2555 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2562 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2563 SmallVector<int, 8> M;
2565 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2568 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2569 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2571 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2572 int NumElems = VT.getVectorNumElements();
2573 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2576 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2578 int BitI1 = Mask[i+1];
2579 if (!isUndefOrEqual(BitI, j))
2581 if (!isUndefOrEqual(BitI1, j))
2587 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2588 SmallVector<int, 8> M;
2590 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2593 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2594 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2596 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2597 int NumElems = VT.getVectorNumElements();
2598 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2601 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2603 int BitI1 = Mask[i+1];
2604 if (!isUndefOrEqual(BitI, j))
2606 if (!isUndefOrEqual(BitI1, j))
2612 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2613 SmallVector<int, 8> M;
2615 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2618 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2619 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2620 /// MOVSD, and MOVD, i.e. setting the lowest element.
2621 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2622 if (VT.getVectorElementType().getSizeInBits() < 32)
2625 int NumElts = VT.getVectorNumElements();
2627 if (!isUndefOrEqual(Mask[0], NumElts))
2630 for (int i = 1; i < NumElts; ++i)
2631 if (!isUndefOrEqual(Mask[i], i))
2637 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2638 SmallVector<int, 8> M;
2640 return ::isMOVLMask(M, N->getValueType(0));
2643 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2644 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2645 /// element of vector 2 and the other elements to come from vector 1 in order.
2646 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2647 bool V2IsSplat = false, bool V2IsUndef = false) {
2648 int NumOps = VT.getVectorNumElements();
2649 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2652 if (!isUndefOrEqual(Mask[0], 0))
2655 for (int i = 1; i < NumOps; ++i)
2656 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2657 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2658 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2664 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2665 bool V2IsUndef = false) {
2666 SmallVector<int, 8> M;
2668 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2671 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2672 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2673 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2674 if (N->getValueType(0).getVectorNumElements() != 4)
2677 // Expect 1, 1, 3, 3
2678 for (unsigned i = 0; i < 2; ++i) {
2679 int Elt = N->getMaskElt(i);
2680 if (Elt >= 0 && Elt != 1)
2685 for (unsigned i = 2; i < 4; ++i) {
2686 int Elt = N->getMaskElt(i);
2687 if (Elt >= 0 && Elt != 3)
2692 // Don't use movshdup if it can be done with a shufps.
2693 // FIXME: verify that matching u, u, 3, 3 is what we want.
2697 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2698 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2699 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2700 if (N->getValueType(0).getVectorNumElements() != 4)
2703 // Expect 0, 0, 2, 2
2704 for (unsigned i = 0; i < 2; ++i)
2705 if (N->getMaskElt(i) > 0)
2709 for (unsigned i = 2; i < 4; ++i) {
2710 int Elt = N->getMaskElt(i);
2711 if (Elt >= 0 && Elt != 2)
2716 // Don't use movsldup if it can be done with a shufps.
2720 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2721 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2722 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2723 int e = N->getValueType(0).getVectorNumElements() / 2;
2725 for (int i = 0; i < e; ++i)
2726 if (!isUndefOrEqual(N->getMaskElt(i), i))
2728 for (int i = 0; i < e; ++i)
2729 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2734 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2735 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2737 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2738 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2739 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2741 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2743 for (int i = 0; i < NumOperands; ++i) {
2744 int Val = SVOp->getMaskElt(NumOperands-i-1);
2745 if (Val < 0) Val = 0;
2746 if (Val >= NumOperands) Val -= NumOperands;
2748 if (i != NumOperands - 1)
2754 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2755 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2757 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2758 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2760 // 8 nodes, but we only care about the last 4.
2761 for (unsigned i = 7; i >= 4; --i) {
2762 int Val = SVOp->getMaskElt(i);
2771 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2772 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2774 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2775 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2777 // 8 nodes, but we only care about the first 4.
2778 for (int i = 3; i >= 0; --i) {
2779 int Val = SVOp->getMaskElt(i);
2788 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2790 bool X86::isZeroNode(SDValue Elt) {
2791 return ((isa<ConstantSDNode>(Elt) &&
2792 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2793 (isa<ConstantFPSDNode>(Elt) &&
2794 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2797 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
2798 /// their permute mask.
2799 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
2800 SelectionDAG &DAG) {
2801 EVT VT = SVOp->getValueType(0);
2802 unsigned NumElems = VT.getVectorNumElements();
2803 SmallVector<int, 8> MaskVec;
2805 for (unsigned i = 0; i != NumElems; ++i) {
2806 int idx = SVOp->getMaskElt(i);
2808 MaskVec.push_back(idx);
2809 else if (idx < (int)NumElems)
2810 MaskVec.push_back(idx + NumElems);
2812 MaskVec.push_back(idx - NumElems);
2814 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
2815 SVOp->getOperand(0), &MaskVec[0]);
2818 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2819 /// the two vector operands have swapped position.
2820 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
2821 unsigned NumElems = VT.getVectorNumElements();
2822 for (unsigned i = 0; i != NumElems; ++i) {
2826 else if (idx < (int)NumElems)
2827 Mask[i] = idx + NumElems;
2829 Mask[i] = idx - NumElems;
2833 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2834 /// match movhlps. The lower half elements should come from upper half of
2835 /// V1 (and in order), and the upper half elements should come from the upper
2836 /// half of V2 (and in order).
2837 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
2838 if (Op->getValueType(0).getVectorNumElements() != 4)
2840 for (unsigned i = 0, e = 2; i != e; ++i)
2841 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
2843 for (unsigned i = 2; i != 4; ++i)
2844 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
2849 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2850 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2852 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2853 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2855 N = N->getOperand(0).getNode();
2856 if (!ISD::isNON_EXTLoad(N))
2859 *LD = cast<LoadSDNode>(N);
2863 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2864 /// match movlp{s|d}. The lower half elements should come from lower half of
2865 /// V1 (and in order), and the upper half elements should come from the upper
2866 /// half of V2 (and in order). And since V1 will become the source of the
2867 /// MOVLP, it must be either a vector load or a scalar load to vector.
2868 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
2869 ShuffleVectorSDNode *Op) {
2870 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2872 // Is V2 is a vector load, don't do this transformation. We will try to use
2873 // load folding shufps op.
2874 if (ISD::isNON_EXTLoad(V2))
2877 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
2879 if (NumElems != 2 && NumElems != 4)
2881 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2882 if (!isUndefOrEqual(Op->getMaskElt(i), i))
2884 for (unsigned i = NumElems/2; i != NumElems; ++i)
2885 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
2890 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2892 static bool isSplatVector(SDNode *N) {
2893 if (N->getOpcode() != ISD::BUILD_VECTOR)
2896 SDValue SplatValue = N->getOperand(0);
2897 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2898 if (N->getOperand(i) != SplatValue)
2903 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2904 /// to an zero vector.
2905 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
2906 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
2907 SDValue V1 = N->getOperand(0);
2908 SDValue V2 = N->getOperand(1);
2909 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2910 for (unsigned i = 0; i != NumElems; ++i) {
2911 int Idx = N->getMaskElt(i);
2912 if (Idx >= (int)NumElems) {
2913 unsigned Opc = V2.getOpcode();
2914 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2916 if (Opc != ISD::BUILD_VECTOR ||
2917 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
2919 } else if (Idx >= 0) {
2920 unsigned Opc = V1.getOpcode();
2921 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2923 if (Opc != ISD::BUILD_VECTOR ||
2924 !X86::isZeroNode(V1.getOperand(Idx)))
2931 /// getZeroVector - Returns a vector of specified type with all zero elements.
2933 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
2935 assert(VT.isVector() && "Expected a vector type");
2937 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2938 // type. This ensures they get CSE'd.
2940 if (VT.getSizeInBits() == 64) { // MMX
2941 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2942 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2943 } else if (HasSSE2) { // SSE2
2944 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2945 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2947 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2948 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
2950 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2953 /// getOnesVector - Returns a vector of specified type with all bits set.
2955 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2956 assert(VT.isVector() && "Expected a vector type");
2958 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2959 // type. This ensures they get CSE'd.
2960 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2962 if (VT.getSizeInBits() == 64) // MMX
2963 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2965 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2966 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2970 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2971 /// that point to V2 points to its first element.
2972 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
2973 EVT VT = SVOp->getValueType(0);
2974 unsigned NumElems = VT.getVectorNumElements();
2976 bool Changed = false;
2977 SmallVector<int, 8> MaskVec;
2978 SVOp->getMask(MaskVec);
2980 for (unsigned i = 0; i != NumElems; ++i) {
2981 if (MaskVec[i] > (int)NumElems) {
2982 MaskVec[i] = NumElems;
2987 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
2988 SVOp->getOperand(1), &MaskVec[0]);
2989 return SDValue(SVOp, 0);
2992 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2993 /// operation of specified width.
2994 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
2996 unsigned NumElems = VT.getVectorNumElements();
2997 SmallVector<int, 8> Mask;
2998 Mask.push_back(NumElems);
2999 for (unsigned i = 1; i != NumElems; ++i)
3001 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3004 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3005 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3007 unsigned NumElems = VT.getVectorNumElements();
3008 SmallVector<int, 8> Mask;
3009 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3011 Mask.push_back(i + NumElems);
3013 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3016 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3017 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3019 unsigned NumElems = VT.getVectorNumElements();
3020 unsigned Half = NumElems/2;
3021 SmallVector<int, 8> Mask;
3022 for (unsigned i = 0; i != Half; ++i) {
3023 Mask.push_back(i + Half);
3024 Mask.push_back(i + NumElems + Half);
3026 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3029 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3030 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3032 if (SV->getValueType(0).getVectorNumElements() <= 4)
3033 return SDValue(SV, 0);
3035 EVT PVT = MVT::v4f32;
3036 EVT VT = SV->getValueType(0);
3037 DebugLoc dl = SV->getDebugLoc();
3038 SDValue V1 = SV->getOperand(0);
3039 int NumElems = VT.getVectorNumElements();
3040 int EltNo = SV->getSplatIndex();
3042 // unpack elements to the correct location
3043 while (NumElems > 4) {
3044 if (EltNo < NumElems/2) {
3045 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3047 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3048 EltNo -= NumElems/2;
3053 // Perform the splat.
3054 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3055 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3056 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3057 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3060 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3061 /// vector of zero or undef vector. This produces a shuffle where the low
3062 /// element of V2 is swizzled into the zero/undef vector, landing at element
3063 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3064 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3065 bool isZero, bool HasSSE2,
3066 SelectionDAG &DAG) {
3067 EVT VT = V2.getValueType();
3069 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3070 unsigned NumElems = VT.getVectorNumElements();
3071 SmallVector<int, 16> MaskVec;
3072 for (unsigned i = 0; i != NumElems; ++i)
3073 // If this is the insertion idx, put the low elt of V2 here.
3074 MaskVec.push_back(i == Idx ? NumElems : i);
3075 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3078 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3079 /// a shuffle that is zero.
3081 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3082 bool Low, SelectionDAG &DAG) {
3083 unsigned NumZeros = 0;
3084 for (int i = 0; i < NumElems; ++i) {
3085 unsigned Index = Low ? i : NumElems-i-1;
3086 int Idx = SVOp->getMaskElt(Index);
3091 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3092 if (Elt.getNode() && X86::isZeroNode(Elt))
3100 /// isVectorShift - Returns true if the shuffle can be implemented as a
3101 /// logical left or right shift of a vector.
3102 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3103 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3104 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3105 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3108 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3111 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3115 bool SeenV1 = false;
3116 bool SeenV2 = false;
3117 for (int i = NumZeros; i < NumElems; ++i) {
3118 int Val = isLeft ? (i - NumZeros) : i;
3119 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3131 if (SeenV1 && SeenV2)
3134 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3140 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3142 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3143 unsigned NumNonZero, unsigned NumZero,
3144 SelectionDAG &DAG, TargetLowering &TLI) {
3148 DebugLoc dl = Op.getDebugLoc();
3151 for (unsigned i = 0; i < 16; ++i) {
3152 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3153 if (ThisIsNonZero && First) {
3155 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3157 V = DAG.getUNDEF(MVT::v8i16);
3162 SDValue ThisElt(0, 0), LastElt(0, 0);
3163 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3164 if (LastIsNonZero) {
3165 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3166 MVT::i16, Op.getOperand(i-1));
3168 if (ThisIsNonZero) {
3169 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3170 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3171 ThisElt, DAG.getConstant(8, MVT::i8));
3173 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3177 if (ThisElt.getNode())
3178 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3179 DAG.getIntPtrConstant(i/2));
3183 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3186 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3188 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3189 unsigned NumNonZero, unsigned NumZero,
3190 SelectionDAG &DAG, TargetLowering &TLI) {
3194 DebugLoc dl = Op.getDebugLoc();
3197 for (unsigned i = 0; i < 8; ++i) {
3198 bool isNonZero = (NonZeros & (1 << i)) != 0;
3202 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3204 V = DAG.getUNDEF(MVT::v8i16);
3207 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3208 MVT::v8i16, V, Op.getOperand(i),
3209 DAG.getIntPtrConstant(i));
3216 /// getVShift - Return a vector logical shift node.
3218 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3219 unsigned NumBits, SelectionDAG &DAG,
3220 const TargetLowering &TLI, DebugLoc dl) {
3221 bool isMMX = VT.getSizeInBits() == 64;
3222 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3223 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3224 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3225 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3226 DAG.getNode(Opc, dl, ShVT, SrcOp,
3227 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3231 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3232 DebugLoc dl = Op.getDebugLoc();
3233 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3234 if (ISD::isBuildVectorAllZeros(Op.getNode())
3235 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3236 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3237 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3238 // eliminated on x86-32 hosts.
3239 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3242 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3243 return getOnesVector(Op.getValueType(), DAG, dl);
3244 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3247 EVT VT = Op.getValueType();
3248 EVT ExtVT = VT.getVectorElementType();
3249 unsigned EVTBits = ExtVT.getSizeInBits();
3251 unsigned NumElems = Op.getNumOperands();
3252 unsigned NumZero = 0;
3253 unsigned NumNonZero = 0;
3254 unsigned NonZeros = 0;
3255 bool IsAllConstants = true;
3256 SmallSet<SDValue, 8> Values;
3257 for (unsigned i = 0; i < NumElems; ++i) {
3258 SDValue Elt = Op.getOperand(i);
3259 if (Elt.getOpcode() == ISD::UNDEF)
3262 if (Elt.getOpcode() != ISD::Constant &&
3263 Elt.getOpcode() != ISD::ConstantFP)
3264 IsAllConstants = false;
3265 if (X86::isZeroNode(Elt))
3268 NonZeros |= (1 << i);
3273 if (NumNonZero == 0) {
3274 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3275 return DAG.getUNDEF(VT);
3278 // Special case for single non-zero, non-undef, element.
3279 if (NumNonZero == 1) {
3280 unsigned Idx = CountTrailingZeros_32(NonZeros);
3281 SDValue Item = Op.getOperand(Idx);
3283 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3284 // the value are obviously zero, truncate the value to i32 and do the
3285 // insertion that way. Only do this if the value is non-constant or if the
3286 // value is a constant being inserted into element 0. It is cheaper to do
3287 // a constant pool load than it is to do a movd + shuffle.
3288 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3289 (!IsAllConstants || Idx == 0)) {
3290 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3291 // Handle MMX and SSE both.
3292 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3293 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3295 // Truncate the value (which may itself be a constant) to i32, and
3296 // convert it to a vector with movd (S2V+shuffle to zero extend).
3297 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3298 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3299 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3300 Subtarget->hasSSE2(), DAG);
3302 // Now we have our 32-bit value zero extended in the low element of
3303 // a vector. If Idx != 0, swizzle it into place.
3305 SmallVector<int, 4> Mask;
3306 Mask.push_back(Idx);
3307 for (unsigned i = 1; i != VecElts; ++i)
3309 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3310 DAG.getUNDEF(Item.getValueType()),
3313 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3317 // If we have a constant or non-constant insertion into the low element of
3318 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3319 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3320 // depending on what the source datatype is.
3323 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3324 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3325 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3326 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3327 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3328 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3330 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3331 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3332 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3333 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3334 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3335 Subtarget->hasSSE2(), DAG);
3336 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3340 // Is it a vector logical left shift?
3341 if (NumElems == 2 && Idx == 1 &&
3342 X86::isZeroNode(Op.getOperand(0)) &&
3343 !X86::isZeroNode(Op.getOperand(1))) {
3344 unsigned NumBits = VT.getSizeInBits();
3345 return getVShift(true, VT,
3346 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3347 VT, Op.getOperand(1)),
3348 NumBits/2, DAG, *this, dl);
3351 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3354 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3355 // is a non-constant being inserted into an element other than the low one,
3356 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3357 // movd/movss) to move this into the low element, then shuffle it into
3359 if (EVTBits == 32) {
3360 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3362 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3363 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3364 Subtarget->hasSSE2(), DAG);
3365 SmallVector<int, 8> MaskVec;
3366 for (unsigned i = 0; i < NumElems; i++)
3367 MaskVec.push_back(i == Idx ? 0 : 1);
3368 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3372 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3373 if (Values.size() == 1)
3376 // A vector full of immediates; various special cases are already
3377 // handled, so this is best done with a single constant-pool load.
3381 // Let legalizer expand 2-wide build_vectors.
3382 if (EVTBits == 64) {
3383 if (NumNonZero == 1) {
3384 // One half is zero or undef.
3385 unsigned Idx = CountTrailingZeros_32(NonZeros);
3386 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3387 Op.getOperand(Idx));
3388 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3389 Subtarget->hasSSE2(), DAG);
3394 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3395 if (EVTBits == 8 && NumElems == 16) {
3396 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3398 if (V.getNode()) return V;
3401 if (EVTBits == 16 && NumElems == 8) {
3402 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3404 if (V.getNode()) return V;
3407 // If element VT is == 32 bits, turn it into a number of shuffles.
3408 SmallVector<SDValue, 8> V;
3410 if (NumElems == 4 && NumZero > 0) {
3411 for (unsigned i = 0; i < 4; ++i) {
3412 bool isZero = !(NonZeros & (1 << i));
3414 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3416 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3419 for (unsigned i = 0; i < 2; ++i) {
3420 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3423 V[i] = V[i*2]; // Must be a zero vector.
3426 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3429 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3432 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3437 SmallVector<int, 8> MaskVec;
3438 bool Reverse = (NonZeros & 0x3) == 2;
3439 for (unsigned i = 0; i < 2; ++i)
3440 MaskVec.push_back(Reverse ? 1-i : i);
3441 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3442 for (unsigned i = 0; i < 2; ++i)
3443 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3444 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3447 if (Values.size() > 2) {
3448 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3449 // values to be inserted is equal to the number of elements, in which case
3450 // use the unpack code below in the hopes of matching the consecutive elts
3451 // load merge pattern for shuffles.
3452 // FIXME: We could probably just check that here directly.
3453 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3454 getSubtarget()->hasSSE41()) {
3455 V[0] = DAG.getUNDEF(VT);
3456 for (unsigned i = 0; i < NumElems; ++i)
3457 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3458 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3459 Op.getOperand(i), DAG.getIntPtrConstant(i));
3462 // Expand into a number of unpckl*.
3464 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3465 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3466 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3467 for (unsigned i = 0; i < NumElems; ++i)
3468 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3470 while (NumElems != 0) {
3471 for (unsigned i = 0; i < NumElems; ++i)
3472 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3481 // v8i16 shuffles - Prefer shuffles in the following order:
3482 // 1. [all] pshuflw, pshufhw, optional move
3483 // 2. [ssse3] 1 x pshufb
3484 // 3. [ssse3] 2 x pshufb + 1 x por
3485 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3487 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3488 SelectionDAG &DAG, X86TargetLowering &TLI) {
3489 SDValue V1 = SVOp->getOperand(0);
3490 SDValue V2 = SVOp->getOperand(1);
3491 DebugLoc dl = SVOp->getDebugLoc();
3492 SmallVector<int, 8> MaskVals;
3494 // Determine if more than 1 of the words in each of the low and high quadwords
3495 // of the result come from the same quadword of one of the two inputs. Undef
3496 // mask values count as coming from any quadword, for better codegen.
3497 SmallVector<unsigned, 4> LoQuad(4);
3498 SmallVector<unsigned, 4> HiQuad(4);
3499 BitVector InputQuads(4);
3500 for (unsigned i = 0; i < 8; ++i) {
3501 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3502 int EltIdx = SVOp->getMaskElt(i);
3503 MaskVals.push_back(EltIdx);
3512 InputQuads.set(EltIdx / 4);
3515 int BestLoQuad = -1;
3516 unsigned MaxQuad = 1;
3517 for (unsigned i = 0; i < 4; ++i) {
3518 if (LoQuad[i] > MaxQuad) {
3520 MaxQuad = LoQuad[i];
3524 int BestHiQuad = -1;
3526 for (unsigned i = 0; i < 4; ++i) {
3527 if (HiQuad[i] > MaxQuad) {
3529 MaxQuad = HiQuad[i];
3533 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3534 // of the two input vectors, shuffle them into one input vector so only a
3535 // single pshufb instruction is necessary. If There are more than 2 input
3536 // quads, disable the next transformation since it does not help SSSE3.
3537 bool V1Used = InputQuads[0] || InputQuads[1];
3538 bool V2Used = InputQuads[2] || InputQuads[3];
3539 if (TLI.getSubtarget()->hasSSSE3()) {
3540 if (InputQuads.count() == 2 && V1Used && V2Used) {
3541 BestLoQuad = InputQuads.find_first();
3542 BestHiQuad = InputQuads.find_next(BestLoQuad);
3544 if (InputQuads.count() > 2) {
3550 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3551 // the shuffle mask. If a quad is scored as -1, that means that it contains
3552 // words from all 4 input quadwords.
3554 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3555 SmallVector<int, 8> MaskV;
3556 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3557 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3558 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3559 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3560 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3561 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3563 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3564 // source words for the shuffle, to aid later transformations.
3565 bool AllWordsInNewV = true;
3566 bool InOrder[2] = { true, true };
3567 for (unsigned i = 0; i != 8; ++i) {
3568 int idx = MaskVals[i];
3570 InOrder[i/4] = false;
3571 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3573 AllWordsInNewV = false;
3577 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3578 if (AllWordsInNewV) {
3579 for (int i = 0; i != 8; ++i) {
3580 int idx = MaskVals[i];
3583 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3584 if ((idx != i) && idx < 4)
3586 if ((idx != i) && idx > 3)
3595 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3596 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3597 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3598 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3599 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3603 // If we have SSSE3, and all words of the result are from 1 input vector,
3604 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3605 // is present, fall back to case 4.
3606 if (TLI.getSubtarget()->hasSSSE3()) {
3607 SmallVector<SDValue,16> pshufbMask;
3609 // If we have elements from both input vectors, set the high bit of the
3610 // shuffle mask element to zero out elements that come from V2 in the V1
3611 // mask, and elements that come from V1 in the V2 mask, so that the two
3612 // results can be OR'd together.
3613 bool TwoInputs = V1Used && V2Used;
3614 for (unsigned i = 0; i != 8; ++i) {
3615 int EltIdx = MaskVals[i] * 2;
3616 if (TwoInputs && (EltIdx >= 16)) {
3617 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3618 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3621 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3622 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3624 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3625 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3626 DAG.getNode(ISD::BUILD_VECTOR, dl,
3627 MVT::v16i8, &pshufbMask[0], 16));
3629 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3631 // Calculate the shuffle mask for the second input, shuffle it, and
3632 // OR it with the first shuffled input.
3634 for (unsigned i = 0; i != 8; ++i) {
3635 int EltIdx = MaskVals[i] * 2;
3637 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3638 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3641 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3642 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3644 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3645 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3646 DAG.getNode(ISD::BUILD_VECTOR, dl,
3647 MVT::v16i8, &pshufbMask[0], 16));
3648 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3649 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3652 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3653 // and update MaskVals with new element order.
3654 BitVector InOrder(8);
3655 if (BestLoQuad >= 0) {
3656 SmallVector<int, 8> MaskV;
3657 for (int i = 0; i != 4; ++i) {
3658 int idx = MaskVals[i];
3660 MaskV.push_back(-1);
3662 } else if ((idx / 4) == BestLoQuad) {
3663 MaskV.push_back(idx & 3);
3666 MaskV.push_back(-1);
3669 for (unsigned i = 4; i != 8; ++i)
3671 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3675 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
3676 // and update MaskVals with the new element order.
3677 if (BestHiQuad >= 0) {
3678 SmallVector<int, 8> MaskV;
3679 for (unsigned i = 0; i != 4; ++i)
3681 for (unsigned i = 4; i != 8; ++i) {
3682 int idx = MaskVals[i];
3684 MaskV.push_back(-1);
3686 } else if ((idx / 4) == BestHiQuad) {
3687 MaskV.push_back((idx & 3) + 4);
3690 MaskV.push_back(-1);
3693 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3697 // In case BestHi & BestLo were both -1, which means each quadword has a word
3698 // from each of the four input quadwords, calculate the InOrder bitvector now
3699 // before falling through to the insert/extract cleanup.
3700 if (BestLoQuad == -1 && BestHiQuad == -1) {
3702 for (int i = 0; i != 8; ++i)
3703 if (MaskVals[i] < 0 || MaskVals[i] == i)
3707 // The other elements are put in the right place using pextrw and pinsrw.
3708 for (unsigned i = 0; i != 8; ++i) {
3711 int EltIdx = MaskVals[i];
3714 SDValue ExtOp = (EltIdx < 8)
3715 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
3716 DAG.getIntPtrConstant(EltIdx))
3717 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
3718 DAG.getIntPtrConstant(EltIdx - 8));
3719 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3720 DAG.getIntPtrConstant(i));
3725 // v16i8 shuffles - Prefer shuffles in the following order:
3726 // 1. [ssse3] 1 x pshufb
3727 // 2. [ssse3] 2 x pshufb + 1 x por
3728 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
3730 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
3731 SelectionDAG &DAG, X86TargetLowering &TLI) {
3732 SDValue V1 = SVOp->getOperand(0);
3733 SDValue V2 = SVOp->getOperand(1);
3734 DebugLoc dl = SVOp->getDebugLoc();
3735 SmallVector<int, 16> MaskVals;
3736 SVOp->getMask(MaskVals);
3738 // If we have SSSE3, case 1 is generated when all result bytes come from
3739 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
3740 // present, fall back to case 3.
3741 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
3744 for (unsigned i = 0; i < 16; ++i) {
3745 int EltIdx = MaskVals[i];
3754 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
3755 if (TLI.getSubtarget()->hasSSSE3()) {
3756 SmallVector<SDValue,16> pshufbMask;
3758 // If all result elements are from one input vector, then only translate
3759 // undef mask values to 0x80 (zero out result) in the pshufb mask.
3761 // Otherwise, we have elements from both input vectors, and must zero out
3762 // elements that come from V2 in the first mask, and V1 in the second mask
3763 // so that we can OR them together.
3764 bool TwoInputs = !(V1Only || V2Only);
3765 for (unsigned i = 0; i != 16; ++i) {
3766 int EltIdx = MaskVals[i];
3767 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
3768 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3771 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3773 // If all the elements are from V2, assign it to V1 and return after
3774 // building the first pshufb.
3777 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3778 DAG.getNode(ISD::BUILD_VECTOR, dl,
3779 MVT::v16i8, &pshufbMask[0], 16));
3783 // Calculate the shuffle mask for the second input, shuffle it, and
3784 // OR it with the first shuffled input.
3786 for (unsigned i = 0; i != 16; ++i) {
3787 int EltIdx = MaskVals[i];
3789 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3792 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3794 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3795 DAG.getNode(ISD::BUILD_VECTOR, dl,
3796 MVT::v16i8, &pshufbMask[0], 16));
3797 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3800 // No SSSE3 - Calculate in place words and then fix all out of place words
3801 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
3802 // the 16 different words that comprise the two doublequadword input vectors.
3803 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3804 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
3805 SDValue NewV = V2Only ? V2 : V1;
3806 for (int i = 0; i != 8; ++i) {
3807 int Elt0 = MaskVals[i*2];
3808 int Elt1 = MaskVals[i*2+1];
3810 // This word of the result is all undef, skip it.
3811 if (Elt0 < 0 && Elt1 < 0)
3814 // This word of the result is already in the correct place, skip it.
3815 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
3817 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
3820 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
3821 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
3824 // If Elt0 and Elt1 are defined, are consecutive, and can be load
3825 // using a single extract together, load it and store it.
3826 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
3827 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3828 DAG.getIntPtrConstant(Elt1 / 2));
3829 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3830 DAG.getIntPtrConstant(i));
3834 // If Elt1 is defined, extract it from the appropriate source. If the
3835 // source byte is not also odd, shift the extracted word left 8 bits
3836 // otherwise clear the bottom 8 bits if we need to do an or.
3838 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3839 DAG.getIntPtrConstant(Elt1 / 2));
3840 if ((Elt1 & 1) == 0)
3841 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
3842 DAG.getConstant(8, TLI.getShiftAmountTy()));
3844 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
3845 DAG.getConstant(0xFF00, MVT::i16));
3847 // If Elt0 is defined, extract it from the appropriate source. If the
3848 // source byte is not also even, shift the extracted word right 8 bits. If
3849 // Elt1 was also defined, OR the extracted values together before
3850 // inserting them in the result.
3852 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
3853 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
3854 if ((Elt0 & 1) != 0)
3855 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
3856 DAG.getConstant(8, TLI.getShiftAmountTy()));
3858 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
3859 DAG.getConstant(0x00FF, MVT::i16));
3860 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
3863 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3864 DAG.getIntPtrConstant(i));
3866 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
3869 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3870 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3871 /// done when every pair / quad of shuffle mask elements point to elements in
3872 /// the right sequence. e.g.
3873 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3875 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
3877 TargetLowering &TLI, DebugLoc dl) {
3878 EVT VT = SVOp->getValueType(0);
3879 SDValue V1 = SVOp->getOperand(0);
3880 SDValue V2 = SVOp->getOperand(1);
3881 unsigned NumElems = VT.getVectorNumElements();
3882 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3883 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3884 EVT MaskEltVT = MaskVT.getVectorElementType();
3886 switch (VT.getSimpleVT().SimpleTy) {
3887 default: assert(false && "Unexpected!");
3888 case MVT::v4f32: NewVT = MVT::v2f64; break;
3889 case MVT::v4i32: NewVT = MVT::v2i64; break;
3890 case MVT::v8i16: NewVT = MVT::v4i32; break;
3891 case MVT::v16i8: NewVT = MVT::v4i32; break;
3894 if (NewWidth == 2) {
3900 int Scale = NumElems / NewWidth;
3901 SmallVector<int, 8> MaskVec;
3902 for (unsigned i = 0; i < NumElems; i += Scale) {
3904 for (int j = 0; j < Scale; ++j) {
3905 int EltIdx = SVOp->getMaskElt(i+j);
3909 StartIdx = EltIdx - (EltIdx % Scale);
3910 if (EltIdx != StartIdx + j)
3914 MaskVec.push_back(-1);
3916 MaskVec.push_back(StartIdx / Scale);
3919 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
3920 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
3921 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
3924 /// getVZextMovL - Return a zero-extending vector move low node.
3926 static SDValue getVZextMovL(EVT VT, EVT OpVT,
3927 SDValue SrcOp, SelectionDAG &DAG,
3928 const X86Subtarget *Subtarget, DebugLoc dl) {
3929 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
3930 LoadSDNode *LD = NULL;
3931 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
3932 LD = dyn_cast<LoadSDNode>(SrcOp);
3934 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
3936 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
3937 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
3938 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
3939 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
3940 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
3942 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
3943 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3944 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3945 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3953 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3954 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3955 DAG.getNode(ISD::BIT_CONVERT, dl,
3959 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
3962 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3963 SDValue V1 = SVOp->getOperand(0);
3964 SDValue V2 = SVOp->getOperand(1);
3965 DebugLoc dl = SVOp->getDebugLoc();
3966 EVT VT = SVOp->getValueType(0);
3968 SmallVector<std::pair<int, int>, 8> Locs;
3970 SmallVector<int, 8> Mask1(4U, -1);
3971 SmallVector<int, 8> PermMask;
3972 SVOp->getMask(PermMask);
3976 for (unsigned i = 0; i != 4; ++i) {
3977 int Idx = PermMask[i];
3979 Locs[i] = std::make_pair(-1, -1);
3981 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
3983 Locs[i] = std::make_pair(0, NumLo);
3987 Locs[i] = std::make_pair(1, NumHi);
3989 Mask1[2+NumHi] = Idx;
3995 if (NumLo <= 2 && NumHi <= 2) {
3996 // If no more than two elements come from either vector. This can be
3997 // implemented with two shuffles. First shuffle gather the elements.
3998 // The second shuffle, which takes the first shuffle as both of its
3999 // vector operands, put the elements into the right order.
4000 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4002 SmallVector<int, 8> Mask2(4U, -1);
4004 for (unsigned i = 0; i != 4; ++i) {
4005 if (Locs[i].first == -1)
4008 unsigned Idx = (i < 2) ? 0 : 4;
4009 Idx += Locs[i].first * 2 + Locs[i].second;
4014 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4015 } else if (NumLo == 3 || NumHi == 3) {
4016 // Otherwise, we must have three elements from one vector, call it X, and
4017 // one element from the other, call it Y. First, use a shufps to build an
4018 // intermediate vector with the one element from Y and the element from X
4019 // that will be in the same half in the final destination (the indexes don't
4020 // matter). Then, use a shufps to build the final vector, taking the half
4021 // containing the element from Y from the intermediate, and the other half
4024 // Normalize it so the 3 elements come from V1.
4025 CommuteVectorShuffleMask(PermMask, VT);
4029 // Find the element from V2.
4031 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4032 int Val = PermMask[HiIndex];
4039 Mask1[0] = PermMask[HiIndex];
4041 Mask1[2] = PermMask[HiIndex^1];
4043 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4046 Mask1[0] = PermMask[0];
4047 Mask1[1] = PermMask[1];
4048 Mask1[2] = HiIndex & 1 ? 6 : 4;
4049 Mask1[3] = HiIndex & 1 ? 4 : 6;
4050 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4052 Mask1[0] = HiIndex & 1 ? 2 : 0;
4053 Mask1[1] = HiIndex & 1 ? 0 : 2;
4054 Mask1[2] = PermMask[2];
4055 Mask1[3] = PermMask[3];
4060 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4064 // Break it into (shuffle shuffle_hi, shuffle_lo).
4066 SmallVector<int,8> LoMask(4U, -1);
4067 SmallVector<int,8> HiMask(4U, -1);
4069 SmallVector<int,8> *MaskPtr = &LoMask;
4070 unsigned MaskIdx = 0;
4073 for (unsigned i = 0; i != 4; ++i) {
4080 int Idx = PermMask[i];
4082 Locs[i] = std::make_pair(-1, -1);
4083 } else if (Idx < 4) {
4084 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4085 (*MaskPtr)[LoIdx] = Idx;
4088 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4089 (*MaskPtr)[HiIdx] = Idx;
4094 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4095 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4096 SmallVector<int, 8> MaskOps;
4097 for (unsigned i = 0; i != 4; ++i) {
4098 if (Locs[i].first == -1) {
4099 MaskOps.push_back(-1);
4101 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4102 MaskOps.push_back(Idx);
4105 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4109 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4110 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4111 SDValue V1 = Op.getOperand(0);
4112 SDValue V2 = Op.getOperand(1);
4113 EVT VT = Op.getValueType();
4114 DebugLoc dl = Op.getDebugLoc();
4115 unsigned NumElems = VT.getVectorNumElements();
4116 bool isMMX = VT.getSizeInBits() == 64;
4117 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4118 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4119 bool V1IsSplat = false;
4120 bool V2IsSplat = false;
4122 if (isZeroShuffle(SVOp))
4123 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4125 // Promote splats to v4f32.
4126 if (SVOp->isSplat()) {
4127 if (isMMX || NumElems < 4)
4129 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4132 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4134 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4135 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4136 if (NewOp.getNode())
4137 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4138 LowerVECTOR_SHUFFLE(NewOp, DAG));
4139 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4140 // FIXME: Figure out a cleaner way to do this.
4141 // Try to make use of movq to zero out the top part.
4142 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4143 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4144 if (NewOp.getNode()) {
4145 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4146 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4147 DAG, Subtarget, dl);
4149 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4150 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4151 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4152 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4153 DAG, Subtarget, dl);
4157 if (X86::isPSHUFDMask(SVOp))
4160 // Check if this can be converted into a logical shift.
4161 bool isLeft = false;
4164 bool isShift = getSubtarget()->hasSSE2() &&
4165 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4166 if (isShift && ShVal.hasOneUse()) {
4167 // If the shifted value has multiple uses, it may be cheaper to use
4168 // v_set0 + movlhps or movhlps, etc.
4169 EVT EltVT = VT.getVectorElementType();
4170 ShAmt *= EltVT.getSizeInBits();
4171 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4174 if (X86::isMOVLMask(SVOp)) {
4177 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4178 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4183 // FIXME: fold these into legal mask.
4184 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4185 X86::isMOVSLDUPMask(SVOp) ||
4186 X86::isMOVHLPSMask(SVOp) ||
4187 X86::isMOVHPMask(SVOp) ||
4188 X86::isMOVLPMask(SVOp)))
4191 if (ShouldXformToMOVHLPS(SVOp) ||
4192 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4193 return CommuteVectorShuffle(SVOp, DAG);
4196 // No better options. Use a vshl / vsrl.
4197 EVT EltVT = VT.getVectorElementType();
4198 ShAmt *= EltVT.getSizeInBits();
4199 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4202 bool Commuted = false;
4203 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4204 // 1,1,1,1 -> v8i16 though.
4205 V1IsSplat = isSplatVector(V1.getNode());
4206 V2IsSplat = isSplatVector(V2.getNode());
4208 // Canonicalize the splat or undef, if present, to be on the RHS.
4209 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4210 Op = CommuteVectorShuffle(SVOp, DAG);
4211 SVOp = cast<ShuffleVectorSDNode>(Op);
4212 V1 = SVOp->getOperand(0);
4213 V2 = SVOp->getOperand(1);
4214 std::swap(V1IsSplat, V2IsSplat);
4215 std::swap(V1IsUndef, V2IsUndef);
4219 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4220 // Shuffling low element of v1 into undef, just return v1.
4223 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4224 // the instruction selector will not match, so get a canonical MOVL with
4225 // swapped operands to undo the commute.
4226 return getMOVL(DAG, dl, VT, V2, V1);
4229 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4230 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4231 X86::isUNPCKLMask(SVOp) ||
4232 X86::isUNPCKHMask(SVOp))
4236 // Normalize mask so all entries that point to V2 points to its first
4237 // element then try to match unpck{h|l} again. If match, return a
4238 // new vector_shuffle with the corrected mask.
4239 SDValue NewMask = NormalizeMask(SVOp, DAG);
4240 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4241 if (NSVOp != SVOp) {
4242 if (X86::isUNPCKLMask(NSVOp, true)) {
4244 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4251 // Commute is back and try unpck* again.
4252 // FIXME: this seems wrong.
4253 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4254 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4255 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4256 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4257 X86::isUNPCKLMask(NewSVOp) ||
4258 X86::isUNPCKHMask(NewSVOp))
4262 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4264 // Normalize the node to match x86 shuffle ops if needed
4265 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4266 return CommuteVectorShuffle(SVOp, DAG);
4268 // Check for legal shuffle and return?
4269 SmallVector<int, 16> PermMask;
4270 SVOp->getMask(PermMask);
4271 if (isShuffleMaskLegal(PermMask, VT))
4274 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4275 if (VT == MVT::v8i16) {
4276 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4277 if (NewOp.getNode())
4281 if (VT == MVT::v16i8) {
4282 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4283 if (NewOp.getNode())
4287 // Handle all 4 wide cases with a number of shuffles except for MMX.
4288 if (NumElems == 4 && !isMMX)
4289 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4295 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4296 SelectionDAG &DAG) {
4297 EVT VT = Op.getValueType();
4298 DebugLoc dl = Op.getDebugLoc();
4299 if (VT.getSizeInBits() == 8) {
4300 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4301 Op.getOperand(0), Op.getOperand(1));
4302 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4303 DAG.getValueType(VT));
4304 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4305 } else if (VT.getSizeInBits() == 16) {
4306 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4307 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4309 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4310 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4311 DAG.getNode(ISD::BIT_CONVERT, dl,
4315 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4316 Op.getOperand(0), Op.getOperand(1));
4317 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4318 DAG.getValueType(VT));
4319 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4320 } else if (VT == MVT::f32) {
4321 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4322 // the result back to FR32 register. It's only worth matching if the
4323 // result has a single use which is a store or a bitcast to i32. And in
4324 // the case of a store, it's not worth it if the index is a constant 0,
4325 // because a MOVSSmr can be used instead, which is smaller and faster.
4326 if (!Op.hasOneUse())
4328 SDNode *User = *Op.getNode()->use_begin();
4329 if ((User->getOpcode() != ISD::STORE ||
4330 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4331 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4332 (User->getOpcode() != ISD::BIT_CONVERT ||
4333 User->getValueType(0) != MVT::i32))
4335 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4336 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4339 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4340 } else if (VT == MVT::i32) {
4341 // ExtractPS works with constant index.
4342 if (isa<ConstantSDNode>(Op.getOperand(1)))
4350 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4351 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4354 if (Subtarget->hasSSE41()) {
4355 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4360 EVT VT = Op.getValueType();
4361 DebugLoc dl = Op.getDebugLoc();
4362 // TODO: handle v16i8.
4363 if (VT.getSizeInBits() == 16) {
4364 SDValue Vec = Op.getOperand(0);
4365 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4367 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4368 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4369 DAG.getNode(ISD::BIT_CONVERT, dl,
4372 // Transform it so it match pextrw which produces a 32-bit result.
4373 EVT EltVT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy+1);
4374 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4375 Op.getOperand(0), Op.getOperand(1));
4376 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4377 DAG.getValueType(VT));
4378 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4379 } else if (VT.getSizeInBits() == 32) {
4380 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4384 // SHUFPS the element to the lowest double word, then movss.
4385 int Mask[4] = { Idx, -1, -1, -1 };
4386 EVT VVT = Op.getOperand(0).getValueType();
4387 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4388 DAG.getUNDEF(VVT), Mask);
4389 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4390 DAG.getIntPtrConstant(0));
4391 } else if (VT.getSizeInBits() == 64) {
4392 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4393 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4394 // to match extract_elt for f64.
4395 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4399 // UNPCKHPD the element to the lowest double word, then movsd.
4400 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4401 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4402 int Mask[2] = { 1, -1 };
4403 EVT VVT = Op.getOperand(0).getValueType();
4404 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4405 DAG.getUNDEF(VVT), Mask);
4406 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4407 DAG.getIntPtrConstant(0));
4414 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4415 EVT VT = Op.getValueType();
4416 EVT EltVT = VT.getVectorElementType();
4417 DebugLoc dl = Op.getDebugLoc();
4419 SDValue N0 = Op.getOperand(0);
4420 SDValue N1 = Op.getOperand(1);
4421 SDValue N2 = Op.getOperand(2);
4423 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4424 isa<ConstantSDNode>(N2)) {
4425 unsigned Opc = (EltVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4427 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4429 if (N1.getValueType() != MVT::i32)
4430 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4431 if (N2.getValueType() != MVT::i32)
4432 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4433 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4434 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4435 // Bits [7:6] of the constant are the source select. This will always be
4436 // zero here. The DAG Combiner may combine an extract_elt index into these
4437 // bits. For example (insert (extract, 3), 2) could be matched by putting
4438 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4439 // Bits [5:4] of the constant are the destination select. This is the
4440 // value of the incoming immediate.
4441 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4442 // combine either bitwise AND or insert of float 0.0 to set these bits.
4443 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4444 // Create this as a scalar to vector..
4445 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4446 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4447 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4448 // PINSR* works with constant index.
4455 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4456 EVT VT = Op.getValueType();
4457 EVT EltVT = VT.getVectorElementType();
4459 if (Subtarget->hasSSE41())
4460 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4462 if (EltVT == MVT::i8)
4465 DebugLoc dl = Op.getDebugLoc();
4466 SDValue N0 = Op.getOperand(0);
4467 SDValue N1 = Op.getOperand(1);
4468 SDValue N2 = Op.getOperand(2);
4470 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4471 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4472 // as its second argument.
4473 if (N1.getValueType() != MVT::i32)
4474 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4475 if (N2.getValueType() != MVT::i32)
4476 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4477 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4483 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4484 DebugLoc dl = Op.getDebugLoc();
4485 if (Op.getValueType() == MVT::v2f32)
4486 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4487 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4488 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4489 Op.getOperand(0))));
4491 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4492 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4494 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4495 EVT VT = MVT::v2i32;
4496 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4503 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4504 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4507 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4508 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4509 // one of the above mentioned nodes. It has to be wrapped because otherwise
4510 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4511 // be used to form addressing mode. These wrapped nodes will be selected
4514 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4515 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4517 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4519 unsigned char OpFlag = 0;
4520 unsigned WrapperKind = X86ISD::Wrapper;
4521 CodeModel::Model M = getTargetMachine().getCodeModel();
4523 if (Subtarget->isPICStyleRIPRel() &&
4524 (M == CodeModel::Small || M == CodeModel::Kernel))
4525 WrapperKind = X86ISD::WrapperRIP;
4526 else if (Subtarget->isPICStyleGOT())
4527 OpFlag = X86II::MO_GOTOFF;
4528 else if (Subtarget->isPICStyleStubPIC())
4529 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4531 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4533 CP->getOffset(), OpFlag);
4534 DebugLoc DL = CP->getDebugLoc();
4535 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4536 // With PIC, the address is actually $g + Offset.
4538 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4539 DAG.getNode(X86ISD::GlobalBaseReg,
4540 DebugLoc::getUnknownLoc(), getPointerTy()),
4547 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4548 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4550 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4552 unsigned char OpFlag = 0;
4553 unsigned WrapperKind = X86ISD::Wrapper;
4554 CodeModel::Model M = getTargetMachine().getCodeModel();
4556 if (Subtarget->isPICStyleRIPRel() &&
4557 (M == CodeModel::Small || M == CodeModel::Kernel))
4558 WrapperKind = X86ISD::WrapperRIP;
4559 else if (Subtarget->isPICStyleGOT())
4560 OpFlag = X86II::MO_GOTOFF;
4561 else if (Subtarget->isPICStyleStubPIC())
4562 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4564 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4566 DebugLoc DL = JT->getDebugLoc();
4567 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4569 // With PIC, the address is actually $g + Offset.
4571 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4572 DAG.getNode(X86ISD::GlobalBaseReg,
4573 DebugLoc::getUnknownLoc(), getPointerTy()),
4581 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4582 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4584 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4586 unsigned char OpFlag = 0;
4587 unsigned WrapperKind = X86ISD::Wrapper;
4588 CodeModel::Model M = getTargetMachine().getCodeModel();
4590 if (Subtarget->isPICStyleRIPRel() &&
4591 (M == CodeModel::Small || M == CodeModel::Kernel))
4592 WrapperKind = X86ISD::WrapperRIP;
4593 else if (Subtarget->isPICStyleGOT())
4594 OpFlag = X86II::MO_GOTOFF;
4595 else if (Subtarget->isPICStyleStubPIC())
4596 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4598 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4600 DebugLoc DL = Op.getDebugLoc();
4601 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4604 // With PIC, the address is actually $g + Offset.
4605 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4606 !Subtarget->is64Bit()) {
4607 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4608 DAG.getNode(X86ISD::GlobalBaseReg,
4609 DebugLoc::getUnknownLoc(),
4618 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4620 SelectionDAG &DAG) const {
4621 // Create the TargetGlobalAddress node, folding in the constant
4622 // offset if it is legal.
4623 unsigned char OpFlags =
4624 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
4625 CodeModel::Model M = getTargetMachine().getCodeModel();
4627 if (OpFlags == X86II::MO_NO_FLAG &&
4628 X86::isOffsetSuitableForCodeModel(Offset, M)) {
4629 // A direct static reference to a global.
4630 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4633 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
4636 if (Subtarget->isPICStyleRIPRel() &&
4637 (M == CodeModel::Small || M == CodeModel::Kernel))
4638 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4640 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4642 // With PIC, the address is actually $g + Offset.
4643 if (isGlobalRelativeToPICBase(OpFlags)) {
4644 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4645 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4649 // For globals that require a load from a stub to get the address, emit the
4651 if (isGlobalStubReference(OpFlags))
4652 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4653 PseudoSourceValue::getGOT(), 0);
4655 // If there was a non-zero offset that we didn't fold, create an explicit
4658 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
4659 DAG.getConstant(Offset, getPointerTy()));
4665 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4666 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4667 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4668 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
4672 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
4673 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
4674 unsigned char OperandFlags) {
4675 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4676 DebugLoc dl = GA->getDebugLoc();
4677 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4678 GA->getValueType(0),
4682 SDValue Ops[] = { Chain, TGA, *InFlag };
4683 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
4685 SDValue Ops[] = { Chain, TGA };
4686 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
4688 SDValue Flag = Chain.getValue(1);
4689 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
4692 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4694 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4697 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4698 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
4699 DAG.getNode(X86ISD::GlobalBaseReg,
4700 DebugLoc::getUnknownLoc(),
4702 InFlag = Chain.getValue(1);
4704 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
4707 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4709 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4711 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
4712 X86::RAX, X86II::MO_TLSGD);
4715 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4716 // "local exec" model.
4717 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4718 const EVT PtrVT, TLSModel::Model model,
4720 DebugLoc dl = GA->getDebugLoc();
4721 // Get the Thread Pointer
4722 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
4723 DebugLoc::getUnknownLoc(), PtrVT,
4724 DAG.getRegister(is64Bit? X86::FS : X86::GS,
4727 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
4730 unsigned char OperandFlags = 0;
4731 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
4733 unsigned WrapperKind = X86ISD::Wrapper;
4734 if (model == TLSModel::LocalExec) {
4735 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
4736 } else if (is64Bit) {
4737 assert(model == TLSModel::InitialExec);
4738 OperandFlags = X86II::MO_GOTTPOFF;
4739 WrapperKind = X86ISD::WrapperRIP;
4741 assert(model == TLSModel::InitialExec);
4742 OperandFlags = X86II::MO_INDNTPOFF;
4745 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4747 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
4748 GA->getOffset(), OperandFlags);
4749 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
4751 if (model == TLSModel::InitialExec)
4752 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
4753 PseudoSourceValue::getGOT(), 0);
4755 // The address of the thread local variable is the add of the thread
4756 // pointer with the offset of the variable.
4757 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
4761 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4762 // TODO: implement the "local dynamic" model
4763 // TODO: implement the "initial exec"model for pic executables
4764 assert(Subtarget->isTargetELF() &&
4765 "TLS not implemented for non-ELF targets");
4766 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4767 const GlobalValue *GV = GA->getGlobal();
4769 // If GV is an alias then use the aliasee for determining
4770 // thread-localness.
4771 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
4772 GV = GA->resolveAliasedGlobal(false);
4774 TLSModel::Model model = getTLSModel(GV,
4775 getTargetMachine().getRelocationModel());
4778 case TLSModel::GeneralDynamic:
4779 case TLSModel::LocalDynamic: // not implemented
4780 if (Subtarget->is64Bit())
4781 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4782 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4784 case TLSModel::InitialExec:
4785 case TLSModel::LocalExec:
4786 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
4787 Subtarget->is64Bit());
4790 llvm_unreachable("Unreachable");
4795 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4796 /// take a 2 x i32 value to shift plus a shift amount.
4797 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4798 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4799 EVT VT = Op.getValueType();
4800 unsigned VTBits = VT.getSizeInBits();
4801 DebugLoc dl = Op.getDebugLoc();
4802 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4803 SDValue ShOpLo = Op.getOperand(0);
4804 SDValue ShOpHi = Op.getOperand(1);
4805 SDValue ShAmt = Op.getOperand(2);
4806 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
4807 DAG.getConstant(VTBits - 1, MVT::i8))
4808 : DAG.getConstant(0, VT);
4811 if (Op.getOpcode() == ISD::SHL_PARTS) {
4812 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
4813 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4815 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
4816 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
4819 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
4820 DAG.getConstant(VTBits, MVT::i8));
4821 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
4822 AndNode, DAG.getConstant(0, MVT::i8));
4825 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4826 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4827 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4829 if (Op.getOpcode() == ISD::SHL_PARTS) {
4830 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4831 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4833 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4834 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4837 SDValue Ops[2] = { Lo, Hi };
4838 return DAG.getMergeValues(Ops, 2, dl);
4841 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4842 EVT SrcVT = Op.getOperand(0).getValueType();
4844 if (SrcVT.isVector()) {
4845 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
4851 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4852 "Unknown SINT_TO_FP to lower!");
4854 // These are really Legal; return the operand so the caller accepts it as
4856 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4858 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
4859 Subtarget->is64Bit()) {
4863 DebugLoc dl = Op.getDebugLoc();
4864 unsigned Size = SrcVT.getSizeInBits()/8;
4865 MachineFunction &MF = DAG.getMachineFunction();
4866 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4867 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4868 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
4870 PseudoSourceValue::getStack(), 0);
4871 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
4874 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
4876 SelectionDAG &DAG) {
4878 DebugLoc dl = Op.getDebugLoc();
4880 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4882 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4884 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4885 SmallVector<SDValue, 8> Ops;
4886 Ops.push_back(Chain);
4887 Ops.push_back(StackSlot);
4888 Ops.push_back(DAG.getValueType(SrcVT));
4889 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
4890 Tys, &Ops[0], Ops.size());
4893 Chain = Result.getValue(1);
4894 SDValue InFlag = Result.getValue(2);
4896 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4897 // shouldn't be necessary except that RFP cannot be live across
4898 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4899 MachineFunction &MF = DAG.getMachineFunction();
4900 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4901 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4902 Tys = DAG.getVTList(MVT::Other);
4903 SmallVector<SDValue, 8> Ops;
4904 Ops.push_back(Chain);
4905 Ops.push_back(Result);
4906 Ops.push_back(StackSlot);
4907 Ops.push_back(DAG.getValueType(Op.getValueType()));
4908 Ops.push_back(InFlag);
4909 Chain = DAG.getNode(X86ISD::FST, dl, Tys, &Ops[0], Ops.size());
4910 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
4911 PseudoSourceValue::getStack(), 0);
4917 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
4918 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
4919 // This algorithm is not obvious. Here it is in C code, more or less:
4921 double uint64_to_double( uint32_t hi, uint32_t lo ) {
4922 static const __m128i exp = { 0x4330000045300000ULL, 0 };
4923 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
4925 // Copy ints to xmm registers.
4926 __m128i xh = _mm_cvtsi32_si128( hi );
4927 __m128i xl = _mm_cvtsi32_si128( lo );
4929 // Combine into low half of a single xmm register.
4930 __m128i x = _mm_unpacklo_epi32( xh, xl );
4934 // Merge in appropriate exponents to give the integer bits the right
4936 x = _mm_unpacklo_epi32( x, exp );
4938 // Subtract away the biases to deal with the IEEE-754 double precision
4940 d = _mm_sub_pd( (__m128d) x, bias );
4942 // All conversions up to here are exact. The correctly rounded result is
4943 // calculated using the current rounding mode using the following
4945 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
4946 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
4947 // store doesn't really need to be here (except
4948 // maybe to zero the other double)
4953 DebugLoc dl = Op.getDebugLoc();
4954 LLVMContext *Context = DAG.getContext();
4956 // Build some magic constants.
4957 std::vector<Constant*> CV0;
4958 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
4959 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
4960 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
4961 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
4962 Constant *C0 = ConstantVector::get(CV0);
4963 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
4965 std::vector<Constant*> CV1;
4967 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
4969 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
4970 Constant *C1 = ConstantVector::get(CV1);
4971 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
4973 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4974 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4976 DAG.getIntPtrConstant(1)));
4977 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4978 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4980 DAG.getIntPtrConstant(0)));
4981 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
4982 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
4983 PseudoSourceValue::getConstantPool(), 0,
4985 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
4986 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
4987 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
4988 PseudoSourceValue::getConstantPool(), 0,
4990 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
4992 // Add the halves; easiest way is to swap them into another reg first.
4993 int ShufMask[2] = { 1, -1 };
4994 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
4995 DAG.getUNDEF(MVT::v2f64), ShufMask);
4996 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
4997 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
4998 DAG.getIntPtrConstant(0));
5001 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5002 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5003 DebugLoc dl = Op.getDebugLoc();
5004 // FP constant to bias correct the final result.
5005 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5008 // Load the 32-bit value into an XMM register.
5009 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5010 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5012 DAG.getIntPtrConstant(0)));
5014 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5015 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5016 DAG.getIntPtrConstant(0));
5018 // Or the load with the bias.
5019 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5020 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5021 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5023 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5024 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5025 MVT::v2f64, Bias)));
5026 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5027 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5028 DAG.getIntPtrConstant(0));
5030 // Subtract the bias.
5031 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5033 // Handle final rounding.
5034 EVT DestVT = Op.getValueType();
5036 if (DestVT.bitsLT(MVT::f64)) {
5037 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5038 DAG.getIntPtrConstant(0));
5039 } else if (DestVT.bitsGT(MVT::f64)) {
5040 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5043 // Handle final rounding.
5047 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5048 SDValue N0 = Op.getOperand(0);
5049 DebugLoc dl = Op.getDebugLoc();
5051 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5052 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5053 // the optimization here.
5054 if (DAG.SignBitIsZero(N0))
5055 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5057 EVT SrcVT = N0.getValueType();
5058 if (SrcVT == MVT::i64) {
5059 // We only handle SSE2 f64 target here; caller can expand the rest.
5060 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5063 return LowerUINT_TO_FP_i64(Op, DAG);
5064 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5065 return LowerUINT_TO_FP_i32(Op, DAG);
5068 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5070 // Make a 64-bit buffer, and use it to build an FILD.
5071 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5072 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5073 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5074 getPointerTy(), StackSlot, WordOff);
5075 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5076 StackSlot, NULL, 0);
5077 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5078 OffsetSlot, NULL, 0);
5079 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5082 std::pair<SDValue,SDValue> X86TargetLowering::
5083 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5084 DebugLoc dl = Op.getDebugLoc();
5086 EVT DstTy = Op.getValueType();
5089 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5093 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5094 DstTy.getSimpleVT() >= MVT::i16 &&
5095 "Unknown FP_TO_SINT to lower!");
5097 // These are really Legal.
5098 if (DstTy == MVT::i32 &&
5099 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5100 return std::make_pair(SDValue(), SDValue());
5101 if (Subtarget->is64Bit() &&
5102 DstTy == MVT::i64 &&
5103 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5104 return std::make_pair(SDValue(), SDValue());
5106 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5108 MachineFunction &MF = DAG.getMachineFunction();
5109 unsigned MemSize = DstTy.getSizeInBits()/8;
5110 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5111 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5114 switch (DstTy.getSimpleVT().SimpleTy) {
5115 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5116 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5117 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5118 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5121 SDValue Chain = DAG.getEntryNode();
5122 SDValue Value = Op.getOperand(0);
5123 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5124 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5125 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5126 PseudoSourceValue::getStack(), 0);
5127 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5129 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5131 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5132 Chain = Value.getValue(1);
5133 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5134 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5137 // Build the FP_TO_INT*_IN_MEM
5138 SDValue Ops[] = { Chain, Value, StackSlot };
5139 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5141 return std::make_pair(FIST, StackSlot);
5144 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5145 if (Op.getValueType().isVector()) {
5146 if (Op.getValueType() == MVT::v2i32 &&
5147 Op.getOperand(0).getValueType() == MVT::v2f64) {
5153 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5154 SDValue FIST = Vals.first, StackSlot = Vals.second;
5155 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5156 if (FIST.getNode() == 0) return Op;
5159 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5160 FIST, StackSlot, NULL, 0);
5163 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5164 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5165 SDValue FIST = Vals.first, StackSlot = Vals.second;
5166 assert(FIST.getNode() && "Unexpected failure");
5169 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5170 FIST, StackSlot, NULL, 0);
5173 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5174 LLVMContext *Context = DAG.getContext();
5175 DebugLoc dl = Op.getDebugLoc();
5176 EVT VT = Op.getValueType();
5179 EltVT = VT.getVectorElementType();
5180 std::vector<Constant*> CV;
5181 if (EltVT == MVT::f64) {
5182 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5186 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5192 Constant *C = ConstantVector::get(CV);
5193 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5194 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5195 PseudoSourceValue::getConstantPool(), 0,
5197 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5200 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5201 LLVMContext *Context = DAG.getContext();
5202 DebugLoc dl = Op.getDebugLoc();
5203 EVT VT = Op.getValueType();
5206 EltVT = VT.getVectorElementType();
5207 std::vector<Constant*> CV;
5208 if (EltVT == MVT::f64) {
5209 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5213 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5219 Constant *C = ConstantVector::get(CV);
5220 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5221 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5222 PseudoSourceValue::getConstantPool(), 0,
5224 if (VT.isVector()) {
5225 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5226 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5227 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5229 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5231 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5235 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5236 LLVMContext *Context = DAG.getContext();
5237 SDValue Op0 = Op.getOperand(0);
5238 SDValue Op1 = Op.getOperand(1);
5239 DebugLoc dl = Op.getDebugLoc();
5240 EVT VT = Op.getValueType();
5241 EVT SrcVT = Op1.getValueType();
5243 // If second operand is smaller, extend it first.
5244 if (SrcVT.bitsLT(VT)) {
5245 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5248 // And if it is bigger, shrink it first.
5249 if (SrcVT.bitsGT(VT)) {
5250 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5254 // At this point the operands and the result should have the same
5255 // type, and that won't be f80 since that is not custom lowered.
5257 // First get the sign bit of second operand.
5258 std::vector<Constant*> CV;
5259 if (SrcVT == MVT::f64) {
5260 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5261 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5263 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5264 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5265 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5266 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5268 Constant *C = ConstantVector::get(CV);
5269 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5270 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5271 PseudoSourceValue::getConstantPool(), 0,
5273 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5275 // Shift sign bit right or left if the two operands have different types.
5276 if (SrcVT.bitsGT(VT)) {
5277 // Op0 is MVT::f32, Op1 is MVT::f64.
5278 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5279 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5280 DAG.getConstant(32, MVT::i32));
5281 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5282 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5283 DAG.getIntPtrConstant(0));
5286 // Clear first operand sign bit.
5288 if (VT == MVT::f64) {
5289 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5290 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5292 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5293 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5294 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5295 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5297 C = ConstantVector::get(CV);
5298 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5299 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5300 PseudoSourceValue::getConstantPool(), 0,
5302 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5304 // Or the value with the sign bit.
5305 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5308 /// Emit nodes that will be selected as "test Op0,Op0", or something
5310 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5311 SelectionDAG &DAG) {
5312 DebugLoc dl = Op.getDebugLoc();
5314 // CF and OF aren't always set the way we want. Determine which
5315 // of these we need.
5316 bool NeedCF = false;
5317 bool NeedOF = false;
5319 case X86::COND_A: case X86::COND_AE:
5320 case X86::COND_B: case X86::COND_BE:
5323 case X86::COND_G: case X86::COND_GE:
5324 case X86::COND_L: case X86::COND_LE:
5325 case X86::COND_O: case X86::COND_NO:
5331 // See if we can use the EFLAGS value from the operand instead of
5332 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5333 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5334 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5335 unsigned Opcode = 0;
5336 unsigned NumOperands = 0;
5337 switch (Op.getNode()->getOpcode()) {
5339 // Due to an isel shortcoming, be conservative if this add is likely to
5340 // be selected as part of a load-modify-store instruction. When the root
5341 // node in a match is a store, isel doesn't know how to remap non-chain
5342 // non-flag uses of other nodes in the match, such as the ADD in this
5343 // case. This leads to the ADD being left around and reselected, with
5344 // the result being two adds in the output.
5345 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5346 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5347 if (UI->getOpcode() == ISD::STORE)
5349 if (ConstantSDNode *C =
5350 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5351 // An add of one will be selected as an INC.
5352 if (C->getAPIntValue() == 1) {
5353 Opcode = X86ISD::INC;
5357 // An add of negative one (subtract of one) will be selected as a DEC.
5358 if (C->getAPIntValue().isAllOnesValue()) {
5359 Opcode = X86ISD::DEC;
5364 // Otherwise use a regular EFLAGS-setting add.
5365 Opcode = X86ISD::ADD;
5369 // If the primary and result isn't used, don't bother using X86ISD::AND,
5370 // because a TEST instruction will be better.
5371 bool NonFlagUse = false;
5372 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5373 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5374 if (UI->getOpcode() != ISD::BRCOND &&
5375 UI->getOpcode() != ISD::SELECT &&
5376 UI->getOpcode() != ISD::SETCC) {
5387 // Due to the ISEL shortcoming noted above, be conservative if this op is
5388 // likely to be selected as part of a load-modify-store instruction.
5389 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5390 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5391 if (UI->getOpcode() == ISD::STORE)
5393 // Otherwise use a regular EFLAGS-setting instruction.
5394 switch (Op.getNode()->getOpcode()) {
5395 case ISD::SUB: Opcode = X86ISD::SUB; break;
5396 case ISD::OR: Opcode = X86ISD::OR; break;
5397 case ISD::XOR: Opcode = X86ISD::XOR; break;
5398 case ISD::AND: Opcode = X86ISD::AND; break;
5399 default: llvm_unreachable("unexpected operator!");
5410 return SDValue(Op.getNode(), 1);
5416 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5417 SmallVector<SDValue, 4> Ops;
5418 for (unsigned i = 0; i != NumOperands; ++i)
5419 Ops.push_back(Op.getOperand(i));
5420 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5421 DAG.ReplaceAllUsesWith(Op, New);
5422 return SDValue(New.getNode(), 1);
5426 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5427 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5428 DAG.getConstant(0, Op.getValueType()));
5431 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5433 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5434 SelectionDAG &DAG) {
5435 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5436 if (C->getAPIntValue() == 0)
5437 return EmitTest(Op0, X86CC, DAG);
5439 DebugLoc dl = Op0.getDebugLoc();
5440 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5443 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5444 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5445 SDValue Op0 = Op.getOperand(0);
5446 SDValue Op1 = Op.getOperand(1);
5447 DebugLoc dl = Op.getDebugLoc();
5448 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5450 // Lower (X & (1 << N)) == 0 to BT(X, N).
5451 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5452 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5453 if (Op0.getOpcode() == ISD::AND &&
5455 Op1.getOpcode() == ISD::Constant &&
5456 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5457 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5459 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5460 if (ConstantSDNode *Op010C =
5461 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5462 if (Op010C->getZExtValue() == 1) {
5463 LHS = Op0.getOperand(0);
5464 RHS = Op0.getOperand(1).getOperand(1);
5466 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5467 if (ConstantSDNode *Op000C =
5468 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5469 if (Op000C->getZExtValue() == 1) {
5470 LHS = Op0.getOperand(1);
5471 RHS = Op0.getOperand(0).getOperand(1);
5473 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5474 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5475 SDValue AndLHS = Op0.getOperand(0);
5476 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5477 LHS = AndLHS.getOperand(0);
5478 RHS = AndLHS.getOperand(1);
5482 if (LHS.getNode()) {
5483 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5484 // instruction. Since the shift amount is in-range-or-undefined, we know
5485 // that doing a bittest on the i16 value is ok. We extend to i32 because
5486 // the encoding for the i16 version is larger than the i32 version.
5487 if (LHS.getValueType() == MVT::i8)
5488 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5490 // If the operand types disagree, extend the shift amount to match. Since
5491 // BT ignores high bits (like shifts) we can use anyextend.
5492 if (LHS.getValueType() != RHS.getValueType())
5493 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5495 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5496 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5497 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5498 DAG.getConstant(Cond, MVT::i8), BT);
5502 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5503 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5505 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5506 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5507 DAG.getConstant(X86CC, MVT::i8), Cond);
5510 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5512 SDValue Op0 = Op.getOperand(0);
5513 SDValue Op1 = Op.getOperand(1);
5514 SDValue CC = Op.getOperand(2);
5515 EVT VT = Op.getValueType();
5516 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5517 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5518 DebugLoc dl = Op.getDebugLoc();
5522 EVT VT0 = Op0.getValueType();
5523 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5524 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5527 switch (SetCCOpcode) {
5530 case ISD::SETEQ: SSECC = 0; break;
5532 case ISD::SETGT: Swap = true; // Fallthrough
5534 case ISD::SETOLT: SSECC = 1; break;
5536 case ISD::SETGE: Swap = true; // Fallthrough
5538 case ISD::SETOLE: SSECC = 2; break;
5539 case ISD::SETUO: SSECC = 3; break;
5541 case ISD::SETNE: SSECC = 4; break;
5542 case ISD::SETULE: Swap = true;
5543 case ISD::SETUGE: SSECC = 5; break;
5544 case ISD::SETULT: Swap = true;
5545 case ISD::SETUGT: SSECC = 6; break;
5546 case ISD::SETO: SSECC = 7; break;
5549 std::swap(Op0, Op1);
5551 // In the two special cases we can't handle, emit two comparisons.
5553 if (SetCCOpcode == ISD::SETUEQ) {
5555 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5556 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5557 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5559 else if (SetCCOpcode == ISD::SETONE) {
5561 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5562 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5563 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5565 llvm_unreachable("Illegal FP comparison");
5567 // Handle all other FP comparisons here.
5568 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5571 // We are handling one of the integer comparisons here. Since SSE only has
5572 // GT and EQ comparisons for integer, swapping operands and multiple
5573 // operations may be required for some comparisons.
5574 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5575 bool Swap = false, Invert = false, FlipSigns = false;
5577 switch (VT.getSimpleVT().SimpleTy) {
5580 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5582 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5584 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5585 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5588 switch (SetCCOpcode) {
5590 case ISD::SETNE: Invert = true;
5591 case ISD::SETEQ: Opc = EQOpc; break;
5592 case ISD::SETLT: Swap = true;
5593 case ISD::SETGT: Opc = GTOpc; break;
5594 case ISD::SETGE: Swap = true;
5595 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5596 case ISD::SETULT: Swap = true;
5597 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5598 case ISD::SETUGE: Swap = true;
5599 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5602 std::swap(Op0, Op1);
5604 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5605 // bits of the inputs before performing those operations.
5607 EVT EltVT = VT.getVectorElementType();
5608 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5610 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5611 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5613 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5614 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5617 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5619 // If the logical-not of the result is required, perform that now.
5621 Result = DAG.getNOT(dl, Result, VT);
5626 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5627 static bool isX86LogicalCmp(SDValue Op) {
5628 unsigned Opc = Op.getNode()->getOpcode();
5629 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
5631 if (Op.getResNo() == 1 &&
5632 (Opc == X86ISD::ADD ||
5633 Opc == X86ISD::SUB ||
5634 Opc == X86ISD::SMUL ||
5635 Opc == X86ISD::UMUL ||
5636 Opc == X86ISD::INC ||
5637 Opc == X86ISD::DEC ||
5638 Opc == X86ISD::OR ||
5639 Opc == X86ISD::XOR ||
5640 Opc == X86ISD::AND))
5646 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5647 bool addTest = true;
5648 SDValue Cond = Op.getOperand(0);
5649 DebugLoc dl = Op.getDebugLoc();
5652 if (Cond.getOpcode() == ISD::SETCC)
5653 Cond = LowerSETCC(Cond, DAG);
5655 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5656 // setting operand in place of the X86ISD::SETCC.
5657 if (Cond.getOpcode() == X86ISD::SETCC) {
5658 CC = Cond.getOperand(0);
5660 SDValue Cmp = Cond.getOperand(1);
5661 unsigned Opc = Cmp.getOpcode();
5662 EVT VT = Op.getValueType();
5664 bool IllegalFPCMov = false;
5665 if (VT.isFloatingPoint() && !VT.isVector() &&
5666 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5667 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5669 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
5670 Opc == X86ISD::BT) { // FIXME
5677 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5678 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5681 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
5682 SmallVector<SDValue, 4> Ops;
5683 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5684 // condition is true.
5685 Ops.push_back(Op.getOperand(2));
5686 Ops.push_back(Op.getOperand(1));
5688 Ops.push_back(Cond);
5689 return DAG.getNode(X86ISD::CMOV, dl, VTs, &Ops[0], Ops.size());
5692 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
5693 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
5694 // from the AND / OR.
5695 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
5696 Opc = Op.getOpcode();
5697 if (Opc != ISD::OR && Opc != ISD::AND)
5699 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5700 Op.getOperand(0).hasOneUse() &&
5701 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
5702 Op.getOperand(1).hasOneUse());
5705 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
5706 // 1 and that the SETCC node has a single use.
5707 static bool isXor1OfSetCC(SDValue Op) {
5708 if (Op.getOpcode() != ISD::XOR)
5710 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5711 if (N1C && N1C->getAPIntValue() == 1) {
5712 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5713 Op.getOperand(0).hasOneUse();
5718 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5719 bool addTest = true;
5720 SDValue Chain = Op.getOperand(0);
5721 SDValue Cond = Op.getOperand(1);
5722 SDValue Dest = Op.getOperand(2);
5723 DebugLoc dl = Op.getDebugLoc();
5726 if (Cond.getOpcode() == ISD::SETCC)
5727 Cond = LowerSETCC(Cond, DAG);
5729 // FIXME: LowerXALUO doesn't handle these!!
5730 else if (Cond.getOpcode() == X86ISD::ADD ||
5731 Cond.getOpcode() == X86ISD::SUB ||
5732 Cond.getOpcode() == X86ISD::SMUL ||
5733 Cond.getOpcode() == X86ISD::UMUL)
5734 Cond = LowerXALUO(Cond, DAG);
5737 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5738 // setting operand in place of the X86ISD::SETCC.
5739 if (Cond.getOpcode() == X86ISD::SETCC) {
5740 CC = Cond.getOperand(0);
5742 SDValue Cmp = Cond.getOperand(1);
5743 unsigned Opc = Cmp.getOpcode();
5744 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
5745 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
5749 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
5753 // These can only come from an arithmetic instruction with overflow,
5754 // e.g. SADDO, UADDO.
5755 Cond = Cond.getNode()->getOperand(1);
5762 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
5763 SDValue Cmp = Cond.getOperand(0).getOperand(1);
5764 if (CondOpc == ISD::OR) {
5765 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
5766 // two branches instead of an explicit OR instruction with a
5768 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5769 isX86LogicalCmp(Cmp)) {
5770 CC = Cond.getOperand(0).getOperand(0);
5771 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5772 Chain, Dest, CC, Cmp);
5773 CC = Cond.getOperand(1).getOperand(0);
5777 } else { // ISD::AND
5778 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
5779 // two branches instead of an explicit AND instruction with a
5780 // separate test. However, we only do this if this block doesn't
5781 // have a fall-through edge, because this requires an explicit
5782 // jmp when the condition is false.
5783 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5784 isX86LogicalCmp(Cmp) &&
5785 Op.getNode()->hasOneUse()) {
5786 X86::CondCode CCode =
5787 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5788 CCode = X86::GetOppositeBranchCondition(CCode);
5789 CC = DAG.getConstant(CCode, MVT::i8);
5790 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
5791 // Look for an unconditional branch following this conditional branch.
5792 // We need this because we need to reverse the successors in order
5793 // to implement FCMP_OEQ.
5794 if (User.getOpcode() == ISD::BR) {
5795 SDValue FalseBB = User.getOperand(1);
5797 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
5798 assert(NewBR == User);
5801 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5802 Chain, Dest, CC, Cmp);
5803 X86::CondCode CCode =
5804 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
5805 CCode = X86::GetOppositeBranchCondition(CCode);
5806 CC = DAG.getConstant(CCode, MVT::i8);
5812 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
5813 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
5814 // It should be transformed during dag combiner except when the condition
5815 // is set by a arithmetics with overflow node.
5816 X86::CondCode CCode =
5817 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5818 CCode = X86::GetOppositeBranchCondition(CCode);
5819 CC = DAG.getConstant(CCode, MVT::i8);
5820 Cond = Cond.getOperand(0).getOperand(1);
5826 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5827 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5829 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5830 Chain, Dest, CC, Cond);
5834 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5835 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5836 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5837 // that the guard pages used by the OS virtual memory manager are allocated in
5838 // correct sequence.
5840 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5841 SelectionDAG &DAG) {
5842 assert(Subtarget->isTargetCygMing() &&
5843 "This should be used only on Cygwin/Mingw targets");
5844 DebugLoc dl = Op.getDebugLoc();
5847 SDValue Chain = Op.getOperand(0);
5848 SDValue Size = Op.getOperand(1);
5849 // FIXME: Ensure alignment here
5853 EVT IntPtr = getPointerTy();
5854 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5856 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
5858 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
5859 Flag = Chain.getValue(1);
5861 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5862 SDValue Ops[] = { Chain,
5863 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5864 DAG.getRegister(X86::EAX, IntPtr),
5865 DAG.getRegister(X86StackPtr, SPTy),
5867 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
5868 Flag = Chain.getValue(1);
5870 Chain = DAG.getCALLSEQ_END(Chain,
5871 DAG.getIntPtrConstant(0, true),
5872 DAG.getIntPtrConstant(0, true),
5875 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
5877 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5878 return DAG.getMergeValues(Ops1, 2, dl);
5882 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
5884 SDValue Dst, SDValue Src,
5885 SDValue Size, unsigned Align,
5887 uint64_t DstSVOff) {
5888 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5890 // If not DWORD aligned or size is more than the threshold, call the library.
5891 // The libc version is likely to be faster for these cases. It can use the
5892 // address value and run time information about the CPU.
5893 if ((Align & 3) != 0 ||
5895 ConstantSize->getZExtValue() >
5896 getSubtarget()->getMaxInlineSizeThreshold()) {
5897 SDValue InFlag(0, 0);
5899 // Check to see if there is a specialized entry-point for memory zeroing.
5900 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5902 if (const char *bzeroEntry = V &&
5903 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5904 EVT IntPtr = getPointerTy();
5905 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
5906 TargetLowering::ArgListTy Args;
5907 TargetLowering::ArgListEntry Entry;
5909 Entry.Ty = IntPtrTy;
5910 Args.push_back(Entry);
5912 Args.push_back(Entry);
5913 std::pair<SDValue,SDValue> CallResult =
5914 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
5915 false, false, false, false,
5916 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
5917 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
5918 return CallResult.second;
5921 // Otherwise have the target-independent code call memset.
5925 uint64_t SizeVal = ConstantSize->getZExtValue();
5926 SDValue InFlag(0, 0);
5929 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5930 unsigned BytesLeft = 0;
5931 bool TwoRepStos = false;
5934 uint64_t Val = ValC->getZExtValue() & 255;
5936 // If the value is a constant, then we can potentially use larger sets.
5937 switch (Align & 3) {
5938 case 2: // WORD aligned
5941 Val = (Val << 8) | Val;
5943 case 0: // DWORD aligned
5946 Val = (Val << 8) | Val;
5947 Val = (Val << 16) | Val;
5948 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
5951 Val = (Val << 32) | Val;
5954 default: // Byte aligned
5957 Count = DAG.getIntPtrConstant(SizeVal);
5961 if (AVT.bitsGT(MVT::i8)) {
5962 unsigned UBytes = AVT.getSizeInBits() / 8;
5963 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
5964 BytesLeft = SizeVal % UBytes;
5967 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
5969 InFlag = Chain.getValue(1);
5972 Count = DAG.getIntPtrConstant(SizeVal);
5973 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
5974 InFlag = Chain.getValue(1);
5977 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5980 InFlag = Chain.getValue(1);
5981 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5984 InFlag = Chain.getValue(1);
5986 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5987 SmallVector<SDValue, 8> Ops;
5988 Ops.push_back(Chain);
5989 Ops.push_back(DAG.getValueType(AVT));
5990 Ops.push_back(InFlag);
5991 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5994 InFlag = Chain.getValue(1);
5996 EVT CVT = Count.getValueType();
5997 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
5998 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
5999 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6002 InFlag = Chain.getValue(1);
6003 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6005 Ops.push_back(Chain);
6006 Ops.push_back(DAG.getValueType(MVT::i8));
6007 Ops.push_back(InFlag);
6008 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
6009 } else if (BytesLeft) {
6010 // Handle the last 1 - 7 bytes.
6011 unsigned Offset = SizeVal - BytesLeft;
6012 EVT AddrVT = Dst.getValueType();
6013 EVT SizeVT = Size.getValueType();
6015 Chain = DAG.getMemset(Chain, dl,
6016 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6017 DAG.getConstant(Offset, AddrVT)),
6019 DAG.getConstant(BytesLeft, SizeVT),
6020 Align, DstSV, DstSVOff + Offset);
6023 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6028 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6029 SDValue Chain, SDValue Dst, SDValue Src,
6030 SDValue Size, unsigned Align,
6032 const Value *DstSV, uint64_t DstSVOff,
6033 const Value *SrcSV, uint64_t SrcSVOff) {
6034 // This requires the copy size to be a constant, preferrably
6035 // within a subtarget-specific limit.
6036 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6039 uint64_t SizeVal = ConstantSize->getZExtValue();
6040 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6043 /// If not DWORD aligned, call the library.
6044 if ((Align & 3) != 0)
6049 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6052 unsigned UBytes = AVT.getSizeInBits() / 8;
6053 unsigned CountVal = SizeVal / UBytes;
6054 SDValue Count = DAG.getIntPtrConstant(CountVal);
6055 unsigned BytesLeft = SizeVal % UBytes;
6057 SDValue InFlag(0, 0);
6058 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6061 InFlag = Chain.getValue(1);
6062 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6065 InFlag = Chain.getValue(1);
6066 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6069 InFlag = Chain.getValue(1);
6071 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6072 SmallVector<SDValue, 8> Ops;
6073 Ops.push_back(Chain);
6074 Ops.push_back(DAG.getValueType(AVT));
6075 Ops.push_back(InFlag);
6076 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, &Ops[0], Ops.size());
6078 SmallVector<SDValue, 4> Results;
6079 Results.push_back(RepMovs);
6081 // Handle the last 1 - 7 bytes.
6082 unsigned Offset = SizeVal - BytesLeft;
6083 EVT DstVT = Dst.getValueType();
6084 EVT SrcVT = Src.getValueType();
6085 EVT SizeVT = Size.getValueType();
6086 Results.push_back(DAG.getMemcpy(Chain, dl,
6087 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6088 DAG.getConstant(Offset, DstVT)),
6089 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6090 DAG.getConstant(Offset, SrcVT)),
6091 DAG.getConstant(BytesLeft, SizeVT),
6092 Align, AlwaysInline,
6093 DstSV, DstSVOff + Offset,
6094 SrcSV, SrcSVOff + Offset));
6097 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6098 &Results[0], Results.size());
6101 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6102 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6103 DebugLoc dl = Op.getDebugLoc();
6105 if (!Subtarget->is64Bit()) {
6106 // vastart just stores the address of the VarArgsFrameIndex slot into the
6107 // memory location argument.
6108 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6109 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6113 // gp_offset (0 - 6 * 8)
6114 // fp_offset (48 - 48 + 8 * 16)
6115 // overflow_arg_area (point to parameters coming in memory).
6117 SmallVector<SDValue, 8> MemOps;
6118 SDValue FIN = Op.getOperand(1);
6120 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6121 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6123 MemOps.push_back(Store);
6126 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6127 FIN, DAG.getIntPtrConstant(4));
6128 Store = DAG.getStore(Op.getOperand(0), dl,
6129 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6131 MemOps.push_back(Store);
6133 // Store ptr to overflow_arg_area
6134 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6135 FIN, DAG.getIntPtrConstant(4));
6136 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6137 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6138 MemOps.push_back(Store);
6140 // Store ptr to reg_save_area.
6141 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6142 FIN, DAG.getIntPtrConstant(8));
6143 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6144 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6145 MemOps.push_back(Store);
6146 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6147 &MemOps[0], MemOps.size());
6150 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6151 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6152 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6153 SDValue Chain = Op.getOperand(0);
6154 SDValue SrcPtr = Op.getOperand(1);
6155 SDValue SrcSV = Op.getOperand(2);
6157 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6161 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6162 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6163 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6164 SDValue Chain = Op.getOperand(0);
6165 SDValue DstPtr = Op.getOperand(1);
6166 SDValue SrcPtr = Op.getOperand(2);
6167 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6168 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6169 DebugLoc dl = Op.getDebugLoc();
6171 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6172 DAG.getIntPtrConstant(24), 8, false,
6173 DstSV, 0, SrcSV, 0);
6177 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6178 DebugLoc dl = Op.getDebugLoc();
6179 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6181 default: return SDValue(); // Don't custom lower most intrinsics.
6182 // Comparison intrinsics.
6183 case Intrinsic::x86_sse_comieq_ss:
6184 case Intrinsic::x86_sse_comilt_ss:
6185 case Intrinsic::x86_sse_comile_ss:
6186 case Intrinsic::x86_sse_comigt_ss:
6187 case Intrinsic::x86_sse_comige_ss:
6188 case Intrinsic::x86_sse_comineq_ss:
6189 case Intrinsic::x86_sse_ucomieq_ss:
6190 case Intrinsic::x86_sse_ucomilt_ss:
6191 case Intrinsic::x86_sse_ucomile_ss:
6192 case Intrinsic::x86_sse_ucomigt_ss:
6193 case Intrinsic::x86_sse_ucomige_ss:
6194 case Intrinsic::x86_sse_ucomineq_ss:
6195 case Intrinsic::x86_sse2_comieq_sd:
6196 case Intrinsic::x86_sse2_comilt_sd:
6197 case Intrinsic::x86_sse2_comile_sd:
6198 case Intrinsic::x86_sse2_comigt_sd:
6199 case Intrinsic::x86_sse2_comige_sd:
6200 case Intrinsic::x86_sse2_comineq_sd:
6201 case Intrinsic::x86_sse2_ucomieq_sd:
6202 case Intrinsic::x86_sse2_ucomilt_sd:
6203 case Intrinsic::x86_sse2_ucomile_sd:
6204 case Intrinsic::x86_sse2_ucomigt_sd:
6205 case Intrinsic::x86_sse2_ucomige_sd:
6206 case Intrinsic::x86_sse2_ucomineq_sd: {
6208 ISD::CondCode CC = ISD::SETCC_INVALID;
6211 case Intrinsic::x86_sse_comieq_ss:
6212 case Intrinsic::x86_sse2_comieq_sd:
6216 case Intrinsic::x86_sse_comilt_ss:
6217 case Intrinsic::x86_sse2_comilt_sd:
6221 case Intrinsic::x86_sse_comile_ss:
6222 case Intrinsic::x86_sse2_comile_sd:
6226 case Intrinsic::x86_sse_comigt_ss:
6227 case Intrinsic::x86_sse2_comigt_sd:
6231 case Intrinsic::x86_sse_comige_ss:
6232 case Intrinsic::x86_sse2_comige_sd:
6236 case Intrinsic::x86_sse_comineq_ss:
6237 case Intrinsic::x86_sse2_comineq_sd:
6241 case Intrinsic::x86_sse_ucomieq_ss:
6242 case Intrinsic::x86_sse2_ucomieq_sd:
6243 Opc = X86ISD::UCOMI;
6246 case Intrinsic::x86_sse_ucomilt_ss:
6247 case Intrinsic::x86_sse2_ucomilt_sd:
6248 Opc = X86ISD::UCOMI;
6251 case Intrinsic::x86_sse_ucomile_ss:
6252 case Intrinsic::x86_sse2_ucomile_sd:
6253 Opc = X86ISD::UCOMI;
6256 case Intrinsic::x86_sse_ucomigt_ss:
6257 case Intrinsic::x86_sse2_ucomigt_sd:
6258 Opc = X86ISD::UCOMI;
6261 case Intrinsic::x86_sse_ucomige_ss:
6262 case Intrinsic::x86_sse2_ucomige_sd:
6263 Opc = X86ISD::UCOMI;
6266 case Intrinsic::x86_sse_ucomineq_ss:
6267 case Intrinsic::x86_sse2_ucomineq_sd:
6268 Opc = X86ISD::UCOMI;
6273 SDValue LHS = Op.getOperand(1);
6274 SDValue RHS = Op.getOperand(2);
6275 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6276 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6277 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6278 DAG.getConstant(X86CC, MVT::i8), Cond);
6279 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6281 // ptest intrinsics. The intrinsic these come from are designed to return
6282 // an integer value, not just an instruction so lower it to the ptest
6283 // pattern and a setcc for the result.
6284 case Intrinsic::x86_sse41_ptestz:
6285 case Intrinsic::x86_sse41_ptestc:
6286 case Intrinsic::x86_sse41_ptestnzc:{
6289 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6290 case Intrinsic::x86_sse41_ptestz:
6292 X86CC = X86::COND_E;
6294 case Intrinsic::x86_sse41_ptestc:
6296 X86CC = X86::COND_B;
6298 case Intrinsic::x86_sse41_ptestnzc:
6300 X86CC = X86::COND_A;
6304 SDValue LHS = Op.getOperand(1);
6305 SDValue RHS = Op.getOperand(2);
6306 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6307 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6308 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6309 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6312 // Fix vector shift instructions where the last operand is a non-immediate
6314 case Intrinsic::x86_sse2_pslli_w:
6315 case Intrinsic::x86_sse2_pslli_d:
6316 case Intrinsic::x86_sse2_pslli_q:
6317 case Intrinsic::x86_sse2_psrli_w:
6318 case Intrinsic::x86_sse2_psrli_d:
6319 case Intrinsic::x86_sse2_psrli_q:
6320 case Intrinsic::x86_sse2_psrai_w:
6321 case Intrinsic::x86_sse2_psrai_d:
6322 case Intrinsic::x86_mmx_pslli_w:
6323 case Intrinsic::x86_mmx_pslli_d:
6324 case Intrinsic::x86_mmx_pslli_q:
6325 case Intrinsic::x86_mmx_psrli_w:
6326 case Intrinsic::x86_mmx_psrli_d:
6327 case Intrinsic::x86_mmx_psrli_q:
6328 case Intrinsic::x86_mmx_psrai_w:
6329 case Intrinsic::x86_mmx_psrai_d: {
6330 SDValue ShAmt = Op.getOperand(2);
6331 if (isa<ConstantSDNode>(ShAmt))
6334 unsigned NewIntNo = 0;
6335 EVT ShAmtVT = MVT::v4i32;
6337 case Intrinsic::x86_sse2_pslli_w:
6338 NewIntNo = Intrinsic::x86_sse2_psll_w;
6340 case Intrinsic::x86_sse2_pslli_d:
6341 NewIntNo = Intrinsic::x86_sse2_psll_d;
6343 case Intrinsic::x86_sse2_pslli_q:
6344 NewIntNo = Intrinsic::x86_sse2_psll_q;
6346 case Intrinsic::x86_sse2_psrli_w:
6347 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6349 case Intrinsic::x86_sse2_psrli_d:
6350 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6352 case Intrinsic::x86_sse2_psrli_q:
6353 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6355 case Intrinsic::x86_sse2_psrai_w:
6356 NewIntNo = Intrinsic::x86_sse2_psra_w;
6358 case Intrinsic::x86_sse2_psrai_d:
6359 NewIntNo = Intrinsic::x86_sse2_psra_d;
6362 ShAmtVT = MVT::v2i32;
6364 case Intrinsic::x86_mmx_pslli_w:
6365 NewIntNo = Intrinsic::x86_mmx_psll_w;
6367 case Intrinsic::x86_mmx_pslli_d:
6368 NewIntNo = Intrinsic::x86_mmx_psll_d;
6370 case Intrinsic::x86_mmx_pslli_q:
6371 NewIntNo = Intrinsic::x86_mmx_psll_q;
6373 case Intrinsic::x86_mmx_psrli_w:
6374 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6376 case Intrinsic::x86_mmx_psrli_d:
6377 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6379 case Intrinsic::x86_mmx_psrli_q:
6380 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6382 case Intrinsic::x86_mmx_psrai_w:
6383 NewIntNo = Intrinsic::x86_mmx_psra_w;
6385 case Intrinsic::x86_mmx_psrai_d:
6386 NewIntNo = Intrinsic::x86_mmx_psra_d;
6388 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6394 // The vector shift intrinsics with scalars uses 32b shift amounts but
6395 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
6399 ShOps[1] = DAG.getConstant(0, MVT::i32);
6400 if (ShAmtVT == MVT::v4i32) {
6401 ShOps[2] = DAG.getUNDEF(MVT::i32);
6402 ShOps[3] = DAG.getUNDEF(MVT::i32);
6403 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
6405 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
6408 EVT VT = Op.getValueType();
6409 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
6410 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6411 DAG.getConstant(NewIntNo, MVT::i32),
6412 Op.getOperand(1), ShAmt);
6417 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6418 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6419 DebugLoc dl = Op.getDebugLoc();
6422 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6424 DAG.getConstant(TD->getPointerSize(),
6425 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6426 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6427 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6432 // Just load the return address.
6433 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6434 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6435 RetAddrFI, NULL, 0);
6438 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6439 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6440 MFI->setFrameAddressIsTaken(true);
6441 EVT VT = Op.getValueType();
6442 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6443 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6444 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6445 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6447 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6451 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6452 SelectionDAG &DAG) {
6453 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6456 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6458 MachineFunction &MF = DAG.getMachineFunction();
6459 SDValue Chain = Op.getOperand(0);
6460 SDValue Offset = Op.getOperand(1);
6461 SDValue Handler = Op.getOperand(2);
6462 DebugLoc dl = Op.getDebugLoc();
6464 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6466 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6468 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6469 DAG.getIntPtrConstant(-TD->getPointerSize()));
6470 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6471 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6472 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6473 MF.getRegInfo().addLiveOut(StoreAddrReg);
6475 return DAG.getNode(X86ISD::EH_RETURN, dl,
6477 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6480 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6481 SelectionDAG &DAG) {
6482 SDValue Root = Op.getOperand(0);
6483 SDValue Trmp = Op.getOperand(1); // trampoline
6484 SDValue FPtr = Op.getOperand(2); // nested function
6485 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6486 DebugLoc dl = Op.getDebugLoc();
6488 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6490 const X86InstrInfo *TII =
6491 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6493 if (Subtarget->is64Bit()) {
6494 SDValue OutChains[6];
6496 // Large code-model.
6498 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6499 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6501 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6502 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6504 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6506 // Load the pointer to the nested function into R11.
6507 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6508 SDValue Addr = Trmp;
6509 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6512 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6513 DAG.getConstant(2, MVT::i64));
6514 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6516 // Load the 'nest' parameter value into R10.
6517 // R10 is specified in X86CallingConv.td
6518 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6519 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6520 DAG.getConstant(10, MVT::i64));
6521 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6522 Addr, TrmpAddr, 10);
6524 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6525 DAG.getConstant(12, MVT::i64));
6526 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6528 // Jump to the nested function.
6529 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6530 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6531 DAG.getConstant(20, MVT::i64));
6532 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6533 Addr, TrmpAddr, 20);
6535 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6536 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6537 DAG.getConstant(22, MVT::i64));
6538 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6542 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6543 return DAG.getMergeValues(Ops, 2, dl);
6545 const Function *Func =
6546 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6547 CallingConv::ID CC = Func->getCallingConv();
6552 llvm_unreachable("Unsupported calling convention");
6553 case CallingConv::C:
6554 case CallingConv::X86_StdCall: {
6555 // Pass 'nest' parameter in ECX.
6556 // Must be kept in sync with X86CallingConv.td
6559 // Check that ECX wasn't needed by an 'inreg' parameter.
6560 const FunctionType *FTy = Func->getFunctionType();
6561 const AttrListPtr &Attrs = Func->getAttributes();
6563 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6564 unsigned InRegCount = 0;
6567 for (FunctionType::param_iterator I = FTy->param_begin(),
6568 E = FTy->param_end(); I != E; ++I, ++Idx)
6569 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6570 // FIXME: should only count parameters that are lowered to integers.
6571 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6573 if (InRegCount > 2) {
6574 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
6579 case CallingConv::X86_FastCall:
6580 case CallingConv::Fast:
6581 // Pass 'nest' parameter in EAX.
6582 // Must be kept in sync with X86CallingConv.td
6587 SDValue OutChains[4];
6590 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6591 DAG.getConstant(10, MVT::i32));
6592 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
6594 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6595 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6596 OutChains[0] = DAG.getStore(Root, dl,
6597 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6600 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6601 DAG.getConstant(1, MVT::i32));
6602 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
6604 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6605 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6606 DAG.getConstant(5, MVT::i32));
6607 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
6608 TrmpAddr, 5, false, 1);
6610 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6611 DAG.getConstant(6, MVT::i32));
6612 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
6615 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
6616 return DAG.getMergeValues(Ops, 2, dl);
6620 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6622 The rounding mode is in bits 11:10 of FPSR, and has the following
6629 FLT_ROUNDS, on the other hand, expects the following:
6636 To perform the conversion, we do:
6637 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6640 MachineFunction &MF = DAG.getMachineFunction();
6641 const TargetMachine &TM = MF.getTarget();
6642 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6643 unsigned StackAlignment = TFI.getStackAlignment();
6644 EVT VT = Op.getValueType();
6645 DebugLoc dl = Op.getDebugLoc();
6647 // Save FP Control Word to stack slot
6648 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
6649 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6651 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
6652 DAG.getEntryNode(), StackSlot);
6654 // Load FP Control Word from stack slot
6655 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
6657 // Transform as necessary
6659 DAG.getNode(ISD::SRL, dl, MVT::i16,
6660 DAG.getNode(ISD::AND, dl, MVT::i16,
6661 CWD, DAG.getConstant(0x800, MVT::i16)),
6662 DAG.getConstant(11, MVT::i8));
6664 DAG.getNode(ISD::SRL, dl, MVT::i16,
6665 DAG.getNode(ISD::AND, dl, MVT::i16,
6666 CWD, DAG.getConstant(0x400, MVT::i16)),
6667 DAG.getConstant(9, MVT::i8));
6670 DAG.getNode(ISD::AND, dl, MVT::i16,
6671 DAG.getNode(ISD::ADD, dl, MVT::i16,
6672 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
6673 DAG.getConstant(1, MVT::i16)),
6674 DAG.getConstant(3, MVT::i16));
6677 return DAG.getNode((VT.getSizeInBits() < 16 ?
6678 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
6681 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
6682 EVT VT = Op.getValueType();
6684 unsigned NumBits = VT.getSizeInBits();
6685 DebugLoc dl = Op.getDebugLoc();
6687 Op = Op.getOperand(0);
6688 if (VT == MVT::i8) {
6689 // Zero extend to i32 since there is not an i8 bsr.
6691 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6694 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
6695 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6696 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
6698 // If src is zero (i.e. bsr sets ZF), returns NumBits.
6699 SmallVector<SDValue, 4> Ops;
6701 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
6702 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6703 Ops.push_back(Op.getValue(1));
6704 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6706 // Finally xor with NumBits-1.
6707 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
6710 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6714 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
6715 EVT VT = Op.getValueType();
6717 unsigned NumBits = VT.getSizeInBits();
6718 DebugLoc dl = Op.getDebugLoc();
6720 Op = Op.getOperand(0);
6721 if (VT == MVT::i8) {
6723 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6726 // Issue a bsf (scan bits forward) which also sets EFLAGS.
6727 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6728 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
6730 // If src is zero (i.e. bsf sets ZF), returns NumBits.
6731 SmallVector<SDValue, 4> Ops;
6733 Ops.push_back(DAG.getConstant(NumBits, OpVT));
6734 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6735 Ops.push_back(Op.getValue(1));
6736 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6739 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6743 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
6744 EVT VT = Op.getValueType();
6745 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
6746 DebugLoc dl = Op.getDebugLoc();
6748 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
6749 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
6750 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
6751 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
6752 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
6754 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
6755 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
6756 // return AloBlo + AloBhi + AhiBlo;
6758 SDValue A = Op.getOperand(0);
6759 SDValue B = Op.getOperand(1);
6761 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6762 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6763 A, DAG.getConstant(32, MVT::i32));
6764 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6765 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6766 B, DAG.getConstant(32, MVT::i32));
6767 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6768 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6770 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6771 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6773 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6774 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6776 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6777 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6778 AloBhi, DAG.getConstant(32, MVT::i32));
6779 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6780 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6781 AhiBlo, DAG.getConstant(32, MVT::i32));
6782 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
6783 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
6788 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
6789 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
6790 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
6791 // looks for this combo and may remove the "setcc" instruction if the "setcc"
6792 // has only one use.
6793 SDNode *N = Op.getNode();
6794 SDValue LHS = N->getOperand(0);
6795 SDValue RHS = N->getOperand(1);
6796 unsigned BaseOp = 0;
6798 DebugLoc dl = Op.getDebugLoc();
6800 switch (Op.getOpcode()) {
6801 default: llvm_unreachable("Unknown ovf instruction!");
6803 // A subtract of one will be selected as a INC. Note that INC doesn't
6804 // set CF, so we can't do this for UADDO.
6805 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6806 if (C->getAPIntValue() == 1) {
6807 BaseOp = X86ISD::INC;
6811 BaseOp = X86ISD::ADD;
6815 BaseOp = X86ISD::ADD;
6819 // A subtract of one will be selected as a DEC. Note that DEC doesn't
6820 // set CF, so we can't do this for USUBO.
6821 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6822 if (C->getAPIntValue() == 1) {
6823 BaseOp = X86ISD::DEC;
6827 BaseOp = X86ISD::SUB;
6831 BaseOp = X86ISD::SUB;
6835 BaseOp = X86ISD::SMUL;
6839 BaseOp = X86ISD::UMUL;
6844 // Also sets EFLAGS.
6845 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
6846 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
6849 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
6850 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
6852 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
6856 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
6857 EVT T = Op.getValueType();
6858 DebugLoc dl = Op.getDebugLoc();
6861 switch(T.getSimpleVT().SimpleTy) {
6863 assert(false && "Invalid value type!");
6864 case MVT::i8: Reg = X86::AL; size = 1; break;
6865 case MVT::i16: Reg = X86::AX; size = 2; break;
6866 case MVT::i32: Reg = X86::EAX; size = 4; break;
6868 assert(Subtarget->is64Bit() && "Node not type legal!");
6869 Reg = X86::RAX; size = 8;
6872 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
6873 Op.getOperand(2), SDValue());
6874 SDValue Ops[] = { cpIn.getValue(0),
6877 DAG.getTargetConstant(size, MVT::i8),
6879 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6880 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
6882 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
6886 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
6887 SelectionDAG &DAG) {
6888 assert(Subtarget->is64Bit() && "Result not type legalized?");
6889 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6890 SDValue TheChain = Op.getOperand(0);
6891 DebugLoc dl = Op.getDebugLoc();
6892 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6893 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
6894 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
6896 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
6897 DAG.getConstant(32, MVT::i8));
6899 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
6902 return DAG.getMergeValues(Ops, 2, dl);
6905 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
6906 SDNode *Node = Op.getNode();
6907 DebugLoc dl = Node->getDebugLoc();
6908 EVT T = Node->getValueType(0);
6909 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
6910 DAG.getConstant(0, T), Node->getOperand(2));
6911 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
6912 cast<AtomicSDNode>(Node)->getMemoryVT(),
6913 Node->getOperand(0),
6914 Node->getOperand(1), negOp,
6915 cast<AtomicSDNode>(Node)->getSrcValue(),
6916 cast<AtomicSDNode>(Node)->getAlignment());
6919 /// LowerOperation - Provide custom lowering hooks for some operations.
6921 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
6922 switch (Op.getOpcode()) {
6923 default: llvm_unreachable("Should not custom lower this!");
6924 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
6925 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
6926 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
6927 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6928 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6929 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6930 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
6931 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6932 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
6933 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6934 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
6935 case ISD::SHL_PARTS:
6936 case ISD::SRA_PARTS:
6937 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
6938 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
6939 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
6940 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
6941 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
6942 case ISD::FABS: return LowerFABS(Op, DAG);
6943 case ISD::FNEG: return LowerFNEG(Op, DAG);
6944 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6945 case ISD::SETCC: return LowerSETCC(Op, DAG);
6946 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
6947 case ISD::SELECT: return LowerSELECT(Op, DAG);
6948 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
6949 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
6950 case ISD::VASTART: return LowerVASTART(Op, DAG);
6951 case ISD::VAARG: return LowerVAARG(Op, DAG);
6952 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
6953 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
6954 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6955 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6956 case ISD::FRAME_TO_ARGS_OFFSET:
6957 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6958 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6959 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6960 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6961 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6962 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6963 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6964 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
6970 case ISD::UMULO: return LowerXALUO(Op, DAG);
6971 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
6975 void X86TargetLowering::
6976 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
6977 SelectionDAG &DAG, unsigned NewOp) {
6978 EVT T = Node->getValueType(0);
6979 DebugLoc dl = Node->getDebugLoc();
6980 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
6982 SDValue Chain = Node->getOperand(0);
6983 SDValue In1 = Node->getOperand(1);
6984 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6985 Node->getOperand(2), DAG.getIntPtrConstant(0));
6986 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6987 Node->getOperand(2), DAG.getIntPtrConstant(1));
6988 SDValue Ops[] = { Chain, In1, In2L, In2H };
6989 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
6991 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
6992 cast<MemSDNode>(Node)->getMemOperand());
6993 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
6994 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6995 Results.push_back(Result.getValue(2));
6998 /// ReplaceNodeResults - Replace a node with an illegal result type
6999 /// with a new node built out of custom code.
7000 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7001 SmallVectorImpl<SDValue>&Results,
7002 SelectionDAG &DAG) {
7003 DebugLoc dl = N->getDebugLoc();
7004 switch (N->getOpcode()) {
7006 assert(false && "Do not know how to custom type legalize this operation!");
7008 case ISD::FP_TO_SINT: {
7009 std::pair<SDValue,SDValue> Vals =
7010 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7011 SDValue FIST = Vals.first, StackSlot = Vals.second;
7012 if (FIST.getNode() != 0) {
7013 EVT VT = N->getValueType(0);
7014 // Return a load from the stack slot.
7015 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
7019 case ISD::READCYCLECOUNTER: {
7020 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7021 SDValue TheChain = N->getOperand(0);
7022 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7023 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7025 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7027 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7028 SDValue Ops[] = { eax, edx };
7029 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7030 Results.push_back(edx.getValue(1));
7033 case ISD::ATOMIC_CMP_SWAP: {
7034 EVT T = N->getValueType(0);
7035 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7036 SDValue cpInL, cpInH;
7037 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7038 DAG.getConstant(0, MVT::i32));
7039 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7040 DAG.getConstant(1, MVT::i32));
7041 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7042 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7044 SDValue swapInL, swapInH;
7045 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7046 DAG.getConstant(0, MVT::i32));
7047 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7048 DAG.getConstant(1, MVT::i32));
7049 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7051 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7052 swapInL.getValue(1));
7053 SDValue Ops[] = { swapInH.getValue(0),
7055 swapInH.getValue(1) };
7056 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7057 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7058 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7059 MVT::i32, Result.getValue(1));
7060 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7061 MVT::i32, cpOutL.getValue(2));
7062 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7063 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7064 Results.push_back(cpOutH.getValue(1));
7067 case ISD::ATOMIC_LOAD_ADD:
7068 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7070 case ISD::ATOMIC_LOAD_AND:
7071 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7073 case ISD::ATOMIC_LOAD_NAND:
7074 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7076 case ISD::ATOMIC_LOAD_OR:
7077 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7079 case ISD::ATOMIC_LOAD_SUB:
7080 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7082 case ISD::ATOMIC_LOAD_XOR:
7083 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7085 case ISD::ATOMIC_SWAP:
7086 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7091 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7093 default: return NULL;
7094 case X86ISD::BSF: return "X86ISD::BSF";
7095 case X86ISD::BSR: return "X86ISD::BSR";
7096 case X86ISD::SHLD: return "X86ISD::SHLD";
7097 case X86ISD::SHRD: return "X86ISD::SHRD";
7098 case X86ISD::FAND: return "X86ISD::FAND";
7099 case X86ISD::FOR: return "X86ISD::FOR";
7100 case X86ISD::FXOR: return "X86ISD::FXOR";
7101 case X86ISD::FSRL: return "X86ISD::FSRL";
7102 case X86ISD::FILD: return "X86ISD::FILD";
7103 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7104 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7105 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7106 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7107 case X86ISD::FLD: return "X86ISD::FLD";
7108 case X86ISD::FST: return "X86ISD::FST";
7109 case X86ISD::CALL: return "X86ISD::CALL";
7110 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7111 case X86ISD::BT: return "X86ISD::BT";
7112 case X86ISD::CMP: return "X86ISD::CMP";
7113 case X86ISD::COMI: return "X86ISD::COMI";
7114 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7115 case X86ISD::SETCC: return "X86ISD::SETCC";
7116 case X86ISD::CMOV: return "X86ISD::CMOV";
7117 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7118 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7119 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7120 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7121 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7122 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7123 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7124 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7125 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7126 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7127 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7128 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7129 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7130 case X86ISD::FMAX: return "X86ISD::FMAX";
7131 case X86ISD::FMIN: return "X86ISD::FMIN";
7132 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7133 case X86ISD::FRCP: return "X86ISD::FRCP";
7134 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7135 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7136 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7137 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7138 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7139 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7140 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7141 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7142 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7143 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7144 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7145 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7146 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7147 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7148 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7149 case X86ISD::VSHL: return "X86ISD::VSHL";
7150 case X86ISD::VSRL: return "X86ISD::VSRL";
7151 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7152 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7153 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7154 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7155 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7156 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7157 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7158 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7159 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7160 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7161 case X86ISD::ADD: return "X86ISD::ADD";
7162 case X86ISD::SUB: return "X86ISD::SUB";
7163 case X86ISD::SMUL: return "X86ISD::SMUL";
7164 case X86ISD::UMUL: return "X86ISD::UMUL";
7165 case X86ISD::INC: return "X86ISD::INC";
7166 case X86ISD::DEC: return "X86ISD::DEC";
7167 case X86ISD::OR: return "X86ISD::OR";
7168 case X86ISD::XOR: return "X86ISD::XOR";
7169 case X86ISD::AND: return "X86ISD::AND";
7170 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7171 case X86ISD::PTEST: return "X86ISD::PTEST";
7172 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7176 // isLegalAddressingMode - Return true if the addressing mode represented
7177 // by AM is legal for this target, for a load/store of the specified type.
7178 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7179 const Type *Ty) const {
7180 // X86 supports extremely general addressing modes.
7181 CodeModel::Model M = getTargetMachine().getCodeModel();
7183 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7184 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7189 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7191 // If a reference to this global requires an extra load, we can't fold it.
7192 if (isGlobalStubReference(GVFlags))
7195 // If BaseGV requires a register for the PIC base, we cannot also have a
7196 // BaseReg specified.
7197 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7200 // If lower 4G is not available, then we must use rip-relative addressing.
7201 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7211 // These scales always work.
7216 // These scales are formed with basereg+scalereg. Only accept if there is
7221 default: // Other stuff never works.
7229 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7230 if (!Ty1->isInteger() || !Ty2->isInteger())
7232 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7233 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7234 if (NumBits1 <= NumBits2)
7236 return Subtarget->is64Bit() || NumBits1 < 64;
7239 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7240 if (!VT1.isInteger() || !VT2.isInteger())
7242 unsigned NumBits1 = VT1.getSizeInBits();
7243 unsigned NumBits2 = VT2.getSizeInBits();
7244 if (NumBits1 <= NumBits2)
7246 return Subtarget->is64Bit() || NumBits1 < 64;
7249 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7250 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7251 return Ty1 == Type::getInt32Ty(Ty1->getContext()) &&
7252 Ty2 == Type::getInt64Ty(Ty1->getContext()) && Subtarget->is64Bit();
7255 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7256 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7257 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7260 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7261 // i16 instructions are longer (0x66 prefix) and potentially slower.
7262 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7265 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7266 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7267 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7268 /// are assumed to be legal.
7270 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7272 // Only do shuffles on 128-bit vector types for now.
7273 if (VT.getSizeInBits() == 64)
7276 // FIXME: pshufb, blends, palignr, shifts.
7277 return (VT.getVectorNumElements() == 2 ||
7278 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7279 isMOVLMask(M, VT) ||
7280 isSHUFPMask(M, VT) ||
7281 isPSHUFDMask(M, VT) ||
7282 isPSHUFHWMask(M, VT) ||
7283 isPSHUFLWMask(M, VT) ||
7284 isUNPCKLMask(M, VT) ||
7285 isUNPCKHMask(M, VT) ||
7286 isUNPCKL_v_undef_Mask(M, VT) ||
7287 isUNPCKH_v_undef_Mask(M, VT));
7291 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7293 unsigned NumElts = VT.getVectorNumElements();
7294 // FIXME: This collection of masks seems suspect.
7297 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7298 return (isMOVLMask(Mask, VT) ||
7299 isCommutedMOVLMask(Mask, VT, true) ||
7300 isSHUFPMask(Mask, VT) ||
7301 isCommutedSHUFPMask(Mask, VT));
7306 //===----------------------------------------------------------------------===//
7307 // X86 Scheduler Hooks
7308 //===----------------------------------------------------------------------===//
7310 // private utility function
7312 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7313 MachineBasicBlock *MBB,
7321 TargetRegisterClass *RC,
7322 bool invSrc) const {
7323 // For the atomic bitwise operator, we generate
7326 // ld t1 = [bitinstr.addr]
7327 // op t2 = t1, [bitinstr.val]
7329 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7331 // fallthrough -->nextMBB
7332 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7333 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7334 MachineFunction::iterator MBBIter = MBB;
7337 /// First build the CFG
7338 MachineFunction *F = MBB->getParent();
7339 MachineBasicBlock *thisMBB = MBB;
7340 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7341 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7342 F->insert(MBBIter, newMBB);
7343 F->insert(MBBIter, nextMBB);
7345 // Move all successors to thisMBB to nextMBB
7346 nextMBB->transferSuccessors(thisMBB);
7348 // Update thisMBB to fall through to newMBB
7349 thisMBB->addSuccessor(newMBB);
7351 // newMBB jumps to itself and fall through to nextMBB
7352 newMBB->addSuccessor(nextMBB);
7353 newMBB->addSuccessor(newMBB);
7355 // Insert instructions into newMBB based on incoming instruction
7356 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7357 "unexpected number of operands");
7358 DebugLoc dl = bInstr->getDebugLoc();
7359 MachineOperand& destOper = bInstr->getOperand(0);
7360 MachineOperand* argOpers[2 + X86AddrNumOperands];
7361 int numArgs = bInstr->getNumOperands() - 1;
7362 for (int i=0; i < numArgs; ++i)
7363 argOpers[i] = &bInstr->getOperand(i+1);
7365 // x86 address has 4 operands: base, index, scale, and displacement
7366 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7367 int valArgIndx = lastAddrIndx + 1;
7369 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7370 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7371 for (int i=0; i <= lastAddrIndx; ++i)
7372 (*MIB).addOperand(*argOpers[i]);
7374 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7376 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7381 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7382 assert((argOpers[valArgIndx]->isReg() ||
7383 argOpers[valArgIndx]->isImm()) &&
7385 if (argOpers[valArgIndx]->isReg())
7386 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7388 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7390 (*MIB).addOperand(*argOpers[valArgIndx]);
7392 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7395 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7396 for (int i=0; i <= lastAddrIndx; ++i)
7397 (*MIB).addOperand(*argOpers[i]);
7399 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7400 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7401 bInstr->memoperands_end());
7403 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7407 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7409 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7413 // private utility function: 64 bit atomics on 32 bit host.
7415 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7416 MachineBasicBlock *MBB,
7421 bool invSrc) const {
7422 // For the atomic bitwise operator, we generate
7423 // thisMBB (instructions are in pairs, except cmpxchg8b)
7424 // ld t1,t2 = [bitinstr.addr]
7426 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7427 // op t5, t6 <- out1, out2, [bitinstr.val]
7428 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7429 // mov ECX, EBX <- t5, t6
7430 // mov EAX, EDX <- t1, t2
7431 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7432 // mov t3, t4 <- EAX, EDX
7434 // result in out1, out2
7435 // fallthrough -->nextMBB
7437 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7438 const unsigned LoadOpc = X86::MOV32rm;
7439 const unsigned copyOpc = X86::MOV32rr;
7440 const unsigned NotOpc = X86::NOT32r;
7441 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7442 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7443 MachineFunction::iterator MBBIter = MBB;
7446 /// First build the CFG
7447 MachineFunction *F = MBB->getParent();
7448 MachineBasicBlock *thisMBB = MBB;
7449 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7450 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7451 F->insert(MBBIter, newMBB);
7452 F->insert(MBBIter, nextMBB);
7454 // Move all successors to thisMBB to nextMBB
7455 nextMBB->transferSuccessors(thisMBB);
7457 // Update thisMBB to fall through to newMBB
7458 thisMBB->addSuccessor(newMBB);
7460 // newMBB jumps to itself and fall through to nextMBB
7461 newMBB->addSuccessor(nextMBB);
7462 newMBB->addSuccessor(newMBB);
7464 DebugLoc dl = bInstr->getDebugLoc();
7465 // Insert instructions into newMBB based on incoming instruction
7466 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7467 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7468 "unexpected number of operands");
7469 MachineOperand& dest1Oper = bInstr->getOperand(0);
7470 MachineOperand& dest2Oper = bInstr->getOperand(1);
7471 MachineOperand* argOpers[2 + X86AddrNumOperands];
7472 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7473 argOpers[i] = &bInstr->getOperand(i+2);
7475 // x86 address has 4 operands: base, index, scale, and displacement
7476 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7478 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7479 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7480 for (int i=0; i <= lastAddrIndx; ++i)
7481 (*MIB).addOperand(*argOpers[i]);
7482 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7483 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7484 // add 4 to displacement.
7485 for (int i=0; i <= lastAddrIndx-2; ++i)
7486 (*MIB).addOperand(*argOpers[i]);
7487 MachineOperand newOp3 = *(argOpers[3]);
7489 newOp3.setImm(newOp3.getImm()+4);
7491 newOp3.setOffset(newOp3.getOffset()+4);
7492 (*MIB).addOperand(newOp3);
7493 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7495 // t3/4 are defined later, at the bottom of the loop
7496 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7497 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7498 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7499 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7500 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7501 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7503 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
7504 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
7506 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt1).addReg(t1);
7507 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt2).addReg(t2);
7513 int valArgIndx = lastAddrIndx + 1;
7514 assert((argOpers[valArgIndx]->isReg() ||
7515 argOpers[valArgIndx]->isImm()) &&
7517 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7518 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7519 if (argOpers[valArgIndx]->isReg())
7520 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7522 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7523 if (regOpcL != X86::MOV32rr)
7525 (*MIB).addOperand(*argOpers[valArgIndx]);
7526 assert(argOpers[valArgIndx + 1]->isReg() ==
7527 argOpers[valArgIndx]->isReg());
7528 assert(argOpers[valArgIndx + 1]->isImm() ==
7529 argOpers[valArgIndx]->isImm());
7530 if (argOpers[valArgIndx + 1]->isReg())
7531 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7533 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7534 if (regOpcH != X86::MOV32rr)
7536 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
7538 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
7540 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
7543 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
7545 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
7548 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
7549 for (int i=0; i <= lastAddrIndx; ++i)
7550 (*MIB).addOperand(*argOpers[i]);
7552 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7553 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7554 bInstr->memoperands_end());
7556 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
7557 MIB.addReg(X86::EAX);
7558 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
7559 MIB.addReg(X86::EDX);
7562 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7564 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7568 // private utility function
7570 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
7571 MachineBasicBlock *MBB,
7572 unsigned cmovOpc) const {
7573 // For the atomic min/max operator, we generate
7576 // ld t1 = [min/max.addr]
7577 // mov t2 = [min/max.val]
7579 // cmov[cond] t2 = t1
7581 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7583 // fallthrough -->nextMBB
7585 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7586 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7587 MachineFunction::iterator MBBIter = MBB;
7590 /// First build the CFG
7591 MachineFunction *F = MBB->getParent();
7592 MachineBasicBlock *thisMBB = MBB;
7593 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7594 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7595 F->insert(MBBIter, newMBB);
7596 F->insert(MBBIter, nextMBB);
7598 // Move all successors of thisMBB to nextMBB
7599 nextMBB->transferSuccessors(thisMBB);
7601 // Update thisMBB to fall through to newMBB
7602 thisMBB->addSuccessor(newMBB);
7604 // newMBB jumps to newMBB and fall through to nextMBB
7605 newMBB->addSuccessor(nextMBB);
7606 newMBB->addSuccessor(newMBB);
7608 DebugLoc dl = mInstr->getDebugLoc();
7609 // Insert instructions into newMBB based on incoming instruction
7610 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7611 "unexpected number of operands");
7612 MachineOperand& destOper = mInstr->getOperand(0);
7613 MachineOperand* argOpers[2 + X86AddrNumOperands];
7614 int numArgs = mInstr->getNumOperands() - 1;
7615 for (int i=0; i < numArgs; ++i)
7616 argOpers[i] = &mInstr->getOperand(i+1);
7618 // x86 address has 4 operands: base, index, scale, and displacement
7619 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7620 int valArgIndx = lastAddrIndx + 1;
7622 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7623 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
7624 for (int i=0; i <= lastAddrIndx; ++i)
7625 (*MIB).addOperand(*argOpers[i]);
7627 // We only support register and immediate values
7628 assert((argOpers[valArgIndx]->isReg() ||
7629 argOpers[valArgIndx]->isImm()) &&
7632 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7633 if (argOpers[valArgIndx]->isReg())
7634 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7636 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7637 (*MIB).addOperand(*argOpers[valArgIndx]);
7639 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
7642 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
7647 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7648 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
7652 // Cmp and exchange if none has modified the memory location
7653 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
7654 for (int i=0; i <= lastAddrIndx; ++i)
7655 (*MIB).addOperand(*argOpers[i]);
7657 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7658 (*MIB).setMemRefs(mInstr->memoperands_begin(),
7659 mInstr->memoperands_end());
7661 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
7662 MIB.addReg(X86::EAX);
7665 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7667 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
7671 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
7672 // all of this code can be replaced with that in the .td file.
7674 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
7675 unsigned numArgs, bool memArg) const {
7677 MachineFunction *F = BB->getParent();
7678 DebugLoc dl = MI->getDebugLoc();
7679 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7683 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
7685 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
7687 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
7689 for (unsigned i = 0; i < numArgs; ++i) {
7690 MachineOperand &Op = MI->getOperand(i+1);
7692 if (!(Op.isReg() && Op.isImplicit()))
7696 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
7699 F->DeleteMachineInstr(MI);
7705 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
7707 MachineBasicBlock *MBB) const {
7708 // Emit code to save XMM registers to the stack. The ABI says that the
7709 // number of registers to save is given in %al, so it's theoretically
7710 // possible to do an indirect jump trick to avoid saving all of them,
7711 // however this code takes a simpler approach and just executes all
7712 // of the stores if %al is non-zero. It's less code, and it's probably
7713 // easier on the hardware branch predictor, and stores aren't all that
7714 // expensive anyway.
7716 // Create the new basic blocks. One block contains all the XMM stores,
7717 // and one block is the final destination regardless of whether any
7718 // stores were performed.
7719 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7720 MachineFunction *F = MBB->getParent();
7721 MachineFrameInfo *MFI = F->getFrameInfo();
7722 MachineFunction::iterator MBBIter = MBB;
7724 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
7725 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
7726 F->insert(MBBIter, XMMSaveMBB);
7727 F->insert(MBBIter, EndMBB);
7730 // Move any original successors of MBB to the end block.
7731 EndMBB->transferSuccessors(MBB);
7732 // The original block will now fall through to the XMM save block.
7733 MBB->addSuccessor(XMMSaveMBB);
7734 // The XMMSaveMBB will fall through to the end block.
7735 XMMSaveMBB->addSuccessor(EndMBB);
7737 // Now add the instructions.
7738 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7739 DebugLoc DL = MI->getDebugLoc();
7741 unsigned CountReg = MI->getOperand(0).getReg();
7742 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
7743 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
7745 if (!Subtarget->isTargetWin64()) {
7746 // If %al is 0, branch around the XMM save block.
7747 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
7748 BuildMI(MBB, DL, TII->get(X86::JE)).addMBB(EndMBB);
7749 MBB->addSuccessor(EndMBB);
7752 // In the XMM save block, save all the XMM argument registers.
7753 const Value *SV = MFI->isFixedObjectIndex(RegSaveFrameIndex)
7754 ? PseudoSourceValue::getFixedStack(RegSaveFrameIndex)
7755 : PseudoSourceValue::getStack();
7756 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
7757 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
7758 MachineMemOperand *MMO =
7759 F->getMachineMemOperand(SV, MachineMemOperand::MOStore, Offset,
7760 /*Size=*/16, /*Align=*/16);
7761 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
7762 .addFrameIndex(RegSaveFrameIndex)
7763 .addImm(/*Scale=*/1)
7764 .addReg(/*IndexReg=*/0)
7765 .addImm(/*Disp=*/Offset)
7766 .addReg(/*Segment=*/0)
7767 .addReg(MI->getOperand(i).getReg())
7768 .addMemOperand(MMO);
7771 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7777 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
7778 MachineBasicBlock *BB,
7779 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
7780 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7781 DebugLoc DL = MI->getDebugLoc();
7783 // To "insert" a SELECT_CC instruction, we actually have to insert the
7784 // diamond control-flow pattern. The incoming instruction knows the
7785 // destination vreg to set, the condition code register to branch on, the
7786 // true/false values to select between, and a branch opcode to use.
7787 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7788 MachineFunction::iterator It = BB;
7794 // cmpTY ccX, r1, r2
7796 // fallthrough --> copy0MBB
7797 MachineBasicBlock *thisMBB = BB;
7798 MachineFunction *F = BB->getParent();
7799 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7800 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7802 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
7803 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
7804 F->insert(It, copy0MBB);
7805 F->insert(It, sinkMBB);
7806 // Update machine-CFG edges by first adding all successors of the current
7807 // block to the new block which will contain the Phi node for the select.
7808 // Also inform sdisel of the edge changes.
7809 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
7810 E = BB->succ_end(); I != E; ++I) {
7811 EM->insert(std::make_pair(*I, sinkMBB));
7812 sinkMBB->addSuccessor(*I);
7814 // Next, remove all successors of the current block, and add the true
7815 // and fallthrough blocks as its successors.
7816 while (!BB->succ_empty())
7817 BB->removeSuccessor(BB->succ_begin());
7818 // Add the true and fallthrough blocks as its successors.
7819 BB->addSuccessor(copy0MBB);
7820 BB->addSuccessor(sinkMBB);
7823 // %FalseValue = ...
7824 // # fallthrough to sinkMBB
7827 // Update machine-CFG edges
7828 BB->addSuccessor(sinkMBB);
7831 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7834 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
7835 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7836 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7838 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7844 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7845 MachineBasicBlock *BB,
7846 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
7847 switch (MI->getOpcode()) {
7848 default: assert(false && "Unexpected instr type to insert");
7850 case X86::CMOV_V1I64:
7851 case X86::CMOV_FR32:
7852 case X86::CMOV_FR64:
7853 case X86::CMOV_V4F32:
7854 case X86::CMOV_V2F64:
7855 case X86::CMOV_V2I64:
7856 return EmitLoweredSelect(MI, BB, EM);
7858 case X86::FP32_TO_INT16_IN_MEM:
7859 case X86::FP32_TO_INT32_IN_MEM:
7860 case X86::FP32_TO_INT64_IN_MEM:
7861 case X86::FP64_TO_INT16_IN_MEM:
7862 case X86::FP64_TO_INT32_IN_MEM:
7863 case X86::FP64_TO_INT64_IN_MEM:
7864 case X86::FP80_TO_INT16_IN_MEM:
7865 case X86::FP80_TO_INT32_IN_MEM:
7866 case X86::FP80_TO_INT64_IN_MEM: {
7867 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7868 DebugLoc DL = MI->getDebugLoc();
7870 // Change the floating point control register to use "round towards zero"
7871 // mode when truncating to an integer value.
7872 MachineFunction *F = BB->getParent();
7873 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
7874 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
7876 // Load the old value of the high byte of the control word...
7878 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
7879 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
7882 // Set the high part to be round to zero...
7883 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
7886 // Reload the modified control word now...
7887 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
7889 // Restore the memory image of control word to original value
7890 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
7893 // Get the X86 opcode to use.
7895 switch (MI->getOpcode()) {
7896 default: llvm_unreachable("illegal opcode!");
7897 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
7898 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
7899 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
7900 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
7901 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
7902 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
7903 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
7904 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
7905 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
7909 MachineOperand &Op = MI->getOperand(0);
7911 AM.BaseType = X86AddressMode::RegBase;
7912 AM.Base.Reg = Op.getReg();
7914 AM.BaseType = X86AddressMode::FrameIndexBase;
7915 AM.Base.FrameIndex = Op.getIndex();
7917 Op = MI->getOperand(1);
7919 AM.Scale = Op.getImm();
7920 Op = MI->getOperand(2);
7922 AM.IndexReg = Op.getImm();
7923 Op = MI->getOperand(3);
7924 if (Op.isGlobal()) {
7925 AM.GV = Op.getGlobal();
7927 AM.Disp = Op.getImm();
7929 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
7930 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
7932 // Reload the original control word now.
7933 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
7935 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7938 // String/text processing lowering.
7939 case X86::PCMPISTRM128REG:
7940 return EmitPCMP(MI, BB, 3, false /* in-mem */);
7941 case X86::PCMPISTRM128MEM:
7942 return EmitPCMP(MI, BB, 3, true /* in-mem */);
7943 case X86::PCMPESTRM128REG:
7944 return EmitPCMP(MI, BB, 5, false /* in mem */);
7945 case X86::PCMPESTRM128MEM:
7946 return EmitPCMP(MI, BB, 5, true /* in mem */);
7949 case X86::ATOMAND32:
7950 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7951 X86::AND32ri, X86::MOV32rm,
7952 X86::LCMPXCHG32, X86::MOV32rr,
7953 X86::NOT32r, X86::EAX,
7954 X86::GR32RegisterClass);
7956 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
7957 X86::OR32ri, X86::MOV32rm,
7958 X86::LCMPXCHG32, X86::MOV32rr,
7959 X86::NOT32r, X86::EAX,
7960 X86::GR32RegisterClass);
7961 case X86::ATOMXOR32:
7962 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
7963 X86::XOR32ri, X86::MOV32rm,
7964 X86::LCMPXCHG32, X86::MOV32rr,
7965 X86::NOT32r, X86::EAX,
7966 X86::GR32RegisterClass);
7967 case X86::ATOMNAND32:
7968 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7969 X86::AND32ri, X86::MOV32rm,
7970 X86::LCMPXCHG32, X86::MOV32rr,
7971 X86::NOT32r, X86::EAX,
7972 X86::GR32RegisterClass, true);
7973 case X86::ATOMMIN32:
7974 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
7975 case X86::ATOMMAX32:
7976 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
7977 case X86::ATOMUMIN32:
7978 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
7979 case X86::ATOMUMAX32:
7980 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
7982 case X86::ATOMAND16:
7983 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7984 X86::AND16ri, X86::MOV16rm,
7985 X86::LCMPXCHG16, X86::MOV16rr,
7986 X86::NOT16r, X86::AX,
7987 X86::GR16RegisterClass);
7989 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
7990 X86::OR16ri, X86::MOV16rm,
7991 X86::LCMPXCHG16, X86::MOV16rr,
7992 X86::NOT16r, X86::AX,
7993 X86::GR16RegisterClass);
7994 case X86::ATOMXOR16:
7995 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
7996 X86::XOR16ri, X86::MOV16rm,
7997 X86::LCMPXCHG16, X86::MOV16rr,
7998 X86::NOT16r, X86::AX,
7999 X86::GR16RegisterClass);
8000 case X86::ATOMNAND16:
8001 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8002 X86::AND16ri, X86::MOV16rm,
8003 X86::LCMPXCHG16, X86::MOV16rr,
8004 X86::NOT16r, X86::AX,
8005 X86::GR16RegisterClass, true);
8006 case X86::ATOMMIN16:
8007 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8008 case X86::ATOMMAX16:
8009 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8010 case X86::ATOMUMIN16:
8011 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8012 case X86::ATOMUMAX16:
8013 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8016 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8017 X86::AND8ri, X86::MOV8rm,
8018 X86::LCMPXCHG8, X86::MOV8rr,
8019 X86::NOT8r, X86::AL,
8020 X86::GR8RegisterClass);
8022 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8023 X86::OR8ri, X86::MOV8rm,
8024 X86::LCMPXCHG8, X86::MOV8rr,
8025 X86::NOT8r, X86::AL,
8026 X86::GR8RegisterClass);
8028 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8029 X86::XOR8ri, X86::MOV8rm,
8030 X86::LCMPXCHG8, X86::MOV8rr,
8031 X86::NOT8r, X86::AL,
8032 X86::GR8RegisterClass);
8033 case X86::ATOMNAND8:
8034 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8035 X86::AND8ri, X86::MOV8rm,
8036 X86::LCMPXCHG8, X86::MOV8rr,
8037 X86::NOT8r, X86::AL,
8038 X86::GR8RegisterClass, true);
8039 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8040 // This group is for 64-bit host.
8041 case X86::ATOMAND64:
8042 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8043 X86::AND64ri32, X86::MOV64rm,
8044 X86::LCMPXCHG64, X86::MOV64rr,
8045 X86::NOT64r, X86::RAX,
8046 X86::GR64RegisterClass);
8048 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8049 X86::OR64ri32, X86::MOV64rm,
8050 X86::LCMPXCHG64, X86::MOV64rr,
8051 X86::NOT64r, X86::RAX,
8052 X86::GR64RegisterClass);
8053 case X86::ATOMXOR64:
8054 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8055 X86::XOR64ri32, X86::MOV64rm,
8056 X86::LCMPXCHG64, X86::MOV64rr,
8057 X86::NOT64r, X86::RAX,
8058 X86::GR64RegisterClass);
8059 case X86::ATOMNAND64:
8060 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8061 X86::AND64ri32, X86::MOV64rm,
8062 X86::LCMPXCHG64, X86::MOV64rr,
8063 X86::NOT64r, X86::RAX,
8064 X86::GR64RegisterClass, true);
8065 case X86::ATOMMIN64:
8066 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8067 case X86::ATOMMAX64:
8068 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8069 case X86::ATOMUMIN64:
8070 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8071 case X86::ATOMUMAX64:
8072 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8074 // This group does 64-bit operations on a 32-bit host.
8075 case X86::ATOMAND6432:
8076 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8077 X86::AND32rr, X86::AND32rr,
8078 X86::AND32ri, X86::AND32ri,
8080 case X86::ATOMOR6432:
8081 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8082 X86::OR32rr, X86::OR32rr,
8083 X86::OR32ri, X86::OR32ri,
8085 case X86::ATOMXOR6432:
8086 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8087 X86::XOR32rr, X86::XOR32rr,
8088 X86::XOR32ri, X86::XOR32ri,
8090 case X86::ATOMNAND6432:
8091 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8092 X86::AND32rr, X86::AND32rr,
8093 X86::AND32ri, X86::AND32ri,
8095 case X86::ATOMADD6432:
8096 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8097 X86::ADD32rr, X86::ADC32rr,
8098 X86::ADD32ri, X86::ADC32ri,
8100 case X86::ATOMSUB6432:
8101 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8102 X86::SUB32rr, X86::SBB32rr,
8103 X86::SUB32ri, X86::SBB32ri,
8105 case X86::ATOMSWAP6432:
8106 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8107 X86::MOV32rr, X86::MOV32rr,
8108 X86::MOV32ri, X86::MOV32ri,
8110 case X86::VASTART_SAVE_XMM_REGS:
8111 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8115 //===----------------------------------------------------------------------===//
8116 // X86 Optimization Hooks
8117 //===----------------------------------------------------------------------===//
8119 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8123 const SelectionDAG &DAG,
8124 unsigned Depth) const {
8125 unsigned Opc = Op.getOpcode();
8126 assert((Opc >= ISD::BUILTIN_OP_END ||
8127 Opc == ISD::INTRINSIC_WO_CHAIN ||
8128 Opc == ISD::INTRINSIC_W_CHAIN ||
8129 Opc == ISD::INTRINSIC_VOID) &&
8130 "Should use MaskedValueIsZero if you don't know whether Op"
8131 " is a target node!");
8133 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8145 // These nodes' second result is a boolean.
8146 if (Op.getResNo() == 0)
8150 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8151 Mask.getBitWidth() - 1);
8156 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8157 /// node is a GlobalAddress + offset.
8158 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8159 GlobalValue* &GA, int64_t &Offset) const{
8160 if (N->getOpcode() == X86ISD::Wrapper) {
8161 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8162 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8163 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8167 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8170 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
8171 const TargetLowering &TLI) {
8174 if (TLI.isGAPlusOffset(Base, GV, Offset))
8175 return (GV->getAlignment() >= N && (Offset % N) == 0);
8176 // DAG combine handles the stack object case.
8180 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8181 EVT EltVT, LoadSDNode *&LDBase,
8182 unsigned &LastLoadedElt,
8183 SelectionDAG &DAG, MachineFrameInfo *MFI,
8184 const TargetLowering &TLI) {
8186 LastLoadedElt = -1U;
8187 for (unsigned i = 0; i < NumElems; ++i) {
8188 if (N->getMaskElt(i) < 0) {
8194 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8195 if (!Elt.getNode() ||
8196 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8199 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8201 LDBase = cast<LoadSDNode>(Elt.getNode());
8205 if (Elt.getOpcode() == ISD::UNDEF)
8208 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8209 if (!TLI.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i, MFI))
8216 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8217 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8218 /// if the load addresses are consecutive, non-overlapping, and in the right
8219 /// order. In the case of v2i64, it will see if it can rewrite the
8220 /// shuffle to be an appropriate build vector so it can take advantage of
8221 // performBuildVectorCombine.
8222 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8223 const TargetLowering &TLI) {
8224 DebugLoc dl = N->getDebugLoc();
8225 EVT VT = N->getValueType(0);
8226 EVT EltVT = VT.getVectorElementType();
8227 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8228 unsigned NumElems = VT.getVectorNumElements();
8230 if (VT.getSizeInBits() != 128)
8233 // Try to combine a vector_shuffle into a 128-bit load.
8234 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8235 LoadSDNode *LD = NULL;
8236 unsigned LastLoadedElt;
8237 if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
8241 if (LastLoadedElt == NumElems - 1) {
8242 if (isBaseAlignmentOfN(16, LD->getBasePtr().getNode(), TLI))
8243 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8244 LD->getSrcValue(), LD->getSrcValueOffset(),
8246 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8247 LD->getSrcValue(), LD->getSrcValueOffset(),
8248 LD->isVolatile(), LD->getAlignment());
8249 } else if (NumElems == 4 && LastLoadedElt == 1) {
8250 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8251 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8252 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8253 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8258 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8259 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8260 const X86Subtarget *Subtarget) {
8261 DebugLoc DL = N->getDebugLoc();
8262 SDValue Cond = N->getOperand(0);
8263 // Get the LHS/RHS of the select.
8264 SDValue LHS = N->getOperand(1);
8265 SDValue RHS = N->getOperand(2);
8267 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8268 // instructions have the peculiarity that if either operand is a NaN,
8269 // they chose what we call the RHS operand (and as such are not symmetric).
8270 // It happens that this matches the semantics of the common C idiom
8271 // x<y?x:y and related forms, so we can recognize these cases.
8272 if (Subtarget->hasSSE2() &&
8273 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8274 Cond.getOpcode() == ISD::SETCC) {
8275 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8277 unsigned Opcode = 0;
8278 // Check for x CC y ? x : y.
8279 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8283 // This can be a min if we can prove that at least one of the operands
8285 if (!FiniteOnlyFPMath()) {
8286 if (DAG.isKnownNeverNaN(RHS)) {
8287 // Put the potential NaN in the RHS so that SSE will preserve it.
8288 std::swap(LHS, RHS);
8289 } else if (!DAG.isKnownNeverNaN(LHS))
8292 Opcode = X86ISD::FMIN;
8295 // This can be a min if we can prove that at least one of the operands
8297 if (!FiniteOnlyFPMath()) {
8298 if (DAG.isKnownNeverNaN(LHS)) {
8299 // Put the potential NaN in the RHS so that SSE will preserve it.
8300 std::swap(LHS, RHS);
8301 } else if (!DAG.isKnownNeverNaN(RHS))
8304 Opcode = X86ISD::FMIN;
8307 // This can be a min, but if either operand is a NaN we need it to
8308 // preserve the original LHS.
8309 std::swap(LHS, RHS);
8313 Opcode = X86ISD::FMIN;
8317 // This can be a max if we can prove that at least one of the operands
8319 if (!FiniteOnlyFPMath()) {
8320 if (DAG.isKnownNeverNaN(LHS)) {
8321 // Put the potential NaN in the RHS so that SSE will preserve it.
8322 std::swap(LHS, RHS);
8323 } else if (!DAG.isKnownNeverNaN(RHS))
8326 Opcode = X86ISD::FMAX;
8329 // This can be a max if we can prove that at least one of the operands
8331 if (!FiniteOnlyFPMath()) {
8332 if (DAG.isKnownNeverNaN(RHS)) {
8333 // Put the potential NaN in the RHS so that SSE will preserve it.
8334 std::swap(LHS, RHS);
8335 } else if (!DAG.isKnownNeverNaN(LHS))
8338 Opcode = X86ISD::FMAX;
8341 // This can be a max, but if either operand is a NaN we need it to
8342 // preserve the original LHS.
8343 std::swap(LHS, RHS);
8347 Opcode = X86ISD::FMAX;
8350 // Check for x CC y ? y : x -- a min/max with reversed arms.
8351 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8355 // This can be a min if we can prove that at least one of the operands
8357 if (!FiniteOnlyFPMath()) {
8358 if (DAG.isKnownNeverNaN(RHS)) {
8359 // Put the potential NaN in the RHS so that SSE will preserve it.
8360 std::swap(LHS, RHS);
8361 } else if (!DAG.isKnownNeverNaN(LHS))
8364 Opcode = X86ISD::FMIN;
8367 // This can be a min if we can prove that at least one of the operands
8369 if (!FiniteOnlyFPMath()) {
8370 if (DAG.isKnownNeverNaN(LHS)) {
8371 // Put the potential NaN in the RHS so that SSE will preserve it.
8372 std::swap(LHS, RHS);
8373 } else if (!DAG.isKnownNeverNaN(RHS))
8376 Opcode = X86ISD::FMIN;
8379 // This can be a min, but if either operand is a NaN we need it to
8380 // preserve the original LHS.
8381 std::swap(LHS, RHS);
8385 Opcode = X86ISD::FMIN;
8389 // This can be a max if we can prove that at least one of the operands
8391 if (!FiniteOnlyFPMath()) {
8392 if (DAG.isKnownNeverNaN(LHS)) {
8393 // Put the potential NaN in the RHS so that SSE will preserve it.
8394 std::swap(LHS, RHS);
8395 } else if (!DAG.isKnownNeverNaN(RHS))
8398 Opcode = X86ISD::FMAX;
8401 // This can be a max if we can prove that at least one of the operands
8403 if (!FiniteOnlyFPMath()) {
8404 if (DAG.isKnownNeverNaN(RHS)) {
8405 // Put the potential NaN in the RHS so that SSE will preserve it.
8406 std::swap(LHS, RHS);
8407 } else if (!DAG.isKnownNeverNaN(LHS))
8410 Opcode = X86ISD::FMAX;
8413 // This can be a max, but if either operand is a NaN we need it to
8414 // preserve the original LHS.
8415 std::swap(LHS, RHS);
8419 Opcode = X86ISD::FMAX;
8425 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8428 // If this is a select between two integer constants, try to do some
8430 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8431 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8432 // Don't do this for crazy integer types.
8433 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8434 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8435 // so that TrueC (the true value) is larger than FalseC.
8436 bool NeedsCondInvert = false;
8438 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8439 // Efficiently invertible.
8440 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8441 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8442 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8443 NeedsCondInvert = true;
8444 std::swap(TrueC, FalseC);
8447 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8448 if (FalseC->getAPIntValue() == 0 &&
8449 TrueC->getAPIntValue().isPowerOf2()) {
8450 if (NeedsCondInvert) // Invert the condition if needed.
8451 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8452 DAG.getConstant(1, Cond.getValueType()));
8454 // Zero extend the condition if needed.
8455 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8457 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8458 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8459 DAG.getConstant(ShAmt, MVT::i8));
8462 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8463 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8464 if (NeedsCondInvert) // Invert the condition if needed.
8465 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8466 DAG.getConstant(1, Cond.getValueType()));
8468 // Zero extend the condition if needed.
8469 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8470 FalseC->getValueType(0), Cond);
8471 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8472 SDValue(FalseC, 0));
8475 // Optimize cases that will turn into an LEA instruction. This requires
8476 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8477 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8478 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8479 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8481 bool isFastMultiplier = false;
8483 switch ((unsigned char)Diff) {
8485 case 1: // result = add base, cond
8486 case 2: // result = lea base( , cond*2)
8487 case 3: // result = lea base(cond, cond*2)
8488 case 4: // result = lea base( , cond*4)
8489 case 5: // result = lea base(cond, cond*4)
8490 case 8: // result = lea base( , cond*8)
8491 case 9: // result = lea base(cond, cond*8)
8492 isFastMultiplier = true;
8497 if (isFastMultiplier) {
8498 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8499 if (NeedsCondInvert) // Invert the condition if needed.
8500 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8501 DAG.getConstant(1, Cond.getValueType()));
8503 // Zero extend the condition if needed.
8504 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8506 // Scale the condition by the difference.
8508 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8509 DAG.getConstant(Diff, Cond.getValueType()));
8511 // Add the base if non-zero.
8512 if (FalseC->getAPIntValue() != 0)
8513 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8514 SDValue(FalseC, 0));
8524 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
8525 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8526 TargetLowering::DAGCombinerInfo &DCI) {
8527 DebugLoc DL = N->getDebugLoc();
8529 // If the flag operand isn't dead, don't touch this CMOV.
8530 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8533 // If this is a select between two integer constants, try to do some
8534 // optimizations. Note that the operands are ordered the opposite of SELECT
8536 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
8537 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8538 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
8539 // larger than FalseC (the false value).
8540 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
8542 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
8543 CC = X86::GetOppositeBranchCondition(CC);
8544 std::swap(TrueC, FalseC);
8547 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
8548 // This is efficient for any integer data type (including i8/i16) and
8550 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
8551 SDValue Cond = N->getOperand(3);
8552 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8553 DAG.getConstant(CC, MVT::i8), Cond);
8555 // Zero extend the condition if needed.
8556 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
8558 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8559 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
8560 DAG.getConstant(ShAmt, MVT::i8));
8561 if (N->getNumValues() == 2) // Dead flag value?
8562 return DCI.CombineTo(N, Cond, SDValue());
8566 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
8567 // for any integer data type, including i8/i16.
8568 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8569 SDValue Cond = N->getOperand(3);
8570 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8571 DAG.getConstant(CC, MVT::i8), Cond);
8573 // Zero extend the condition if needed.
8574 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8575 FalseC->getValueType(0), Cond);
8576 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8577 SDValue(FalseC, 0));
8579 if (N->getNumValues() == 2) // Dead flag value?
8580 return DCI.CombineTo(N, Cond, SDValue());
8584 // Optimize cases that will turn into an LEA instruction. This requires
8585 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8586 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8587 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8588 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8590 bool isFastMultiplier = false;
8592 switch ((unsigned char)Diff) {
8594 case 1: // result = add base, cond
8595 case 2: // result = lea base( , cond*2)
8596 case 3: // result = lea base(cond, cond*2)
8597 case 4: // result = lea base( , cond*4)
8598 case 5: // result = lea base(cond, cond*4)
8599 case 8: // result = lea base( , cond*8)
8600 case 9: // result = lea base(cond, cond*8)
8601 isFastMultiplier = true;
8606 if (isFastMultiplier) {
8607 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8608 SDValue Cond = N->getOperand(3);
8609 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8610 DAG.getConstant(CC, MVT::i8), Cond);
8611 // Zero extend the condition if needed.
8612 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8614 // Scale the condition by the difference.
8616 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8617 DAG.getConstant(Diff, Cond.getValueType()));
8619 // Add the base if non-zero.
8620 if (FalseC->getAPIntValue() != 0)
8621 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8622 SDValue(FalseC, 0));
8623 if (N->getNumValues() == 2) // Dead flag value?
8624 return DCI.CombineTo(N, Cond, SDValue());
8634 /// PerformMulCombine - Optimize a single multiply with constant into two
8635 /// in order to implement it with two cheaper instructions, e.g.
8636 /// LEA + SHL, LEA + LEA.
8637 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
8638 TargetLowering::DAGCombinerInfo &DCI) {
8639 if (DAG.getMachineFunction().
8640 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
8643 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8646 EVT VT = N->getValueType(0);
8650 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8653 uint64_t MulAmt = C->getZExtValue();
8654 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
8657 uint64_t MulAmt1 = 0;
8658 uint64_t MulAmt2 = 0;
8659 if ((MulAmt % 9) == 0) {
8661 MulAmt2 = MulAmt / 9;
8662 } else if ((MulAmt % 5) == 0) {
8664 MulAmt2 = MulAmt / 5;
8665 } else if ((MulAmt % 3) == 0) {
8667 MulAmt2 = MulAmt / 3;
8670 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
8671 DebugLoc DL = N->getDebugLoc();
8673 if (isPowerOf2_64(MulAmt2) &&
8674 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
8675 // If second multiplifer is pow2, issue it first. We want the multiply by
8676 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
8678 std::swap(MulAmt1, MulAmt2);
8681 if (isPowerOf2_64(MulAmt1))
8682 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
8683 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
8685 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
8686 DAG.getConstant(MulAmt1, VT));
8688 if (isPowerOf2_64(MulAmt2))
8689 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
8690 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
8692 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
8693 DAG.getConstant(MulAmt2, VT));
8695 // Do not add new nodes to DAG combiner worklist.
8696 DCI.CombineTo(N, NewMul, false);
8702 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
8704 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
8705 const X86Subtarget *Subtarget) {
8706 // On X86 with SSE2 support, we can transform this to a vector shift if
8707 // all elements are shifted by the same amount. We can't do this in legalize
8708 // because the a constant vector is typically transformed to a constant pool
8709 // so we have no knowledge of the shift amount.
8710 if (!Subtarget->hasSSE2())
8713 EVT VT = N->getValueType(0);
8714 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
8717 SDValue ShAmtOp = N->getOperand(1);
8718 EVT EltVT = VT.getVectorElementType();
8719 DebugLoc DL = N->getDebugLoc();
8720 SDValue BaseShAmt = SDValue();
8721 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
8722 unsigned NumElts = VT.getVectorNumElements();
8724 for (; i != NumElts; ++i) {
8725 SDValue Arg = ShAmtOp.getOperand(i);
8726 if (Arg.getOpcode() == ISD::UNDEF) continue;
8730 for (; i != NumElts; ++i) {
8731 SDValue Arg = ShAmtOp.getOperand(i);
8732 if (Arg.getOpcode() == ISD::UNDEF) continue;
8733 if (Arg != BaseShAmt) {
8737 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
8738 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
8739 SDValue InVec = ShAmtOp.getOperand(0);
8740 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
8741 unsigned NumElts = InVec.getValueType().getVectorNumElements();
8743 for (; i != NumElts; ++i) {
8744 SDValue Arg = InVec.getOperand(i);
8745 if (Arg.getOpcode() == ISD::UNDEF) continue;
8749 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
8750 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
8751 unsigned SplatIdx = cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
8752 if (C->getZExtValue() == SplatIdx)
8753 BaseShAmt = InVec.getOperand(1);
8756 if (BaseShAmt.getNode() == 0)
8757 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
8758 DAG.getIntPtrConstant(0));
8762 // The shift amount is an i32.
8763 if (EltVT.bitsGT(MVT::i32))
8764 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
8765 else if (EltVT.bitsLT(MVT::i32))
8766 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
8768 // The shift amount is identical so we can do a vector shift.
8769 SDValue ValOp = N->getOperand(0);
8770 switch (N->getOpcode()) {
8772 llvm_unreachable("Unknown shift opcode!");
8775 if (VT == MVT::v2i64)
8776 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8777 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8779 if (VT == MVT::v4i32)
8780 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8781 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8783 if (VT == MVT::v8i16)
8784 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8785 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8789 if (VT == MVT::v4i32)
8790 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8791 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
8793 if (VT == MVT::v8i16)
8794 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8795 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
8799 if (VT == MVT::v2i64)
8800 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8801 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8803 if (VT == MVT::v4i32)
8804 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8805 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
8807 if (VT == MVT::v8i16)
8808 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8809 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
8816 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
8817 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
8818 const X86Subtarget *Subtarget) {
8819 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
8820 // the FP state in cases where an emms may be missing.
8821 // A preferable solution to the general problem is to figure out the right
8822 // places to insert EMMS. This qualifies as a quick hack.
8824 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
8825 StoreSDNode *St = cast<StoreSDNode>(N);
8826 EVT VT = St->getValue().getValueType();
8827 if (VT.getSizeInBits() != 64)
8830 const Function *F = DAG.getMachineFunction().getFunction();
8831 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
8832 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
8833 && Subtarget->hasSSE2();
8834 if ((VT.isVector() ||
8835 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
8836 isa<LoadSDNode>(St->getValue()) &&
8837 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
8838 St->getChain().hasOneUse() && !St->isVolatile()) {
8839 SDNode* LdVal = St->getValue().getNode();
8841 int TokenFactorIndex = -1;
8842 SmallVector<SDValue, 8> Ops;
8843 SDNode* ChainVal = St->getChain().getNode();
8844 // Must be a store of a load. We currently handle two cases: the load
8845 // is a direct child, and it's under an intervening TokenFactor. It is
8846 // possible to dig deeper under nested TokenFactors.
8847 if (ChainVal == LdVal)
8848 Ld = cast<LoadSDNode>(St->getChain());
8849 else if (St->getValue().hasOneUse() &&
8850 ChainVal->getOpcode() == ISD::TokenFactor) {
8851 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
8852 if (ChainVal->getOperand(i).getNode() == LdVal) {
8853 TokenFactorIndex = i;
8854 Ld = cast<LoadSDNode>(St->getValue());
8856 Ops.push_back(ChainVal->getOperand(i));
8860 if (!Ld || !ISD::isNormalLoad(Ld))
8863 // If this is not the MMX case, i.e. we are just turning i64 load/store
8864 // into f64 load/store, avoid the transformation if there are multiple
8865 // uses of the loaded value.
8866 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
8869 DebugLoc LdDL = Ld->getDebugLoc();
8870 DebugLoc StDL = N->getDebugLoc();
8871 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
8872 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
8874 if (Subtarget->is64Bit() || F64IsLegal) {
8875 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
8876 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
8877 Ld->getBasePtr(), Ld->getSrcValue(),
8878 Ld->getSrcValueOffset(), Ld->isVolatile(),
8879 Ld->getAlignment());
8880 SDValue NewChain = NewLd.getValue(1);
8881 if (TokenFactorIndex != -1) {
8882 Ops.push_back(NewChain);
8883 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8886 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
8887 St->getSrcValue(), St->getSrcValueOffset(),
8888 St->isVolatile(), St->getAlignment());
8891 // Otherwise, lower to two pairs of 32-bit loads / stores.
8892 SDValue LoAddr = Ld->getBasePtr();
8893 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
8894 DAG.getConstant(4, MVT::i32));
8896 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
8897 Ld->getSrcValue(), Ld->getSrcValueOffset(),
8898 Ld->isVolatile(), Ld->getAlignment());
8899 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
8900 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
8902 MinAlign(Ld->getAlignment(), 4));
8904 SDValue NewChain = LoLd.getValue(1);
8905 if (TokenFactorIndex != -1) {
8906 Ops.push_back(LoLd);
8907 Ops.push_back(HiLd);
8908 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8912 LoAddr = St->getBasePtr();
8913 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
8914 DAG.getConstant(4, MVT::i32));
8916 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
8917 St->getSrcValue(), St->getSrcValueOffset(),
8918 St->isVolatile(), St->getAlignment());
8919 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
8921 St->getSrcValueOffset() + 4,
8923 MinAlign(St->getAlignment(), 4));
8924 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
8929 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
8930 /// X86ISD::FXOR nodes.
8931 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
8932 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
8933 // F[X]OR(0.0, x) -> x
8934 // F[X]OR(x, 0.0) -> x
8935 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8936 if (C->getValueAPF().isPosZero())
8937 return N->getOperand(1);
8938 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8939 if (C->getValueAPF().isPosZero())
8940 return N->getOperand(0);
8944 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
8945 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
8946 // FAND(0.0, x) -> 0.0
8947 // FAND(x, 0.0) -> 0.0
8948 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8949 if (C->getValueAPF().isPosZero())
8950 return N->getOperand(0);
8951 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8952 if (C->getValueAPF().isPosZero())
8953 return N->getOperand(1);
8957 static SDValue PerformBTCombine(SDNode *N,
8959 TargetLowering::DAGCombinerInfo &DCI) {
8960 // BT ignores high bits in the bit index operand.
8961 SDValue Op1 = N->getOperand(1);
8962 if (Op1.hasOneUse()) {
8963 unsigned BitWidth = Op1.getValueSizeInBits();
8964 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
8965 APInt KnownZero, KnownOne;
8966 TargetLowering::TargetLoweringOpt TLO(DAG);
8967 TargetLowering &TLI = DAG.getTargetLoweringInfo();
8968 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
8969 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
8970 DCI.CommitTargetLoweringOpt(TLO);
8975 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
8976 SDValue Op = N->getOperand(0);
8977 if (Op.getOpcode() == ISD::BIT_CONVERT)
8978 Op = Op.getOperand(0);
8979 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
8980 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
8981 VT.getVectorElementType().getSizeInBits() ==
8982 OpVT.getVectorElementType().getSizeInBits()) {
8983 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
8988 // On X86 and X86-64, atomic operations are lowered to locked instructions.
8989 // Locked instructions, in turn, have implicit fence semantics (all memory
8990 // operations are flushed before issuing the locked instruction, and the
8991 // are not buffered), so we can fold away the common pattern of
8992 // fence-atomic-fence.
8993 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
8994 SDValue atomic = N->getOperand(0);
8995 switch (atomic.getOpcode()) {
8996 case ISD::ATOMIC_CMP_SWAP:
8997 case ISD::ATOMIC_SWAP:
8998 case ISD::ATOMIC_LOAD_ADD:
8999 case ISD::ATOMIC_LOAD_SUB:
9000 case ISD::ATOMIC_LOAD_AND:
9001 case ISD::ATOMIC_LOAD_OR:
9002 case ISD::ATOMIC_LOAD_XOR:
9003 case ISD::ATOMIC_LOAD_NAND:
9004 case ISD::ATOMIC_LOAD_MIN:
9005 case ISD::ATOMIC_LOAD_MAX:
9006 case ISD::ATOMIC_LOAD_UMIN:
9007 case ISD::ATOMIC_LOAD_UMAX:
9013 SDValue fence = atomic.getOperand(0);
9014 if (fence.getOpcode() != ISD::MEMBARRIER)
9017 switch (atomic.getOpcode()) {
9018 case ISD::ATOMIC_CMP_SWAP:
9019 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9020 atomic.getOperand(1), atomic.getOperand(2),
9021 atomic.getOperand(3));
9022 case ISD::ATOMIC_SWAP:
9023 case ISD::ATOMIC_LOAD_ADD:
9024 case ISD::ATOMIC_LOAD_SUB:
9025 case ISD::ATOMIC_LOAD_AND:
9026 case ISD::ATOMIC_LOAD_OR:
9027 case ISD::ATOMIC_LOAD_XOR:
9028 case ISD::ATOMIC_LOAD_NAND:
9029 case ISD::ATOMIC_LOAD_MIN:
9030 case ISD::ATOMIC_LOAD_MAX:
9031 case ISD::ATOMIC_LOAD_UMIN:
9032 case ISD::ATOMIC_LOAD_UMAX:
9033 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9034 atomic.getOperand(1), atomic.getOperand(2));
9040 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9041 DAGCombinerInfo &DCI) const {
9042 SelectionDAG &DAG = DCI.DAG;
9043 switch (N->getOpcode()) {
9045 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9046 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9047 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9048 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9051 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9052 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9054 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9055 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9056 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9057 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9058 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9064 //===----------------------------------------------------------------------===//
9065 // X86 Inline Assembly Support
9066 //===----------------------------------------------------------------------===//
9068 static bool LowerToBSwap(CallInst *CI) {
9069 // FIXME: this should verify that we are targetting a 486 or better. If not,
9070 // we will turn this bswap into something that will be lowered to logical ops
9071 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9072 // so don't worry about this.
9074 // Verify this is a simple bswap.
9075 if (CI->getNumOperands() != 2 ||
9076 CI->getType() != CI->getOperand(1)->getType() ||
9077 !CI->getType()->isInteger())
9080 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9081 if (!Ty || Ty->getBitWidth() % 16 != 0)
9084 // Okay, we can do this xform, do so now.
9085 const Type *Tys[] = { Ty };
9086 Module *M = CI->getParent()->getParent()->getParent();
9087 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9089 Value *Op = CI->getOperand(1);
9090 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9092 CI->replaceAllUsesWith(Op);
9093 CI->eraseFromParent();
9097 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9098 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9099 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9101 std::string AsmStr = IA->getAsmString();
9103 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9104 std::vector<std::string> AsmPieces;
9105 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9107 switch (AsmPieces.size()) {
9108 default: return false;
9110 AsmStr = AsmPieces[0];
9112 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9115 if (AsmPieces.size() == 2 &&
9116 (AsmPieces[0] == "bswap" ||
9117 AsmPieces[0] == "bswapq" ||
9118 AsmPieces[0] == "bswapl") &&
9119 (AsmPieces[1] == "$0" ||
9120 AsmPieces[1] == "${0:q}")) {
9121 // No need to check constraints, nothing other than the equivalent of
9122 // "=r,0" would be valid here.
9123 return LowerToBSwap(CI);
9125 // rorw $$8, ${0:w} --> llvm.bswap.i16
9126 if (CI->getType() == Type::getInt16Ty(CI->getContext()) &&
9127 AsmPieces.size() == 3 &&
9128 AsmPieces[0] == "rorw" &&
9129 AsmPieces[1] == "$$8," &&
9130 AsmPieces[2] == "${0:w}" &&
9131 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
9132 return LowerToBSwap(CI);
9136 if (CI->getType() == Type::getInt64Ty(CI->getContext()) &&
9137 Constraints.size() >= 2 &&
9138 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9139 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9140 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9141 std::vector<std::string> Words;
9142 SplitString(AsmPieces[0], Words, " \t");
9143 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9145 SplitString(AsmPieces[1], Words, " \t");
9146 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9148 SplitString(AsmPieces[2], Words, " \t,");
9149 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9150 Words[2] == "%edx") {
9151 return LowerToBSwap(CI);
9163 /// getConstraintType - Given a constraint letter, return the type of
9164 /// constraint it is for this target.
9165 X86TargetLowering::ConstraintType
9166 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
9167 if (Constraint.size() == 1) {
9168 switch (Constraint[0]) {
9180 return C_RegisterClass;
9188 return TargetLowering::getConstraintType(Constraint);
9191 /// LowerXConstraint - try to replace an X constraint, which matches anything,
9192 /// with another that has more specific requirements based on the type of the
9193 /// corresponding operand.
9194 const char *X86TargetLowering::
9195 LowerXConstraint(EVT ConstraintVT) const {
9196 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
9197 // 'f' like normal targets.
9198 if (ConstraintVT.isFloatingPoint()) {
9199 if (Subtarget->hasSSE2())
9201 if (Subtarget->hasSSE1())
9205 return TargetLowering::LowerXConstraint(ConstraintVT);
9208 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9209 /// vector. If it is invalid, don't add anything to Ops.
9210 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9213 std::vector<SDValue>&Ops,
9214 SelectionDAG &DAG) const {
9215 SDValue Result(0, 0);
9217 switch (Constraint) {
9220 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9221 if (C->getZExtValue() <= 31) {
9222 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9228 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9229 if (C->getZExtValue() <= 63) {
9230 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9236 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9237 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9238 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9244 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9245 if (C->getZExtValue() <= 255) {
9246 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9252 // 32-bit signed value
9253 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9254 const ConstantInt *CI = C->getConstantIntValue();
9255 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9256 C->getSExtValue())) {
9257 // Widen to 64 bits here to get it sign extended.
9258 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9261 // FIXME gcc accepts some relocatable values here too, but only in certain
9262 // memory models; it's complicated.
9267 // 32-bit unsigned value
9268 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9269 const ConstantInt *CI = C->getConstantIntValue();
9270 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9271 C->getZExtValue())) {
9272 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9276 // FIXME gcc accepts some relocatable values here too, but only in certain
9277 // memory models; it's complicated.
9281 // Literal immediates are always ok.
9282 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
9283 // Widen to 64 bits here to get it sign extended.
9284 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
9288 // If we are in non-pic codegen mode, we allow the address of a global (with
9289 // an optional displacement) to be used with 'i'.
9290 GlobalAddressSDNode *GA = 0;
9293 // Match either (GA), (GA+C), (GA+C1+C2), etc.
9295 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
9296 Offset += GA->getOffset();
9298 } else if (Op.getOpcode() == ISD::ADD) {
9299 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9300 Offset += C->getZExtValue();
9301 Op = Op.getOperand(0);
9304 } else if (Op.getOpcode() == ISD::SUB) {
9305 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9306 Offset += -C->getZExtValue();
9307 Op = Op.getOperand(0);
9312 // Otherwise, this isn't something we can handle, reject it.
9316 GlobalValue *GV = GA->getGlobal();
9317 // If we require an extra load to get this address, as in PIC mode, we
9319 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
9320 getTargetMachine())))
9324 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
9326 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
9332 if (Result.getNode()) {
9333 Ops.push_back(Result);
9336 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
9340 std::vector<unsigned> X86TargetLowering::
9341 getRegClassForInlineAsmConstraint(const std::string &Constraint,
9343 if (Constraint.size() == 1) {
9344 // FIXME: not handling fp-stack yet!
9345 switch (Constraint[0]) { // GCC X86 Constraint Letters
9346 default: break; // Unknown constraint letter
9347 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
9348 if (Subtarget->is64Bit()) {
9350 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
9351 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
9352 X86::R10D,X86::R11D,X86::R12D,
9353 X86::R13D,X86::R14D,X86::R15D,
9354 X86::EBP, X86::ESP, 0);
9355 else if (VT == MVT::i16)
9356 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
9357 X86::SI, X86::DI, X86::R8W,X86::R9W,
9358 X86::R10W,X86::R11W,X86::R12W,
9359 X86::R13W,X86::R14W,X86::R15W,
9360 X86::BP, X86::SP, 0);
9361 else if (VT == MVT::i8)
9362 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
9363 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
9364 X86::R10B,X86::R11B,X86::R12B,
9365 X86::R13B,X86::R14B,X86::R15B,
9366 X86::BPL, X86::SPL, 0);
9368 else if (VT == MVT::i64)
9369 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
9370 X86::RSI, X86::RDI, X86::R8, X86::R9,
9371 X86::R10, X86::R11, X86::R12,
9372 X86::R13, X86::R14, X86::R15,
9373 X86::RBP, X86::RSP, 0);
9377 // 32-bit fallthrough
9380 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
9381 else if (VT == MVT::i16)
9382 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
9383 else if (VT == MVT::i8)
9384 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
9385 else if (VT == MVT::i64)
9386 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
9391 return std::vector<unsigned>();
9394 std::pair<unsigned, const TargetRegisterClass*>
9395 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9397 // First, see if this is a constraint that directly corresponds to an LLVM
9399 if (Constraint.size() == 1) {
9400 // GCC Constraint Letters
9401 switch (Constraint[0]) {
9403 case 'r': // GENERAL_REGS
9404 case 'l': // INDEX_REGS
9406 return std::make_pair(0U, X86::GR8RegisterClass);
9408 return std::make_pair(0U, X86::GR16RegisterClass);
9409 if (VT == MVT::i32 || !Subtarget->is64Bit())
9410 return std::make_pair(0U, X86::GR32RegisterClass);
9411 return std::make_pair(0U, X86::GR64RegisterClass);
9412 case 'R': // LEGACY_REGS
9414 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
9416 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
9417 if (VT == MVT::i32 || !Subtarget->is64Bit())
9418 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
9419 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
9420 case 'f': // FP Stack registers.
9421 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
9422 // value to the correct fpstack register class.
9423 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
9424 return std::make_pair(0U, X86::RFP32RegisterClass);
9425 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
9426 return std::make_pair(0U, X86::RFP64RegisterClass);
9427 return std::make_pair(0U, X86::RFP80RegisterClass);
9428 case 'y': // MMX_REGS if MMX allowed.
9429 if (!Subtarget->hasMMX()) break;
9430 return std::make_pair(0U, X86::VR64RegisterClass);
9431 case 'Y': // SSE_REGS if SSE2 allowed
9432 if (!Subtarget->hasSSE2()) break;
9434 case 'x': // SSE_REGS if SSE1 allowed
9435 if (!Subtarget->hasSSE1()) break;
9437 switch (VT.getSimpleVT().SimpleTy) {
9439 // Scalar SSE types.
9442 return std::make_pair(0U, X86::FR32RegisterClass);
9445 return std::make_pair(0U, X86::FR64RegisterClass);
9453 return std::make_pair(0U, X86::VR128RegisterClass);
9459 // Use the default implementation in TargetLowering to convert the register
9460 // constraint into a member of a register class.
9461 std::pair<unsigned, const TargetRegisterClass*> Res;
9462 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9464 // Not found as a standard register?
9465 if (Res.second == 0) {
9466 // Map st(0) -> st(7) -> ST0
9467 if (Constraint.size() == 7 && Constraint[0] == '{' &&
9468 tolower(Constraint[1]) == 's' &&
9469 tolower(Constraint[2]) == 't' &&
9470 Constraint[3] == '(' &&
9471 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
9472 Constraint[5] == ')' &&
9473 Constraint[6] == '}') {
9475 Res.first = X86::ST0+Constraint[4]-'0';
9476 Res.second = X86::RFP80RegisterClass;
9480 // GCC allows "st(0)" to be called just plain "st".
9481 if (StringsEqualNoCase("{st}", Constraint)) {
9482 Res.first = X86::ST0;
9483 Res.second = X86::RFP80RegisterClass;
9488 if (StringsEqualNoCase("{flags}", Constraint)) {
9489 Res.first = X86::EFLAGS;
9490 Res.second = X86::CCRRegisterClass;
9494 // 'A' means EAX + EDX.
9495 if (Constraint == "A") {
9496 Res.first = X86::EAX;
9497 Res.second = X86::GR32_ADRegisterClass;
9503 // Otherwise, check to see if this is a register class of the wrong value
9504 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
9505 // turn into {ax},{dx}.
9506 if (Res.second->hasType(VT))
9507 return Res; // Correct type already, nothing to do.
9509 // All of the single-register GCC register classes map their values onto
9510 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
9511 // really want an 8-bit or 32-bit register, map to the appropriate register
9512 // class and return the appropriate register.
9513 if (Res.second == X86::GR16RegisterClass) {
9514 if (VT == MVT::i8) {
9515 unsigned DestReg = 0;
9516 switch (Res.first) {
9518 case X86::AX: DestReg = X86::AL; break;
9519 case X86::DX: DestReg = X86::DL; break;
9520 case X86::CX: DestReg = X86::CL; break;
9521 case X86::BX: DestReg = X86::BL; break;
9524 Res.first = DestReg;
9525 Res.second = X86::GR8RegisterClass;
9527 } else if (VT == MVT::i32) {
9528 unsigned DestReg = 0;
9529 switch (Res.first) {
9531 case X86::AX: DestReg = X86::EAX; break;
9532 case X86::DX: DestReg = X86::EDX; break;
9533 case X86::CX: DestReg = X86::ECX; break;
9534 case X86::BX: DestReg = X86::EBX; break;
9535 case X86::SI: DestReg = X86::ESI; break;
9536 case X86::DI: DestReg = X86::EDI; break;
9537 case X86::BP: DestReg = X86::EBP; break;
9538 case X86::SP: DestReg = X86::ESP; break;
9541 Res.first = DestReg;
9542 Res.second = X86::GR32RegisterClass;
9544 } else if (VT == MVT::i64) {
9545 unsigned DestReg = 0;
9546 switch (Res.first) {
9548 case X86::AX: DestReg = X86::RAX; break;
9549 case X86::DX: DestReg = X86::RDX; break;
9550 case X86::CX: DestReg = X86::RCX; break;
9551 case X86::BX: DestReg = X86::RBX; break;
9552 case X86::SI: DestReg = X86::RSI; break;
9553 case X86::DI: DestReg = X86::RDI; break;
9554 case X86::BP: DestReg = X86::RBP; break;
9555 case X86::SP: DestReg = X86::RSP; break;
9558 Res.first = DestReg;
9559 Res.second = X86::GR64RegisterClass;
9562 } else if (Res.second == X86::FR32RegisterClass ||
9563 Res.second == X86::FR64RegisterClass ||
9564 Res.second == X86::VR128RegisterClass) {
9565 // Handle references to XMM physical registers that got mapped into the
9566 // wrong class. This can happen with constraints like {xmm0} where the
9567 // target independent register mapper will just pick the first match it can
9568 // find, ignoring the required type.
9570 Res.second = X86::FR32RegisterClass;
9571 else if (VT == MVT::f64)
9572 Res.second = X86::FR64RegisterClass;
9573 else if (X86::VR128RegisterClass->hasType(VT))
9574 Res.second = X86::VR128RegisterClass;
9580 //===----------------------------------------------------------------------===//
9581 // X86 Widen vector type
9582 //===----------------------------------------------------------------------===//
9584 /// getWidenVectorType: given a vector type, returns the type to widen
9585 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
9586 /// If there is no vector type that we want to widen to, returns MVT::Other
9587 /// When and where to widen is target dependent based on the cost of
9588 /// scalarizing vs using the wider vector type.
9590 EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
9591 assert(VT.isVector());
9592 if (isTypeLegal(VT))
9595 // TODO: In computeRegisterProperty, we can compute the list of legal vector
9596 // type based on element type. This would speed up our search (though
9597 // it may not be worth it since the size of the list is relatively
9599 EVT EltVT = VT.getVectorElementType();
9600 unsigned NElts = VT.getVectorNumElements();
9602 // On X86, it make sense to widen any vector wider than 1
9606 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
9607 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
9608 EVT SVT = (MVT::SimpleValueType)nVT;
9610 if (isTypeLegal(SVT) &&
9611 SVT.getVectorElementType() == EltVT &&
9612 SVT.getVectorNumElements() > NElts)