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 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
332 if (Subtarget->is64Bit()) {
333 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
334 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
335 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
336 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
337 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
339 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
340 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
341 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
342 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
343 if (Subtarget->is64Bit()) {
344 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
345 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
346 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
349 if (Subtarget->hasSSE1())
350 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
352 if (!Subtarget->hasSSE2())
353 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
355 // Expand certain atomics
356 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
357 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
358 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
359 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
361 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
362 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
363 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
364 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
366 if (!Subtarget->is64Bit()) {
367 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
368 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
369 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
370 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
371 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
372 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
373 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
376 // Use the default ISD::DBG_STOPPOINT.
377 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
378 // FIXME - use subtarget debug flags
379 if (!Subtarget->isTargetDarwin() &&
380 !Subtarget->isTargetELF() &&
381 !Subtarget->isTargetCygMing()) {
382 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
383 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
386 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
387 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
388 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
389 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
390 if (Subtarget->is64Bit()) {
391 setExceptionPointerRegister(X86::RAX);
392 setExceptionSelectorRegister(X86::RDX);
394 setExceptionPointerRegister(X86::EAX);
395 setExceptionSelectorRegister(X86::EDX);
397 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
398 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
400 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
402 setOperationAction(ISD::TRAP, MVT::Other, Legal);
404 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
405 setOperationAction(ISD::VASTART , MVT::Other, Custom);
406 setOperationAction(ISD::VAEND , MVT::Other, Expand);
407 if (Subtarget->is64Bit()) {
408 setOperationAction(ISD::VAARG , MVT::Other, Custom);
409 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
411 setOperationAction(ISD::VAARG , MVT::Other, Expand);
412 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
415 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
416 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
417 if (Subtarget->is64Bit())
418 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
419 if (Subtarget->isTargetCygMing())
420 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
422 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
424 if (!UseSoftFloat && X86ScalarSSEf64) {
425 // f32 and f64 use SSE.
426 // Set up the FP register classes.
427 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
428 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
430 // Use ANDPD to simulate FABS.
431 setOperationAction(ISD::FABS , MVT::f64, Custom);
432 setOperationAction(ISD::FABS , MVT::f32, Custom);
434 // Use XORP to simulate FNEG.
435 setOperationAction(ISD::FNEG , MVT::f64, Custom);
436 setOperationAction(ISD::FNEG , MVT::f32, Custom);
438 // Use ANDPD and ORPD to simulate FCOPYSIGN.
439 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
440 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
442 // We don't support sin/cos/fmod
443 setOperationAction(ISD::FSIN , MVT::f64, Expand);
444 setOperationAction(ISD::FCOS , MVT::f64, Expand);
445 setOperationAction(ISD::FSIN , MVT::f32, Expand);
446 setOperationAction(ISD::FCOS , MVT::f32, Expand);
448 // Expand FP immediates into loads from the stack, except for the special
450 addLegalFPImmediate(APFloat(+0.0)); // xorpd
451 addLegalFPImmediate(APFloat(+0.0f)); // xorps
452 } else if (!UseSoftFloat && X86ScalarSSEf32) {
453 // Use SSE for f32, x87 for f64.
454 // Set up the FP register classes.
455 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
456 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
458 // Use ANDPS to simulate FABS.
459 setOperationAction(ISD::FABS , MVT::f32, Custom);
461 // Use XORP to simulate FNEG.
462 setOperationAction(ISD::FNEG , MVT::f32, Custom);
464 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
466 // Use ANDPS and ORPS to simulate FCOPYSIGN.
467 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
468 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
470 // We don't support sin/cos/fmod
471 setOperationAction(ISD::FSIN , MVT::f32, Expand);
472 setOperationAction(ISD::FCOS , MVT::f32, Expand);
474 // Special cases we handle for FP constants.
475 addLegalFPImmediate(APFloat(+0.0f)); // xorps
476 addLegalFPImmediate(APFloat(+0.0)); // FLD0
477 addLegalFPImmediate(APFloat(+1.0)); // FLD1
478 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
479 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
482 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
483 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
485 } else if (!UseSoftFloat) {
486 // f32 and f64 in x87.
487 // Set up the FP register classes.
488 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
489 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
491 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
492 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
493 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
494 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
497 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
498 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
500 addLegalFPImmediate(APFloat(+0.0)); // FLD0
501 addLegalFPImmediate(APFloat(+1.0)); // FLD1
502 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
503 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
504 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
505 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
506 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
507 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
510 // Long double always uses X87.
512 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
513 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
514 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
517 APFloat TmpFlt(+0.0);
518 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
520 addLegalFPImmediate(TmpFlt); // FLD0
522 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
523 APFloat TmpFlt2(+1.0);
524 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
526 addLegalFPImmediate(TmpFlt2); // FLD1
527 TmpFlt2.changeSign();
528 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
532 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
533 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
537 // Always use a library call for pow.
538 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
539 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
540 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
542 setOperationAction(ISD::FLOG, MVT::f80, Expand);
543 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
544 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
545 setOperationAction(ISD::FEXP, MVT::f80, Expand);
546 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
548 // First set operation action for all vector types to either promote
549 // (for widening) or expand (for scalarization). Then we will selectively
550 // turn on ones that can be effectively codegen'd.
551 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
552 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
553 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
568 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
569 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
599 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
603 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
604 // with -msoft-float, disable use of MMX as well.
605 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
606 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
607 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
608 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
609 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
610 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
612 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
613 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
614 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
615 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
617 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
618 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
619 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
620 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
622 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
623 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
625 setOperationAction(ISD::AND, MVT::v8i8, Promote);
626 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
627 setOperationAction(ISD::AND, MVT::v4i16, Promote);
628 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
629 setOperationAction(ISD::AND, MVT::v2i32, Promote);
630 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
631 setOperationAction(ISD::AND, MVT::v1i64, Legal);
633 setOperationAction(ISD::OR, MVT::v8i8, Promote);
634 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
635 setOperationAction(ISD::OR, MVT::v4i16, Promote);
636 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
637 setOperationAction(ISD::OR, MVT::v2i32, Promote);
638 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
639 setOperationAction(ISD::OR, MVT::v1i64, Legal);
641 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
642 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
643 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
644 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
645 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
646 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
647 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
649 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
650 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
651 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
652 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
653 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
654 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
655 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
656 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
657 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
659 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
660 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
661 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
662 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
663 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
665 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
666 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
667 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
668 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
670 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
671 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
672 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
673 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
675 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
677 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Expand);
678 setOperationAction(ISD::TRUNCATE, MVT::v8i8, Expand);
679 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
680 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
681 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
682 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
683 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
684 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
685 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
688 if (!UseSoftFloat && Subtarget->hasSSE1()) {
689 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
691 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
692 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
693 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
694 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
695 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
696 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
697 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
698 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
699 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
700 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
701 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
702 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
705 if (!UseSoftFloat && Subtarget->hasSSE2()) {
706 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
708 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
709 // registers cannot be used even for integer operations.
710 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
711 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
712 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
713 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
715 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
716 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
717 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
718 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
719 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
720 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
721 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
722 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
723 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
724 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
725 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
726 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
727 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
728 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
729 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
730 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
732 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
733 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
734 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
735 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
737 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
738 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
739 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
740 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
741 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
743 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
744 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
745 EVT VT = (MVT::SimpleValueType)i;
746 // Do not attempt to custom lower non-power-of-2 vectors
747 if (!isPowerOf2_32(VT.getVectorNumElements()))
749 // Do not attempt to custom lower non-128-bit vectors
750 if (!VT.is128BitVector())
752 setOperationAction(ISD::BUILD_VECTOR,
753 VT.getSimpleVT().SimpleTy, Custom);
754 setOperationAction(ISD::VECTOR_SHUFFLE,
755 VT.getSimpleVT().SimpleTy, Custom);
756 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
757 VT.getSimpleVT().SimpleTy, Custom);
760 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
761 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
762 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
763 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
764 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
765 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
767 if (Subtarget->is64Bit()) {
768 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
769 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
772 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
773 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
774 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
777 // Do not attempt to promote non-128-bit vectors
778 if (!VT.is128BitVector()) {
781 setOperationAction(ISD::AND, SVT, Promote);
782 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
783 setOperationAction(ISD::OR, SVT, Promote);
784 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
785 setOperationAction(ISD::XOR, SVT, Promote);
786 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
787 setOperationAction(ISD::LOAD, SVT, Promote);
788 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
789 setOperationAction(ISD::SELECT, SVT, Promote);
790 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
793 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
795 // Custom lower v2i64 and v2f64 selects.
796 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
797 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
798 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
799 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
801 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
802 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
803 if (!DisableMMX && Subtarget->hasMMX()) {
804 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
805 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
809 if (Subtarget->hasSSE41()) {
810 // FIXME: Do we need to handle scalar-to-vector here?
811 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
813 // i8 and i16 vectors are custom , because the source register and source
814 // source memory operand types are not the same width. f32 vectors are
815 // custom since the immediate controlling the insert encodes additional
817 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
818 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
819 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
820 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
822 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
823 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
824 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
825 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
827 if (Subtarget->is64Bit()) {
828 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
829 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
833 if (Subtarget->hasSSE42()) {
834 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
837 if (!UseSoftFloat && Subtarget->hasAVX()) {
838 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
839 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
840 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
841 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
843 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
844 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
845 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
846 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
847 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
848 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
849 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
850 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
851 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
852 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
853 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
854 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
855 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
856 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
857 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
859 // Operations to consider commented out -v16i16 v32i8
860 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
861 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
862 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
863 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
864 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
865 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
866 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
867 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
868 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
869 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
870 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
871 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
872 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
873 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
875 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
876 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
877 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
878 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
880 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
881 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
882 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
883 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
884 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
886 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
887 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
888 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
889 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
890 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
891 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
894 // Not sure we want to do this since there are no 256-bit integer
897 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
898 // This includes 256-bit vectors
899 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
900 EVT VT = (MVT::SimpleValueType)i;
902 // Do not attempt to custom lower non-power-of-2 vectors
903 if (!isPowerOf2_32(VT.getVectorNumElements()))
906 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
907 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
908 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
911 if (Subtarget->is64Bit()) {
912 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
913 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
918 // Not sure we want to do this since there are no 256-bit integer
921 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
922 // Including 256-bit vectors
923 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
924 EVT VT = (MVT::SimpleValueType)i;
926 if (!VT.is256BitVector()) {
929 setOperationAction(ISD::AND, VT, Promote);
930 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
931 setOperationAction(ISD::OR, VT, Promote);
932 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
933 setOperationAction(ISD::XOR, VT, Promote);
934 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
935 setOperationAction(ISD::LOAD, VT, Promote);
936 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
937 setOperationAction(ISD::SELECT, VT, Promote);
938 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
941 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
945 // We want to custom lower some of our intrinsics.
946 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
948 // Add/Sub/Mul with overflow operations are custom lowered.
949 setOperationAction(ISD::SADDO, MVT::i32, Custom);
950 setOperationAction(ISD::SADDO, MVT::i64, Custom);
951 setOperationAction(ISD::UADDO, MVT::i32, Custom);
952 setOperationAction(ISD::UADDO, MVT::i64, Custom);
953 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
954 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
955 setOperationAction(ISD::USUBO, MVT::i32, Custom);
956 setOperationAction(ISD::USUBO, MVT::i64, Custom);
957 setOperationAction(ISD::SMULO, MVT::i32, Custom);
958 setOperationAction(ISD::SMULO, MVT::i64, Custom);
960 if (!Subtarget->is64Bit()) {
961 // These libcalls are not available in 32-bit.
962 setLibcallName(RTLIB::SHL_I128, 0);
963 setLibcallName(RTLIB::SRL_I128, 0);
964 setLibcallName(RTLIB::SRA_I128, 0);
967 // We have target-specific dag combine patterns for the following nodes:
968 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
969 setTargetDAGCombine(ISD::BUILD_VECTOR);
970 setTargetDAGCombine(ISD::SELECT);
971 setTargetDAGCombine(ISD::SHL);
972 setTargetDAGCombine(ISD::SRA);
973 setTargetDAGCombine(ISD::SRL);
974 setTargetDAGCombine(ISD::STORE);
975 setTargetDAGCombine(ISD::MEMBARRIER);
976 if (Subtarget->is64Bit())
977 setTargetDAGCombine(ISD::MUL);
979 computeRegisterProperties();
981 // FIXME: These should be based on subtarget info. Plus, the values should
982 // be smaller when we are in optimizing for size mode.
983 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
984 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
985 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
986 setPrefLoopAlignment(16);
987 benefitFromCodePlacementOpt = true;
991 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
996 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
997 /// the desired ByVal argument alignment.
998 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1001 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1002 if (VTy->getBitWidth() == 128)
1004 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1005 unsigned EltAlign = 0;
1006 getMaxByValAlign(ATy->getElementType(), EltAlign);
1007 if (EltAlign > MaxAlign)
1008 MaxAlign = EltAlign;
1009 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1010 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1011 unsigned EltAlign = 0;
1012 getMaxByValAlign(STy->getElementType(i), EltAlign);
1013 if (EltAlign > MaxAlign)
1014 MaxAlign = EltAlign;
1022 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1023 /// function arguments in the caller parameter area. For X86, aggregates
1024 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1025 /// are at 4-byte boundaries.
1026 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1027 if (Subtarget->is64Bit()) {
1028 // Max of 8 and alignment of type.
1029 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1036 if (Subtarget->hasSSE1())
1037 getMaxByValAlign(Ty, Align);
1041 /// getOptimalMemOpType - Returns the target specific optimal type for load
1042 /// and store operations as a result of memset, memcpy, and memmove
1043 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
1046 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1047 bool isSrcConst, bool isSrcStr,
1048 SelectionDAG &DAG) const {
1049 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1050 // linux. This is because the stack realignment code can't handle certain
1051 // cases like PR2962. This should be removed when PR2962 is fixed.
1052 const Function *F = DAG.getMachineFunction().getFunction();
1053 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1054 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1055 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1057 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1060 if (Subtarget->is64Bit() && Size >= 8)
1065 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1067 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1068 SelectionDAG &DAG) const {
1069 if (usesGlobalOffsetTable())
1070 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
1071 if (!Subtarget->is64Bit())
1072 // This doesn't have DebugLoc associated with it, but is not really the
1073 // same as a Register.
1074 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1079 /// getFunctionAlignment - Return the Log2 alignment of this function.
1080 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1081 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1084 //===----------------------------------------------------------------------===//
1085 // Return Value Calling Convention Implementation
1086 //===----------------------------------------------------------------------===//
1088 #include "X86GenCallingConv.inc"
1091 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1092 const SmallVectorImpl<EVT> &OutTys,
1093 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1094 SelectionDAG &DAG) {
1095 SmallVector<CCValAssign, 16> RVLocs;
1096 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1097 RVLocs, *DAG.getContext());
1098 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1102 X86TargetLowering::LowerReturn(SDValue Chain,
1103 CallingConv::ID CallConv, bool isVarArg,
1104 const SmallVectorImpl<ISD::OutputArg> &Outs,
1105 DebugLoc dl, SelectionDAG &DAG) {
1107 SmallVector<CCValAssign, 16> RVLocs;
1108 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1109 RVLocs, *DAG.getContext());
1110 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1112 // If this is the first return lowered for this function, add the regs to the
1113 // liveout set for the function.
1114 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1115 for (unsigned i = 0; i != RVLocs.size(); ++i)
1116 if (RVLocs[i].isRegLoc())
1117 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1122 SmallVector<SDValue, 6> RetOps;
1123 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1124 // Operand #1 = Bytes To Pop
1125 RetOps.push_back(DAG.getTargetConstant(getBytesToPopOnReturn(), MVT::i16));
1127 // Copy the result values into the output registers.
1128 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1129 CCValAssign &VA = RVLocs[i];
1130 assert(VA.isRegLoc() && "Can only return in registers!");
1131 SDValue ValToCopy = Outs[i].Val;
1133 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1134 // the RET instruction and handled by the FP Stackifier.
1135 if (VA.getLocReg() == X86::ST0 ||
1136 VA.getLocReg() == X86::ST1) {
1137 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1138 // change the value to the FP stack register class.
1139 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1140 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1141 RetOps.push_back(ValToCopy);
1142 // Don't emit a copytoreg.
1146 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1147 // which is returned in RAX / RDX.
1148 if (Subtarget->is64Bit()) {
1149 EVT ValVT = ValToCopy.getValueType();
1150 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1151 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1152 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1153 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1157 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1158 Flag = Chain.getValue(1);
1161 // The x86-64 ABI for returning structs by value requires that we copy
1162 // the sret argument into %rax for the return. We saved the argument into
1163 // a virtual register in the entry block, so now we copy the value out
1165 if (Subtarget->is64Bit() &&
1166 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1167 MachineFunction &MF = DAG.getMachineFunction();
1168 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1169 unsigned Reg = FuncInfo->getSRetReturnReg();
1171 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1172 FuncInfo->setSRetReturnReg(Reg);
1174 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1176 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1177 Flag = Chain.getValue(1);
1179 // RAX now acts like a return value.
1180 MF.getRegInfo().addLiveOut(X86::RAX);
1183 RetOps[0] = Chain; // Update chain.
1185 // Add the flag if we have it.
1187 RetOps.push_back(Flag);
1189 return DAG.getNode(X86ISD::RET_FLAG, dl,
1190 MVT::Other, &RetOps[0], RetOps.size());
1193 /// LowerCallResult - Lower the result values of a call into the
1194 /// appropriate copies out of appropriate physical registers.
1197 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1198 CallingConv::ID CallConv, bool isVarArg,
1199 const SmallVectorImpl<ISD::InputArg> &Ins,
1200 DebugLoc dl, SelectionDAG &DAG,
1201 SmallVectorImpl<SDValue> &InVals) {
1203 // Assign locations to each value returned by this call.
1204 SmallVector<CCValAssign, 16> RVLocs;
1205 bool Is64Bit = Subtarget->is64Bit();
1206 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1207 RVLocs, *DAG.getContext());
1208 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1210 // Copy all of the result registers out of their specified physreg.
1211 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1212 CCValAssign &VA = RVLocs[i];
1213 EVT CopyVT = VA.getValVT();
1215 // If this is x86-64, and we disabled SSE, we can't return FP values
1216 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1217 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1218 llvm_report_error("SSE register return with SSE disabled");
1221 // If this is a call to a function that returns an fp value on the floating
1222 // point stack, but where we prefer to use the value in xmm registers, copy
1223 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1224 if ((VA.getLocReg() == X86::ST0 ||
1225 VA.getLocReg() == X86::ST1) &&
1226 isScalarFPTypeInSSEReg(VA.getValVT())) {
1231 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1232 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1233 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1234 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1235 MVT::v2i64, InFlag).getValue(1);
1236 Val = Chain.getValue(0);
1237 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1238 Val, DAG.getConstant(0, MVT::i64));
1240 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1241 MVT::i64, InFlag).getValue(1);
1242 Val = Chain.getValue(0);
1244 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1246 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1247 CopyVT, InFlag).getValue(1);
1248 Val = Chain.getValue(0);
1250 InFlag = Chain.getValue(2);
1252 if (CopyVT != VA.getValVT()) {
1253 // Round the F80 the right size, which also moves to the appropriate xmm
1255 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1256 // This truncation won't change the value.
1257 DAG.getIntPtrConstant(1));
1260 InVals.push_back(Val);
1267 //===----------------------------------------------------------------------===//
1268 // C & StdCall & Fast Calling Convention implementation
1269 //===----------------------------------------------------------------------===//
1270 // StdCall calling convention seems to be standard for many Windows' API
1271 // routines and around. It differs from C calling convention just a little:
1272 // callee should clean up the stack, not caller. Symbols should be also
1273 // decorated in some fancy way :) It doesn't support any vector arguments.
1274 // For info on fast calling convention see Fast Calling Convention (tail call)
1275 // implementation LowerX86_32FastCCCallTo.
1277 /// CallIsStructReturn - Determines whether a call uses struct return
1279 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1283 return Outs[0].Flags.isSRet();
1286 /// ArgsAreStructReturn - Determines whether a function uses struct
1287 /// return semantics.
1289 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1293 return Ins[0].Flags.isSRet();
1296 /// IsCalleePop - Determines whether the callee is required to pop its
1297 /// own arguments. Callee pop is necessary to support tail calls.
1298 bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
1302 switch (CallingConv) {
1305 case CallingConv::X86_StdCall:
1306 return !Subtarget->is64Bit();
1307 case CallingConv::X86_FastCall:
1308 return !Subtarget->is64Bit();
1309 case CallingConv::Fast:
1310 return PerformTailCallOpt;
1314 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1315 /// given CallingConvention value.
1316 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1317 if (Subtarget->is64Bit()) {
1318 if (Subtarget->isTargetWin64())
1319 return CC_X86_Win64_C;
1324 if (CC == CallingConv::X86_FastCall)
1325 return CC_X86_32_FastCall;
1326 else if (CC == CallingConv::Fast)
1327 return CC_X86_32_FastCC;
1332 /// NameDecorationForCallConv - Selects the appropriate decoration to
1333 /// apply to a MachineFunction containing a given calling convention.
1335 X86TargetLowering::NameDecorationForCallConv(CallingConv::ID CallConv) {
1336 if (CallConv == CallingConv::X86_FastCall)
1338 else if (CallConv == CallingConv::X86_StdCall)
1344 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1345 /// by "Src" to address "Dst" with size and alignment information specified by
1346 /// the specific parameter attribute. The copy will be passed as a byval
1347 /// function parameter.
1349 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1350 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1352 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1353 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1354 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1358 X86TargetLowering::LowerMemArgument(SDValue Chain,
1359 CallingConv::ID CallConv,
1360 const SmallVectorImpl<ISD::InputArg> &Ins,
1361 DebugLoc dl, SelectionDAG &DAG,
1362 const CCValAssign &VA,
1363 MachineFrameInfo *MFI,
1366 // Create the nodes corresponding to a load from this parameter slot.
1367 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1368 bool AlwaysUseMutable = (CallConv==CallingConv::Fast) && PerformTailCallOpt;
1369 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1372 // If value is passed by pointer we have address passed instead of the value
1374 if (VA.getLocInfo() == CCValAssign::Indirect)
1375 ValVT = VA.getLocVT();
1377 ValVT = VA.getValVT();
1379 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1380 // changed with more analysis.
1381 // In case of tail call optimization mark all arguments mutable. Since they
1382 // could be overwritten by lowering of arguments in case of a tail call.
1383 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1384 VA.getLocMemOffset(), isImmutable, false);
1385 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1386 if (Flags.isByVal())
1388 return DAG.getLoad(ValVT, dl, Chain, FIN,
1389 PseudoSourceValue::getFixedStack(FI), 0);
1393 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1394 CallingConv::ID CallConv,
1396 const SmallVectorImpl<ISD::InputArg> &Ins,
1399 SmallVectorImpl<SDValue> &InVals) {
1401 MachineFunction &MF = DAG.getMachineFunction();
1402 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1404 const Function* Fn = MF.getFunction();
1405 if (Fn->hasExternalLinkage() &&
1406 Subtarget->isTargetCygMing() &&
1407 Fn->getName() == "main")
1408 FuncInfo->setForceFramePointer(true);
1410 // Decorate the function name.
1411 FuncInfo->setDecorationStyle(NameDecorationForCallConv(CallConv));
1413 MachineFrameInfo *MFI = MF.getFrameInfo();
1414 bool Is64Bit = Subtarget->is64Bit();
1415 bool IsWin64 = Subtarget->isTargetWin64();
1417 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1418 "Var args not supported with calling convention fastcc");
1420 // Assign locations to all of the incoming arguments.
1421 SmallVector<CCValAssign, 16> ArgLocs;
1422 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1423 ArgLocs, *DAG.getContext());
1424 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1426 unsigned LastVal = ~0U;
1428 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1429 CCValAssign &VA = ArgLocs[i];
1430 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1432 assert(VA.getValNo() != LastVal &&
1433 "Don't support value assigned to multiple locs yet");
1434 LastVal = VA.getValNo();
1436 if (VA.isRegLoc()) {
1437 EVT RegVT = VA.getLocVT();
1438 TargetRegisterClass *RC = NULL;
1439 if (RegVT == MVT::i32)
1440 RC = X86::GR32RegisterClass;
1441 else if (Is64Bit && RegVT == MVT::i64)
1442 RC = X86::GR64RegisterClass;
1443 else if (RegVT == MVT::f32)
1444 RC = X86::FR32RegisterClass;
1445 else if (RegVT == MVT::f64)
1446 RC = X86::FR64RegisterClass;
1447 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1448 RC = X86::VR128RegisterClass;
1449 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1450 RC = X86::VR64RegisterClass;
1452 llvm_unreachable("Unknown argument type!");
1454 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1455 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1457 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1458 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1460 if (VA.getLocInfo() == CCValAssign::SExt)
1461 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1462 DAG.getValueType(VA.getValVT()));
1463 else if (VA.getLocInfo() == CCValAssign::ZExt)
1464 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1465 DAG.getValueType(VA.getValVT()));
1466 else if (VA.getLocInfo() == CCValAssign::BCvt)
1467 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1469 if (VA.isExtInLoc()) {
1470 // Handle MMX values passed in XMM regs.
1471 if (RegVT.isVector()) {
1472 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1473 ArgValue, DAG.getConstant(0, MVT::i64));
1474 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1476 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1479 assert(VA.isMemLoc());
1480 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1483 // If value is passed via pointer - do a load.
1484 if (VA.getLocInfo() == CCValAssign::Indirect)
1485 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0);
1487 InVals.push_back(ArgValue);
1490 // The x86-64 ABI for returning structs by value requires that we copy
1491 // the sret argument into %rax for the return. Save the argument into
1492 // a virtual register so that we can access it from the return points.
1493 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1494 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1495 unsigned Reg = FuncInfo->getSRetReturnReg();
1497 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1498 FuncInfo->setSRetReturnReg(Reg);
1500 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1501 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1504 unsigned StackSize = CCInfo.getNextStackOffset();
1505 // align stack specially for tail calls
1506 if (PerformTailCallOpt && CallConv == CallingConv::Fast)
1507 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1509 // If the function takes variable number of arguments, make a frame index for
1510 // the start of the first vararg value... for expansion of llvm.va_start.
1512 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1513 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize, true, false);
1516 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1518 // FIXME: We should really autogenerate these arrays
1519 static const unsigned GPR64ArgRegsWin64[] = {
1520 X86::RCX, X86::RDX, X86::R8, X86::R9
1522 static const unsigned XMMArgRegsWin64[] = {
1523 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1525 static const unsigned GPR64ArgRegs64Bit[] = {
1526 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1528 static const unsigned XMMArgRegs64Bit[] = {
1529 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1530 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1532 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1535 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1536 GPR64ArgRegs = GPR64ArgRegsWin64;
1537 XMMArgRegs = XMMArgRegsWin64;
1539 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1540 GPR64ArgRegs = GPR64ArgRegs64Bit;
1541 XMMArgRegs = XMMArgRegs64Bit;
1543 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1545 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1548 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1549 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1550 "SSE register cannot be used when SSE is disabled!");
1551 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1552 "SSE register cannot be used when SSE is disabled!");
1553 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1554 // Kernel mode asks for SSE to be disabled, so don't push them
1556 TotalNumXMMRegs = 0;
1558 // For X86-64, if there are vararg parameters that are passed via
1559 // registers, then we must store them to their spots on the stack so they
1560 // may be loaded by deferencing the result of va_next.
1561 VarArgsGPOffset = NumIntRegs * 8;
1562 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1563 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1564 TotalNumXMMRegs * 16, 16,
1567 // Store the integer parameter registers.
1568 SmallVector<SDValue, 8> MemOps;
1569 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1570 unsigned Offset = VarArgsGPOffset;
1571 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1572 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1573 DAG.getIntPtrConstant(Offset));
1574 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1575 X86::GR64RegisterClass);
1576 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1578 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1579 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1581 MemOps.push_back(Store);
1585 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1586 // Now store the XMM (fp + vector) parameter registers.
1587 SmallVector<SDValue, 11> SaveXMMOps;
1588 SaveXMMOps.push_back(Chain);
1590 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1591 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1592 SaveXMMOps.push_back(ALVal);
1594 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1595 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1597 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1598 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1599 X86::VR128RegisterClass);
1600 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1601 SaveXMMOps.push_back(Val);
1603 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1605 &SaveXMMOps[0], SaveXMMOps.size()));
1608 if (!MemOps.empty())
1609 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1610 &MemOps[0], MemOps.size());
1614 // Some CCs need callee pop.
1615 if (IsCalleePop(isVarArg, CallConv)) {
1616 BytesToPopOnReturn = StackSize; // Callee pops everything.
1617 BytesCallerReserves = 0;
1619 BytesToPopOnReturn = 0; // Callee pops nothing.
1620 // If this is an sret function, the return should pop the hidden pointer.
1621 if (!Is64Bit && CallConv != CallingConv::Fast && ArgsAreStructReturn(Ins))
1622 BytesToPopOnReturn = 4;
1623 BytesCallerReserves = StackSize;
1627 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1628 if (CallConv == CallingConv::X86_FastCall)
1629 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1632 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1638 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1639 SDValue StackPtr, SDValue Arg,
1640 DebugLoc dl, SelectionDAG &DAG,
1641 const CCValAssign &VA,
1642 ISD::ArgFlagsTy Flags) {
1643 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1644 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1645 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1646 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1647 if (Flags.isByVal()) {
1648 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1650 return DAG.getStore(Chain, dl, Arg, PtrOff,
1651 PseudoSourceValue::getStack(), LocMemOffset);
1654 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1655 /// optimization is performed and it is required.
1657 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1658 SDValue &OutRetAddr,
1664 if (!IsTailCall || FPDiff==0) return Chain;
1666 // Adjust the Return address stack slot.
1667 EVT VT = getPointerTy();
1668 OutRetAddr = getReturnAddressFrameIndex(DAG);
1670 // Load the "old" Return address.
1671 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1672 return SDValue(OutRetAddr.getNode(), 1);
1675 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1676 /// optimization is performed and it is required (FPDiff!=0).
1678 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1679 SDValue Chain, SDValue RetAddrFrIdx,
1680 bool Is64Bit, int FPDiff, DebugLoc dl) {
1681 // Store the return address to the appropriate stack slot.
1682 if (!FPDiff) return Chain;
1683 // Calculate the new stack slot for the return address.
1684 int SlotSize = Is64Bit ? 8 : 4;
1685 int NewReturnAddrFI =
1686 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize,
1688 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1689 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1690 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1691 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1696 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1697 CallingConv::ID CallConv, bool isVarArg,
1699 const SmallVectorImpl<ISD::OutputArg> &Outs,
1700 const SmallVectorImpl<ISD::InputArg> &Ins,
1701 DebugLoc dl, SelectionDAG &DAG,
1702 SmallVectorImpl<SDValue> &InVals) {
1704 MachineFunction &MF = DAG.getMachineFunction();
1705 bool Is64Bit = Subtarget->is64Bit();
1706 bool IsStructRet = CallIsStructReturn(Outs);
1708 assert((!isTailCall ||
1709 (CallConv == CallingConv::Fast && PerformTailCallOpt)) &&
1710 "IsEligibleForTailCallOptimization missed a case!");
1711 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1712 "Var args not supported with calling convention fastcc");
1714 // Analyze operands of the call, assigning locations to each operand.
1715 SmallVector<CCValAssign, 16> ArgLocs;
1716 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1717 ArgLocs, *DAG.getContext());
1718 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1720 // Get a count of how many bytes are to be pushed on the stack.
1721 unsigned NumBytes = CCInfo.getNextStackOffset();
1722 if (PerformTailCallOpt && CallConv == CallingConv::Fast)
1723 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1727 // Lower arguments at fp - stackoffset + fpdiff.
1728 unsigned NumBytesCallerPushed =
1729 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1730 FPDiff = NumBytesCallerPushed - NumBytes;
1732 // Set the delta of movement of the returnaddr stackslot.
1733 // But only set if delta is greater than previous delta.
1734 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1735 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1738 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1740 SDValue RetAddrFrIdx;
1741 // Load return adress for tail calls.
1742 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall, Is64Bit,
1745 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1746 SmallVector<SDValue, 8> MemOpChains;
1749 // Walk the register/memloc assignments, inserting copies/loads. In the case
1750 // of tail call optimization arguments are handle later.
1751 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1752 CCValAssign &VA = ArgLocs[i];
1753 EVT RegVT = VA.getLocVT();
1754 SDValue Arg = Outs[i].Val;
1755 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1756 bool isByVal = Flags.isByVal();
1758 // Promote the value if needed.
1759 switch (VA.getLocInfo()) {
1760 default: llvm_unreachable("Unknown loc info!");
1761 case CCValAssign::Full: break;
1762 case CCValAssign::SExt:
1763 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1765 case CCValAssign::ZExt:
1766 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1768 case CCValAssign::AExt:
1769 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1770 // Special case: passing MMX values in XMM registers.
1771 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1772 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1773 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1775 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1777 case CCValAssign::BCvt:
1778 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1780 case CCValAssign::Indirect: {
1781 // Store the argument.
1782 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1783 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1784 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1785 PseudoSourceValue::getFixedStack(FI), 0);
1791 if (VA.isRegLoc()) {
1792 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1794 if (!isTailCall || (isTailCall && isByVal)) {
1795 assert(VA.isMemLoc());
1796 if (StackPtr.getNode() == 0)
1797 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1799 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1800 dl, DAG, VA, Flags));
1805 if (!MemOpChains.empty())
1806 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1807 &MemOpChains[0], MemOpChains.size());
1809 // Build a sequence of copy-to-reg nodes chained together with token chain
1810 // and flag operands which copy the outgoing args into registers.
1812 // Tail call byval lowering might overwrite argument registers so in case of
1813 // tail call optimization the copies to registers are lowered later.
1815 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1816 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1817 RegsToPass[i].second, InFlag);
1818 InFlag = Chain.getValue(1);
1822 if (Subtarget->isPICStyleGOT()) {
1823 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1826 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1827 DAG.getNode(X86ISD::GlobalBaseReg,
1828 DebugLoc::getUnknownLoc(),
1831 InFlag = Chain.getValue(1);
1833 // If we are tail calling and generating PIC/GOT style code load the
1834 // address of the callee into ECX. The value in ecx is used as target of
1835 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1836 // for tail calls on PIC/GOT architectures. Normally we would just put the
1837 // address of GOT into ebx and then call target@PLT. But for tail calls
1838 // ebx would be restored (since ebx is callee saved) before jumping to the
1841 // Note: The actual moving to ECX is done further down.
1842 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1843 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1844 !G->getGlobal()->hasProtectedVisibility())
1845 Callee = LowerGlobalAddress(Callee, DAG);
1846 else if (isa<ExternalSymbolSDNode>(Callee))
1847 Callee = LowerExternalSymbol(Callee, DAG);
1851 if (Is64Bit && isVarArg) {
1852 // From AMD64 ABI document:
1853 // For calls that may call functions that use varargs or stdargs
1854 // (prototype-less calls or calls to functions containing ellipsis (...) in
1855 // the declaration) %al is used as hidden argument to specify the number
1856 // of SSE registers used. The contents of %al do not need to match exactly
1857 // the number of registers, but must be an ubound on the number of SSE
1858 // registers used and is in the range 0 - 8 inclusive.
1860 // FIXME: Verify this on Win64
1861 // Count the number of XMM registers allocated.
1862 static const unsigned XMMArgRegs[] = {
1863 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1864 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1866 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1867 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1868 && "SSE registers cannot be used when SSE is disabled");
1870 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1871 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1872 InFlag = Chain.getValue(1);
1876 // For tail calls lower the arguments to the 'real' stack slot.
1878 // Force all the incoming stack arguments to be loaded from the stack
1879 // before any new outgoing arguments are stored to the stack, because the
1880 // outgoing stack slots may alias the incoming argument stack slots, and
1881 // the alias isn't otherwise explicit. This is slightly more conservative
1882 // than necessary, because it means that each store effectively depends
1883 // on every argument instead of just those arguments it would clobber.
1884 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1886 SmallVector<SDValue, 8> MemOpChains2;
1889 // Do not flag preceeding copytoreg stuff together with the following stuff.
1891 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1892 CCValAssign &VA = ArgLocs[i];
1893 if (!VA.isRegLoc()) {
1894 assert(VA.isMemLoc());
1895 SDValue Arg = Outs[i].Val;
1896 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1897 // Create frame index.
1898 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1899 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1900 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
1901 FIN = DAG.getFrameIndex(FI, getPointerTy());
1903 if (Flags.isByVal()) {
1904 // Copy relative to framepointer.
1905 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1906 if (StackPtr.getNode() == 0)
1907 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1909 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1911 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
1915 // Store relative to framepointer.
1916 MemOpChains2.push_back(
1917 DAG.getStore(ArgChain, dl, Arg, FIN,
1918 PseudoSourceValue::getFixedStack(FI), 0));
1923 if (!MemOpChains2.empty())
1924 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1925 &MemOpChains2[0], MemOpChains2.size());
1927 // Copy arguments to their registers.
1928 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1929 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1930 RegsToPass[i].second, InFlag);
1931 InFlag = Chain.getValue(1);
1935 // Store the return address to the appropriate stack slot.
1936 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1940 bool WasGlobalOrExternal = false;
1941 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
1942 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
1943 // In the 64-bit large code model, we have to make all calls
1944 // through a register, since the call instruction's 32-bit
1945 // pc-relative offset may not be large enough to hold the whole
1947 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1948 WasGlobalOrExternal = true;
1949 // If the callee is a GlobalAddress node (quite common, every direct call
1950 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
1953 // We should use extra load for direct calls to dllimported functions in
1955 GlobalValue *GV = G->getGlobal();
1956 if (!GV->hasDLLImportLinkage()) {
1957 unsigned char OpFlags = 0;
1959 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
1960 // external symbols most go through the PLT in PIC mode. If the symbol
1961 // has hidden or protected visibility, or if it is static or local, then
1962 // we don't need to use the PLT - we can directly call it.
1963 if (Subtarget->isTargetELF() &&
1964 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1965 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
1966 OpFlags = X86II::MO_PLT;
1967 } else if (Subtarget->isPICStyleStubAny() &&
1968 (GV->isDeclaration() || GV->isWeakForLinker()) &&
1969 Subtarget->getDarwinVers() < 9) {
1970 // PC-relative references to external symbols should go through $stub,
1971 // unless we're building with the leopard linker or later, which
1972 // automatically synthesizes these stubs.
1973 OpFlags = X86II::MO_DARWIN_STUB;
1976 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
1977 G->getOffset(), OpFlags);
1979 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1980 WasGlobalOrExternal = true;
1981 unsigned char OpFlags = 0;
1983 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
1984 // symbols should go through the PLT.
1985 if (Subtarget->isTargetELF() &&
1986 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1987 OpFlags = X86II::MO_PLT;
1988 } else if (Subtarget->isPICStyleStubAny() &&
1989 Subtarget->getDarwinVers() < 9) {
1990 // PC-relative references to external symbols should go through $stub,
1991 // unless we're building with the leopard linker or later, which
1992 // automatically synthesizes these stubs.
1993 OpFlags = X86II::MO_DARWIN_STUB;
1996 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2000 if (isTailCall && !WasGlobalOrExternal) {
2001 unsigned Opc = Is64Bit ? X86::R11 : X86::EAX;
2003 Chain = DAG.getCopyToReg(Chain, dl,
2004 DAG.getRegister(Opc, getPointerTy()),
2006 Callee = DAG.getRegister(Opc, getPointerTy());
2007 // Add register as live out.
2008 MF.getRegInfo().addLiveOut(Opc);
2011 // Returns a chain & a flag for retval copy to use.
2012 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2013 SmallVector<SDValue, 8> Ops;
2016 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2017 DAG.getIntPtrConstant(0, true), InFlag);
2018 InFlag = Chain.getValue(1);
2021 Ops.push_back(Chain);
2022 Ops.push_back(Callee);
2025 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2027 // Add argument registers to the end of the list so that they are known live
2029 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2030 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2031 RegsToPass[i].second.getValueType()));
2033 // Add an implicit use GOT pointer in EBX.
2034 if (!isTailCall && Subtarget->isPICStyleGOT())
2035 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2037 // Add an implicit use of AL for x86 vararg functions.
2038 if (Is64Bit && isVarArg)
2039 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2041 if (InFlag.getNode())
2042 Ops.push_back(InFlag);
2045 // If this is the first return lowered for this function, add the regs
2046 // to the liveout set for the function.
2047 if (MF.getRegInfo().liveout_empty()) {
2048 SmallVector<CCValAssign, 16> RVLocs;
2049 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2051 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2052 for (unsigned i = 0; i != RVLocs.size(); ++i)
2053 if (RVLocs[i].isRegLoc())
2054 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2057 assert(((Callee.getOpcode() == ISD::Register &&
2058 (cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
2059 cast<RegisterSDNode>(Callee)->getReg() == X86::R9)) ||
2060 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2061 Callee.getOpcode() == ISD::TargetGlobalAddress) &&
2062 "Expecting an global address, external symbol, or register");
2064 return DAG.getNode(X86ISD::TC_RETURN, dl,
2065 NodeTys, &Ops[0], Ops.size());
2068 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2069 InFlag = Chain.getValue(1);
2071 // Create the CALLSEQ_END node.
2072 unsigned NumBytesForCalleeToPush;
2073 if (IsCalleePop(isVarArg, CallConv))
2074 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2075 else if (!Is64Bit && CallConv != CallingConv::Fast && IsStructRet)
2076 // If this is is a call to a struct-return function, the callee
2077 // pops the hidden struct pointer, so we have to push it back.
2078 // This is common for Darwin/X86, Linux & Mingw32 targets.
2079 NumBytesForCalleeToPush = 4;
2081 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2083 // Returns a flag for retval copy to use.
2084 Chain = DAG.getCALLSEQ_END(Chain,
2085 DAG.getIntPtrConstant(NumBytes, true),
2086 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2089 InFlag = Chain.getValue(1);
2091 // Handle result values, copying them out of physregs into vregs that we
2093 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2094 Ins, dl, DAG, InVals);
2098 //===----------------------------------------------------------------------===//
2099 // Fast Calling Convention (tail call) implementation
2100 //===----------------------------------------------------------------------===//
2102 // Like std call, callee cleans arguments, convention except that ECX is
2103 // reserved for storing the tail called function address. Only 2 registers are
2104 // free for argument passing (inreg). Tail call optimization is performed
2106 // * tailcallopt is enabled
2107 // * caller/callee are fastcc
2108 // On X86_64 architecture with GOT-style position independent code only local
2109 // (within module) calls are supported at the moment.
2110 // To keep the stack aligned according to platform abi the function
2111 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2112 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2113 // If a tail called function callee has more arguments than the caller the
2114 // caller needs to make sure that there is room to move the RETADDR to. This is
2115 // achieved by reserving an area the size of the argument delta right after the
2116 // original REtADDR, but before the saved framepointer or the spilled registers
2117 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2129 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2130 /// for a 16 byte align requirement.
2131 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2132 SelectionDAG& DAG) {
2133 MachineFunction &MF = DAG.getMachineFunction();
2134 const TargetMachine &TM = MF.getTarget();
2135 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2136 unsigned StackAlignment = TFI.getStackAlignment();
2137 uint64_t AlignMask = StackAlignment - 1;
2138 int64_t Offset = StackSize;
2139 uint64_t SlotSize = TD->getPointerSize();
2140 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2141 // Number smaller than 12 so just add the difference.
2142 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2144 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2145 Offset = ((~AlignMask) & Offset) + StackAlignment +
2146 (StackAlignment-SlotSize);
2151 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2152 /// for tail call optimization. Targets which want to do tail call
2153 /// optimization should implement this function.
2155 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2156 CallingConv::ID CalleeCC,
2158 const SmallVectorImpl<ISD::InputArg> &Ins,
2159 SelectionDAG& DAG) const {
2160 MachineFunction &MF = DAG.getMachineFunction();
2161 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2162 return CalleeCC == CallingConv::Fast && CallerCC == CalleeCC;
2166 X86TargetLowering::createFastISel(MachineFunction &mf,
2167 MachineModuleInfo *mmo,
2169 DenseMap<const Value *, unsigned> &vm,
2170 DenseMap<const BasicBlock *,
2171 MachineBasicBlock *> &bm,
2172 DenseMap<const AllocaInst *, int> &am
2174 , SmallSet<Instruction*, 8> &cil
2177 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2185 //===----------------------------------------------------------------------===//
2186 // Other Lowering Hooks
2187 //===----------------------------------------------------------------------===//
2190 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2191 MachineFunction &MF = DAG.getMachineFunction();
2192 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2193 int ReturnAddrIndex = FuncInfo->getRAIndex();
2195 if (ReturnAddrIndex == 0) {
2196 // Set up a frame object for the return address.
2197 uint64_t SlotSize = TD->getPointerSize();
2198 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2200 FuncInfo->setRAIndex(ReturnAddrIndex);
2203 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2207 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2208 bool hasSymbolicDisplacement) {
2209 // Offset should fit into 32 bit immediate field.
2210 if (!isInt32(Offset))
2213 // If we don't have a symbolic displacement - we don't have any extra
2215 if (!hasSymbolicDisplacement)
2218 // FIXME: Some tweaks might be needed for medium code model.
2219 if (M != CodeModel::Small && M != CodeModel::Kernel)
2222 // For small code model we assume that latest object is 16MB before end of 31
2223 // bits boundary. We may also accept pretty large negative constants knowing
2224 // that all objects are in the positive half of address space.
2225 if (M == CodeModel::Small && Offset < 16*1024*1024)
2228 // For kernel code model we know that all object resist in the negative half
2229 // of 32bits address space. We may not accept negative offsets, since they may
2230 // be just off and we may accept pretty large positive ones.
2231 if (M == CodeModel::Kernel && Offset > 0)
2237 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2238 /// specific condition code, returning the condition code and the LHS/RHS of the
2239 /// comparison to make.
2240 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2241 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2243 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2244 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2245 // X > -1 -> X == 0, jump !sign.
2246 RHS = DAG.getConstant(0, RHS.getValueType());
2247 return X86::COND_NS;
2248 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2249 // X < 0 -> X == 0, jump on sign.
2251 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2253 RHS = DAG.getConstant(0, RHS.getValueType());
2254 return X86::COND_LE;
2258 switch (SetCCOpcode) {
2259 default: llvm_unreachable("Invalid integer condition!");
2260 case ISD::SETEQ: return X86::COND_E;
2261 case ISD::SETGT: return X86::COND_G;
2262 case ISD::SETGE: return X86::COND_GE;
2263 case ISD::SETLT: return X86::COND_L;
2264 case ISD::SETLE: return X86::COND_LE;
2265 case ISD::SETNE: return X86::COND_NE;
2266 case ISD::SETULT: return X86::COND_B;
2267 case ISD::SETUGT: return X86::COND_A;
2268 case ISD::SETULE: return X86::COND_BE;
2269 case ISD::SETUGE: return X86::COND_AE;
2273 // First determine if it is required or is profitable to flip the operands.
2275 // If LHS is a foldable load, but RHS is not, flip the condition.
2276 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2277 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2278 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2279 std::swap(LHS, RHS);
2282 switch (SetCCOpcode) {
2288 std::swap(LHS, RHS);
2292 // On a floating point condition, the flags are set as follows:
2294 // 0 | 0 | 0 | X > Y
2295 // 0 | 0 | 1 | X < Y
2296 // 1 | 0 | 0 | X == Y
2297 // 1 | 1 | 1 | unordered
2298 switch (SetCCOpcode) {
2299 default: llvm_unreachable("Condcode should be pre-legalized away");
2301 case ISD::SETEQ: return X86::COND_E;
2302 case ISD::SETOLT: // flipped
2304 case ISD::SETGT: return X86::COND_A;
2305 case ISD::SETOLE: // flipped
2307 case ISD::SETGE: return X86::COND_AE;
2308 case ISD::SETUGT: // flipped
2310 case ISD::SETLT: return X86::COND_B;
2311 case ISD::SETUGE: // flipped
2313 case ISD::SETLE: return X86::COND_BE;
2315 case ISD::SETNE: return X86::COND_NE;
2316 case ISD::SETUO: return X86::COND_P;
2317 case ISD::SETO: return X86::COND_NP;
2319 case ISD::SETUNE: return X86::COND_INVALID;
2323 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2324 /// code. Current x86 isa includes the following FP cmov instructions:
2325 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2326 static bool hasFPCMov(unsigned X86CC) {
2342 /// isFPImmLegal - Returns true if the target can instruction select the
2343 /// specified FP immediate natively. If false, the legalizer will
2344 /// materialize the FP immediate as a load from a constant pool.
2345 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2346 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2347 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2353 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2354 /// the specified range (L, H].
2355 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2356 return (Val < 0) || (Val >= Low && Val < Hi);
2359 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2360 /// specified value.
2361 static bool isUndefOrEqual(int Val, int CmpVal) {
2362 if (Val < 0 || Val == CmpVal)
2367 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2368 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2369 /// the second operand.
2370 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2371 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2372 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2373 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2374 return (Mask[0] < 2 && Mask[1] < 2);
2378 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2379 SmallVector<int, 8> M;
2381 return ::isPSHUFDMask(M, N->getValueType(0));
2384 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2385 /// is suitable for input to PSHUFHW.
2386 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2387 if (VT != MVT::v8i16)
2390 // Lower quadword copied in order or undef.
2391 for (int i = 0; i != 4; ++i)
2392 if (Mask[i] >= 0 && Mask[i] != i)
2395 // Upper quadword shuffled.
2396 for (int i = 4; i != 8; ++i)
2397 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2403 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2404 SmallVector<int, 8> M;
2406 return ::isPSHUFHWMask(M, N->getValueType(0));
2409 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2410 /// is suitable for input to PSHUFLW.
2411 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2412 if (VT != MVT::v8i16)
2415 // Upper quadword copied in order.
2416 for (int i = 4; i != 8; ++i)
2417 if (Mask[i] >= 0 && Mask[i] != i)
2420 // Lower quadword shuffled.
2421 for (int i = 0; i != 4; ++i)
2428 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2429 SmallVector<int, 8> M;
2431 return ::isPSHUFLWMask(M, N->getValueType(0));
2434 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2435 /// is suitable for input to PALIGNR.
2436 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2438 int i, e = VT.getVectorNumElements();
2440 // Do not handle v2i64 / v2f64 shuffles with palignr.
2441 if (e < 4 || !hasSSSE3)
2444 for (i = 0; i != e; ++i)
2448 // All undef, not a palignr.
2452 // Determine if it's ok to perform a palignr with only the LHS, since we
2453 // don't have access to the actual shuffle elements to see if RHS is undef.
2454 bool Unary = Mask[i] < (int)e;
2455 bool NeedsUnary = false;
2457 int s = Mask[i] - i;
2459 // Check the rest of the elements to see if they are consecutive.
2460 for (++i; i != e; ++i) {
2465 Unary = Unary && (m < (int)e);
2466 NeedsUnary = NeedsUnary || (m < s);
2468 if (NeedsUnary && !Unary)
2470 if (Unary && m != ((s+i) & (e-1)))
2472 if (!Unary && m != (s+i))
2478 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2479 SmallVector<int, 8> M;
2481 return ::isPALIGNRMask(M, N->getValueType(0), true);
2484 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2485 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2486 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2487 int NumElems = VT.getVectorNumElements();
2488 if (NumElems != 2 && NumElems != 4)
2491 int Half = NumElems / 2;
2492 for (int i = 0; i < Half; ++i)
2493 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2495 for (int i = Half; i < NumElems; ++i)
2496 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2502 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2503 SmallVector<int, 8> M;
2505 return ::isSHUFPMask(M, N->getValueType(0));
2508 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2509 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2510 /// half elements to come from vector 1 (which would equal the dest.) and
2511 /// the upper half to come from vector 2.
2512 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2513 int NumElems = VT.getVectorNumElements();
2515 if (NumElems != 2 && NumElems != 4)
2518 int Half = NumElems / 2;
2519 for (int i = 0; i < Half; ++i)
2520 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2522 for (int i = Half; i < NumElems; ++i)
2523 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2528 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2529 SmallVector<int, 8> M;
2531 return isCommutedSHUFPMask(M, N->getValueType(0));
2534 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2535 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2536 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2537 if (N->getValueType(0).getVectorNumElements() != 4)
2540 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2541 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2542 isUndefOrEqual(N->getMaskElt(1), 7) &&
2543 isUndefOrEqual(N->getMaskElt(2), 2) &&
2544 isUndefOrEqual(N->getMaskElt(3), 3);
2547 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2548 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2550 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2551 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2556 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2557 isUndefOrEqual(N->getMaskElt(1), 3) &&
2558 isUndefOrEqual(N->getMaskElt(2), 2) &&
2559 isUndefOrEqual(N->getMaskElt(3), 3);
2562 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2563 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2564 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2565 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2567 if (NumElems != 2 && NumElems != 4)
2570 for (unsigned i = 0; i < NumElems/2; ++i)
2571 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2574 for (unsigned i = NumElems/2; i < NumElems; ++i)
2575 if (!isUndefOrEqual(N->getMaskElt(i), i))
2581 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2582 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2583 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2584 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2586 if (NumElems != 2 && NumElems != 4)
2589 for (unsigned i = 0; i < NumElems/2; ++i)
2590 if (!isUndefOrEqual(N->getMaskElt(i), i))
2593 for (unsigned i = 0; i < NumElems/2; ++i)
2594 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2600 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2601 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2602 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2603 bool V2IsSplat = false) {
2604 int NumElts = VT.getVectorNumElements();
2605 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2608 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2610 int BitI1 = Mask[i+1];
2611 if (!isUndefOrEqual(BitI, j))
2614 if (!isUndefOrEqual(BitI1, NumElts))
2617 if (!isUndefOrEqual(BitI1, j + NumElts))
2624 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2625 SmallVector<int, 8> M;
2627 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2630 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2631 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2632 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2633 bool V2IsSplat = false) {
2634 int NumElts = VT.getVectorNumElements();
2635 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2638 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2640 int BitI1 = Mask[i+1];
2641 if (!isUndefOrEqual(BitI, j + NumElts/2))
2644 if (isUndefOrEqual(BitI1, NumElts))
2647 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2654 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2655 SmallVector<int, 8> M;
2657 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2660 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2661 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2663 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2664 int NumElems = VT.getVectorNumElements();
2665 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2668 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2670 int BitI1 = Mask[i+1];
2671 if (!isUndefOrEqual(BitI, j))
2673 if (!isUndefOrEqual(BitI1, j))
2679 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2680 SmallVector<int, 8> M;
2682 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2685 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2686 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2688 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2689 int NumElems = VT.getVectorNumElements();
2690 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2693 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2695 int BitI1 = Mask[i+1];
2696 if (!isUndefOrEqual(BitI, j))
2698 if (!isUndefOrEqual(BitI1, j))
2704 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2705 SmallVector<int, 8> M;
2707 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2710 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2711 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2712 /// MOVSD, and MOVD, i.e. setting the lowest element.
2713 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2714 if (VT.getVectorElementType().getSizeInBits() < 32)
2717 int NumElts = VT.getVectorNumElements();
2719 if (!isUndefOrEqual(Mask[0], NumElts))
2722 for (int i = 1; i < NumElts; ++i)
2723 if (!isUndefOrEqual(Mask[i], i))
2729 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2730 SmallVector<int, 8> M;
2732 return ::isMOVLMask(M, N->getValueType(0));
2735 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2736 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2737 /// element of vector 2 and the other elements to come from vector 1 in order.
2738 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2739 bool V2IsSplat = false, bool V2IsUndef = false) {
2740 int NumOps = VT.getVectorNumElements();
2741 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2744 if (!isUndefOrEqual(Mask[0], 0))
2747 for (int i = 1; i < NumOps; ++i)
2748 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2749 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2750 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2756 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2757 bool V2IsUndef = false) {
2758 SmallVector<int, 8> M;
2760 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2763 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2764 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2765 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2766 if (N->getValueType(0).getVectorNumElements() != 4)
2769 // Expect 1, 1, 3, 3
2770 for (unsigned i = 0; i < 2; ++i) {
2771 int Elt = N->getMaskElt(i);
2772 if (Elt >= 0 && Elt != 1)
2777 for (unsigned i = 2; i < 4; ++i) {
2778 int Elt = N->getMaskElt(i);
2779 if (Elt >= 0 && Elt != 3)
2784 // Don't use movshdup if it can be done with a shufps.
2785 // FIXME: verify that matching u, u, 3, 3 is what we want.
2789 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2790 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2791 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2792 if (N->getValueType(0).getVectorNumElements() != 4)
2795 // Expect 0, 0, 2, 2
2796 for (unsigned i = 0; i < 2; ++i)
2797 if (N->getMaskElt(i) > 0)
2801 for (unsigned i = 2; i < 4; ++i) {
2802 int Elt = N->getMaskElt(i);
2803 if (Elt >= 0 && Elt != 2)
2808 // Don't use movsldup if it can be done with a shufps.
2812 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2813 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2814 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2815 int e = N->getValueType(0).getVectorNumElements() / 2;
2817 for (int i = 0; i < e; ++i)
2818 if (!isUndefOrEqual(N->getMaskElt(i), i))
2820 for (int i = 0; i < e; ++i)
2821 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2826 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2827 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
2828 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2829 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2830 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2832 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2834 for (int i = 0; i < NumOperands; ++i) {
2835 int Val = SVOp->getMaskElt(NumOperands-i-1);
2836 if (Val < 0) Val = 0;
2837 if (Val >= NumOperands) Val -= NumOperands;
2839 if (i != NumOperands - 1)
2845 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2846 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
2847 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2848 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2850 // 8 nodes, but we only care about the last 4.
2851 for (unsigned i = 7; i >= 4; --i) {
2852 int Val = SVOp->getMaskElt(i);
2861 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2862 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
2863 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2864 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2866 // 8 nodes, but we only care about the first 4.
2867 for (int i = 3; i >= 0; --i) {
2868 int Val = SVOp->getMaskElt(i);
2877 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
2878 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
2879 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
2880 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2881 EVT VVT = N->getValueType(0);
2882 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
2886 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
2887 Val = SVOp->getMaskElt(i);
2891 return (Val - i) * EltSize;
2894 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2896 bool X86::isZeroNode(SDValue Elt) {
2897 return ((isa<ConstantSDNode>(Elt) &&
2898 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2899 (isa<ConstantFPSDNode>(Elt) &&
2900 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2903 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
2904 /// their permute mask.
2905 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
2906 SelectionDAG &DAG) {
2907 EVT VT = SVOp->getValueType(0);
2908 unsigned NumElems = VT.getVectorNumElements();
2909 SmallVector<int, 8> MaskVec;
2911 for (unsigned i = 0; i != NumElems; ++i) {
2912 int idx = SVOp->getMaskElt(i);
2914 MaskVec.push_back(idx);
2915 else if (idx < (int)NumElems)
2916 MaskVec.push_back(idx + NumElems);
2918 MaskVec.push_back(idx - NumElems);
2920 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
2921 SVOp->getOperand(0), &MaskVec[0]);
2924 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2925 /// the two vector operands have swapped position.
2926 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
2927 unsigned NumElems = VT.getVectorNumElements();
2928 for (unsigned i = 0; i != NumElems; ++i) {
2932 else if (idx < (int)NumElems)
2933 Mask[i] = idx + NumElems;
2935 Mask[i] = idx - NumElems;
2939 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2940 /// match movhlps. The lower half elements should come from upper half of
2941 /// V1 (and in order), and the upper half elements should come from the upper
2942 /// half of V2 (and in order).
2943 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
2944 if (Op->getValueType(0).getVectorNumElements() != 4)
2946 for (unsigned i = 0, e = 2; i != e; ++i)
2947 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
2949 for (unsigned i = 2; i != 4; ++i)
2950 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
2955 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2956 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2958 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2959 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2961 N = N->getOperand(0).getNode();
2962 if (!ISD::isNON_EXTLoad(N))
2965 *LD = cast<LoadSDNode>(N);
2969 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2970 /// match movlp{s|d}. The lower half elements should come from lower half of
2971 /// V1 (and in order), and the upper half elements should come from the upper
2972 /// half of V2 (and in order). And since V1 will become the source of the
2973 /// MOVLP, it must be either a vector load or a scalar load to vector.
2974 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
2975 ShuffleVectorSDNode *Op) {
2976 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2978 // Is V2 is a vector load, don't do this transformation. We will try to use
2979 // load folding shufps op.
2980 if (ISD::isNON_EXTLoad(V2))
2983 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
2985 if (NumElems != 2 && NumElems != 4)
2987 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2988 if (!isUndefOrEqual(Op->getMaskElt(i), i))
2990 for (unsigned i = NumElems/2; i != NumElems; ++i)
2991 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
2996 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2998 static bool isSplatVector(SDNode *N) {
2999 if (N->getOpcode() != ISD::BUILD_VECTOR)
3002 SDValue SplatValue = N->getOperand(0);
3003 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3004 if (N->getOperand(i) != SplatValue)
3009 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3010 /// to an zero vector.
3011 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3012 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3013 SDValue V1 = N->getOperand(0);
3014 SDValue V2 = N->getOperand(1);
3015 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3016 for (unsigned i = 0; i != NumElems; ++i) {
3017 int Idx = N->getMaskElt(i);
3018 if (Idx >= (int)NumElems) {
3019 unsigned Opc = V2.getOpcode();
3020 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3022 if (Opc != ISD::BUILD_VECTOR ||
3023 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3025 } else if (Idx >= 0) {
3026 unsigned Opc = V1.getOpcode();
3027 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3029 if (Opc != ISD::BUILD_VECTOR ||
3030 !X86::isZeroNode(V1.getOperand(Idx)))
3037 /// getZeroVector - Returns a vector of specified type with all zero elements.
3039 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3041 assert(VT.isVector() && "Expected a vector type");
3043 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3044 // type. This ensures they get CSE'd.
3046 if (VT.getSizeInBits() == 64) { // MMX
3047 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3048 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3049 } else if (HasSSE2) { // SSE2
3050 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3051 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3053 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3054 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3056 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3059 /// getOnesVector - Returns a vector of specified type with all bits set.
3061 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3062 assert(VT.isVector() && "Expected a vector type");
3064 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3065 // type. This ensures they get CSE'd.
3066 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3068 if (VT.getSizeInBits() == 64) // MMX
3069 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3071 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3072 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3076 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3077 /// that point to V2 points to its first element.
3078 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3079 EVT VT = SVOp->getValueType(0);
3080 unsigned NumElems = VT.getVectorNumElements();
3082 bool Changed = false;
3083 SmallVector<int, 8> MaskVec;
3084 SVOp->getMask(MaskVec);
3086 for (unsigned i = 0; i != NumElems; ++i) {
3087 if (MaskVec[i] > (int)NumElems) {
3088 MaskVec[i] = NumElems;
3093 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3094 SVOp->getOperand(1), &MaskVec[0]);
3095 return SDValue(SVOp, 0);
3098 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3099 /// operation of specified width.
3100 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3102 unsigned NumElems = VT.getVectorNumElements();
3103 SmallVector<int, 8> Mask;
3104 Mask.push_back(NumElems);
3105 for (unsigned i = 1; i != NumElems; ++i)
3107 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3110 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3111 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3113 unsigned NumElems = VT.getVectorNumElements();
3114 SmallVector<int, 8> Mask;
3115 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3117 Mask.push_back(i + NumElems);
3119 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3122 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3123 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3125 unsigned NumElems = VT.getVectorNumElements();
3126 unsigned Half = NumElems/2;
3127 SmallVector<int, 8> Mask;
3128 for (unsigned i = 0; i != Half; ++i) {
3129 Mask.push_back(i + Half);
3130 Mask.push_back(i + NumElems + Half);
3132 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3135 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3136 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3138 if (SV->getValueType(0).getVectorNumElements() <= 4)
3139 return SDValue(SV, 0);
3141 EVT PVT = MVT::v4f32;
3142 EVT VT = SV->getValueType(0);
3143 DebugLoc dl = SV->getDebugLoc();
3144 SDValue V1 = SV->getOperand(0);
3145 int NumElems = VT.getVectorNumElements();
3146 int EltNo = SV->getSplatIndex();
3148 // unpack elements to the correct location
3149 while (NumElems > 4) {
3150 if (EltNo < NumElems/2) {
3151 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3153 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3154 EltNo -= NumElems/2;
3159 // Perform the splat.
3160 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3161 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3162 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3163 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3166 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3167 /// vector of zero or undef vector. This produces a shuffle where the low
3168 /// element of V2 is swizzled into the zero/undef vector, landing at element
3169 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3170 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3171 bool isZero, bool HasSSE2,
3172 SelectionDAG &DAG) {
3173 EVT VT = V2.getValueType();
3175 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3176 unsigned NumElems = VT.getVectorNumElements();
3177 SmallVector<int, 16> MaskVec;
3178 for (unsigned i = 0; i != NumElems; ++i)
3179 // If this is the insertion idx, put the low elt of V2 here.
3180 MaskVec.push_back(i == Idx ? NumElems : i);
3181 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3184 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3185 /// a shuffle that is zero.
3187 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3188 bool Low, SelectionDAG &DAG) {
3189 unsigned NumZeros = 0;
3190 for (int i = 0; i < NumElems; ++i) {
3191 unsigned Index = Low ? i : NumElems-i-1;
3192 int Idx = SVOp->getMaskElt(Index);
3197 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3198 if (Elt.getNode() && X86::isZeroNode(Elt))
3206 /// isVectorShift - Returns true if the shuffle can be implemented as a
3207 /// logical left or right shift of a vector.
3208 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3209 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3210 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3211 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3214 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3217 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3221 bool SeenV1 = false;
3222 bool SeenV2 = false;
3223 for (int i = NumZeros; i < NumElems; ++i) {
3224 int Val = isLeft ? (i - NumZeros) : i;
3225 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3237 if (SeenV1 && SeenV2)
3240 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3246 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3248 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3249 unsigned NumNonZero, unsigned NumZero,
3250 SelectionDAG &DAG, TargetLowering &TLI) {
3254 DebugLoc dl = Op.getDebugLoc();
3257 for (unsigned i = 0; i < 16; ++i) {
3258 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3259 if (ThisIsNonZero && First) {
3261 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3263 V = DAG.getUNDEF(MVT::v8i16);
3268 SDValue ThisElt(0, 0), LastElt(0, 0);
3269 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3270 if (LastIsNonZero) {
3271 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3272 MVT::i16, Op.getOperand(i-1));
3274 if (ThisIsNonZero) {
3275 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3276 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3277 ThisElt, DAG.getConstant(8, MVT::i8));
3279 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3283 if (ThisElt.getNode())
3284 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3285 DAG.getIntPtrConstant(i/2));
3289 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3292 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3294 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3295 unsigned NumNonZero, unsigned NumZero,
3296 SelectionDAG &DAG, TargetLowering &TLI) {
3300 DebugLoc dl = Op.getDebugLoc();
3303 for (unsigned i = 0; i < 8; ++i) {
3304 bool isNonZero = (NonZeros & (1 << i)) != 0;
3308 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3310 V = DAG.getUNDEF(MVT::v8i16);
3313 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3314 MVT::v8i16, V, Op.getOperand(i),
3315 DAG.getIntPtrConstant(i));
3322 /// getVShift - Return a vector logical shift node.
3324 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3325 unsigned NumBits, SelectionDAG &DAG,
3326 const TargetLowering &TLI, DebugLoc dl) {
3327 bool isMMX = VT.getSizeInBits() == 64;
3328 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3329 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3330 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3331 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3332 DAG.getNode(Opc, dl, ShVT, SrcOp,
3333 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3337 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3338 DebugLoc dl = Op.getDebugLoc();
3339 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3340 if (ISD::isBuildVectorAllZeros(Op.getNode())
3341 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3342 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3343 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3344 // eliminated on x86-32 hosts.
3345 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3348 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3349 return getOnesVector(Op.getValueType(), DAG, dl);
3350 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3353 EVT VT = Op.getValueType();
3354 EVT ExtVT = VT.getVectorElementType();
3355 unsigned EVTBits = ExtVT.getSizeInBits();
3357 unsigned NumElems = Op.getNumOperands();
3358 unsigned NumZero = 0;
3359 unsigned NumNonZero = 0;
3360 unsigned NonZeros = 0;
3361 bool IsAllConstants = true;
3362 SmallSet<SDValue, 8> Values;
3363 for (unsigned i = 0; i < NumElems; ++i) {
3364 SDValue Elt = Op.getOperand(i);
3365 if (Elt.getOpcode() == ISD::UNDEF)
3368 if (Elt.getOpcode() != ISD::Constant &&
3369 Elt.getOpcode() != ISD::ConstantFP)
3370 IsAllConstants = false;
3371 if (X86::isZeroNode(Elt))
3374 NonZeros |= (1 << i);
3379 if (NumNonZero == 0) {
3380 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3381 return DAG.getUNDEF(VT);
3384 // Special case for single non-zero, non-undef, element.
3385 if (NumNonZero == 1) {
3386 unsigned Idx = CountTrailingZeros_32(NonZeros);
3387 SDValue Item = Op.getOperand(Idx);
3389 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3390 // the value are obviously zero, truncate the value to i32 and do the
3391 // insertion that way. Only do this if the value is non-constant or if the
3392 // value is a constant being inserted into element 0. It is cheaper to do
3393 // a constant pool load than it is to do a movd + shuffle.
3394 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3395 (!IsAllConstants || Idx == 0)) {
3396 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3397 // Handle MMX and SSE both.
3398 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3399 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3401 // Truncate the value (which may itself be a constant) to i32, and
3402 // convert it to a vector with movd (S2V+shuffle to zero extend).
3403 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3404 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3405 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3406 Subtarget->hasSSE2(), DAG);
3408 // Now we have our 32-bit value zero extended in the low element of
3409 // a vector. If Idx != 0, swizzle it into place.
3411 SmallVector<int, 4> Mask;
3412 Mask.push_back(Idx);
3413 for (unsigned i = 1; i != VecElts; ++i)
3415 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3416 DAG.getUNDEF(Item.getValueType()),
3419 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3423 // If we have a constant or non-constant insertion into the low element of
3424 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3425 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3426 // depending on what the source datatype is.
3429 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3430 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3431 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3432 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3433 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3434 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3436 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3437 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3438 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3439 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3440 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3441 Subtarget->hasSSE2(), DAG);
3442 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3446 // Is it a vector logical left shift?
3447 if (NumElems == 2 && Idx == 1 &&
3448 X86::isZeroNode(Op.getOperand(0)) &&
3449 !X86::isZeroNode(Op.getOperand(1))) {
3450 unsigned NumBits = VT.getSizeInBits();
3451 return getVShift(true, VT,
3452 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3453 VT, Op.getOperand(1)),
3454 NumBits/2, DAG, *this, dl);
3457 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3460 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3461 // is a non-constant being inserted into an element other than the low one,
3462 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3463 // movd/movss) to move this into the low element, then shuffle it into
3465 if (EVTBits == 32) {
3466 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3468 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3469 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3470 Subtarget->hasSSE2(), DAG);
3471 SmallVector<int, 8> MaskVec;
3472 for (unsigned i = 0; i < NumElems; i++)
3473 MaskVec.push_back(i == Idx ? 0 : 1);
3474 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3478 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3479 if (Values.size() == 1)
3482 // A vector full of immediates; various special cases are already
3483 // handled, so this is best done with a single constant-pool load.
3487 // Let legalizer expand 2-wide build_vectors.
3488 if (EVTBits == 64) {
3489 if (NumNonZero == 1) {
3490 // One half is zero or undef.
3491 unsigned Idx = CountTrailingZeros_32(NonZeros);
3492 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3493 Op.getOperand(Idx));
3494 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3495 Subtarget->hasSSE2(), DAG);
3500 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3501 if (EVTBits == 8 && NumElems == 16) {
3502 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3504 if (V.getNode()) return V;
3507 if (EVTBits == 16 && NumElems == 8) {
3508 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3510 if (V.getNode()) return V;
3513 // If element VT is == 32 bits, turn it into a number of shuffles.
3514 SmallVector<SDValue, 8> V;
3516 if (NumElems == 4 && NumZero > 0) {
3517 for (unsigned i = 0; i < 4; ++i) {
3518 bool isZero = !(NonZeros & (1 << i));
3520 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3522 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3525 for (unsigned i = 0; i < 2; ++i) {
3526 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3529 V[i] = V[i*2]; // Must be a zero vector.
3532 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3535 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3538 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3543 SmallVector<int, 8> MaskVec;
3544 bool Reverse = (NonZeros & 0x3) == 2;
3545 for (unsigned i = 0; i < 2; ++i)
3546 MaskVec.push_back(Reverse ? 1-i : i);
3547 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3548 for (unsigned i = 0; i < 2; ++i)
3549 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3550 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3553 if (Values.size() > 2) {
3554 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3555 // values to be inserted is equal to the number of elements, in which case
3556 // use the unpack code below in the hopes of matching the consecutive elts
3557 // load merge pattern for shuffles.
3558 // FIXME: We could probably just check that here directly.
3559 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3560 getSubtarget()->hasSSE41()) {
3561 V[0] = DAG.getUNDEF(VT);
3562 for (unsigned i = 0; i < NumElems; ++i)
3563 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3564 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3565 Op.getOperand(i), DAG.getIntPtrConstant(i));
3568 // Expand into a number of unpckl*.
3570 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3571 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3572 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3573 for (unsigned i = 0; i < NumElems; ++i)
3574 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3576 while (NumElems != 0) {
3577 for (unsigned i = 0; i < NumElems; ++i)
3578 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3587 // v8i16 shuffles - Prefer shuffles in the following order:
3588 // 1. [all] pshuflw, pshufhw, optional move
3589 // 2. [ssse3] 1 x pshufb
3590 // 3. [ssse3] 2 x pshufb + 1 x por
3591 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3593 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3594 SelectionDAG &DAG, X86TargetLowering &TLI) {
3595 SDValue V1 = SVOp->getOperand(0);
3596 SDValue V2 = SVOp->getOperand(1);
3597 DebugLoc dl = SVOp->getDebugLoc();
3598 SmallVector<int, 8> MaskVals;
3600 // Determine if more than 1 of the words in each of the low and high quadwords
3601 // of the result come from the same quadword of one of the two inputs. Undef
3602 // mask values count as coming from any quadword, for better codegen.
3603 SmallVector<unsigned, 4> LoQuad(4);
3604 SmallVector<unsigned, 4> HiQuad(4);
3605 BitVector InputQuads(4);
3606 for (unsigned i = 0; i < 8; ++i) {
3607 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3608 int EltIdx = SVOp->getMaskElt(i);
3609 MaskVals.push_back(EltIdx);
3618 InputQuads.set(EltIdx / 4);
3621 int BestLoQuad = -1;
3622 unsigned MaxQuad = 1;
3623 for (unsigned i = 0; i < 4; ++i) {
3624 if (LoQuad[i] > MaxQuad) {
3626 MaxQuad = LoQuad[i];
3630 int BestHiQuad = -1;
3632 for (unsigned i = 0; i < 4; ++i) {
3633 if (HiQuad[i] > MaxQuad) {
3635 MaxQuad = HiQuad[i];
3639 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3640 // of the two input vectors, shuffle them into one input vector so only a
3641 // single pshufb instruction is necessary. If There are more than 2 input
3642 // quads, disable the next transformation since it does not help SSSE3.
3643 bool V1Used = InputQuads[0] || InputQuads[1];
3644 bool V2Used = InputQuads[2] || InputQuads[3];
3645 if (TLI.getSubtarget()->hasSSSE3()) {
3646 if (InputQuads.count() == 2 && V1Used && V2Used) {
3647 BestLoQuad = InputQuads.find_first();
3648 BestHiQuad = InputQuads.find_next(BestLoQuad);
3650 if (InputQuads.count() > 2) {
3656 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3657 // the shuffle mask. If a quad is scored as -1, that means that it contains
3658 // words from all 4 input quadwords.
3660 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3661 SmallVector<int, 8> MaskV;
3662 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3663 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3664 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3665 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3666 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3667 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3669 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3670 // source words for the shuffle, to aid later transformations.
3671 bool AllWordsInNewV = true;
3672 bool InOrder[2] = { true, true };
3673 for (unsigned i = 0; i != 8; ++i) {
3674 int idx = MaskVals[i];
3676 InOrder[i/4] = false;
3677 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3679 AllWordsInNewV = false;
3683 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3684 if (AllWordsInNewV) {
3685 for (int i = 0; i != 8; ++i) {
3686 int idx = MaskVals[i];
3689 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3690 if ((idx != i) && idx < 4)
3692 if ((idx != i) && idx > 3)
3701 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3702 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3703 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3704 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3705 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3709 // If we have SSSE3, and all words of the result are from 1 input vector,
3710 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3711 // is present, fall back to case 4.
3712 if (TLI.getSubtarget()->hasSSSE3()) {
3713 SmallVector<SDValue,16> pshufbMask;
3715 // If we have elements from both input vectors, set the high bit of the
3716 // shuffle mask element to zero out elements that come from V2 in the V1
3717 // mask, and elements that come from V1 in the V2 mask, so that the two
3718 // results can be OR'd together.
3719 bool TwoInputs = V1Used && V2Used;
3720 for (unsigned i = 0; i != 8; ++i) {
3721 int EltIdx = MaskVals[i] * 2;
3722 if (TwoInputs && (EltIdx >= 16)) {
3723 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3724 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3727 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3728 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3730 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3731 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3732 DAG.getNode(ISD::BUILD_VECTOR, dl,
3733 MVT::v16i8, &pshufbMask[0], 16));
3735 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3737 // Calculate the shuffle mask for the second input, shuffle it, and
3738 // OR it with the first shuffled input.
3740 for (unsigned i = 0; i != 8; ++i) {
3741 int EltIdx = MaskVals[i] * 2;
3743 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3744 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3747 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3748 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3750 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3751 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3752 DAG.getNode(ISD::BUILD_VECTOR, dl,
3753 MVT::v16i8, &pshufbMask[0], 16));
3754 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3755 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3758 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3759 // and update MaskVals with new element order.
3760 BitVector InOrder(8);
3761 if (BestLoQuad >= 0) {
3762 SmallVector<int, 8> MaskV;
3763 for (int i = 0; i != 4; ++i) {
3764 int idx = MaskVals[i];
3766 MaskV.push_back(-1);
3768 } else if ((idx / 4) == BestLoQuad) {
3769 MaskV.push_back(idx & 3);
3772 MaskV.push_back(-1);
3775 for (unsigned i = 4; i != 8; ++i)
3777 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3781 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
3782 // and update MaskVals with the new element order.
3783 if (BestHiQuad >= 0) {
3784 SmallVector<int, 8> MaskV;
3785 for (unsigned i = 0; i != 4; ++i)
3787 for (unsigned i = 4; i != 8; ++i) {
3788 int idx = MaskVals[i];
3790 MaskV.push_back(-1);
3792 } else if ((idx / 4) == BestHiQuad) {
3793 MaskV.push_back((idx & 3) + 4);
3796 MaskV.push_back(-1);
3799 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3803 // In case BestHi & BestLo were both -1, which means each quadword has a word
3804 // from each of the four input quadwords, calculate the InOrder bitvector now
3805 // before falling through to the insert/extract cleanup.
3806 if (BestLoQuad == -1 && BestHiQuad == -1) {
3808 for (int i = 0; i != 8; ++i)
3809 if (MaskVals[i] < 0 || MaskVals[i] == i)
3813 // The other elements are put in the right place using pextrw and pinsrw.
3814 for (unsigned i = 0; i != 8; ++i) {
3817 int EltIdx = MaskVals[i];
3820 SDValue ExtOp = (EltIdx < 8)
3821 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
3822 DAG.getIntPtrConstant(EltIdx))
3823 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
3824 DAG.getIntPtrConstant(EltIdx - 8));
3825 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3826 DAG.getIntPtrConstant(i));
3831 // v16i8 shuffles - Prefer shuffles in the following order:
3832 // 1. [ssse3] 1 x pshufb
3833 // 2. [ssse3] 2 x pshufb + 1 x por
3834 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
3836 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
3837 SelectionDAG &DAG, X86TargetLowering &TLI) {
3838 SDValue V1 = SVOp->getOperand(0);
3839 SDValue V2 = SVOp->getOperand(1);
3840 DebugLoc dl = SVOp->getDebugLoc();
3841 SmallVector<int, 16> MaskVals;
3842 SVOp->getMask(MaskVals);
3844 // If we have SSSE3, case 1 is generated when all result bytes come from
3845 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
3846 // present, fall back to case 3.
3847 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
3850 for (unsigned i = 0; i < 16; ++i) {
3851 int EltIdx = MaskVals[i];
3860 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
3861 if (TLI.getSubtarget()->hasSSSE3()) {
3862 SmallVector<SDValue,16> pshufbMask;
3864 // If all result elements are from one input vector, then only translate
3865 // undef mask values to 0x80 (zero out result) in the pshufb mask.
3867 // Otherwise, we have elements from both input vectors, and must zero out
3868 // elements that come from V2 in the first mask, and V1 in the second mask
3869 // so that we can OR them together.
3870 bool TwoInputs = !(V1Only || V2Only);
3871 for (unsigned i = 0; i != 16; ++i) {
3872 int EltIdx = MaskVals[i];
3873 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
3874 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3877 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3879 // If all the elements are from V2, assign it to V1 and return after
3880 // building the first pshufb.
3883 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3884 DAG.getNode(ISD::BUILD_VECTOR, dl,
3885 MVT::v16i8, &pshufbMask[0], 16));
3889 // Calculate the shuffle mask for the second input, shuffle it, and
3890 // OR it with the first shuffled input.
3892 for (unsigned i = 0; i != 16; ++i) {
3893 int EltIdx = MaskVals[i];
3895 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3898 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3900 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3901 DAG.getNode(ISD::BUILD_VECTOR, dl,
3902 MVT::v16i8, &pshufbMask[0], 16));
3903 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3906 // No SSSE3 - Calculate in place words and then fix all out of place words
3907 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
3908 // the 16 different words that comprise the two doublequadword input vectors.
3909 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3910 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
3911 SDValue NewV = V2Only ? V2 : V1;
3912 for (int i = 0; i != 8; ++i) {
3913 int Elt0 = MaskVals[i*2];
3914 int Elt1 = MaskVals[i*2+1];
3916 // This word of the result is all undef, skip it.
3917 if (Elt0 < 0 && Elt1 < 0)
3920 // This word of the result is already in the correct place, skip it.
3921 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
3923 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
3926 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
3927 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
3930 // If Elt0 and Elt1 are defined, are consecutive, and can be load
3931 // using a single extract together, load it and store it.
3932 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
3933 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3934 DAG.getIntPtrConstant(Elt1 / 2));
3935 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3936 DAG.getIntPtrConstant(i));
3940 // If Elt1 is defined, extract it from the appropriate source. If the
3941 // source byte is not also odd, shift the extracted word left 8 bits
3942 // otherwise clear the bottom 8 bits if we need to do an or.
3944 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3945 DAG.getIntPtrConstant(Elt1 / 2));
3946 if ((Elt1 & 1) == 0)
3947 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
3948 DAG.getConstant(8, TLI.getShiftAmountTy()));
3950 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
3951 DAG.getConstant(0xFF00, MVT::i16));
3953 // If Elt0 is defined, extract it from the appropriate source. If the
3954 // source byte is not also even, shift the extracted word right 8 bits. If
3955 // Elt1 was also defined, OR the extracted values together before
3956 // inserting them in the result.
3958 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
3959 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
3960 if ((Elt0 & 1) != 0)
3961 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
3962 DAG.getConstant(8, TLI.getShiftAmountTy()));
3964 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
3965 DAG.getConstant(0x00FF, MVT::i16));
3966 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
3969 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3970 DAG.getIntPtrConstant(i));
3972 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
3975 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3976 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3977 /// done when every pair / quad of shuffle mask elements point to elements in
3978 /// the right sequence. e.g.
3979 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3981 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
3983 TargetLowering &TLI, DebugLoc dl) {
3984 EVT VT = SVOp->getValueType(0);
3985 SDValue V1 = SVOp->getOperand(0);
3986 SDValue V2 = SVOp->getOperand(1);
3987 unsigned NumElems = VT.getVectorNumElements();
3988 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3989 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3990 EVT MaskEltVT = MaskVT.getVectorElementType();
3992 switch (VT.getSimpleVT().SimpleTy) {
3993 default: assert(false && "Unexpected!");
3994 case MVT::v4f32: NewVT = MVT::v2f64; break;
3995 case MVT::v4i32: NewVT = MVT::v2i64; break;
3996 case MVT::v8i16: NewVT = MVT::v4i32; break;
3997 case MVT::v16i8: NewVT = MVT::v4i32; break;
4000 if (NewWidth == 2) {
4006 int Scale = NumElems / NewWidth;
4007 SmallVector<int, 8> MaskVec;
4008 for (unsigned i = 0; i < NumElems; i += Scale) {
4010 for (int j = 0; j < Scale; ++j) {
4011 int EltIdx = SVOp->getMaskElt(i+j);
4015 StartIdx = EltIdx - (EltIdx % Scale);
4016 if (EltIdx != StartIdx + j)
4020 MaskVec.push_back(-1);
4022 MaskVec.push_back(StartIdx / Scale);
4025 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4026 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4027 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4030 /// getVZextMovL - Return a zero-extending vector move low node.
4032 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4033 SDValue SrcOp, SelectionDAG &DAG,
4034 const X86Subtarget *Subtarget, DebugLoc dl) {
4035 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4036 LoadSDNode *LD = NULL;
4037 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4038 LD = dyn_cast<LoadSDNode>(SrcOp);
4040 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4042 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4043 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4044 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4045 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4046 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4048 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4049 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4050 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4051 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4059 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4060 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4061 DAG.getNode(ISD::BIT_CONVERT, dl,
4065 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4068 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4069 SDValue V1 = SVOp->getOperand(0);
4070 SDValue V2 = SVOp->getOperand(1);
4071 DebugLoc dl = SVOp->getDebugLoc();
4072 EVT VT = SVOp->getValueType(0);
4074 SmallVector<std::pair<int, int>, 8> Locs;
4076 SmallVector<int, 8> Mask1(4U, -1);
4077 SmallVector<int, 8> PermMask;
4078 SVOp->getMask(PermMask);
4082 for (unsigned i = 0; i != 4; ++i) {
4083 int Idx = PermMask[i];
4085 Locs[i] = std::make_pair(-1, -1);
4087 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4089 Locs[i] = std::make_pair(0, NumLo);
4093 Locs[i] = std::make_pair(1, NumHi);
4095 Mask1[2+NumHi] = Idx;
4101 if (NumLo <= 2 && NumHi <= 2) {
4102 // If no more than two elements come from either vector. This can be
4103 // implemented with two shuffles. First shuffle gather the elements.
4104 // The second shuffle, which takes the first shuffle as both of its
4105 // vector operands, put the elements into the right order.
4106 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4108 SmallVector<int, 8> Mask2(4U, -1);
4110 for (unsigned i = 0; i != 4; ++i) {
4111 if (Locs[i].first == -1)
4114 unsigned Idx = (i < 2) ? 0 : 4;
4115 Idx += Locs[i].first * 2 + Locs[i].second;
4120 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4121 } else if (NumLo == 3 || NumHi == 3) {
4122 // Otherwise, we must have three elements from one vector, call it X, and
4123 // one element from the other, call it Y. First, use a shufps to build an
4124 // intermediate vector with the one element from Y and the element from X
4125 // that will be in the same half in the final destination (the indexes don't
4126 // matter). Then, use a shufps to build the final vector, taking the half
4127 // containing the element from Y from the intermediate, and the other half
4130 // Normalize it so the 3 elements come from V1.
4131 CommuteVectorShuffleMask(PermMask, VT);
4135 // Find the element from V2.
4137 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4138 int Val = PermMask[HiIndex];
4145 Mask1[0] = PermMask[HiIndex];
4147 Mask1[2] = PermMask[HiIndex^1];
4149 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4152 Mask1[0] = PermMask[0];
4153 Mask1[1] = PermMask[1];
4154 Mask1[2] = HiIndex & 1 ? 6 : 4;
4155 Mask1[3] = HiIndex & 1 ? 4 : 6;
4156 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4158 Mask1[0] = HiIndex & 1 ? 2 : 0;
4159 Mask1[1] = HiIndex & 1 ? 0 : 2;
4160 Mask1[2] = PermMask[2];
4161 Mask1[3] = PermMask[3];
4166 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4170 // Break it into (shuffle shuffle_hi, shuffle_lo).
4172 SmallVector<int,8> LoMask(4U, -1);
4173 SmallVector<int,8> HiMask(4U, -1);
4175 SmallVector<int,8> *MaskPtr = &LoMask;
4176 unsigned MaskIdx = 0;
4179 for (unsigned i = 0; i != 4; ++i) {
4186 int Idx = PermMask[i];
4188 Locs[i] = std::make_pair(-1, -1);
4189 } else if (Idx < 4) {
4190 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4191 (*MaskPtr)[LoIdx] = Idx;
4194 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4195 (*MaskPtr)[HiIdx] = Idx;
4200 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4201 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4202 SmallVector<int, 8> MaskOps;
4203 for (unsigned i = 0; i != 4; ++i) {
4204 if (Locs[i].first == -1) {
4205 MaskOps.push_back(-1);
4207 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4208 MaskOps.push_back(Idx);
4211 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4215 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4216 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4217 SDValue V1 = Op.getOperand(0);
4218 SDValue V2 = Op.getOperand(1);
4219 EVT VT = Op.getValueType();
4220 DebugLoc dl = Op.getDebugLoc();
4221 unsigned NumElems = VT.getVectorNumElements();
4222 bool isMMX = VT.getSizeInBits() == 64;
4223 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4224 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4225 bool V1IsSplat = false;
4226 bool V2IsSplat = false;
4228 if (isZeroShuffle(SVOp))
4229 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4231 // Promote splats to v4f32.
4232 if (SVOp->isSplat()) {
4233 if (isMMX || NumElems < 4)
4235 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4238 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4240 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4241 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4242 if (NewOp.getNode())
4243 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4244 LowerVECTOR_SHUFFLE(NewOp, DAG));
4245 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4246 // FIXME: Figure out a cleaner way to do this.
4247 // Try to make use of movq to zero out the top part.
4248 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4249 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4250 if (NewOp.getNode()) {
4251 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4252 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4253 DAG, Subtarget, dl);
4255 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4256 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4257 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4258 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4259 DAG, Subtarget, dl);
4263 if (X86::isPSHUFDMask(SVOp))
4266 // Check if this can be converted into a logical shift.
4267 bool isLeft = false;
4270 bool isShift = getSubtarget()->hasSSE2() &&
4271 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4272 if (isShift && ShVal.hasOneUse()) {
4273 // If the shifted value has multiple uses, it may be cheaper to use
4274 // v_set0 + movlhps or movhlps, etc.
4275 EVT EltVT = VT.getVectorElementType();
4276 ShAmt *= EltVT.getSizeInBits();
4277 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4280 if (X86::isMOVLMask(SVOp)) {
4283 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4284 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4289 // FIXME: fold these into legal mask.
4290 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4291 X86::isMOVSLDUPMask(SVOp) ||
4292 X86::isMOVHLPSMask(SVOp) ||
4293 X86::isMOVLHPSMask(SVOp) ||
4294 X86::isMOVLPMask(SVOp)))
4297 if (ShouldXformToMOVHLPS(SVOp) ||
4298 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4299 return CommuteVectorShuffle(SVOp, DAG);
4302 // No better options. Use a vshl / vsrl.
4303 EVT EltVT = VT.getVectorElementType();
4304 ShAmt *= EltVT.getSizeInBits();
4305 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4308 bool Commuted = false;
4309 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4310 // 1,1,1,1 -> v8i16 though.
4311 V1IsSplat = isSplatVector(V1.getNode());
4312 V2IsSplat = isSplatVector(V2.getNode());
4314 // Canonicalize the splat or undef, if present, to be on the RHS.
4315 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4316 Op = CommuteVectorShuffle(SVOp, DAG);
4317 SVOp = cast<ShuffleVectorSDNode>(Op);
4318 V1 = SVOp->getOperand(0);
4319 V2 = SVOp->getOperand(1);
4320 std::swap(V1IsSplat, V2IsSplat);
4321 std::swap(V1IsUndef, V2IsUndef);
4325 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4326 // Shuffling low element of v1 into undef, just return v1.
4329 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4330 // the instruction selector will not match, so get a canonical MOVL with
4331 // swapped operands to undo the commute.
4332 return getMOVL(DAG, dl, VT, V2, V1);
4335 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4336 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4337 X86::isUNPCKLMask(SVOp) ||
4338 X86::isUNPCKHMask(SVOp))
4342 // Normalize mask so all entries that point to V2 points to its first
4343 // element then try to match unpck{h|l} again. If match, return a
4344 // new vector_shuffle with the corrected mask.
4345 SDValue NewMask = NormalizeMask(SVOp, DAG);
4346 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4347 if (NSVOp != SVOp) {
4348 if (X86::isUNPCKLMask(NSVOp, true)) {
4350 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4357 // Commute is back and try unpck* again.
4358 // FIXME: this seems wrong.
4359 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4360 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4361 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4362 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4363 X86::isUNPCKLMask(NewSVOp) ||
4364 X86::isUNPCKHMask(NewSVOp))
4368 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4370 // Normalize the node to match x86 shuffle ops if needed
4371 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4372 return CommuteVectorShuffle(SVOp, DAG);
4374 // Check for legal shuffle and return?
4375 SmallVector<int, 16> PermMask;
4376 SVOp->getMask(PermMask);
4377 if (isShuffleMaskLegal(PermMask, VT))
4380 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4381 if (VT == MVT::v8i16) {
4382 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4383 if (NewOp.getNode())
4387 if (VT == MVT::v16i8) {
4388 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4389 if (NewOp.getNode())
4393 // Handle all 4 wide cases with a number of shuffles except for MMX.
4394 if (NumElems == 4 && !isMMX)
4395 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4401 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4402 SelectionDAG &DAG) {
4403 EVT VT = Op.getValueType();
4404 DebugLoc dl = Op.getDebugLoc();
4405 if (VT.getSizeInBits() == 8) {
4406 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4407 Op.getOperand(0), Op.getOperand(1));
4408 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4409 DAG.getValueType(VT));
4410 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4411 } else if (VT.getSizeInBits() == 16) {
4412 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4413 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4415 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4416 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4417 DAG.getNode(ISD::BIT_CONVERT, dl,
4421 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4422 Op.getOperand(0), Op.getOperand(1));
4423 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4424 DAG.getValueType(VT));
4425 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4426 } else if (VT == MVT::f32) {
4427 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4428 // the result back to FR32 register. It's only worth matching if the
4429 // result has a single use which is a store or a bitcast to i32. And in
4430 // the case of a store, it's not worth it if the index is a constant 0,
4431 // because a MOVSSmr can be used instead, which is smaller and faster.
4432 if (!Op.hasOneUse())
4434 SDNode *User = *Op.getNode()->use_begin();
4435 if ((User->getOpcode() != ISD::STORE ||
4436 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4437 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4438 (User->getOpcode() != ISD::BIT_CONVERT ||
4439 User->getValueType(0) != MVT::i32))
4441 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4442 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4445 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4446 } else if (VT == MVT::i32) {
4447 // ExtractPS works with constant index.
4448 if (isa<ConstantSDNode>(Op.getOperand(1)))
4456 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4457 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4460 if (Subtarget->hasSSE41()) {
4461 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4466 EVT VT = Op.getValueType();
4467 DebugLoc dl = Op.getDebugLoc();
4468 // TODO: handle v16i8.
4469 if (VT.getSizeInBits() == 16) {
4470 SDValue Vec = Op.getOperand(0);
4471 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4473 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4474 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4475 DAG.getNode(ISD::BIT_CONVERT, dl,
4478 // Transform it so it match pextrw which produces a 32-bit result.
4479 EVT EltVT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy+1);
4480 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4481 Op.getOperand(0), Op.getOperand(1));
4482 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4483 DAG.getValueType(VT));
4484 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4485 } else if (VT.getSizeInBits() == 32) {
4486 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4490 // SHUFPS the element to the lowest double word, then movss.
4491 int Mask[4] = { Idx, -1, -1, -1 };
4492 EVT VVT = Op.getOperand(0).getValueType();
4493 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4494 DAG.getUNDEF(VVT), Mask);
4495 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4496 DAG.getIntPtrConstant(0));
4497 } else if (VT.getSizeInBits() == 64) {
4498 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4499 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4500 // to match extract_elt for f64.
4501 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4505 // UNPCKHPD the element to the lowest double word, then movsd.
4506 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4507 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4508 int Mask[2] = { 1, -1 };
4509 EVT VVT = Op.getOperand(0).getValueType();
4510 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4511 DAG.getUNDEF(VVT), Mask);
4512 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4513 DAG.getIntPtrConstant(0));
4520 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4521 EVT VT = Op.getValueType();
4522 EVT EltVT = VT.getVectorElementType();
4523 DebugLoc dl = Op.getDebugLoc();
4525 SDValue N0 = Op.getOperand(0);
4526 SDValue N1 = Op.getOperand(1);
4527 SDValue N2 = Op.getOperand(2);
4529 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4530 isa<ConstantSDNode>(N2)) {
4531 unsigned Opc = (EltVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4533 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4535 if (N1.getValueType() != MVT::i32)
4536 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4537 if (N2.getValueType() != MVT::i32)
4538 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4539 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4540 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4541 // Bits [7:6] of the constant are the source select. This will always be
4542 // zero here. The DAG Combiner may combine an extract_elt index into these
4543 // bits. For example (insert (extract, 3), 2) could be matched by putting
4544 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4545 // Bits [5:4] of the constant are the destination select. This is the
4546 // value of the incoming immediate.
4547 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4548 // combine either bitwise AND or insert of float 0.0 to set these bits.
4549 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4550 // Create this as a scalar to vector..
4551 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4552 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4553 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4554 // PINSR* works with constant index.
4561 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4562 EVT VT = Op.getValueType();
4563 EVT EltVT = VT.getVectorElementType();
4565 if (Subtarget->hasSSE41())
4566 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4568 if (EltVT == MVT::i8)
4571 DebugLoc dl = Op.getDebugLoc();
4572 SDValue N0 = Op.getOperand(0);
4573 SDValue N1 = Op.getOperand(1);
4574 SDValue N2 = Op.getOperand(2);
4576 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4577 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4578 // as its second argument.
4579 if (N1.getValueType() != MVT::i32)
4580 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4581 if (N2.getValueType() != MVT::i32)
4582 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4583 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4589 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4590 DebugLoc dl = Op.getDebugLoc();
4591 if (Op.getValueType() == MVT::v2f32)
4592 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4593 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4594 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4595 Op.getOperand(0))));
4597 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4598 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4600 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4601 EVT VT = MVT::v2i32;
4602 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4609 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4610 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4613 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4614 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4615 // one of the above mentioned nodes. It has to be wrapped because otherwise
4616 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4617 // be used to form addressing mode. These wrapped nodes will be selected
4620 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4621 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4623 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4625 unsigned char OpFlag = 0;
4626 unsigned WrapperKind = X86ISD::Wrapper;
4627 CodeModel::Model M = getTargetMachine().getCodeModel();
4629 if (Subtarget->isPICStyleRIPRel() &&
4630 (M == CodeModel::Small || M == CodeModel::Kernel))
4631 WrapperKind = X86ISD::WrapperRIP;
4632 else if (Subtarget->isPICStyleGOT())
4633 OpFlag = X86II::MO_GOTOFF;
4634 else if (Subtarget->isPICStyleStubPIC())
4635 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4637 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4639 CP->getOffset(), OpFlag);
4640 DebugLoc DL = CP->getDebugLoc();
4641 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4642 // With PIC, the address is actually $g + Offset.
4644 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4645 DAG.getNode(X86ISD::GlobalBaseReg,
4646 DebugLoc::getUnknownLoc(), getPointerTy()),
4653 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4654 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4656 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4658 unsigned char OpFlag = 0;
4659 unsigned WrapperKind = X86ISD::Wrapper;
4660 CodeModel::Model M = getTargetMachine().getCodeModel();
4662 if (Subtarget->isPICStyleRIPRel() &&
4663 (M == CodeModel::Small || M == CodeModel::Kernel))
4664 WrapperKind = X86ISD::WrapperRIP;
4665 else if (Subtarget->isPICStyleGOT())
4666 OpFlag = X86II::MO_GOTOFF;
4667 else if (Subtarget->isPICStyleStubPIC())
4668 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4670 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4672 DebugLoc DL = JT->getDebugLoc();
4673 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4675 // With PIC, the address is actually $g + Offset.
4677 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4678 DAG.getNode(X86ISD::GlobalBaseReg,
4679 DebugLoc::getUnknownLoc(), getPointerTy()),
4687 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4688 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4690 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4692 unsigned char OpFlag = 0;
4693 unsigned WrapperKind = X86ISD::Wrapper;
4694 CodeModel::Model M = getTargetMachine().getCodeModel();
4696 if (Subtarget->isPICStyleRIPRel() &&
4697 (M == CodeModel::Small || M == CodeModel::Kernel))
4698 WrapperKind = X86ISD::WrapperRIP;
4699 else if (Subtarget->isPICStyleGOT())
4700 OpFlag = X86II::MO_GOTOFF;
4701 else if (Subtarget->isPICStyleStubPIC())
4702 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4704 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4706 DebugLoc DL = Op.getDebugLoc();
4707 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4710 // With PIC, the address is actually $g + Offset.
4711 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4712 !Subtarget->is64Bit()) {
4713 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4714 DAG.getNode(X86ISD::GlobalBaseReg,
4715 DebugLoc::getUnknownLoc(),
4724 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
4725 unsigned WrapperKind = X86ISD::Wrapper;
4726 CodeModel::Model M = getTargetMachine().getCodeModel();
4727 if (Subtarget->isPICStyleRIPRel() &&
4728 (M == CodeModel::Small || M == CodeModel::Kernel))
4729 WrapperKind = X86ISD::WrapperRIP;
4731 DebugLoc DL = Op.getDebugLoc();
4733 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
4734 SDValue Result = DAG.getBlockAddress(BA, DL, /*isTarget=*/true);
4736 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4742 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4744 SelectionDAG &DAG) const {
4745 // Create the TargetGlobalAddress node, folding in the constant
4746 // offset if it is legal.
4747 unsigned char OpFlags =
4748 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
4749 CodeModel::Model M = getTargetMachine().getCodeModel();
4751 if (OpFlags == X86II::MO_NO_FLAG &&
4752 X86::isOffsetSuitableForCodeModel(Offset, M)) {
4753 // A direct static reference to a global.
4754 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4757 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
4760 if (Subtarget->isPICStyleRIPRel() &&
4761 (M == CodeModel::Small || M == CodeModel::Kernel))
4762 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4764 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4766 // With PIC, the address is actually $g + Offset.
4767 if (isGlobalRelativeToPICBase(OpFlags)) {
4768 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4769 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4773 // For globals that require a load from a stub to get the address, emit the
4775 if (isGlobalStubReference(OpFlags))
4776 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4777 PseudoSourceValue::getGOT(), 0);
4779 // If there was a non-zero offset that we didn't fold, create an explicit
4782 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
4783 DAG.getConstant(Offset, getPointerTy()));
4789 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4790 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4791 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4792 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
4796 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
4797 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
4798 unsigned char OperandFlags) {
4799 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4800 DebugLoc dl = GA->getDebugLoc();
4801 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4802 GA->getValueType(0),
4806 SDValue Ops[] = { Chain, TGA, *InFlag };
4807 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
4809 SDValue Ops[] = { Chain, TGA };
4810 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
4812 SDValue Flag = Chain.getValue(1);
4813 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
4816 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4818 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4821 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4822 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
4823 DAG.getNode(X86ISD::GlobalBaseReg,
4824 DebugLoc::getUnknownLoc(),
4826 InFlag = Chain.getValue(1);
4828 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
4831 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4833 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4835 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
4836 X86::RAX, X86II::MO_TLSGD);
4839 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4840 // "local exec" model.
4841 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4842 const EVT PtrVT, TLSModel::Model model,
4844 DebugLoc dl = GA->getDebugLoc();
4845 // Get the Thread Pointer
4846 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
4847 DebugLoc::getUnknownLoc(), PtrVT,
4848 DAG.getRegister(is64Bit? X86::FS : X86::GS,
4851 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
4854 unsigned char OperandFlags = 0;
4855 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
4857 unsigned WrapperKind = X86ISD::Wrapper;
4858 if (model == TLSModel::LocalExec) {
4859 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
4860 } else if (is64Bit) {
4861 assert(model == TLSModel::InitialExec);
4862 OperandFlags = X86II::MO_GOTTPOFF;
4863 WrapperKind = X86ISD::WrapperRIP;
4865 assert(model == TLSModel::InitialExec);
4866 OperandFlags = X86II::MO_INDNTPOFF;
4869 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4871 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
4872 GA->getOffset(), OperandFlags);
4873 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
4875 if (model == TLSModel::InitialExec)
4876 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
4877 PseudoSourceValue::getGOT(), 0);
4879 // The address of the thread local variable is the add of the thread
4880 // pointer with the offset of the variable.
4881 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
4885 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4886 // TODO: implement the "local dynamic" model
4887 // TODO: implement the "initial exec"model for pic executables
4888 assert(Subtarget->isTargetELF() &&
4889 "TLS not implemented for non-ELF targets");
4890 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4891 const GlobalValue *GV = GA->getGlobal();
4893 // If GV is an alias then use the aliasee for determining
4894 // thread-localness.
4895 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
4896 GV = GA->resolveAliasedGlobal(false);
4898 TLSModel::Model model = getTLSModel(GV,
4899 getTargetMachine().getRelocationModel());
4902 case TLSModel::GeneralDynamic:
4903 case TLSModel::LocalDynamic: // not implemented
4904 if (Subtarget->is64Bit())
4905 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4906 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4908 case TLSModel::InitialExec:
4909 case TLSModel::LocalExec:
4910 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
4911 Subtarget->is64Bit());
4914 llvm_unreachable("Unreachable");
4919 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4920 /// take a 2 x i32 value to shift plus a shift amount.
4921 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4922 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4923 EVT VT = Op.getValueType();
4924 unsigned VTBits = VT.getSizeInBits();
4925 DebugLoc dl = Op.getDebugLoc();
4926 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4927 SDValue ShOpLo = Op.getOperand(0);
4928 SDValue ShOpHi = Op.getOperand(1);
4929 SDValue ShAmt = Op.getOperand(2);
4930 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
4931 DAG.getConstant(VTBits - 1, MVT::i8))
4932 : DAG.getConstant(0, VT);
4935 if (Op.getOpcode() == ISD::SHL_PARTS) {
4936 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
4937 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4939 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
4940 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
4943 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
4944 DAG.getConstant(VTBits, MVT::i8));
4945 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
4946 AndNode, DAG.getConstant(0, MVT::i8));
4949 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4950 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4951 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4953 if (Op.getOpcode() == ISD::SHL_PARTS) {
4954 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4955 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4957 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4958 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4961 SDValue Ops[2] = { Lo, Hi };
4962 return DAG.getMergeValues(Ops, 2, dl);
4965 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4966 EVT SrcVT = Op.getOperand(0).getValueType();
4968 if (SrcVT.isVector()) {
4969 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
4975 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4976 "Unknown SINT_TO_FP to lower!");
4978 // These are really Legal; return the operand so the caller accepts it as
4980 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4982 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
4983 Subtarget->is64Bit()) {
4987 DebugLoc dl = Op.getDebugLoc();
4988 unsigned Size = SrcVT.getSizeInBits()/8;
4989 MachineFunction &MF = DAG.getMachineFunction();
4990 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
4991 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4992 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
4994 PseudoSourceValue::getFixedStack(SSFI), 0);
4995 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
4998 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5000 SelectionDAG &DAG) {
5002 DebugLoc dl = Op.getDebugLoc();
5004 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5006 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5008 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5009 SmallVector<SDValue, 8> Ops;
5010 Ops.push_back(Chain);
5011 Ops.push_back(StackSlot);
5012 Ops.push_back(DAG.getValueType(SrcVT));
5013 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5014 Tys, &Ops[0], Ops.size());
5017 Chain = Result.getValue(1);
5018 SDValue InFlag = Result.getValue(2);
5020 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5021 // shouldn't be necessary except that RFP cannot be live across
5022 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5023 MachineFunction &MF = DAG.getMachineFunction();
5024 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5025 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5026 Tys = DAG.getVTList(MVT::Other);
5027 SmallVector<SDValue, 8> Ops;
5028 Ops.push_back(Chain);
5029 Ops.push_back(Result);
5030 Ops.push_back(StackSlot);
5031 Ops.push_back(DAG.getValueType(Op.getValueType()));
5032 Ops.push_back(InFlag);
5033 Chain = DAG.getNode(X86ISD::FST, dl, Tys, &Ops[0], Ops.size());
5034 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5035 PseudoSourceValue::getFixedStack(SSFI), 0);
5041 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5042 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5043 // This algorithm is not obvious. Here it is in C code, more or less:
5045 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5046 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5047 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5049 // Copy ints to xmm registers.
5050 __m128i xh = _mm_cvtsi32_si128( hi );
5051 __m128i xl = _mm_cvtsi32_si128( lo );
5053 // Combine into low half of a single xmm register.
5054 __m128i x = _mm_unpacklo_epi32( xh, xl );
5058 // Merge in appropriate exponents to give the integer bits the right
5060 x = _mm_unpacklo_epi32( x, exp );
5062 // Subtract away the biases to deal with the IEEE-754 double precision
5064 d = _mm_sub_pd( (__m128d) x, bias );
5066 // All conversions up to here are exact. The correctly rounded result is
5067 // calculated using the current rounding mode using the following
5069 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5070 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5071 // store doesn't really need to be here (except
5072 // maybe to zero the other double)
5077 DebugLoc dl = Op.getDebugLoc();
5078 LLVMContext *Context = DAG.getContext();
5080 // Build some magic constants.
5081 std::vector<Constant*> CV0;
5082 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5083 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5084 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5085 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5086 Constant *C0 = ConstantVector::get(CV0);
5087 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5089 std::vector<Constant*> CV1;
5091 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5093 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5094 Constant *C1 = ConstantVector::get(CV1);
5095 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5097 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5098 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5100 DAG.getIntPtrConstant(1)));
5101 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5102 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5104 DAG.getIntPtrConstant(0)));
5105 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5106 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5107 PseudoSourceValue::getConstantPool(), 0,
5109 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5110 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5111 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5112 PseudoSourceValue::getConstantPool(), 0,
5114 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5116 // Add the halves; easiest way is to swap them into another reg first.
5117 int ShufMask[2] = { 1, -1 };
5118 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5119 DAG.getUNDEF(MVT::v2f64), ShufMask);
5120 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5121 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5122 DAG.getIntPtrConstant(0));
5125 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5126 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5127 DebugLoc dl = Op.getDebugLoc();
5128 // FP constant to bias correct the final result.
5129 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5132 // Load the 32-bit value into an XMM register.
5133 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5134 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5136 DAG.getIntPtrConstant(0)));
5138 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5139 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5140 DAG.getIntPtrConstant(0));
5142 // Or the load with the bias.
5143 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5144 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5145 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5147 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5148 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5149 MVT::v2f64, Bias)));
5150 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5151 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5152 DAG.getIntPtrConstant(0));
5154 // Subtract the bias.
5155 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5157 // Handle final rounding.
5158 EVT DestVT = Op.getValueType();
5160 if (DestVT.bitsLT(MVT::f64)) {
5161 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5162 DAG.getIntPtrConstant(0));
5163 } else if (DestVT.bitsGT(MVT::f64)) {
5164 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5167 // Handle final rounding.
5171 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5172 SDValue N0 = Op.getOperand(0);
5173 DebugLoc dl = Op.getDebugLoc();
5175 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5176 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5177 // the optimization here.
5178 if (DAG.SignBitIsZero(N0))
5179 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5181 EVT SrcVT = N0.getValueType();
5182 if (SrcVT == MVT::i64) {
5183 // We only handle SSE2 f64 target here; caller can expand the rest.
5184 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5187 return LowerUINT_TO_FP_i64(Op, DAG);
5188 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5189 return LowerUINT_TO_FP_i32(Op, DAG);
5192 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5194 // Make a 64-bit buffer, and use it to build an FILD.
5195 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5196 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5197 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5198 getPointerTy(), StackSlot, WordOff);
5199 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5200 StackSlot, NULL, 0);
5201 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5202 OffsetSlot, NULL, 0);
5203 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5206 std::pair<SDValue,SDValue> X86TargetLowering::
5207 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5208 DebugLoc dl = Op.getDebugLoc();
5210 EVT DstTy = Op.getValueType();
5213 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5217 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5218 DstTy.getSimpleVT() >= MVT::i16 &&
5219 "Unknown FP_TO_SINT to lower!");
5221 // These are really Legal.
5222 if (DstTy == MVT::i32 &&
5223 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5224 return std::make_pair(SDValue(), SDValue());
5225 if (Subtarget->is64Bit() &&
5226 DstTy == MVT::i64 &&
5227 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5228 return std::make_pair(SDValue(), SDValue());
5230 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5232 MachineFunction &MF = DAG.getMachineFunction();
5233 unsigned MemSize = DstTy.getSizeInBits()/8;
5234 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5235 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5238 switch (DstTy.getSimpleVT().SimpleTy) {
5239 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5240 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5241 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5242 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5245 SDValue Chain = DAG.getEntryNode();
5246 SDValue Value = Op.getOperand(0);
5247 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5248 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5249 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5250 PseudoSourceValue::getFixedStack(SSFI), 0);
5251 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5253 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5255 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5256 Chain = Value.getValue(1);
5257 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5258 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5261 // Build the FP_TO_INT*_IN_MEM
5262 SDValue Ops[] = { Chain, Value, StackSlot };
5263 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5265 return std::make_pair(FIST, StackSlot);
5268 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5269 if (Op.getValueType().isVector()) {
5270 if (Op.getValueType() == MVT::v2i32 &&
5271 Op.getOperand(0).getValueType() == MVT::v2f64) {
5277 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5278 SDValue FIST = Vals.first, StackSlot = Vals.second;
5279 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5280 if (FIST.getNode() == 0) return Op;
5283 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5284 FIST, StackSlot, NULL, 0);
5287 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5288 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5289 SDValue FIST = Vals.first, StackSlot = Vals.second;
5290 assert(FIST.getNode() && "Unexpected failure");
5293 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5294 FIST, StackSlot, NULL, 0);
5297 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5298 LLVMContext *Context = DAG.getContext();
5299 DebugLoc dl = Op.getDebugLoc();
5300 EVT VT = Op.getValueType();
5303 EltVT = VT.getVectorElementType();
5304 std::vector<Constant*> CV;
5305 if (EltVT == MVT::f64) {
5306 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5310 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5316 Constant *C = ConstantVector::get(CV);
5317 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5318 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5319 PseudoSourceValue::getConstantPool(), 0,
5321 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5324 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5325 LLVMContext *Context = DAG.getContext();
5326 DebugLoc dl = Op.getDebugLoc();
5327 EVT VT = Op.getValueType();
5330 EltVT = VT.getVectorElementType();
5331 std::vector<Constant*> CV;
5332 if (EltVT == MVT::f64) {
5333 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5337 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5343 Constant *C = ConstantVector::get(CV);
5344 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5345 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5346 PseudoSourceValue::getConstantPool(), 0,
5348 if (VT.isVector()) {
5349 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5350 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5351 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5353 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5355 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5359 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5360 LLVMContext *Context = DAG.getContext();
5361 SDValue Op0 = Op.getOperand(0);
5362 SDValue Op1 = Op.getOperand(1);
5363 DebugLoc dl = Op.getDebugLoc();
5364 EVT VT = Op.getValueType();
5365 EVT SrcVT = Op1.getValueType();
5367 // If second operand is smaller, extend it first.
5368 if (SrcVT.bitsLT(VT)) {
5369 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5372 // And if it is bigger, shrink it first.
5373 if (SrcVT.bitsGT(VT)) {
5374 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5378 // At this point the operands and the result should have the same
5379 // type, and that won't be f80 since that is not custom lowered.
5381 // First get the sign bit of second operand.
5382 std::vector<Constant*> CV;
5383 if (SrcVT == MVT::f64) {
5384 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5385 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5387 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5388 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5389 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5390 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5392 Constant *C = ConstantVector::get(CV);
5393 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5394 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5395 PseudoSourceValue::getConstantPool(), 0,
5397 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5399 // Shift sign bit right or left if the two operands have different types.
5400 if (SrcVT.bitsGT(VT)) {
5401 // Op0 is MVT::f32, Op1 is MVT::f64.
5402 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5403 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5404 DAG.getConstant(32, MVT::i32));
5405 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5406 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5407 DAG.getIntPtrConstant(0));
5410 // Clear first operand sign bit.
5412 if (VT == MVT::f64) {
5413 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5414 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5416 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5417 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5418 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5419 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5421 C = ConstantVector::get(CV);
5422 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5423 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5424 PseudoSourceValue::getConstantPool(), 0,
5426 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5428 // Or the value with the sign bit.
5429 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5432 /// Emit nodes that will be selected as "test Op0,Op0", or something
5434 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5435 SelectionDAG &DAG) {
5436 DebugLoc dl = Op.getDebugLoc();
5438 // CF and OF aren't always set the way we want. Determine which
5439 // of these we need.
5440 bool NeedCF = false;
5441 bool NeedOF = false;
5443 case X86::COND_A: case X86::COND_AE:
5444 case X86::COND_B: case X86::COND_BE:
5447 case X86::COND_G: case X86::COND_GE:
5448 case X86::COND_L: case X86::COND_LE:
5449 case X86::COND_O: case X86::COND_NO:
5455 // See if we can use the EFLAGS value from the operand instead of
5456 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5457 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5458 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5459 unsigned Opcode = 0;
5460 unsigned NumOperands = 0;
5461 switch (Op.getNode()->getOpcode()) {
5463 // Due to an isel shortcoming, be conservative if this add is likely to
5464 // be selected as part of a load-modify-store instruction. When the root
5465 // node in a match is a store, isel doesn't know how to remap non-chain
5466 // non-flag uses of other nodes in the match, such as the ADD in this
5467 // case. This leads to the ADD being left around and reselected, with
5468 // the result being two adds in the output.
5469 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5470 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5471 if (UI->getOpcode() == ISD::STORE)
5473 if (ConstantSDNode *C =
5474 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5475 // An add of one will be selected as an INC.
5476 if (C->getAPIntValue() == 1) {
5477 Opcode = X86ISD::INC;
5481 // An add of negative one (subtract of one) will be selected as a DEC.
5482 if (C->getAPIntValue().isAllOnesValue()) {
5483 Opcode = X86ISD::DEC;
5488 // Otherwise use a regular EFLAGS-setting add.
5489 Opcode = X86ISD::ADD;
5493 // If the primary and result isn't used, don't bother using X86ISD::AND,
5494 // because a TEST instruction will be better.
5495 bool NonFlagUse = false;
5496 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5497 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5498 if (UI->getOpcode() != ISD::BRCOND &&
5499 UI->getOpcode() != ISD::SELECT &&
5500 UI->getOpcode() != ISD::SETCC) {
5511 // Due to the ISEL shortcoming noted above, be conservative if this op is
5512 // likely to be selected as part of a load-modify-store instruction.
5513 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5514 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5515 if (UI->getOpcode() == ISD::STORE)
5517 // Otherwise use a regular EFLAGS-setting instruction.
5518 switch (Op.getNode()->getOpcode()) {
5519 case ISD::SUB: Opcode = X86ISD::SUB; break;
5520 case ISD::OR: Opcode = X86ISD::OR; break;
5521 case ISD::XOR: Opcode = X86ISD::XOR; break;
5522 case ISD::AND: Opcode = X86ISD::AND; break;
5523 default: llvm_unreachable("unexpected operator!");
5534 return SDValue(Op.getNode(), 1);
5540 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5541 SmallVector<SDValue, 4> Ops;
5542 for (unsigned i = 0; i != NumOperands; ++i)
5543 Ops.push_back(Op.getOperand(i));
5544 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5545 DAG.ReplaceAllUsesWith(Op, New);
5546 return SDValue(New.getNode(), 1);
5550 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5551 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5552 DAG.getConstant(0, Op.getValueType()));
5555 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5557 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5558 SelectionDAG &DAG) {
5559 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5560 if (C->getAPIntValue() == 0)
5561 return EmitTest(Op0, X86CC, DAG);
5563 DebugLoc dl = Op0.getDebugLoc();
5564 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5567 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5568 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5569 SDValue Op0 = Op.getOperand(0);
5570 SDValue Op1 = Op.getOperand(1);
5571 DebugLoc dl = Op.getDebugLoc();
5572 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5574 // Lower (X & (1 << N)) == 0 to BT(X, N).
5575 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5576 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5577 if (Op0.getOpcode() == ISD::AND &&
5579 Op1.getOpcode() == ISD::Constant &&
5580 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5581 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5583 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5584 if (ConstantSDNode *Op010C =
5585 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5586 if (Op010C->getZExtValue() == 1) {
5587 LHS = Op0.getOperand(0);
5588 RHS = Op0.getOperand(1).getOperand(1);
5590 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5591 if (ConstantSDNode *Op000C =
5592 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5593 if (Op000C->getZExtValue() == 1) {
5594 LHS = Op0.getOperand(1);
5595 RHS = Op0.getOperand(0).getOperand(1);
5597 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5598 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5599 SDValue AndLHS = Op0.getOperand(0);
5600 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5601 LHS = AndLHS.getOperand(0);
5602 RHS = AndLHS.getOperand(1);
5606 if (LHS.getNode()) {
5607 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5608 // instruction. Since the shift amount is in-range-or-undefined, we know
5609 // that doing a bittest on the i16 value is ok. We extend to i32 because
5610 // the encoding for the i16 version is larger than the i32 version.
5611 if (LHS.getValueType() == MVT::i8)
5612 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5614 // If the operand types disagree, extend the shift amount to match. Since
5615 // BT ignores high bits (like shifts) we can use anyextend.
5616 if (LHS.getValueType() != RHS.getValueType())
5617 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5619 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5620 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5621 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5622 DAG.getConstant(Cond, MVT::i8), BT);
5626 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5627 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5628 if (X86CC == X86::COND_INVALID)
5631 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5632 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5633 DAG.getConstant(X86CC, MVT::i8), Cond);
5636 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5638 SDValue Op0 = Op.getOperand(0);
5639 SDValue Op1 = Op.getOperand(1);
5640 SDValue CC = Op.getOperand(2);
5641 EVT VT = Op.getValueType();
5642 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5643 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5644 DebugLoc dl = Op.getDebugLoc();
5648 EVT VT0 = Op0.getValueType();
5649 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5650 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5653 switch (SetCCOpcode) {
5656 case ISD::SETEQ: SSECC = 0; break;
5658 case ISD::SETGT: Swap = true; // Fallthrough
5660 case ISD::SETOLT: SSECC = 1; break;
5662 case ISD::SETGE: Swap = true; // Fallthrough
5664 case ISD::SETOLE: SSECC = 2; break;
5665 case ISD::SETUO: SSECC = 3; break;
5667 case ISD::SETNE: SSECC = 4; break;
5668 case ISD::SETULE: Swap = true;
5669 case ISD::SETUGE: SSECC = 5; break;
5670 case ISD::SETULT: Swap = true;
5671 case ISD::SETUGT: SSECC = 6; break;
5672 case ISD::SETO: SSECC = 7; break;
5675 std::swap(Op0, Op1);
5677 // In the two special cases we can't handle, emit two comparisons.
5679 if (SetCCOpcode == ISD::SETUEQ) {
5681 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5682 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5683 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5685 else if (SetCCOpcode == ISD::SETONE) {
5687 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5688 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5689 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5691 llvm_unreachable("Illegal FP comparison");
5693 // Handle all other FP comparisons here.
5694 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5697 // We are handling one of the integer comparisons here. Since SSE only has
5698 // GT and EQ comparisons for integer, swapping operands and multiple
5699 // operations may be required for some comparisons.
5700 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5701 bool Swap = false, Invert = false, FlipSigns = false;
5703 switch (VT.getSimpleVT().SimpleTy) {
5706 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5708 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5710 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5711 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5714 switch (SetCCOpcode) {
5716 case ISD::SETNE: Invert = true;
5717 case ISD::SETEQ: Opc = EQOpc; break;
5718 case ISD::SETLT: Swap = true;
5719 case ISD::SETGT: Opc = GTOpc; break;
5720 case ISD::SETGE: Swap = true;
5721 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5722 case ISD::SETULT: Swap = true;
5723 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5724 case ISD::SETUGE: Swap = true;
5725 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5728 std::swap(Op0, Op1);
5730 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5731 // bits of the inputs before performing those operations.
5733 EVT EltVT = VT.getVectorElementType();
5734 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5736 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5737 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5739 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5740 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5743 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5745 // If the logical-not of the result is required, perform that now.
5747 Result = DAG.getNOT(dl, Result, VT);
5752 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5753 static bool isX86LogicalCmp(SDValue Op) {
5754 unsigned Opc = Op.getNode()->getOpcode();
5755 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
5757 if (Op.getResNo() == 1 &&
5758 (Opc == X86ISD::ADD ||
5759 Opc == X86ISD::SUB ||
5760 Opc == X86ISD::SMUL ||
5761 Opc == X86ISD::UMUL ||
5762 Opc == X86ISD::INC ||
5763 Opc == X86ISD::DEC ||
5764 Opc == X86ISD::OR ||
5765 Opc == X86ISD::XOR ||
5766 Opc == X86ISD::AND))
5772 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5773 bool addTest = true;
5774 SDValue Cond = Op.getOperand(0);
5775 DebugLoc dl = Op.getDebugLoc();
5778 if (Cond.getOpcode() == ISD::SETCC) {
5779 SDValue NewCond = LowerSETCC(Cond, DAG);
5780 if (NewCond.getNode())
5784 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5785 // setting operand in place of the X86ISD::SETCC.
5786 if (Cond.getOpcode() == X86ISD::SETCC) {
5787 CC = Cond.getOperand(0);
5789 SDValue Cmp = Cond.getOperand(1);
5790 unsigned Opc = Cmp.getOpcode();
5791 EVT VT = Op.getValueType();
5793 bool IllegalFPCMov = false;
5794 if (VT.isFloatingPoint() && !VT.isVector() &&
5795 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5796 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5798 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
5799 Opc == X86ISD::BT) { // FIXME
5806 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5807 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5810 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
5811 SmallVector<SDValue, 4> Ops;
5812 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5813 // condition is true.
5814 Ops.push_back(Op.getOperand(2));
5815 Ops.push_back(Op.getOperand(1));
5817 Ops.push_back(Cond);
5818 return DAG.getNode(X86ISD::CMOV, dl, VTs, &Ops[0], Ops.size());
5821 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
5822 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
5823 // from the AND / OR.
5824 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
5825 Opc = Op.getOpcode();
5826 if (Opc != ISD::OR && Opc != ISD::AND)
5828 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5829 Op.getOperand(0).hasOneUse() &&
5830 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
5831 Op.getOperand(1).hasOneUse());
5834 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
5835 // 1 and that the SETCC node has a single use.
5836 static bool isXor1OfSetCC(SDValue Op) {
5837 if (Op.getOpcode() != ISD::XOR)
5839 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5840 if (N1C && N1C->getAPIntValue() == 1) {
5841 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5842 Op.getOperand(0).hasOneUse();
5847 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5848 bool addTest = true;
5849 SDValue Chain = Op.getOperand(0);
5850 SDValue Cond = Op.getOperand(1);
5851 SDValue Dest = Op.getOperand(2);
5852 DebugLoc dl = Op.getDebugLoc();
5855 if (Cond.getOpcode() == ISD::SETCC) {
5856 SDValue NewCond = LowerSETCC(Cond, DAG);
5857 if (NewCond.getNode())
5861 // FIXME: LowerXALUO doesn't handle these!!
5862 else if (Cond.getOpcode() == X86ISD::ADD ||
5863 Cond.getOpcode() == X86ISD::SUB ||
5864 Cond.getOpcode() == X86ISD::SMUL ||
5865 Cond.getOpcode() == X86ISD::UMUL)
5866 Cond = LowerXALUO(Cond, DAG);
5869 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5870 // setting operand in place of the X86ISD::SETCC.
5871 if (Cond.getOpcode() == X86ISD::SETCC) {
5872 CC = Cond.getOperand(0);
5874 SDValue Cmp = Cond.getOperand(1);
5875 unsigned Opc = Cmp.getOpcode();
5876 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
5877 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
5881 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
5885 // These can only come from an arithmetic instruction with overflow,
5886 // e.g. SADDO, UADDO.
5887 Cond = Cond.getNode()->getOperand(1);
5894 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
5895 SDValue Cmp = Cond.getOperand(0).getOperand(1);
5896 if (CondOpc == ISD::OR) {
5897 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
5898 // two branches instead of an explicit OR instruction with a
5900 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5901 isX86LogicalCmp(Cmp)) {
5902 CC = Cond.getOperand(0).getOperand(0);
5903 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5904 Chain, Dest, CC, Cmp);
5905 CC = Cond.getOperand(1).getOperand(0);
5909 } else { // ISD::AND
5910 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
5911 // two branches instead of an explicit AND instruction with a
5912 // separate test. However, we only do this if this block doesn't
5913 // have a fall-through edge, because this requires an explicit
5914 // jmp when the condition is false.
5915 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5916 isX86LogicalCmp(Cmp) &&
5917 Op.getNode()->hasOneUse()) {
5918 X86::CondCode CCode =
5919 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5920 CCode = X86::GetOppositeBranchCondition(CCode);
5921 CC = DAG.getConstant(CCode, MVT::i8);
5922 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
5923 // Look for an unconditional branch following this conditional branch.
5924 // We need this because we need to reverse the successors in order
5925 // to implement FCMP_OEQ.
5926 if (User.getOpcode() == ISD::BR) {
5927 SDValue FalseBB = User.getOperand(1);
5929 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
5930 assert(NewBR == User);
5933 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5934 Chain, Dest, CC, Cmp);
5935 X86::CondCode CCode =
5936 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
5937 CCode = X86::GetOppositeBranchCondition(CCode);
5938 CC = DAG.getConstant(CCode, MVT::i8);
5944 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
5945 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
5946 // It should be transformed during dag combiner except when the condition
5947 // is set by a arithmetics with overflow node.
5948 X86::CondCode CCode =
5949 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5950 CCode = X86::GetOppositeBranchCondition(CCode);
5951 CC = DAG.getConstant(CCode, MVT::i8);
5952 Cond = Cond.getOperand(0).getOperand(1);
5958 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5959 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5961 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5962 Chain, Dest, CC, Cond);
5966 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5967 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5968 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5969 // that the guard pages used by the OS virtual memory manager are allocated in
5970 // correct sequence.
5972 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5973 SelectionDAG &DAG) {
5974 assert(Subtarget->isTargetCygMing() &&
5975 "This should be used only on Cygwin/Mingw targets");
5976 DebugLoc dl = Op.getDebugLoc();
5979 SDValue Chain = Op.getOperand(0);
5980 SDValue Size = Op.getOperand(1);
5981 // FIXME: Ensure alignment here
5985 EVT IntPtr = getPointerTy();
5986 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5988 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
5990 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
5991 Flag = Chain.getValue(1);
5993 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5994 SDValue Ops[] = { Chain,
5995 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5996 DAG.getRegister(X86::EAX, IntPtr),
5997 DAG.getRegister(X86StackPtr, SPTy),
5999 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
6000 Flag = Chain.getValue(1);
6002 Chain = DAG.getCALLSEQ_END(Chain,
6003 DAG.getIntPtrConstant(0, true),
6004 DAG.getIntPtrConstant(0, true),
6007 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6009 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6010 return DAG.getMergeValues(Ops1, 2, dl);
6014 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6016 SDValue Dst, SDValue Src,
6017 SDValue Size, unsigned Align,
6019 uint64_t DstSVOff) {
6020 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6022 // If not DWORD aligned or size is more than the threshold, call the library.
6023 // The libc version is likely to be faster for these cases. It can use the
6024 // address value and run time information about the CPU.
6025 if ((Align & 3) != 0 ||
6027 ConstantSize->getZExtValue() >
6028 getSubtarget()->getMaxInlineSizeThreshold()) {
6029 SDValue InFlag(0, 0);
6031 // Check to see if there is a specialized entry-point for memory zeroing.
6032 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6034 if (const char *bzeroEntry = V &&
6035 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6036 EVT IntPtr = getPointerTy();
6037 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6038 TargetLowering::ArgListTy Args;
6039 TargetLowering::ArgListEntry Entry;
6041 Entry.Ty = IntPtrTy;
6042 Args.push_back(Entry);
6044 Args.push_back(Entry);
6045 std::pair<SDValue,SDValue> CallResult =
6046 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6047 false, false, false, false,
6048 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6049 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
6050 return CallResult.second;
6053 // Otherwise have the target-independent code call memset.
6057 uint64_t SizeVal = ConstantSize->getZExtValue();
6058 SDValue InFlag(0, 0);
6061 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6062 unsigned BytesLeft = 0;
6063 bool TwoRepStos = false;
6066 uint64_t Val = ValC->getZExtValue() & 255;
6068 // If the value is a constant, then we can potentially use larger sets.
6069 switch (Align & 3) {
6070 case 2: // WORD aligned
6073 Val = (Val << 8) | Val;
6075 case 0: // DWORD aligned
6078 Val = (Val << 8) | Val;
6079 Val = (Val << 16) | Val;
6080 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6083 Val = (Val << 32) | Val;
6086 default: // Byte aligned
6089 Count = DAG.getIntPtrConstant(SizeVal);
6093 if (AVT.bitsGT(MVT::i8)) {
6094 unsigned UBytes = AVT.getSizeInBits() / 8;
6095 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6096 BytesLeft = SizeVal % UBytes;
6099 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6101 InFlag = Chain.getValue(1);
6104 Count = DAG.getIntPtrConstant(SizeVal);
6105 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6106 InFlag = Chain.getValue(1);
6109 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6112 InFlag = Chain.getValue(1);
6113 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6116 InFlag = Chain.getValue(1);
6118 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6119 SmallVector<SDValue, 8> Ops;
6120 Ops.push_back(Chain);
6121 Ops.push_back(DAG.getValueType(AVT));
6122 Ops.push_back(InFlag);
6123 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
6126 InFlag = Chain.getValue(1);
6128 EVT CVT = Count.getValueType();
6129 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6130 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6131 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6134 InFlag = Chain.getValue(1);
6135 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6137 Ops.push_back(Chain);
6138 Ops.push_back(DAG.getValueType(MVT::i8));
6139 Ops.push_back(InFlag);
6140 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
6141 } else if (BytesLeft) {
6142 // Handle the last 1 - 7 bytes.
6143 unsigned Offset = SizeVal - BytesLeft;
6144 EVT AddrVT = Dst.getValueType();
6145 EVT SizeVT = Size.getValueType();
6147 Chain = DAG.getMemset(Chain, dl,
6148 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6149 DAG.getConstant(Offset, AddrVT)),
6151 DAG.getConstant(BytesLeft, SizeVT),
6152 Align, DstSV, DstSVOff + Offset);
6155 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6160 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6161 SDValue Chain, SDValue Dst, SDValue Src,
6162 SDValue Size, unsigned Align,
6164 const Value *DstSV, uint64_t DstSVOff,
6165 const Value *SrcSV, uint64_t SrcSVOff) {
6166 // This requires the copy size to be a constant, preferrably
6167 // within a subtarget-specific limit.
6168 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6171 uint64_t SizeVal = ConstantSize->getZExtValue();
6172 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6175 /// If not DWORD aligned, call the library.
6176 if ((Align & 3) != 0)
6181 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6184 unsigned UBytes = AVT.getSizeInBits() / 8;
6185 unsigned CountVal = SizeVal / UBytes;
6186 SDValue Count = DAG.getIntPtrConstant(CountVal);
6187 unsigned BytesLeft = SizeVal % UBytes;
6189 SDValue InFlag(0, 0);
6190 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6193 InFlag = Chain.getValue(1);
6194 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6197 InFlag = Chain.getValue(1);
6198 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6201 InFlag = Chain.getValue(1);
6203 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6204 SmallVector<SDValue, 8> Ops;
6205 Ops.push_back(Chain);
6206 Ops.push_back(DAG.getValueType(AVT));
6207 Ops.push_back(InFlag);
6208 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, &Ops[0], Ops.size());
6210 SmallVector<SDValue, 4> Results;
6211 Results.push_back(RepMovs);
6213 // Handle the last 1 - 7 bytes.
6214 unsigned Offset = SizeVal - BytesLeft;
6215 EVT DstVT = Dst.getValueType();
6216 EVT SrcVT = Src.getValueType();
6217 EVT SizeVT = Size.getValueType();
6218 Results.push_back(DAG.getMemcpy(Chain, dl,
6219 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6220 DAG.getConstant(Offset, DstVT)),
6221 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6222 DAG.getConstant(Offset, SrcVT)),
6223 DAG.getConstant(BytesLeft, SizeVT),
6224 Align, AlwaysInline,
6225 DstSV, DstSVOff + Offset,
6226 SrcSV, SrcSVOff + Offset));
6229 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6230 &Results[0], Results.size());
6233 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6234 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6235 DebugLoc dl = Op.getDebugLoc();
6237 if (!Subtarget->is64Bit()) {
6238 // vastart just stores the address of the VarArgsFrameIndex slot into the
6239 // memory location argument.
6240 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6241 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6245 // gp_offset (0 - 6 * 8)
6246 // fp_offset (48 - 48 + 8 * 16)
6247 // overflow_arg_area (point to parameters coming in memory).
6249 SmallVector<SDValue, 8> MemOps;
6250 SDValue FIN = Op.getOperand(1);
6252 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6253 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6255 MemOps.push_back(Store);
6258 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6259 FIN, DAG.getIntPtrConstant(4));
6260 Store = DAG.getStore(Op.getOperand(0), dl,
6261 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6263 MemOps.push_back(Store);
6265 // Store ptr to overflow_arg_area
6266 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6267 FIN, DAG.getIntPtrConstant(4));
6268 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6269 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6270 MemOps.push_back(Store);
6272 // Store ptr to reg_save_area.
6273 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6274 FIN, DAG.getIntPtrConstant(8));
6275 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6276 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6277 MemOps.push_back(Store);
6278 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6279 &MemOps[0], MemOps.size());
6282 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6283 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6284 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6285 SDValue Chain = Op.getOperand(0);
6286 SDValue SrcPtr = Op.getOperand(1);
6287 SDValue SrcSV = Op.getOperand(2);
6289 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6293 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6294 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6295 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6296 SDValue Chain = Op.getOperand(0);
6297 SDValue DstPtr = Op.getOperand(1);
6298 SDValue SrcPtr = Op.getOperand(2);
6299 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6300 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6301 DebugLoc dl = Op.getDebugLoc();
6303 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6304 DAG.getIntPtrConstant(24), 8, false,
6305 DstSV, 0, SrcSV, 0);
6309 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6310 DebugLoc dl = Op.getDebugLoc();
6311 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6313 default: return SDValue(); // Don't custom lower most intrinsics.
6314 // Comparison intrinsics.
6315 case Intrinsic::x86_sse_comieq_ss:
6316 case Intrinsic::x86_sse_comilt_ss:
6317 case Intrinsic::x86_sse_comile_ss:
6318 case Intrinsic::x86_sse_comigt_ss:
6319 case Intrinsic::x86_sse_comige_ss:
6320 case Intrinsic::x86_sse_comineq_ss:
6321 case Intrinsic::x86_sse_ucomieq_ss:
6322 case Intrinsic::x86_sse_ucomilt_ss:
6323 case Intrinsic::x86_sse_ucomile_ss:
6324 case Intrinsic::x86_sse_ucomigt_ss:
6325 case Intrinsic::x86_sse_ucomige_ss:
6326 case Intrinsic::x86_sse_ucomineq_ss:
6327 case Intrinsic::x86_sse2_comieq_sd:
6328 case Intrinsic::x86_sse2_comilt_sd:
6329 case Intrinsic::x86_sse2_comile_sd:
6330 case Intrinsic::x86_sse2_comigt_sd:
6331 case Intrinsic::x86_sse2_comige_sd:
6332 case Intrinsic::x86_sse2_comineq_sd:
6333 case Intrinsic::x86_sse2_ucomieq_sd:
6334 case Intrinsic::x86_sse2_ucomilt_sd:
6335 case Intrinsic::x86_sse2_ucomile_sd:
6336 case Intrinsic::x86_sse2_ucomigt_sd:
6337 case Intrinsic::x86_sse2_ucomige_sd:
6338 case Intrinsic::x86_sse2_ucomineq_sd: {
6340 ISD::CondCode CC = ISD::SETCC_INVALID;
6343 case Intrinsic::x86_sse_comieq_ss:
6344 case Intrinsic::x86_sse2_comieq_sd:
6348 case Intrinsic::x86_sse_comilt_ss:
6349 case Intrinsic::x86_sse2_comilt_sd:
6353 case Intrinsic::x86_sse_comile_ss:
6354 case Intrinsic::x86_sse2_comile_sd:
6358 case Intrinsic::x86_sse_comigt_ss:
6359 case Intrinsic::x86_sse2_comigt_sd:
6363 case Intrinsic::x86_sse_comige_ss:
6364 case Intrinsic::x86_sse2_comige_sd:
6368 case Intrinsic::x86_sse_comineq_ss:
6369 case Intrinsic::x86_sse2_comineq_sd:
6373 case Intrinsic::x86_sse_ucomieq_ss:
6374 case Intrinsic::x86_sse2_ucomieq_sd:
6375 Opc = X86ISD::UCOMI;
6378 case Intrinsic::x86_sse_ucomilt_ss:
6379 case Intrinsic::x86_sse2_ucomilt_sd:
6380 Opc = X86ISD::UCOMI;
6383 case Intrinsic::x86_sse_ucomile_ss:
6384 case Intrinsic::x86_sse2_ucomile_sd:
6385 Opc = X86ISD::UCOMI;
6388 case Intrinsic::x86_sse_ucomigt_ss:
6389 case Intrinsic::x86_sse2_ucomigt_sd:
6390 Opc = X86ISD::UCOMI;
6393 case Intrinsic::x86_sse_ucomige_ss:
6394 case Intrinsic::x86_sse2_ucomige_sd:
6395 Opc = X86ISD::UCOMI;
6398 case Intrinsic::x86_sse_ucomineq_ss:
6399 case Intrinsic::x86_sse2_ucomineq_sd:
6400 Opc = X86ISD::UCOMI;
6405 SDValue LHS = Op.getOperand(1);
6406 SDValue RHS = Op.getOperand(2);
6407 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6408 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6409 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6410 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6411 DAG.getConstant(X86CC, MVT::i8), Cond);
6412 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6414 // ptest intrinsics. The intrinsic these come from are designed to return
6415 // an integer value, not just an instruction so lower it to the ptest
6416 // pattern and a setcc for the result.
6417 case Intrinsic::x86_sse41_ptestz:
6418 case Intrinsic::x86_sse41_ptestc:
6419 case Intrinsic::x86_sse41_ptestnzc:{
6422 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6423 case Intrinsic::x86_sse41_ptestz:
6425 X86CC = X86::COND_E;
6427 case Intrinsic::x86_sse41_ptestc:
6429 X86CC = X86::COND_B;
6431 case Intrinsic::x86_sse41_ptestnzc:
6433 X86CC = X86::COND_A;
6437 SDValue LHS = Op.getOperand(1);
6438 SDValue RHS = Op.getOperand(2);
6439 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6440 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6441 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6442 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6445 // Fix vector shift instructions where the last operand is a non-immediate
6447 case Intrinsic::x86_sse2_pslli_w:
6448 case Intrinsic::x86_sse2_pslli_d:
6449 case Intrinsic::x86_sse2_pslli_q:
6450 case Intrinsic::x86_sse2_psrli_w:
6451 case Intrinsic::x86_sse2_psrli_d:
6452 case Intrinsic::x86_sse2_psrli_q:
6453 case Intrinsic::x86_sse2_psrai_w:
6454 case Intrinsic::x86_sse2_psrai_d:
6455 case Intrinsic::x86_mmx_pslli_w:
6456 case Intrinsic::x86_mmx_pslli_d:
6457 case Intrinsic::x86_mmx_pslli_q:
6458 case Intrinsic::x86_mmx_psrli_w:
6459 case Intrinsic::x86_mmx_psrli_d:
6460 case Intrinsic::x86_mmx_psrli_q:
6461 case Intrinsic::x86_mmx_psrai_w:
6462 case Intrinsic::x86_mmx_psrai_d: {
6463 SDValue ShAmt = Op.getOperand(2);
6464 if (isa<ConstantSDNode>(ShAmt))
6467 unsigned NewIntNo = 0;
6468 EVT ShAmtVT = MVT::v4i32;
6470 case Intrinsic::x86_sse2_pslli_w:
6471 NewIntNo = Intrinsic::x86_sse2_psll_w;
6473 case Intrinsic::x86_sse2_pslli_d:
6474 NewIntNo = Intrinsic::x86_sse2_psll_d;
6476 case Intrinsic::x86_sse2_pslli_q:
6477 NewIntNo = Intrinsic::x86_sse2_psll_q;
6479 case Intrinsic::x86_sse2_psrli_w:
6480 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6482 case Intrinsic::x86_sse2_psrli_d:
6483 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6485 case Intrinsic::x86_sse2_psrli_q:
6486 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6488 case Intrinsic::x86_sse2_psrai_w:
6489 NewIntNo = Intrinsic::x86_sse2_psra_w;
6491 case Intrinsic::x86_sse2_psrai_d:
6492 NewIntNo = Intrinsic::x86_sse2_psra_d;
6495 ShAmtVT = MVT::v2i32;
6497 case Intrinsic::x86_mmx_pslli_w:
6498 NewIntNo = Intrinsic::x86_mmx_psll_w;
6500 case Intrinsic::x86_mmx_pslli_d:
6501 NewIntNo = Intrinsic::x86_mmx_psll_d;
6503 case Intrinsic::x86_mmx_pslli_q:
6504 NewIntNo = Intrinsic::x86_mmx_psll_q;
6506 case Intrinsic::x86_mmx_psrli_w:
6507 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6509 case Intrinsic::x86_mmx_psrli_d:
6510 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6512 case Intrinsic::x86_mmx_psrli_q:
6513 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6515 case Intrinsic::x86_mmx_psrai_w:
6516 NewIntNo = Intrinsic::x86_mmx_psra_w;
6518 case Intrinsic::x86_mmx_psrai_d:
6519 NewIntNo = Intrinsic::x86_mmx_psra_d;
6521 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6527 // The vector shift intrinsics with scalars uses 32b shift amounts but
6528 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
6532 ShOps[1] = DAG.getConstant(0, MVT::i32);
6533 if (ShAmtVT == MVT::v4i32) {
6534 ShOps[2] = DAG.getUNDEF(MVT::i32);
6535 ShOps[3] = DAG.getUNDEF(MVT::i32);
6536 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
6538 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
6541 EVT VT = Op.getValueType();
6542 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
6543 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6544 DAG.getConstant(NewIntNo, MVT::i32),
6545 Op.getOperand(1), ShAmt);
6550 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6551 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6552 DebugLoc dl = Op.getDebugLoc();
6555 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6557 DAG.getConstant(TD->getPointerSize(),
6558 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6559 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6560 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6565 // Just load the return address.
6566 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6567 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6568 RetAddrFI, NULL, 0);
6571 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6572 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6573 MFI->setFrameAddressIsTaken(true);
6574 EVT VT = Op.getValueType();
6575 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6576 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6577 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6578 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6580 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6584 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6585 SelectionDAG &DAG) {
6586 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6589 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6591 MachineFunction &MF = DAG.getMachineFunction();
6592 SDValue Chain = Op.getOperand(0);
6593 SDValue Offset = Op.getOperand(1);
6594 SDValue Handler = Op.getOperand(2);
6595 DebugLoc dl = Op.getDebugLoc();
6597 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6599 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6601 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6602 DAG.getIntPtrConstant(-TD->getPointerSize()));
6603 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6604 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6605 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6606 MF.getRegInfo().addLiveOut(StoreAddrReg);
6608 return DAG.getNode(X86ISD::EH_RETURN, dl,
6610 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6613 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6614 SelectionDAG &DAG) {
6615 SDValue Root = Op.getOperand(0);
6616 SDValue Trmp = Op.getOperand(1); // trampoline
6617 SDValue FPtr = Op.getOperand(2); // nested function
6618 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6619 DebugLoc dl = Op.getDebugLoc();
6621 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6623 const X86InstrInfo *TII =
6624 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6626 if (Subtarget->is64Bit()) {
6627 SDValue OutChains[6];
6629 // Large code-model.
6631 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6632 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6634 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6635 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6637 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6639 // Load the pointer to the nested function into R11.
6640 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6641 SDValue Addr = Trmp;
6642 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6645 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6646 DAG.getConstant(2, MVT::i64));
6647 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6649 // Load the 'nest' parameter value into R10.
6650 // R10 is specified in X86CallingConv.td
6651 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6652 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6653 DAG.getConstant(10, MVT::i64));
6654 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6655 Addr, TrmpAddr, 10);
6657 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6658 DAG.getConstant(12, MVT::i64));
6659 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6661 // Jump to the nested function.
6662 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6663 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6664 DAG.getConstant(20, MVT::i64));
6665 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6666 Addr, TrmpAddr, 20);
6668 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6669 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6670 DAG.getConstant(22, MVT::i64));
6671 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6675 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6676 return DAG.getMergeValues(Ops, 2, dl);
6678 const Function *Func =
6679 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6680 CallingConv::ID CC = Func->getCallingConv();
6685 llvm_unreachable("Unsupported calling convention");
6686 case CallingConv::C:
6687 case CallingConv::X86_StdCall: {
6688 // Pass 'nest' parameter in ECX.
6689 // Must be kept in sync with X86CallingConv.td
6692 // Check that ECX wasn't needed by an 'inreg' parameter.
6693 const FunctionType *FTy = Func->getFunctionType();
6694 const AttrListPtr &Attrs = Func->getAttributes();
6696 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6697 unsigned InRegCount = 0;
6700 for (FunctionType::param_iterator I = FTy->param_begin(),
6701 E = FTy->param_end(); I != E; ++I, ++Idx)
6702 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6703 // FIXME: should only count parameters that are lowered to integers.
6704 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6706 if (InRegCount > 2) {
6707 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
6712 case CallingConv::X86_FastCall:
6713 case CallingConv::Fast:
6714 // Pass 'nest' parameter in EAX.
6715 // Must be kept in sync with X86CallingConv.td
6720 SDValue OutChains[4];
6723 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6724 DAG.getConstant(10, MVT::i32));
6725 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
6727 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6728 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6729 OutChains[0] = DAG.getStore(Root, dl,
6730 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6733 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6734 DAG.getConstant(1, MVT::i32));
6735 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
6737 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6738 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6739 DAG.getConstant(5, MVT::i32));
6740 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
6741 TrmpAddr, 5, false, 1);
6743 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6744 DAG.getConstant(6, MVT::i32));
6745 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
6748 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
6749 return DAG.getMergeValues(Ops, 2, dl);
6753 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6755 The rounding mode is in bits 11:10 of FPSR, and has the following
6762 FLT_ROUNDS, on the other hand, expects the following:
6769 To perform the conversion, we do:
6770 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6773 MachineFunction &MF = DAG.getMachineFunction();
6774 const TargetMachine &TM = MF.getTarget();
6775 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6776 unsigned StackAlignment = TFI.getStackAlignment();
6777 EVT VT = Op.getValueType();
6778 DebugLoc dl = Op.getDebugLoc();
6780 // Save FP Control Word to stack slot
6781 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
6782 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6784 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
6785 DAG.getEntryNode(), StackSlot);
6787 // Load FP Control Word from stack slot
6788 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
6790 // Transform as necessary
6792 DAG.getNode(ISD::SRL, dl, MVT::i16,
6793 DAG.getNode(ISD::AND, dl, MVT::i16,
6794 CWD, DAG.getConstant(0x800, MVT::i16)),
6795 DAG.getConstant(11, MVT::i8));
6797 DAG.getNode(ISD::SRL, dl, MVT::i16,
6798 DAG.getNode(ISD::AND, dl, MVT::i16,
6799 CWD, DAG.getConstant(0x400, MVT::i16)),
6800 DAG.getConstant(9, MVT::i8));
6803 DAG.getNode(ISD::AND, dl, MVT::i16,
6804 DAG.getNode(ISD::ADD, dl, MVT::i16,
6805 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
6806 DAG.getConstant(1, MVT::i16)),
6807 DAG.getConstant(3, MVT::i16));
6810 return DAG.getNode((VT.getSizeInBits() < 16 ?
6811 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
6814 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
6815 EVT VT = Op.getValueType();
6817 unsigned NumBits = VT.getSizeInBits();
6818 DebugLoc dl = Op.getDebugLoc();
6820 Op = Op.getOperand(0);
6821 if (VT == MVT::i8) {
6822 // Zero extend to i32 since there is not an i8 bsr.
6824 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6827 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
6828 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6829 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
6831 // If src is zero (i.e. bsr sets ZF), returns NumBits.
6832 SmallVector<SDValue, 4> Ops;
6834 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
6835 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6836 Ops.push_back(Op.getValue(1));
6837 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6839 // Finally xor with NumBits-1.
6840 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
6843 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6847 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
6848 EVT VT = Op.getValueType();
6850 unsigned NumBits = VT.getSizeInBits();
6851 DebugLoc dl = Op.getDebugLoc();
6853 Op = Op.getOperand(0);
6854 if (VT == MVT::i8) {
6856 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6859 // Issue a bsf (scan bits forward) which also sets EFLAGS.
6860 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6861 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
6863 // If src is zero (i.e. bsf sets ZF), returns NumBits.
6864 SmallVector<SDValue, 4> Ops;
6866 Ops.push_back(DAG.getConstant(NumBits, OpVT));
6867 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6868 Ops.push_back(Op.getValue(1));
6869 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6872 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6876 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
6877 EVT VT = Op.getValueType();
6878 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
6879 DebugLoc dl = Op.getDebugLoc();
6881 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
6882 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
6883 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
6884 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
6885 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
6887 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
6888 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
6889 // return AloBlo + AloBhi + AhiBlo;
6891 SDValue A = Op.getOperand(0);
6892 SDValue B = Op.getOperand(1);
6894 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6895 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6896 A, DAG.getConstant(32, MVT::i32));
6897 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6898 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6899 B, DAG.getConstant(32, MVT::i32));
6900 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6901 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6903 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6904 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6906 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6907 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6909 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6910 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6911 AloBhi, DAG.getConstant(32, MVT::i32));
6912 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6913 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6914 AhiBlo, DAG.getConstant(32, MVT::i32));
6915 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
6916 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
6921 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
6922 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
6923 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
6924 // looks for this combo and may remove the "setcc" instruction if the "setcc"
6925 // has only one use.
6926 SDNode *N = Op.getNode();
6927 SDValue LHS = N->getOperand(0);
6928 SDValue RHS = N->getOperand(1);
6929 unsigned BaseOp = 0;
6931 DebugLoc dl = Op.getDebugLoc();
6933 switch (Op.getOpcode()) {
6934 default: llvm_unreachable("Unknown ovf instruction!");
6936 // A subtract of one will be selected as a INC. Note that INC doesn't
6937 // set CF, so we can't do this for UADDO.
6938 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6939 if (C->getAPIntValue() == 1) {
6940 BaseOp = X86ISD::INC;
6944 BaseOp = X86ISD::ADD;
6948 BaseOp = X86ISD::ADD;
6952 // A subtract of one will be selected as a DEC. Note that DEC doesn't
6953 // set CF, so we can't do this for USUBO.
6954 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6955 if (C->getAPIntValue() == 1) {
6956 BaseOp = X86ISD::DEC;
6960 BaseOp = X86ISD::SUB;
6964 BaseOp = X86ISD::SUB;
6968 BaseOp = X86ISD::SMUL;
6972 BaseOp = X86ISD::UMUL;
6977 // Also sets EFLAGS.
6978 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
6979 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
6982 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
6983 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
6985 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
6989 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
6990 EVT T = Op.getValueType();
6991 DebugLoc dl = Op.getDebugLoc();
6994 switch(T.getSimpleVT().SimpleTy) {
6996 assert(false && "Invalid value type!");
6997 case MVT::i8: Reg = X86::AL; size = 1; break;
6998 case MVT::i16: Reg = X86::AX; size = 2; break;
6999 case MVT::i32: Reg = X86::EAX; size = 4; break;
7001 assert(Subtarget->is64Bit() && "Node not type legal!");
7002 Reg = X86::RAX; size = 8;
7005 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7006 Op.getOperand(2), SDValue());
7007 SDValue Ops[] = { cpIn.getValue(0),
7010 DAG.getTargetConstant(size, MVT::i8),
7012 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7013 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7015 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7019 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7020 SelectionDAG &DAG) {
7021 assert(Subtarget->is64Bit() && "Result not type legalized?");
7022 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7023 SDValue TheChain = Op.getOperand(0);
7024 DebugLoc dl = Op.getDebugLoc();
7025 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7026 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7027 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7029 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7030 DAG.getConstant(32, MVT::i8));
7032 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7035 return DAG.getMergeValues(Ops, 2, dl);
7038 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7039 SDNode *Node = Op.getNode();
7040 DebugLoc dl = Node->getDebugLoc();
7041 EVT T = Node->getValueType(0);
7042 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7043 DAG.getConstant(0, T), Node->getOperand(2));
7044 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7045 cast<AtomicSDNode>(Node)->getMemoryVT(),
7046 Node->getOperand(0),
7047 Node->getOperand(1), negOp,
7048 cast<AtomicSDNode>(Node)->getSrcValue(),
7049 cast<AtomicSDNode>(Node)->getAlignment());
7052 /// LowerOperation - Provide custom lowering hooks for some operations.
7054 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7055 switch (Op.getOpcode()) {
7056 default: llvm_unreachable("Should not custom lower this!");
7057 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7058 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7059 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7060 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7061 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7062 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7063 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7064 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7065 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7066 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7067 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7068 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7069 case ISD::SHL_PARTS:
7070 case ISD::SRA_PARTS:
7071 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7072 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7073 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7074 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7075 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7076 case ISD::FABS: return LowerFABS(Op, DAG);
7077 case ISD::FNEG: return LowerFNEG(Op, DAG);
7078 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7079 case ISD::SETCC: return LowerSETCC(Op, DAG);
7080 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7081 case ISD::SELECT: return LowerSELECT(Op, DAG);
7082 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7083 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7084 case ISD::VASTART: return LowerVASTART(Op, DAG);
7085 case ISD::VAARG: return LowerVAARG(Op, DAG);
7086 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7087 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7088 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7089 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7090 case ISD::FRAME_TO_ARGS_OFFSET:
7091 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7092 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7093 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7094 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7095 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7096 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7097 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7098 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7104 case ISD::UMULO: return LowerXALUO(Op, DAG);
7105 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7109 void X86TargetLowering::
7110 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7111 SelectionDAG &DAG, unsigned NewOp) {
7112 EVT T = Node->getValueType(0);
7113 DebugLoc dl = Node->getDebugLoc();
7114 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7116 SDValue Chain = Node->getOperand(0);
7117 SDValue In1 = Node->getOperand(1);
7118 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7119 Node->getOperand(2), DAG.getIntPtrConstant(0));
7120 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7121 Node->getOperand(2), DAG.getIntPtrConstant(1));
7122 SDValue Ops[] = { Chain, In1, In2L, In2H };
7123 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7125 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7126 cast<MemSDNode>(Node)->getMemOperand());
7127 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7128 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7129 Results.push_back(Result.getValue(2));
7132 /// ReplaceNodeResults - Replace a node with an illegal result type
7133 /// with a new node built out of custom code.
7134 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7135 SmallVectorImpl<SDValue>&Results,
7136 SelectionDAG &DAG) {
7137 DebugLoc dl = N->getDebugLoc();
7138 switch (N->getOpcode()) {
7140 assert(false && "Do not know how to custom type legalize this operation!");
7142 case ISD::FP_TO_SINT: {
7143 std::pair<SDValue,SDValue> Vals =
7144 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7145 SDValue FIST = Vals.first, StackSlot = Vals.second;
7146 if (FIST.getNode() != 0) {
7147 EVT VT = N->getValueType(0);
7148 // Return a load from the stack slot.
7149 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
7153 case ISD::READCYCLECOUNTER: {
7154 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7155 SDValue TheChain = N->getOperand(0);
7156 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7157 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7159 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7161 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7162 SDValue Ops[] = { eax, edx };
7163 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7164 Results.push_back(edx.getValue(1));
7167 case ISD::ATOMIC_CMP_SWAP: {
7168 EVT T = N->getValueType(0);
7169 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7170 SDValue cpInL, cpInH;
7171 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7172 DAG.getConstant(0, MVT::i32));
7173 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7174 DAG.getConstant(1, MVT::i32));
7175 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7176 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7178 SDValue swapInL, swapInH;
7179 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7180 DAG.getConstant(0, MVT::i32));
7181 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7182 DAG.getConstant(1, MVT::i32));
7183 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7185 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7186 swapInL.getValue(1));
7187 SDValue Ops[] = { swapInH.getValue(0),
7189 swapInH.getValue(1) };
7190 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7191 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7192 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7193 MVT::i32, Result.getValue(1));
7194 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7195 MVT::i32, cpOutL.getValue(2));
7196 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7197 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7198 Results.push_back(cpOutH.getValue(1));
7201 case ISD::ATOMIC_LOAD_ADD:
7202 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7204 case ISD::ATOMIC_LOAD_AND:
7205 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7207 case ISD::ATOMIC_LOAD_NAND:
7208 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7210 case ISD::ATOMIC_LOAD_OR:
7211 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7213 case ISD::ATOMIC_LOAD_SUB:
7214 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7216 case ISD::ATOMIC_LOAD_XOR:
7217 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7219 case ISD::ATOMIC_SWAP:
7220 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7225 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7227 default: return NULL;
7228 case X86ISD::BSF: return "X86ISD::BSF";
7229 case X86ISD::BSR: return "X86ISD::BSR";
7230 case X86ISD::SHLD: return "X86ISD::SHLD";
7231 case X86ISD::SHRD: return "X86ISD::SHRD";
7232 case X86ISD::FAND: return "X86ISD::FAND";
7233 case X86ISD::FOR: return "X86ISD::FOR";
7234 case X86ISD::FXOR: return "X86ISD::FXOR";
7235 case X86ISD::FSRL: return "X86ISD::FSRL";
7236 case X86ISD::FILD: return "X86ISD::FILD";
7237 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7238 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7239 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7240 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7241 case X86ISD::FLD: return "X86ISD::FLD";
7242 case X86ISD::FST: return "X86ISD::FST";
7243 case X86ISD::CALL: return "X86ISD::CALL";
7244 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7245 case X86ISD::BT: return "X86ISD::BT";
7246 case X86ISD::CMP: return "X86ISD::CMP";
7247 case X86ISD::COMI: return "X86ISD::COMI";
7248 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7249 case X86ISD::SETCC: return "X86ISD::SETCC";
7250 case X86ISD::CMOV: return "X86ISD::CMOV";
7251 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7252 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7253 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7254 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7255 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7256 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7257 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7258 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7259 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7260 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7261 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7262 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7263 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7264 case X86ISD::FMAX: return "X86ISD::FMAX";
7265 case X86ISD::FMIN: return "X86ISD::FMIN";
7266 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7267 case X86ISD::FRCP: return "X86ISD::FRCP";
7268 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7269 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7270 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7271 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7272 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7273 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7274 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7275 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7276 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7277 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7278 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7279 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7280 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7281 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7282 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7283 case X86ISD::VSHL: return "X86ISD::VSHL";
7284 case X86ISD::VSRL: return "X86ISD::VSRL";
7285 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7286 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7287 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7288 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7289 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7290 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7291 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7292 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7293 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7294 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7295 case X86ISD::ADD: return "X86ISD::ADD";
7296 case X86ISD::SUB: return "X86ISD::SUB";
7297 case X86ISD::SMUL: return "X86ISD::SMUL";
7298 case X86ISD::UMUL: return "X86ISD::UMUL";
7299 case X86ISD::INC: return "X86ISD::INC";
7300 case X86ISD::DEC: return "X86ISD::DEC";
7301 case X86ISD::OR: return "X86ISD::OR";
7302 case X86ISD::XOR: return "X86ISD::XOR";
7303 case X86ISD::AND: return "X86ISD::AND";
7304 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7305 case X86ISD::PTEST: return "X86ISD::PTEST";
7306 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7310 // isLegalAddressingMode - Return true if the addressing mode represented
7311 // by AM is legal for this target, for a load/store of the specified type.
7312 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7313 const Type *Ty) const {
7314 // X86 supports extremely general addressing modes.
7315 CodeModel::Model M = getTargetMachine().getCodeModel();
7317 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7318 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7323 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7325 // If a reference to this global requires an extra load, we can't fold it.
7326 if (isGlobalStubReference(GVFlags))
7329 // If BaseGV requires a register for the PIC base, we cannot also have a
7330 // BaseReg specified.
7331 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7334 // If lower 4G is not available, then we must use rip-relative addressing.
7335 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7345 // These scales always work.
7350 // These scales are formed with basereg+scalereg. Only accept if there is
7355 default: // Other stuff never works.
7363 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7364 if (!Ty1->isInteger() || !Ty2->isInteger())
7366 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7367 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7368 if (NumBits1 <= NumBits2)
7370 return Subtarget->is64Bit() || NumBits1 < 64;
7373 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7374 if (!VT1.isInteger() || !VT2.isInteger())
7376 unsigned NumBits1 = VT1.getSizeInBits();
7377 unsigned NumBits2 = VT2.getSizeInBits();
7378 if (NumBits1 <= NumBits2)
7380 return Subtarget->is64Bit() || NumBits1 < 64;
7383 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7384 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7385 return Ty1 == Type::getInt32Ty(Ty1->getContext()) &&
7386 Ty2 == Type::getInt64Ty(Ty1->getContext()) && Subtarget->is64Bit();
7389 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7390 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7391 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7394 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7395 // i16 instructions are longer (0x66 prefix) and potentially slower.
7396 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7399 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7400 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7401 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7402 /// are assumed to be legal.
7404 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7406 // Only do shuffles on 128-bit vector types for now.
7407 if (VT.getSizeInBits() == 64)
7410 // FIXME: pshufb, blends, shifts.
7411 return (VT.getVectorNumElements() == 2 ||
7412 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7413 isMOVLMask(M, VT) ||
7414 isSHUFPMask(M, VT) ||
7415 isPSHUFDMask(M, VT) ||
7416 isPSHUFHWMask(M, VT) ||
7417 isPSHUFLWMask(M, VT) ||
7418 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7419 isUNPCKLMask(M, VT) ||
7420 isUNPCKHMask(M, VT) ||
7421 isUNPCKL_v_undef_Mask(M, VT) ||
7422 isUNPCKH_v_undef_Mask(M, VT));
7426 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7428 unsigned NumElts = VT.getVectorNumElements();
7429 // FIXME: This collection of masks seems suspect.
7432 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7433 return (isMOVLMask(Mask, VT) ||
7434 isCommutedMOVLMask(Mask, VT, true) ||
7435 isSHUFPMask(Mask, VT) ||
7436 isCommutedSHUFPMask(Mask, VT));
7441 //===----------------------------------------------------------------------===//
7442 // X86 Scheduler Hooks
7443 //===----------------------------------------------------------------------===//
7445 // private utility function
7447 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7448 MachineBasicBlock *MBB,
7456 TargetRegisterClass *RC,
7457 bool invSrc) const {
7458 // For the atomic bitwise operator, we generate
7461 // ld t1 = [bitinstr.addr]
7462 // op t2 = t1, [bitinstr.val]
7464 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7466 // fallthrough -->nextMBB
7467 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7468 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7469 MachineFunction::iterator MBBIter = MBB;
7472 /// First build the CFG
7473 MachineFunction *F = MBB->getParent();
7474 MachineBasicBlock *thisMBB = MBB;
7475 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7476 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7477 F->insert(MBBIter, newMBB);
7478 F->insert(MBBIter, nextMBB);
7480 // Move all successors to thisMBB to nextMBB
7481 nextMBB->transferSuccessors(thisMBB);
7483 // Update thisMBB to fall through to newMBB
7484 thisMBB->addSuccessor(newMBB);
7486 // newMBB jumps to itself and fall through to nextMBB
7487 newMBB->addSuccessor(nextMBB);
7488 newMBB->addSuccessor(newMBB);
7490 // Insert instructions into newMBB based on incoming instruction
7491 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7492 "unexpected number of operands");
7493 DebugLoc dl = bInstr->getDebugLoc();
7494 MachineOperand& destOper = bInstr->getOperand(0);
7495 MachineOperand* argOpers[2 + X86AddrNumOperands];
7496 int numArgs = bInstr->getNumOperands() - 1;
7497 for (int i=0; i < numArgs; ++i)
7498 argOpers[i] = &bInstr->getOperand(i+1);
7500 // x86 address has 4 operands: base, index, scale, and displacement
7501 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7502 int valArgIndx = lastAddrIndx + 1;
7504 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7505 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7506 for (int i=0; i <= lastAddrIndx; ++i)
7507 (*MIB).addOperand(*argOpers[i]);
7509 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7511 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7516 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7517 assert((argOpers[valArgIndx]->isReg() ||
7518 argOpers[valArgIndx]->isImm()) &&
7520 if (argOpers[valArgIndx]->isReg())
7521 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7523 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7525 (*MIB).addOperand(*argOpers[valArgIndx]);
7527 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7530 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7531 for (int i=0; i <= lastAddrIndx; ++i)
7532 (*MIB).addOperand(*argOpers[i]);
7534 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7535 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7536 bInstr->memoperands_end());
7538 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7542 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7544 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7548 // private utility function: 64 bit atomics on 32 bit host.
7550 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7551 MachineBasicBlock *MBB,
7556 bool invSrc) const {
7557 // For the atomic bitwise operator, we generate
7558 // thisMBB (instructions are in pairs, except cmpxchg8b)
7559 // ld t1,t2 = [bitinstr.addr]
7561 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7562 // op t5, t6 <- out1, out2, [bitinstr.val]
7563 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7564 // mov ECX, EBX <- t5, t6
7565 // mov EAX, EDX <- t1, t2
7566 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7567 // mov t3, t4 <- EAX, EDX
7569 // result in out1, out2
7570 // fallthrough -->nextMBB
7572 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7573 const unsigned LoadOpc = X86::MOV32rm;
7574 const unsigned copyOpc = X86::MOV32rr;
7575 const unsigned NotOpc = X86::NOT32r;
7576 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7577 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7578 MachineFunction::iterator MBBIter = MBB;
7581 /// First build the CFG
7582 MachineFunction *F = MBB->getParent();
7583 MachineBasicBlock *thisMBB = MBB;
7584 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7585 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7586 F->insert(MBBIter, newMBB);
7587 F->insert(MBBIter, nextMBB);
7589 // Move all successors to thisMBB to nextMBB
7590 nextMBB->transferSuccessors(thisMBB);
7592 // Update thisMBB to fall through to newMBB
7593 thisMBB->addSuccessor(newMBB);
7595 // newMBB jumps to itself and fall through to nextMBB
7596 newMBB->addSuccessor(nextMBB);
7597 newMBB->addSuccessor(newMBB);
7599 DebugLoc dl = bInstr->getDebugLoc();
7600 // Insert instructions into newMBB based on incoming instruction
7601 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7602 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7603 "unexpected number of operands");
7604 MachineOperand& dest1Oper = bInstr->getOperand(0);
7605 MachineOperand& dest2Oper = bInstr->getOperand(1);
7606 MachineOperand* argOpers[2 + X86AddrNumOperands];
7607 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7608 argOpers[i] = &bInstr->getOperand(i+2);
7610 // x86 address has 4 operands: base, index, scale, and displacement
7611 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7613 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7614 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7615 for (int i=0; i <= lastAddrIndx; ++i)
7616 (*MIB).addOperand(*argOpers[i]);
7617 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7618 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7619 // add 4 to displacement.
7620 for (int i=0; i <= lastAddrIndx-2; ++i)
7621 (*MIB).addOperand(*argOpers[i]);
7622 MachineOperand newOp3 = *(argOpers[3]);
7624 newOp3.setImm(newOp3.getImm()+4);
7626 newOp3.setOffset(newOp3.getOffset()+4);
7627 (*MIB).addOperand(newOp3);
7628 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7630 // t3/4 are defined later, at the bottom of the loop
7631 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7632 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7633 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7634 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7635 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7636 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7638 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
7639 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
7641 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt1).addReg(t1);
7642 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt2).addReg(t2);
7648 int valArgIndx = lastAddrIndx + 1;
7649 assert((argOpers[valArgIndx]->isReg() ||
7650 argOpers[valArgIndx]->isImm()) &&
7652 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7653 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7654 if (argOpers[valArgIndx]->isReg())
7655 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7657 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7658 if (regOpcL != X86::MOV32rr)
7660 (*MIB).addOperand(*argOpers[valArgIndx]);
7661 assert(argOpers[valArgIndx + 1]->isReg() ==
7662 argOpers[valArgIndx]->isReg());
7663 assert(argOpers[valArgIndx + 1]->isImm() ==
7664 argOpers[valArgIndx]->isImm());
7665 if (argOpers[valArgIndx + 1]->isReg())
7666 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7668 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7669 if (regOpcH != X86::MOV32rr)
7671 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
7673 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
7675 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
7678 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
7680 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
7683 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
7684 for (int i=0; i <= lastAddrIndx; ++i)
7685 (*MIB).addOperand(*argOpers[i]);
7687 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7688 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7689 bInstr->memoperands_end());
7691 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
7692 MIB.addReg(X86::EAX);
7693 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
7694 MIB.addReg(X86::EDX);
7697 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7699 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7703 // private utility function
7705 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
7706 MachineBasicBlock *MBB,
7707 unsigned cmovOpc) const {
7708 // For the atomic min/max operator, we generate
7711 // ld t1 = [min/max.addr]
7712 // mov t2 = [min/max.val]
7714 // cmov[cond] t2 = t1
7716 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7718 // fallthrough -->nextMBB
7720 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7721 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7722 MachineFunction::iterator MBBIter = MBB;
7725 /// First build the CFG
7726 MachineFunction *F = MBB->getParent();
7727 MachineBasicBlock *thisMBB = MBB;
7728 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7729 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7730 F->insert(MBBIter, newMBB);
7731 F->insert(MBBIter, nextMBB);
7733 // Move all successors of thisMBB to nextMBB
7734 nextMBB->transferSuccessors(thisMBB);
7736 // Update thisMBB to fall through to newMBB
7737 thisMBB->addSuccessor(newMBB);
7739 // newMBB jumps to newMBB and fall through to nextMBB
7740 newMBB->addSuccessor(nextMBB);
7741 newMBB->addSuccessor(newMBB);
7743 DebugLoc dl = mInstr->getDebugLoc();
7744 // Insert instructions into newMBB based on incoming instruction
7745 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7746 "unexpected number of operands");
7747 MachineOperand& destOper = mInstr->getOperand(0);
7748 MachineOperand* argOpers[2 + X86AddrNumOperands];
7749 int numArgs = mInstr->getNumOperands() - 1;
7750 for (int i=0; i < numArgs; ++i)
7751 argOpers[i] = &mInstr->getOperand(i+1);
7753 // x86 address has 4 operands: base, index, scale, and displacement
7754 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7755 int valArgIndx = lastAddrIndx + 1;
7757 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7758 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
7759 for (int i=0; i <= lastAddrIndx; ++i)
7760 (*MIB).addOperand(*argOpers[i]);
7762 // We only support register and immediate values
7763 assert((argOpers[valArgIndx]->isReg() ||
7764 argOpers[valArgIndx]->isImm()) &&
7767 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7768 if (argOpers[valArgIndx]->isReg())
7769 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7771 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7772 (*MIB).addOperand(*argOpers[valArgIndx]);
7774 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
7777 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
7782 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7783 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
7787 // Cmp and exchange if none has modified the memory location
7788 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
7789 for (int i=0; i <= lastAddrIndx; ++i)
7790 (*MIB).addOperand(*argOpers[i]);
7792 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7793 (*MIB).setMemRefs(mInstr->memoperands_begin(),
7794 mInstr->memoperands_end());
7796 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
7797 MIB.addReg(X86::EAX);
7800 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7802 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
7806 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
7807 // all of this code can be replaced with that in the .td file.
7809 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
7810 unsigned numArgs, bool memArg) const {
7812 MachineFunction *F = BB->getParent();
7813 DebugLoc dl = MI->getDebugLoc();
7814 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7818 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
7820 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
7822 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
7824 for (unsigned i = 0; i < numArgs; ++i) {
7825 MachineOperand &Op = MI->getOperand(i+1);
7827 if (!(Op.isReg() && Op.isImplicit()))
7831 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
7834 F->DeleteMachineInstr(MI);
7840 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
7842 MachineBasicBlock *MBB) const {
7843 // Emit code to save XMM registers to the stack. The ABI says that the
7844 // number of registers to save is given in %al, so it's theoretically
7845 // possible to do an indirect jump trick to avoid saving all of them,
7846 // however this code takes a simpler approach and just executes all
7847 // of the stores if %al is non-zero. It's less code, and it's probably
7848 // easier on the hardware branch predictor, and stores aren't all that
7849 // expensive anyway.
7851 // Create the new basic blocks. One block contains all the XMM stores,
7852 // and one block is the final destination regardless of whether any
7853 // stores were performed.
7854 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7855 MachineFunction *F = MBB->getParent();
7856 MachineFunction::iterator MBBIter = MBB;
7858 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
7859 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
7860 F->insert(MBBIter, XMMSaveMBB);
7861 F->insert(MBBIter, EndMBB);
7864 // Move any original successors of MBB to the end block.
7865 EndMBB->transferSuccessors(MBB);
7866 // The original block will now fall through to the XMM save block.
7867 MBB->addSuccessor(XMMSaveMBB);
7868 // The XMMSaveMBB will fall through to the end block.
7869 XMMSaveMBB->addSuccessor(EndMBB);
7871 // Now add the instructions.
7872 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7873 DebugLoc DL = MI->getDebugLoc();
7875 unsigned CountReg = MI->getOperand(0).getReg();
7876 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
7877 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
7879 if (!Subtarget->isTargetWin64()) {
7880 // If %al is 0, branch around the XMM save block.
7881 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
7882 BuildMI(MBB, DL, TII->get(X86::JE)).addMBB(EndMBB);
7883 MBB->addSuccessor(EndMBB);
7886 // In the XMM save block, save all the XMM argument registers.
7887 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
7888 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
7889 MachineMemOperand *MMO =
7890 F->getMachineMemOperand(
7891 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
7892 MachineMemOperand::MOStore, Offset,
7893 /*Size=*/16, /*Align=*/16);
7894 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
7895 .addFrameIndex(RegSaveFrameIndex)
7896 .addImm(/*Scale=*/1)
7897 .addReg(/*IndexReg=*/0)
7898 .addImm(/*Disp=*/Offset)
7899 .addReg(/*Segment=*/0)
7900 .addReg(MI->getOperand(i).getReg())
7901 .addMemOperand(MMO);
7904 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7910 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
7911 MachineBasicBlock *BB,
7912 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
7913 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7914 DebugLoc DL = MI->getDebugLoc();
7916 // To "insert" a SELECT_CC instruction, we actually have to insert the
7917 // diamond control-flow pattern. The incoming instruction knows the
7918 // destination vreg to set, the condition code register to branch on, the
7919 // true/false values to select between, and a branch opcode to use.
7920 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7921 MachineFunction::iterator It = BB;
7927 // cmpTY ccX, r1, r2
7929 // fallthrough --> copy0MBB
7930 MachineBasicBlock *thisMBB = BB;
7931 MachineFunction *F = BB->getParent();
7932 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7933 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7935 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
7936 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
7937 F->insert(It, copy0MBB);
7938 F->insert(It, sinkMBB);
7939 // Update machine-CFG edges by first adding all successors of the current
7940 // block to the new block which will contain the Phi node for the select.
7941 // Also inform sdisel of the edge changes.
7942 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
7943 E = BB->succ_end(); I != E; ++I) {
7944 EM->insert(std::make_pair(*I, sinkMBB));
7945 sinkMBB->addSuccessor(*I);
7947 // Next, remove all successors of the current block, and add the true
7948 // and fallthrough blocks as its successors.
7949 while (!BB->succ_empty())
7950 BB->removeSuccessor(BB->succ_begin());
7951 // Add the true and fallthrough blocks as its successors.
7952 BB->addSuccessor(copy0MBB);
7953 BB->addSuccessor(sinkMBB);
7956 // %FalseValue = ...
7957 // # fallthrough to sinkMBB
7960 // Update machine-CFG edges
7961 BB->addSuccessor(sinkMBB);
7964 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7967 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
7968 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7969 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7971 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7977 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7978 MachineBasicBlock *BB,
7979 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
7980 switch (MI->getOpcode()) {
7981 default: assert(false && "Unexpected instr type to insert");
7983 case X86::CMOV_V1I64:
7984 case X86::CMOV_FR32:
7985 case X86::CMOV_FR64:
7986 case X86::CMOV_V4F32:
7987 case X86::CMOV_V2F64:
7988 case X86::CMOV_V2I64:
7989 return EmitLoweredSelect(MI, BB, EM);
7991 case X86::FP32_TO_INT16_IN_MEM:
7992 case X86::FP32_TO_INT32_IN_MEM:
7993 case X86::FP32_TO_INT64_IN_MEM:
7994 case X86::FP64_TO_INT16_IN_MEM:
7995 case X86::FP64_TO_INT32_IN_MEM:
7996 case X86::FP64_TO_INT64_IN_MEM:
7997 case X86::FP80_TO_INT16_IN_MEM:
7998 case X86::FP80_TO_INT32_IN_MEM:
7999 case X86::FP80_TO_INT64_IN_MEM: {
8000 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8001 DebugLoc DL = MI->getDebugLoc();
8003 // Change the floating point control register to use "round towards zero"
8004 // mode when truncating to an integer value.
8005 MachineFunction *F = BB->getParent();
8006 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8007 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8009 // Load the old value of the high byte of the control word...
8011 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8012 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8015 // Set the high part to be round to zero...
8016 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8019 // Reload the modified control word now...
8020 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8022 // Restore the memory image of control word to original value
8023 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8026 // Get the X86 opcode to use.
8028 switch (MI->getOpcode()) {
8029 default: llvm_unreachable("illegal opcode!");
8030 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8031 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8032 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8033 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8034 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8035 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8036 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8037 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8038 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8042 MachineOperand &Op = MI->getOperand(0);
8044 AM.BaseType = X86AddressMode::RegBase;
8045 AM.Base.Reg = Op.getReg();
8047 AM.BaseType = X86AddressMode::FrameIndexBase;
8048 AM.Base.FrameIndex = Op.getIndex();
8050 Op = MI->getOperand(1);
8052 AM.Scale = Op.getImm();
8053 Op = MI->getOperand(2);
8055 AM.IndexReg = Op.getImm();
8056 Op = MI->getOperand(3);
8057 if (Op.isGlobal()) {
8058 AM.GV = Op.getGlobal();
8060 AM.Disp = Op.getImm();
8062 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8063 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8065 // Reload the original control word now.
8066 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8068 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8071 // String/text processing lowering.
8072 case X86::PCMPISTRM128REG:
8073 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8074 case X86::PCMPISTRM128MEM:
8075 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8076 case X86::PCMPESTRM128REG:
8077 return EmitPCMP(MI, BB, 5, false /* in mem */);
8078 case X86::PCMPESTRM128MEM:
8079 return EmitPCMP(MI, BB, 5, true /* in mem */);
8082 case X86::ATOMAND32:
8083 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8084 X86::AND32ri, X86::MOV32rm,
8085 X86::LCMPXCHG32, X86::MOV32rr,
8086 X86::NOT32r, X86::EAX,
8087 X86::GR32RegisterClass);
8089 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8090 X86::OR32ri, X86::MOV32rm,
8091 X86::LCMPXCHG32, X86::MOV32rr,
8092 X86::NOT32r, X86::EAX,
8093 X86::GR32RegisterClass);
8094 case X86::ATOMXOR32:
8095 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8096 X86::XOR32ri, X86::MOV32rm,
8097 X86::LCMPXCHG32, X86::MOV32rr,
8098 X86::NOT32r, X86::EAX,
8099 X86::GR32RegisterClass);
8100 case X86::ATOMNAND32:
8101 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8102 X86::AND32ri, X86::MOV32rm,
8103 X86::LCMPXCHG32, X86::MOV32rr,
8104 X86::NOT32r, X86::EAX,
8105 X86::GR32RegisterClass, true);
8106 case X86::ATOMMIN32:
8107 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8108 case X86::ATOMMAX32:
8109 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8110 case X86::ATOMUMIN32:
8111 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8112 case X86::ATOMUMAX32:
8113 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8115 case X86::ATOMAND16:
8116 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8117 X86::AND16ri, X86::MOV16rm,
8118 X86::LCMPXCHG16, X86::MOV16rr,
8119 X86::NOT16r, X86::AX,
8120 X86::GR16RegisterClass);
8122 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8123 X86::OR16ri, X86::MOV16rm,
8124 X86::LCMPXCHG16, X86::MOV16rr,
8125 X86::NOT16r, X86::AX,
8126 X86::GR16RegisterClass);
8127 case X86::ATOMXOR16:
8128 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8129 X86::XOR16ri, X86::MOV16rm,
8130 X86::LCMPXCHG16, X86::MOV16rr,
8131 X86::NOT16r, X86::AX,
8132 X86::GR16RegisterClass);
8133 case X86::ATOMNAND16:
8134 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8135 X86::AND16ri, X86::MOV16rm,
8136 X86::LCMPXCHG16, X86::MOV16rr,
8137 X86::NOT16r, X86::AX,
8138 X86::GR16RegisterClass, true);
8139 case X86::ATOMMIN16:
8140 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8141 case X86::ATOMMAX16:
8142 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8143 case X86::ATOMUMIN16:
8144 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8145 case X86::ATOMUMAX16:
8146 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8149 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8150 X86::AND8ri, X86::MOV8rm,
8151 X86::LCMPXCHG8, X86::MOV8rr,
8152 X86::NOT8r, X86::AL,
8153 X86::GR8RegisterClass);
8155 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8156 X86::OR8ri, X86::MOV8rm,
8157 X86::LCMPXCHG8, X86::MOV8rr,
8158 X86::NOT8r, X86::AL,
8159 X86::GR8RegisterClass);
8161 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8162 X86::XOR8ri, X86::MOV8rm,
8163 X86::LCMPXCHG8, X86::MOV8rr,
8164 X86::NOT8r, X86::AL,
8165 X86::GR8RegisterClass);
8166 case X86::ATOMNAND8:
8167 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8168 X86::AND8ri, X86::MOV8rm,
8169 X86::LCMPXCHG8, X86::MOV8rr,
8170 X86::NOT8r, X86::AL,
8171 X86::GR8RegisterClass, true);
8172 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8173 // This group is for 64-bit host.
8174 case X86::ATOMAND64:
8175 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8176 X86::AND64ri32, X86::MOV64rm,
8177 X86::LCMPXCHG64, X86::MOV64rr,
8178 X86::NOT64r, X86::RAX,
8179 X86::GR64RegisterClass);
8181 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8182 X86::OR64ri32, X86::MOV64rm,
8183 X86::LCMPXCHG64, X86::MOV64rr,
8184 X86::NOT64r, X86::RAX,
8185 X86::GR64RegisterClass);
8186 case X86::ATOMXOR64:
8187 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8188 X86::XOR64ri32, X86::MOV64rm,
8189 X86::LCMPXCHG64, X86::MOV64rr,
8190 X86::NOT64r, X86::RAX,
8191 X86::GR64RegisterClass);
8192 case X86::ATOMNAND64:
8193 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8194 X86::AND64ri32, X86::MOV64rm,
8195 X86::LCMPXCHG64, X86::MOV64rr,
8196 X86::NOT64r, X86::RAX,
8197 X86::GR64RegisterClass, true);
8198 case X86::ATOMMIN64:
8199 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8200 case X86::ATOMMAX64:
8201 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8202 case X86::ATOMUMIN64:
8203 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8204 case X86::ATOMUMAX64:
8205 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8207 // This group does 64-bit operations on a 32-bit host.
8208 case X86::ATOMAND6432:
8209 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8210 X86::AND32rr, X86::AND32rr,
8211 X86::AND32ri, X86::AND32ri,
8213 case X86::ATOMOR6432:
8214 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8215 X86::OR32rr, X86::OR32rr,
8216 X86::OR32ri, X86::OR32ri,
8218 case X86::ATOMXOR6432:
8219 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8220 X86::XOR32rr, X86::XOR32rr,
8221 X86::XOR32ri, X86::XOR32ri,
8223 case X86::ATOMNAND6432:
8224 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8225 X86::AND32rr, X86::AND32rr,
8226 X86::AND32ri, X86::AND32ri,
8228 case X86::ATOMADD6432:
8229 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8230 X86::ADD32rr, X86::ADC32rr,
8231 X86::ADD32ri, X86::ADC32ri,
8233 case X86::ATOMSUB6432:
8234 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8235 X86::SUB32rr, X86::SBB32rr,
8236 X86::SUB32ri, X86::SBB32ri,
8238 case X86::ATOMSWAP6432:
8239 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8240 X86::MOV32rr, X86::MOV32rr,
8241 X86::MOV32ri, X86::MOV32ri,
8243 case X86::VASTART_SAVE_XMM_REGS:
8244 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8248 //===----------------------------------------------------------------------===//
8249 // X86 Optimization Hooks
8250 //===----------------------------------------------------------------------===//
8252 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8256 const SelectionDAG &DAG,
8257 unsigned Depth) const {
8258 unsigned Opc = Op.getOpcode();
8259 assert((Opc >= ISD::BUILTIN_OP_END ||
8260 Opc == ISD::INTRINSIC_WO_CHAIN ||
8261 Opc == ISD::INTRINSIC_W_CHAIN ||
8262 Opc == ISD::INTRINSIC_VOID) &&
8263 "Should use MaskedValueIsZero if you don't know whether Op"
8264 " is a target node!");
8266 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8278 // These nodes' second result is a boolean.
8279 if (Op.getResNo() == 0)
8283 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8284 Mask.getBitWidth() - 1);
8289 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8290 /// node is a GlobalAddress + offset.
8291 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8292 GlobalValue* &GA, int64_t &Offset) const{
8293 if (N->getOpcode() == X86ISD::Wrapper) {
8294 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8295 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8296 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8300 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8303 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
8304 const TargetLowering &TLI) {
8307 if (TLI.isGAPlusOffset(Base, GV, Offset))
8308 return (GV->getAlignment() >= N && (Offset % N) == 0);
8309 // DAG combine handles the stack object case.
8313 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8314 EVT EltVT, LoadSDNode *&LDBase,
8315 unsigned &LastLoadedElt,
8316 SelectionDAG &DAG, MachineFrameInfo *MFI,
8317 const TargetLowering &TLI) {
8319 LastLoadedElt = -1U;
8320 for (unsigned i = 0; i < NumElems; ++i) {
8321 if (N->getMaskElt(i) < 0) {
8327 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8328 if (!Elt.getNode() ||
8329 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8332 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8334 LDBase = cast<LoadSDNode>(Elt.getNode());
8338 if (Elt.getOpcode() == ISD::UNDEF)
8341 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8342 if (!TLI.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i, MFI))
8349 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8350 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8351 /// if the load addresses are consecutive, non-overlapping, and in the right
8352 /// order. In the case of v2i64, it will see if it can rewrite the
8353 /// shuffle to be an appropriate build vector so it can take advantage of
8354 // performBuildVectorCombine.
8355 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8356 const TargetLowering &TLI) {
8357 DebugLoc dl = N->getDebugLoc();
8358 EVT VT = N->getValueType(0);
8359 EVT EltVT = VT.getVectorElementType();
8360 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8361 unsigned NumElems = VT.getVectorNumElements();
8363 if (VT.getSizeInBits() != 128)
8366 // Try to combine a vector_shuffle into a 128-bit load.
8367 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8368 LoadSDNode *LD = NULL;
8369 unsigned LastLoadedElt;
8370 if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
8374 if (LastLoadedElt == NumElems - 1) {
8375 if (isBaseAlignmentOfN(16, LD->getBasePtr().getNode(), TLI))
8376 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8377 LD->getSrcValue(), LD->getSrcValueOffset(),
8379 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8380 LD->getSrcValue(), LD->getSrcValueOffset(),
8381 LD->isVolatile(), LD->getAlignment());
8382 } else if (NumElems == 4 && LastLoadedElt == 1) {
8383 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8384 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8385 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8386 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8391 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8392 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8393 const X86Subtarget *Subtarget) {
8394 DebugLoc DL = N->getDebugLoc();
8395 SDValue Cond = N->getOperand(0);
8396 // Get the LHS/RHS of the select.
8397 SDValue LHS = N->getOperand(1);
8398 SDValue RHS = N->getOperand(2);
8400 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8401 // instructions have the peculiarity that if either operand is a NaN,
8402 // they chose what we call the RHS operand (and as such are not symmetric).
8403 // It happens that this matches the semantics of the common C idiom
8404 // x<y?x:y and related forms, so we can recognize these cases.
8405 if (Subtarget->hasSSE2() &&
8406 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8407 Cond.getOpcode() == ISD::SETCC) {
8408 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8410 unsigned Opcode = 0;
8411 // Check for x CC y ? x : y.
8412 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8416 // This can be a min if we can prove that at least one of the operands
8418 if (!FiniteOnlyFPMath()) {
8419 if (DAG.isKnownNeverNaN(RHS)) {
8420 // Put the potential NaN in the RHS so that SSE will preserve it.
8421 std::swap(LHS, RHS);
8422 } else if (!DAG.isKnownNeverNaN(LHS))
8425 Opcode = X86ISD::FMIN;
8428 // This can be a min if we can prove that at least one of the operands
8430 if (!FiniteOnlyFPMath()) {
8431 if (DAG.isKnownNeverNaN(LHS)) {
8432 // Put the potential NaN in the RHS so that SSE will preserve it.
8433 std::swap(LHS, RHS);
8434 } else if (!DAG.isKnownNeverNaN(RHS))
8437 Opcode = X86ISD::FMIN;
8440 // This can be a min, but if either operand is a NaN we need it to
8441 // preserve the original LHS.
8442 std::swap(LHS, RHS);
8446 Opcode = X86ISD::FMIN;
8450 // This can be a max if we can prove that at least one of the operands
8452 if (!FiniteOnlyFPMath()) {
8453 if (DAG.isKnownNeverNaN(LHS)) {
8454 // Put the potential NaN in the RHS so that SSE will preserve it.
8455 std::swap(LHS, RHS);
8456 } else if (!DAG.isKnownNeverNaN(RHS))
8459 Opcode = X86ISD::FMAX;
8462 // This can be a max if we can prove that at least one of the operands
8464 if (!FiniteOnlyFPMath()) {
8465 if (DAG.isKnownNeverNaN(RHS)) {
8466 // Put the potential NaN in the RHS so that SSE will preserve it.
8467 std::swap(LHS, RHS);
8468 } else if (!DAG.isKnownNeverNaN(LHS))
8471 Opcode = X86ISD::FMAX;
8474 // This can be a max, but if either operand is a NaN we need it to
8475 // preserve the original LHS.
8476 std::swap(LHS, RHS);
8480 Opcode = X86ISD::FMAX;
8483 // Check for x CC y ? y : x -- a min/max with reversed arms.
8484 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8488 // This can be a min if we can prove that at least one of the operands
8490 if (!FiniteOnlyFPMath()) {
8491 if (DAG.isKnownNeverNaN(RHS)) {
8492 // Put the potential NaN in the RHS so that SSE will preserve it.
8493 std::swap(LHS, RHS);
8494 } else if (!DAG.isKnownNeverNaN(LHS))
8497 Opcode = X86ISD::FMIN;
8500 // This can be a min if we can prove that at least one of the operands
8502 if (!FiniteOnlyFPMath()) {
8503 if (DAG.isKnownNeverNaN(LHS)) {
8504 // Put the potential NaN in the RHS so that SSE will preserve it.
8505 std::swap(LHS, RHS);
8506 } else if (!DAG.isKnownNeverNaN(RHS))
8509 Opcode = X86ISD::FMIN;
8512 // This can be a min, but if either operand is a NaN we need it to
8513 // preserve the original LHS.
8514 std::swap(LHS, RHS);
8518 Opcode = X86ISD::FMIN;
8522 // This can be a max if we can prove that at least one of the operands
8524 if (!FiniteOnlyFPMath()) {
8525 if (DAG.isKnownNeverNaN(LHS)) {
8526 // Put the potential NaN in the RHS so that SSE will preserve it.
8527 std::swap(LHS, RHS);
8528 } else if (!DAG.isKnownNeverNaN(RHS))
8531 Opcode = X86ISD::FMAX;
8534 // This can be a max if we can prove that at least one of the operands
8536 if (!FiniteOnlyFPMath()) {
8537 if (DAG.isKnownNeverNaN(RHS)) {
8538 // Put the potential NaN in the RHS so that SSE will preserve it.
8539 std::swap(LHS, RHS);
8540 } else if (!DAG.isKnownNeverNaN(LHS))
8543 Opcode = X86ISD::FMAX;
8546 // This can be a max, but if either operand is a NaN we need it to
8547 // preserve the original LHS.
8548 std::swap(LHS, RHS);
8552 Opcode = X86ISD::FMAX;
8558 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8561 // If this is a select between two integer constants, try to do some
8563 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8564 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8565 // Don't do this for crazy integer types.
8566 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8567 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8568 // so that TrueC (the true value) is larger than FalseC.
8569 bool NeedsCondInvert = false;
8571 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8572 // Efficiently invertible.
8573 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8574 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8575 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8576 NeedsCondInvert = true;
8577 std::swap(TrueC, FalseC);
8580 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8581 if (FalseC->getAPIntValue() == 0 &&
8582 TrueC->getAPIntValue().isPowerOf2()) {
8583 if (NeedsCondInvert) // Invert the condition if needed.
8584 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8585 DAG.getConstant(1, Cond.getValueType()));
8587 // Zero extend the condition if needed.
8588 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8590 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8591 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8592 DAG.getConstant(ShAmt, MVT::i8));
8595 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8596 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8597 if (NeedsCondInvert) // Invert the condition if needed.
8598 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8599 DAG.getConstant(1, Cond.getValueType()));
8601 // Zero extend the condition if needed.
8602 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8603 FalseC->getValueType(0), Cond);
8604 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8605 SDValue(FalseC, 0));
8608 // Optimize cases that will turn into an LEA instruction. This requires
8609 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8610 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8611 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8612 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8614 bool isFastMultiplier = false;
8616 switch ((unsigned char)Diff) {
8618 case 1: // result = add base, cond
8619 case 2: // result = lea base( , cond*2)
8620 case 3: // result = lea base(cond, cond*2)
8621 case 4: // result = lea base( , cond*4)
8622 case 5: // result = lea base(cond, cond*4)
8623 case 8: // result = lea base( , cond*8)
8624 case 9: // result = lea base(cond, cond*8)
8625 isFastMultiplier = true;
8630 if (isFastMultiplier) {
8631 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8632 if (NeedsCondInvert) // Invert the condition if needed.
8633 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8634 DAG.getConstant(1, Cond.getValueType()));
8636 // Zero extend the condition if needed.
8637 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8639 // Scale the condition by the difference.
8641 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8642 DAG.getConstant(Diff, Cond.getValueType()));
8644 // Add the base if non-zero.
8645 if (FalseC->getAPIntValue() != 0)
8646 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8647 SDValue(FalseC, 0));
8657 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
8658 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8659 TargetLowering::DAGCombinerInfo &DCI) {
8660 DebugLoc DL = N->getDebugLoc();
8662 // If the flag operand isn't dead, don't touch this CMOV.
8663 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8666 // If this is a select between two integer constants, try to do some
8667 // optimizations. Note that the operands are ordered the opposite of SELECT
8669 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
8670 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8671 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
8672 // larger than FalseC (the false value).
8673 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
8675 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
8676 CC = X86::GetOppositeBranchCondition(CC);
8677 std::swap(TrueC, FalseC);
8680 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
8681 // This is efficient for any integer data type (including i8/i16) and
8683 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
8684 SDValue Cond = N->getOperand(3);
8685 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8686 DAG.getConstant(CC, MVT::i8), Cond);
8688 // Zero extend the condition if needed.
8689 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
8691 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8692 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
8693 DAG.getConstant(ShAmt, MVT::i8));
8694 if (N->getNumValues() == 2) // Dead flag value?
8695 return DCI.CombineTo(N, Cond, SDValue());
8699 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
8700 // for any integer data type, including i8/i16.
8701 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8702 SDValue Cond = N->getOperand(3);
8703 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8704 DAG.getConstant(CC, MVT::i8), Cond);
8706 // Zero extend the condition if needed.
8707 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8708 FalseC->getValueType(0), Cond);
8709 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8710 SDValue(FalseC, 0));
8712 if (N->getNumValues() == 2) // Dead flag value?
8713 return DCI.CombineTo(N, Cond, SDValue());
8717 // Optimize cases that will turn into an LEA instruction. This requires
8718 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8719 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8720 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8721 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8723 bool isFastMultiplier = false;
8725 switch ((unsigned char)Diff) {
8727 case 1: // result = add base, cond
8728 case 2: // result = lea base( , cond*2)
8729 case 3: // result = lea base(cond, cond*2)
8730 case 4: // result = lea base( , cond*4)
8731 case 5: // result = lea base(cond, cond*4)
8732 case 8: // result = lea base( , cond*8)
8733 case 9: // result = lea base(cond, cond*8)
8734 isFastMultiplier = true;
8739 if (isFastMultiplier) {
8740 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8741 SDValue Cond = N->getOperand(3);
8742 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8743 DAG.getConstant(CC, MVT::i8), Cond);
8744 // Zero extend the condition if needed.
8745 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8747 // Scale the condition by the difference.
8749 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8750 DAG.getConstant(Diff, Cond.getValueType()));
8752 // Add the base if non-zero.
8753 if (FalseC->getAPIntValue() != 0)
8754 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8755 SDValue(FalseC, 0));
8756 if (N->getNumValues() == 2) // Dead flag value?
8757 return DCI.CombineTo(N, Cond, SDValue());
8767 /// PerformMulCombine - Optimize a single multiply with constant into two
8768 /// in order to implement it with two cheaper instructions, e.g.
8769 /// LEA + SHL, LEA + LEA.
8770 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
8771 TargetLowering::DAGCombinerInfo &DCI) {
8772 if (DAG.getMachineFunction().
8773 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
8776 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8779 EVT VT = N->getValueType(0);
8783 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8786 uint64_t MulAmt = C->getZExtValue();
8787 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
8790 uint64_t MulAmt1 = 0;
8791 uint64_t MulAmt2 = 0;
8792 if ((MulAmt % 9) == 0) {
8794 MulAmt2 = MulAmt / 9;
8795 } else if ((MulAmt % 5) == 0) {
8797 MulAmt2 = MulAmt / 5;
8798 } else if ((MulAmt % 3) == 0) {
8800 MulAmt2 = MulAmt / 3;
8803 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
8804 DebugLoc DL = N->getDebugLoc();
8806 if (isPowerOf2_64(MulAmt2) &&
8807 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
8808 // If second multiplifer is pow2, issue it first. We want the multiply by
8809 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
8811 std::swap(MulAmt1, MulAmt2);
8814 if (isPowerOf2_64(MulAmt1))
8815 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
8816 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
8818 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
8819 DAG.getConstant(MulAmt1, VT));
8821 if (isPowerOf2_64(MulAmt2))
8822 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
8823 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
8825 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
8826 DAG.getConstant(MulAmt2, VT));
8828 // Do not add new nodes to DAG combiner worklist.
8829 DCI.CombineTo(N, NewMul, false);
8835 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
8837 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
8838 const X86Subtarget *Subtarget) {
8839 // On X86 with SSE2 support, we can transform this to a vector shift if
8840 // all elements are shifted by the same amount. We can't do this in legalize
8841 // because the a constant vector is typically transformed to a constant pool
8842 // so we have no knowledge of the shift amount.
8843 if (!Subtarget->hasSSE2())
8846 EVT VT = N->getValueType(0);
8847 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
8850 SDValue ShAmtOp = N->getOperand(1);
8851 EVT EltVT = VT.getVectorElementType();
8852 DebugLoc DL = N->getDebugLoc();
8853 SDValue BaseShAmt = SDValue();
8854 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
8855 unsigned NumElts = VT.getVectorNumElements();
8857 for (; i != NumElts; ++i) {
8858 SDValue Arg = ShAmtOp.getOperand(i);
8859 if (Arg.getOpcode() == ISD::UNDEF) continue;
8863 for (; i != NumElts; ++i) {
8864 SDValue Arg = ShAmtOp.getOperand(i);
8865 if (Arg.getOpcode() == ISD::UNDEF) continue;
8866 if (Arg != BaseShAmt) {
8870 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
8871 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
8872 SDValue InVec = ShAmtOp.getOperand(0);
8873 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
8874 unsigned NumElts = InVec.getValueType().getVectorNumElements();
8876 for (; i != NumElts; ++i) {
8877 SDValue Arg = InVec.getOperand(i);
8878 if (Arg.getOpcode() == ISD::UNDEF) continue;
8882 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
8883 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
8884 unsigned SplatIdx = cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
8885 if (C->getZExtValue() == SplatIdx)
8886 BaseShAmt = InVec.getOperand(1);
8889 if (BaseShAmt.getNode() == 0)
8890 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
8891 DAG.getIntPtrConstant(0));
8895 // The shift amount is an i32.
8896 if (EltVT.bitsGT(MVT::i32))
8897 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
8898 else if (EltVT.bitsLT(MVT::i32))
8899 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
8901 // The shift amount is identical so we can do a vector shift.
8902 SDValue ValOp = N->getOperand(0);
8903 switch (N->getOpcode()) {
8905 llvm_unreachable("Unknown shift opcode!");
8908 if (VT == MVT::v2i64)
8909 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8910 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8912 if (VT == MVT::v4i32)
8913 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8914 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8916 if (VT == MVT::v8i16)
8917 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8918 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8922 if (VT == MVT::v4i32)
8923 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8924 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
8926 if (VT == MVT::v8i16)
8927 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8928 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
8932 if (VT == MVT::v2i64)
8933 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8934 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8936 if (VT == MVT::v4i32)
8937 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8938 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
8940 if (VT == MVT::v8i16)
8941 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8942 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
8949 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
8950 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
8951 const X86Subtarget *Subtarget) {
8952 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
8953 // the FP state in cases where an emms may be missing.
8954 // A preferable solution to the general problem is to figure out the right
8955 // places to insert EMMS. This qualifies as a quick hack.
8957 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
8958 StoreSDNode *St = cast<StoreSDNode>(N);
8959 EVT VT = St->getValue().getValueType();
8960 if (VT.getSizeInBits() != 64)
8963 const Function *F = DAG.getMachineFunction().getFunction();
8964 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
8965 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
8966 && Subtarget->hasSSE2();
8967 if ((VT.isVector() ||
8968 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
8969 isa<LoadSDNode>(St->getValue()) &&
8970 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
8971 St->getChain().hasOneUse() && !St->isVolatile()) {
8972 SDNode* LdVal = St->getValue().getNode();
8974 int TokenFactorIndex = -1;
8975 SmallVector<SDValue, 8> Ops;
8976 SDNode* ChainVal = St->getChain().getNode();
8977 // Must be a store of a load. We currently handle two cases: the load
8978 // is a direct child, and it's under an intervening TokenFactor. It is
8979 // possible to dig deeper under nested TokenFactors.
8980 if (ChainVal == LdVal)
8981 Ld = cast<LoadSDNode>(St->getChain());
8982 else if (St->getValue().hasOneUse() &&
8983 ChainVal->getOpcode() == ISD::TokenFactor) {
8984 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
8985 if (ChainVal->getOperand(i).getNode() == LdVal) {
8986 TokenFactorIndex = i;
8987 Ld = cast<LoadSDNode>(St->getValue());
8989 Ops.push_back(ChainVal->getOperand(i));
8993 if (!Ld || !ISD::isNormalLoad(Ld))
8996 // If this is not the MMX case, i.e. we are just turning i64 load/store
8997 // into f64 load/store, avoid the transformation if there are multiple
8998 // uses of the loaded value.
8999 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9002 DebugLoc LdDL = Ld->getDebugLoc();
9003 DebugLoc StDL = N->getDebugLoc();
9004 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9005 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9007 if (Subtarget->is64Bit() || F64IsLegal) {
9008 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9009 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9010 Ld->getBasePtr(), Ld->getSrcValue(),
9011 Ld->getSrcValueOffset(), Ld->isVolatile(),
9012 Ld->getAlignment());
9013 SDValue NewChain = NewLd.getValue(1);
9014 if (TokenFactorIndex != -1) {
9015 Ops.push_back(NewChain);
9016 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9019 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9020 St->getSrcValue(), St->getSrcValueOffset(),
9021 St->isVolatile(), St->getAlignment());
9024 // Otherwise, lower to two pairs of 32-bit loads / stores.
9025 SDValue LoAddr = Ld->getBasePtr();
9026 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9027 DAG.getConstant(4, MVT::i32));
9029 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9030 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9031 Ld->isVolatile(), Ld->getAlignment());
9032 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9033 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9035 MinAlign(Ld->getAlignment(), 4));
9037 SDValue NewChain = LoLd.getValue(1);
9038 if (TokenFactorIndex != -1) {
9039 Ops.push_back(LoLd);
9040 Ops.push_back(HiLd);
9041 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9045 LoAddr = St->getBasePtr();
9046 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9047 DAG.getConstant(4, MVT::i32));
9049 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9050 St->getSrcValue(), St->getSrcValueOffset(),
9051 St->isVolatile(), St->getAlignment());
9052 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9054 St->getSrcValueOffset() + 4,
9056 MinAlign(St->getAlignment(), 4));
9057 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9062 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9063 /// X86ISD::FXOR nodes.
9064 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9065 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9066 // F[X]OR(0.0, x) -> x
9067 // F[X]OR(x, 0.0) -> x
9068 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9069 if (C->getValueAPF().isPosZero())
9070 return N->getOperand(1);
9071 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9072 if (C->getValueAPF().isPosZero())
9073 return N->getOperand(0);
9077 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9078 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9079 // FAND(0.0, x) -> 0.0
9080 // FAND(x, 0.0) -> 0.0
9081 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9082 if (C->getValueAPF().isPosZero())
9083 return N->getOperand(0);
9084 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9085 if (C->getValueAPF().isPosZero())
9086 return N->getOperand(1);
9090 static SDValue PerformBTCombine(SDNode *N,
9092 TargetLowering::DAGCombinerInfo &DCI) {
9093 // BT ignores high bits in the bit index operand.
9094 SDValue Op1 = N->getOperand(1);
9095 if (Op1.hasOneUse()) {
9096 unsigned BitWidth = Op1.getValueSizeInBits();
9097 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9098 APInt KnownZero, KnownOne;
9099 TargetLowering::TargetLoweringOpt TLO(DAG);
9100 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9101 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9102 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9103 DCI.CommitTargetLoweringOpt(TLO);
9108 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9109 SDValue Op = N->getOperand(0);
9110 if (Op.getOpcode() == ISD::BIT_CONVERT)
9111 Op = Op.getOperand(0);
9112 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9113 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9114 VT.getVectorElementType().getSizeInBits() ==
9115 OpVT.getVectorElementType().getSizeInBits()) {
9116 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9121 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9122 // Locked instructions, in turn, have implicit fence semantics (all memory
9123 // operations are flushed before issuing the locked instruction, and the
9124 // are not buffered), so we can fold away the common pattern of
9125 // fence-atomic-fence.
9126 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9127 SDValue atomic = N->getOperand(0);
9128 switch (atomic.getOpcode()) {
9129 case ISD::ATOMIC_CMP_SWAP:
9130 case ISD::ATOMIC_SWAP:
9131 case ISD::ATOMIC_LOAD_ADD:
9132 case ISD::ATOMIC_LOAD_SUB:
9133 case ISD::ATOMIC_LOAD_AND:
9134 case ISD::ATOMIC_LOAD_OR:
9135 case ISD::ATOMIC_LOAD_XOR:
9136 case ISD::ATOMIC_LOAD_NAND:
9137 case ISD::ATOMIC_LOAD_MIN:
9138 case ISD::ATOMIC_LOAD_MAX:
9139 case ISD::ATOMIC_LOAD_UMIN:
9140 case ISD::ATOMIC_LOAD_UMAX:
9146 SDValue fence = atomic.getOperand(0);
9147 if (fence.getOpcode() != ISD::MEMBARRIER)
9150 switch (atomic.getOpcode()) {
9151 case ISD::ATOMIC_CMP_SWAP:
9152 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9153 atomic.getOperand(1), atomic.getOperand(2),
9154 atomic.getOperand(3));
9155 case ISD::ATOMIC_SWAP:
9156 case ISD::ATOMIC_LOAD_ADD:
9157 case ISD::ATOMIC_LOAD_SUB:
9158 case ISD::ATOMIC_LOAD_AND:
9159 case ISD::ATOMIC_LOAD_OR:
9160 case ISD::ATOMIC_LOAD_XOR:
9161 case ISD::ATOMIC_LOAD_NAND:
9162 case ISD::ATOMIC_LOAD_MIN:
9163 case ISD::ATOMIC_LOAD_MAX:
9164 case ISD::ATOMIC_LOAD_UMIN:
9165 case ISD::ATOMIC_LOAD_UMAX:
9166 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9167 atomic.getOperand(1), atomic.getOperand(2));
9173 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9174 DAGCombinerInfo &DCI) const {
9175 SelectionDAG &DAG = DCI.DAG;
9176 switch (N->getOpcode()) {
9178 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9179 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9180 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9181 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9184 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9185 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9187 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9188 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9189 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9190 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9191 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9197 //===----------------------------------------------------------------------===//
9198 // X86 Inline Assembly Support
9199 //===----------------------------------------------------------------------===//
9201 static bool LowerToBSwap(CallInst *CI) {
9202 // FIXME: this should verify that we are targetting a 486 or better. If not,
9203 // we will turn this bswap into something that will be lowered to logical ops
9204 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9205 // so don't worry about this.
9207 // Verify this is a simple bswap.
9208 if (CI->getNumOperands() != 2 ||
9209 CI->getType() != CI->getOperand(1)->getType() ||
9210 !CI->getType()->isInteger())
9213 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9214 if (!Ty || Ty->getBitWidth() % 16 != 0)
9217 // Okay, we can do this xform, do so now.
9218 const Type *Tys[] = { Ty };
9219 Module *M = CI->getParent()->getParent()->getParent();
9220 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9222 Value *Op = CI->getOperand(1);
9223 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9225 CI->replaceAllUsesWith(Op);
9226 CI->eraseFromParent();
9230 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9231 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9232 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9234 std::string AsmStr = IA->getAsmString();
9236 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9237 std::vector<std::string> AsmPieces;
9238 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9240 switch (AsmPieces.size()) {
9241 default: return false;
9243 AsmStr = AsmPieces[0];
9245 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9248 if (AsmPieces.size() == 2 &&
9249 (AsmPieces[0] == "bswap" ||
9250 AsmPieces[0] == "bswapq" ||
9251 AsmPieces[0] == "bswapl") &&
9252 (AsmPieces[1] == "$0" ||
9253 AsmPieces[1] == "${0:q}")) {
9254 // No need to check constraints, nothing other than the equivalent of
9255 // "=r,0" would be valid here.
9256 return LowerToBSwap(CI);
9258 // rorw $$8, ${0:w} --> llvm.bswap.i16
9259 if (CI->getType() == Type::getInt16Ty(CI->getContext()) &&
9260 AsmPieces.size() == 3 &&
9261 AsmPieces[0] == "rorw" &&
9262 AsmPieces[1] == "$$8," &&
9263 AsmPieces[2] == "${0:w}" &&
9264 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
9265 return LowerToBSwap(CI);
9269 if (CI->getType() == Type::getInt64Ty(CI->getContext()) &&
9270 Constraints.size() >= 2 &&
9271 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9272 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9273 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9274 std::vector<std::string> Words;
9275 SplitString(AsmPieces[0], Words, " \t");
9276 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9278 SplitString(AsmPieces[1], Words, " \t");
9279 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9281 SplitString(AsmPieces[2], Words, " \t,");
9282 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9283 Words[2] == "%edx") {
9284 return LowerToBSwap(CI);
9296 /// getConstraintType - Given a constraint letter, return the type of
9297 /// constraint it is for this target.
9298 X86TargetLowering::ConstraintType
9299 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
9300 if (Constraint.size() == 1) {
9301 switch (Constraint[0]) {
9313 return C_RegisterClass;
9321 return TargetLowering::getConstraintType(Constraint);
9324 /// LowerXConstraint - try to replace an X constraint, which matches anything,
9325 /// with another that has more specific requirements based on the type of the
9326 /// corresponding operand.
9327 const char *X86TargetLowering::
9328 LowerXConstraint(EVT ConstraintVT) const {
9329 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
9330 // 'f' like normal targets.
9331 if (ConstraintVT.isFloatingPoint()) {
9332 if (Subtarget->hasSSE2())
9334 if (Subtarget->hasSSE1())
9338 return TargetLowering::LowerXConstraint(ConstraintVT);
9341 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9342 /// vector. If it is invalid, don't add anything to Ops.
9343 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9346 std::vector<SDValue>&Ops,
9347 SelectionDAG &DAG) const {
9348 SDValue Result(0, 0);
9350 switch (Constraint) {
9353 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9354 if (C->getZExtValue() <= 31) {
9355 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9361 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9362 if (C->getZExtValue() <= 63) {
9363 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9369 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9370 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9371 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9377 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9378 if (C->getZExtValue() <= 255) {
9379 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9385 // 32-bit signed value
9386 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9387 const ConstantInt *CI = C->getConstantIntValue();
9388 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9389 C->getSExtValue())) {
9390 // Widen to 64 bits here to get it sign extended.
9391 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9394 // FIXME gcc accepts some relocatable values here too, but only in certain
9395 // memory models; it's complicated.
9400 // 32-bit unsigned value
9401 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9402 const ConstantInt *CI = C->getConstantIntValue();
9403 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9404 C->getZExtValue())) {
9405 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9409 // FIXME gcc accepts some relocatable values here too, but only in certain
9410 // memory models; it's complicated.
9414 // Literal immediates are always ok.
9415 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
9416 // Widen to 64 bits here to get it sign extended.
9417 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
9421 // If we are in non-pic codegen mode, we allow the address of a global (with
9422 // an optional displacement) to be used with 'i'.
9423 GlobalAddressSDNode *GA = 0;
9426 // Match either (GA), (GA+C), (GA+C1+C2), etc.
9428 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
9429 Offset += GA->getOffset();
9431 } else if (Op.getOpcode() == ISD::ADD) {
9432 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9433 Offset += C->getZExtValue();
9434 Op = Op.getOperand(0);
9437 } else if (Op.getOpcode() == ISD::SUB) {
9438 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9439 Offset += -C->getZExtValue();
9440 Op = Op.getOperand(0);
9445 // Otherwise, this isn't something we can handle, reject it.
9449 GlobalValue *GV = GA->getGlobal();
9450 // If we require an extra load to get this address, as in PIC mode, we
9452 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
9453 getTargetMachine())))
9457 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
9459 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
9465 if (Result.getNode()) {
9466 Ops.push_back(Result);
9469 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
9473 std::vector<unsigned> X86TargetLowering::
9474 getRegClassForInlineAsmConstraint(const std::string &Constraint,
9476 if (Constraint.size() == 1) {
9477 // FIXME: not handling fp-stack yet!
9478 switch (Constraint[0]) { // GCC X86 Constraint Letters
9479 default: break; // Unknown constraint letter
9480 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
9481 if (Subtarget->is64Bit()) {
9483 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
9484 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
9485 X86::R10D,X86::R11D,X86::R12D,
9486 X86::R13D,X86::R14D,X86::R15D,
9487 X86::EBP, X86::ESP, 0);
9488 else if (VT == MVT::i16)
9489 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
9490 X86::SI, X86::DI, X86::R8W,X86::R9W,
9491 X86::R10W,X86::R11W,X86::R12W,
9492 X86::R13W,X86::R14W,X86::R15W,
9493 X86::BP, X86::SP, 0);
9494 else if (VT == MVT::i8)
9495 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
9496 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
9497 X86::R10B,X86::R11B,X86::R12B,
9498 X86::R13B,X86::R14B,X86::R15B,
9499 X86::BPL, X86::SPL, 0);
9501 else if (VT == MVT::i64)
9502 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
9503 X86::RSI, X86::RDI, X86::R8, X86::R9,
9504 X86::R10, X86::R11, X86::R12,
9505 X86::R13, X86::R14, X86::R15,
9506 X86::RBP, X86::RSP, 0);
9510 // 32-bit fallthrough
9513 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
9514 else if (VT == MVT::i16)
9515 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
9516 else if (VT == MVT::i8)
9517 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
9518 else if (VT == MVT::i64)
9519 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
9524 return std::vector<unsigned>();
9527 std::pair<unsigned, const TargetRegisterClass*>
9528 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9530 // First, see if this is a constraint that directly corresponds to an LLVM
9532 if (Constraint.size() == 1) {
9533 // GCC Constraint Letters
9534 switch (Constraint[0]) {
9536 case 'r': // GENERAL_REGS
9537 case 'l': // INDEX_REGS
9539 return std::make_pair(0U, X86::GR8RegisterClass);
9541 return std::make_pair(0U, X86::GR16RegisterClass);
9542 if (VT == MVT::i32 || !Subtarget->is64Bit())
9543 return std::make_pair(0U, X86::GR32RegisterClass);
9544 return std::make_pair(0U, X86::GR64RegisterClass);
9545 case 'R': // LEGACY_REGS
9547 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
9549 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
9550 if (VT == MVT::i32 || !Subtarget->is64Bit())
9551 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
9552 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
9553 case 'f': // FP Stack registers.
9554 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
9555 // value to the correct fpstack register class.
9556 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
9557 return std::make_pair(0U, X86::RFP32RegisterClass);
9558 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
9559 return std::make_pair(0U, X86::RFP64RegisterClass);
9560 return std::make_pair(0U, X86::RFP80RegisterClass);
9561 case 'y': // MMX_REGS if MMX allowed.
9562 if (!Subtarget->hasMMX()) break;
9563 return std::make_pair(0U, X86::VR64RegisterClass);
9564 case 'Y': // SSE_REGS if SSE2 allowed
9565 if (!Subtarget->hasSSE2()) break;
9567 case 'x': // SSE_REGS if SSE1 allowed
9568 if (!Subtarget->hasSSE1()) break;
9570 switch (VT.getSimpleVT().SimpleTy) {
9572 // Scalar SSE types.
9575 return std::make_pair(0U, X86::FR32RegisterClass);
9578 return std::make_pair(0U, X86::FR64RegisterClass);
9586 return std::make_pair(0U, X86::VR128RegisterClass);
9592 // Use the default implementation in TargetLowering to convert the register
9593 // constraint into a member of a register class.
9594 std::pair<unsigned, const TargetRegisterClass*> Res;
9595 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9597 // Not found as a standard register?
9598 if (Res.second == 0) {
9599 // Map st(0) -> st(7) -> ST0
9600 if (Constraint.size() == 7 && Constraint[0] == '{' &&
9601 tolower(Constraint[1]) == 's' &&
9602 tolower(Constraint[2]) == 't' &&
9603 Constraint[3] == '(' &&
9604 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
9605 Constraint[5] == ')' &&
9606 Constraint[6] == '}') {
9608 Res.first = X86::ST0+Constraint[4]-'0';
9609 Res.second = X86::RFP80RegisterClass;
9613 // GCC allows "st(0)" to be called just plain "st".
9614 if (StringRef("{st}").equals_lower(Constraint)) {
9615 Res.first = X86::ST0;
9616 Res.second = X86::RFP80RegisterClass;
9621 if (StringRef("{flags}").equals_lower(Constraint)) {
9622 Res.first = X86::EFLAGS;
9623 Res.second = X86::CCRRegisterClass;
9627 // 'A' means EAX + EDX.
9628 if (Constraint == "A") {
9629 Res.first = X86::EAX;
9630 Res.second = X86::GR32_ADRegisterClass;
9636 // Otherwise, check to see if this is a register class of the wrong value
9637 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
9638 // turn into {ax},{dx}.
9639 if (Res.second->hasType(VT))
9640 return Res; // Correct type already, nothing to do.
9642 // All of the single-register GCC register classes map their values onto
9643 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
9644 // really want an 8-bit or 32-bit register, map to the appropriate register
9645 // class and return the appropriate register.
9646 if (Res.second == X86::GR16RegisterClass) {
9647 if (VT == MVT::i8) {
9648 unsigned DestReg = 0;
9649 switch (Res.first) {
9651 case X86::AX: DestReg = X86::AL; break;
9652 case X86::DX: DestReg = X86::DL; break;
9653 case X86::CX: DestReg = X86::CL; break;
9654 case X86::BX: DestReg = X86::BL; break;
9657 Res.first = DestReg;
9658 Res.second = X86::GR8RegisterClass;
9660 } else if (VT == MVT::i32) {
9661 unsigned DestReg = 0;
9662 switch (Res.first) {
9664 case X86::AX: DestReg = X86::EAX; break;
9665 case X86::DX: DestReg = X86::EDX; break;
9666 case X86::CX: DestReg = X86::ECX; break;
9667 case X86::BX: DestReg = X86::EBX; break;
9668 case X86::SI: DestReg = X86::ESI; break;
9669 case X86::DI: DestReg = X86::EDI; break;
9670 case X86::BP: DestReg = X86::EBP; break;
9671 case X86::SP: DestReg = X86::ESP; break;
9674 Res.first = DestReg;
9675 Res.second = X86::GR32RegisterClass;
9677 } else if (VT == MVT::i64) {
9678 unsigned DestReg = 0;
9679 switch (Res.first) {
9681 case X86::AX: DestReg = X86::RAX; break;
9682 case X86::DX: DestReg = X86::RDX; break;
9683 case X86::CX: DestReg = X86::RCX; break;
9684 case X86::BX: DestReg = X86::RBX; break;
9685 case X86::SI: DestReg = X86::RSI; break;
9686 case X86::DI: DestReg = X86::RDI; break;
9687 case X86::BP: DestReg = X86::RBP; break;
9688 case X86::SP: DestReg = X86::RSP; break;
9691 Res.first = DestReg;
9692 Res.second = X86::GR64RegisterClass;
9695 } else if (Res.second == X86::FR32RegisterClass ||
9696 Res.second == X86::FR64RegisterClass ||
9697 Res.second == X86::VR128RegisterClass) {
9698 // Handle references to XMM physical registers that got mapped into the
9699 // wrong class. This can happen with constraints like {xmm0} where the
9700 // target independent register mapper will just pick the first match it can
9701 // find, ignoring the required type.
9703 Res.second = X86::FR32RegisterClass;
9704 else if (VT == MVT::f64)
9705 Res.second = X86::FR64RegisterClass;
9706 else if (X86::VR128RegisterClass->hasType(VT))
9707 Res.second = X86::VR128RegisterClass;
9713 //===----------------------------------------------------------------------===//
9714 // X86 Widen vector type
9715 //===----------------------------------------------------------------------===//
9717 /// getWidenVectorType: given a vector type, returns the type to widen
9718 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
9719 /// If there is no vector type that we want to widen to, returns MVT::Other
9720 /// When and where to widen is target dependent based on the cost of
9721 /// scalarizing vs using the wider vector type.
9723 EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
9724 assert(VT.isVector());
9725 if (isTypeLegal(VT))
9728 // TODO: In computeRegisterProperty, we can compute the list of legal vector
9729 // type based on element type. This would speed up our search (though
9730 // it may not be worth it since the size of the list is relatively
9732 EVT EltVT = VT.getVectorElementType();
9733 unsigned NElts = VT.getVectorNumElements();
9735 // On X86, it make sense to widen any vector wider than 1
9739 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
9740 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
9741 EVT SVT = (MVT::SimpleValueType)nVT;
9743 if (isTypeLegal(SVT) &&
9744 SVT.getVectorElementType() == EltVT &&
9745 SVT.getVectorNumElements() > NElts)