1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "x86-isel"
17 #include "X86InstrBuilder.h"
18 #include "X86ISelLowering.h"
19 #include "X86TargetMachine.h"
20 #include "X86TargetObjectFile.h"
21 #include "llvm/CallingConv.h"
22 #include "llvm/Constants.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/GlobalAlias.h"
25 #include "llvm/GlobalVariable.h"
26 #include "llvm/Function.h"
27 #include "llvm/Instructions.h"
28 #include "llvm/Intrinsics.h"
29 #include "llvm/LLVMContext.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/PseudoSourceValue.h"
37 #include "llvm/MC/MCAsmInfo.h"
38 #include "llvm/MC/MCContext.h"
39 #include "llvm/MC/MCExpr.h"
40 #include "llvm/MC/MCSymbol.h"
41 #include "llvm/ADT/BitVector.h"
42 #include "llvm/ADT/SmallSet.h"
43 #include "llvm/ADT/Statistic.h"
44 #include "llvm/ADT/StringExtras.h"
45 #include "llvm/ADT/VectorExtras.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/Dwarf.h"
49 #include "llvm/Support/ErrorHandling.h"
50 #include "llvm/Support/MathExtras.h"
51 #include "llvm/Support/raw_ostream.h"
53 using namespace dwarf;
55 STATISTIC(NumTailCalls, "Number of tail calls");
58 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
60 // Forward declarations.
61 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
64 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
66 bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
68 if (TM.getSubtarget<X86Subtarget>().isTargetDarwin()) {
69 if (is64Bit) return new X8664_MachoTargetObjectFile();
70 return new TargetLoweringObjectFileMachO();
71 } else if (TM.getSubtarget<X86Subtarget>().isTargetELF() ){
72 if (is64Bit) return new X8664_ELFTargetObjectFile(TM);
73 return new X8632_ELFTargetObjectFile(TM);
74 } else if (TM.getSubtarget<X86Subtarget>().isTargetCOFF()) {
75 return new TargetLoweringObjectFileCOFF();
77 llvm_unreachable("unknown subtarget type");
80 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
81 : TargetLowering(TM, createTLOF(TM)) {
82 Subtarget = &TM.getSubtarget<X86Subtarget>();
83 X86ScalarSSEf64 = Subtarget->hasSSE2();
84 X86ScalarSSEf32 = Subtarget->hasSSE1();
85 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
87 RegInfo = TM.getRegisterInfo();
90 // Set up the TargetLowering object.
92 // X86 is weird, it always uses i8 for shift amounts and setcc results.
93 setShiftAmountType(MVT::i8);
94 setBooleanContents(ZeroOrOneBooleanContent);
95 setSchedulingPreference(Sched::RegPressure);
96 setStackPointerRegisterToSaveRestore(X86StackPtr);
98 if (Subtarget->isTargetDarwin()) {
99 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
100 setUseUnderscoreSetJmp(false);
101 setUseUnderscoreLongJmp(false);
102 } else if (Subtarget->isTargetMingw()) {
103 // MS runtime is weird: it exports _setjmp, but longjmp!
104 setUseUnderscoreSetJmp(true);
105 setUseUnderscoreLongJmp(false);
107 setUseUnderscoreSetJmp(true);
108 setUseUnderscoreLongJmp(true);
111 // Set up the register classes.
112 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);
122 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
123 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
124 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
125 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
126 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
128 // SETOEQ and SETUNE require checking two conditions.
129 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
130 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
131 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
132 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
133 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
134 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
136 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
138 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
139 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
140 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
142 if (Subtarget->is64Bit()) {
143 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
144 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
145 } else if (!UseSoftFloat) {
146 // We have an algorithm for SSE2->double, and we turn this into a
147 // 64-bit FILD followed by conditional FADD for other targets.
148 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
149 // We have an algorithm for SSE2, and we turn this into a 64-bit
150 // FILD for other targets.
151 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
154 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
156 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
157 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
160 // SSE has no i16 to fp conversion, only i32
161 if (X86ScalarSSEf32) {
162 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
163 // f32 and f64 cases are Legal, f80 case is not
164 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
166 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
167 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
170 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
171 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
174 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
175 // are Legal, f80 is custom lowered.
176 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
177 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
179 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
181 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
182 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
184 if (X86ScalarSSEf32) {
185 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
186 // f32 and f64 cases are Legal, f80 case is not
187 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
189 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
190 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
193 // Handle FP_TO_UINT by promoting the destination to a larger signed
195 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
196 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
197 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
199 if (Subtarget->is64Bit()) {
200 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
201 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
202 } else if (!UseSoftFloat) {
203 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
204 // Expand FP_TO_UINT into a select.
205 // FIXME: We would like to use a Custom expander here eventually to do
206 // the optimal thing for SSE vs. the default expansion in the legalizer.
207 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
209 // With SSE3 we can use fisttpll to convert to a signed i64; without
210 // SSE, we're stuck with a fistpll.
211 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
214 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
215 if (!X86ScalarSSEf64) {
216 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
217 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
218 if (Subtarget->is64Bit()) {
219 setOperationAction(ISD::BIT_CONVERT , MVT::f64 , Expand);
220 // Without SSE, i64->f64 goes through memory; i64->MMX is Legal.
221 if (Subtarget->hasMMX() && !DisableMMX)
222 setOperationAction(ISD::BIT_CONVERT , MVT::i64 , Custom);
224 setOperationAction(ISD::BIT_CONVERT , MVT::i64 , Expand);
228 // Scalar integer divide and remainder are lowered to use operations that
229 // produce two results, to match the available instructions. This exposes
230 // the two-result form to trivial CSE, which is able to combine x/y and x%y
231 // into a single instruction.
233 // Scalar integer multiply-high is also lowered to use two-result
234 // operations, to match the available instructions. However, plain multiply
235 // (low) operations are left as Legal, as there are single-result
236 // instructions for this in x86. Using the two-result multiply instructions
237 // when both high and low results are needed must be arranged by dagcombine.
238 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
239 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
240 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
241 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
242 setOperationAction(ISD::SREM , MVT::i8 , Expand);
243 setOperationAction(ISD::UREM , MVT::i8 , Expand);
244 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
245 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
246 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
247 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
248 setOperationAction(ISD::SREM , MVT::i16 , Expand);
249 setOperationAction(ISD::UREM , MVT::i16 , Expand);
250 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
251 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
252 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
253 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
254 setOperationAction(ISD::SREM , MVT::i32 , Expand);
255 setOperationAction(ISD::UREM , MVT::i32 , Expand);
256 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
257 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
258 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
259 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
260 setOperationAction(ISD::SREM , MVT::i64 , Expand);
261 setOperationAction(ISD::UREM , MVT::i64 , Expand);
263 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
264 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
265 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
266 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
267 if (Subtarget->is64Bit())
268 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
269 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
270 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
271 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
272 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
273 setOperationAction(ISD::FREM , MVT::f32 , Expand);
274 setOperationAction(ISD::FREM , MVT::f64 , Expand);
275 setOperationAction(ISD::FREM , MVT::f80 , Expand);
276 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
278 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
279 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
280 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
281 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
282 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
283 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);
300 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
301 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
302 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
303 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
304 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
305 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
306 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
307 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
308 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
309 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
310 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
311 if (Subtarget->is64Bit()) {
312 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
313 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
315 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
318 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
319 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
320 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
321 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
322 if (Subtarget->is64Bit())
323 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
324 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
325 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
326 if (Subtarget->is64Bit()) {
327 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
328 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
329 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
330 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
331 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
333 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
334 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
335 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
336 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
337 if (Subtarget->is64Bit()) {
338 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
339 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
340 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
343 if (Subtarget->hasSSE1())
344 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
346 // We may not have a libcall for MEMBARRIER so we should lower this.
347 setOperationAction(ISD::MEMBARRIER , MVT::Other, Custom);
349 // On X86 and X86-64, atomic operations are lowered to locked instructions.
350 // Locked instructions, in turn, have implicit fence semantics (all memory
351 // operations are flushed before issuing the locked instruction, and they
352 // are not buffered), so we can fold away the common pattern of
353 // fence-atomic-fence.
354 setShouldFoldAtomicFences(true);
356 // Expand certain atomics
357 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
358 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
359 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
360 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
362 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
363 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
364 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
365 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
367 if (!Subtarget->is64Bit()) {
368 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
369 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
370 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
371 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
372 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
373 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
374 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
377 // FIXME - use subtarget debug flags
378 if (!Subtarget->isTargetDarwin() &&
379 !Subtarget->isTargetELF() &&
380 !Subtarget->isTargetCygMing()) {
381 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
384 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
385 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
386 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
387 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
388 if (Subtarget->is64Bit()) {
389 setExceptionPointerRegister(X86::RAX);
390 setExceptionSelectorRegister(X86::RDX);
392 setExceptionPointerRegister(X86::EAX);
393 setExceptionSelectorRegister(X86::EDX);
395 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
396 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
398 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
400 setOperationAction(ISD::TRAP, MVT::Other, Legal);
402 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
403 setOperationAction(ISD::VASTART , MVT::Other, Custom);
404 setOperationAction(ISD::VAEND , MVT::Other, Expand);
405 if (Subtarget->is64Bit()) {
406 setOperationAction(ISD::VAARG , MVT::Other, Custom);
407 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
409 setOperationAction(ISD::VAARG , MVT::Other, Expand);
410 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
413 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
414 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
415 if (Subtarget->is64Bit())
416 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
417 if (Subtarget->isTargetCygMing())
418 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
420 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
422 if (!UseSoftFloat && X86ScalarSSEf64) {
423 // f32 and f64 use SSE.
424 // Set up the FP register classes.
425 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
426 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
428 // Use ANDPD to simulate FABS.
429 setOperationAction(ISD::FABS , MVT::f64, Custom);
430 setOperationAction(ISD::FABS , MVT::f32, Custom);
432 // Use XORP to simulate FNEG.
433 setOperationAction(ISD::FNEG , MVT::f64, Custom);
434 setOperationAction(ISD::FNEG , MVT::f32, Custom);
436 // Use ANDPD and ORPD to simulate FCOPYSIGN.
437 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
438 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
440 // We don't support sin/cos/fmod
441 setOperationAction(ISD::FSIN , MVT::f64, Expand);
442 setOperationAction(ISD::FCOS , MVT::f64, Expand);
443 setOperationAction(ISD::FSIN , MVT::f32, Expand);
444 setOperationAction(ISD::FCOS , MVT::f32, Expand);
446 // Expand FP immediates into loads from the stack, except for the special
448 addLegalFPImmediate(APFloat(+0.0)); // xorpd
449 addLegalFPImmediate(APFloat(+0.0f)); // xorps
450 } else if (!UseSoftFloat && X86ScalarSSEf32) {
451 // Use SSE for f32, x87 for f64.
452 // Set up the FP register classes.
453 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
454 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
456 // Use ANDPS to simulate FABS.
457 setOperationAction(ISD::FABS , MVT::f32, Custom);
459 // Use XORP to simulate FNEG.
460 setOperationAction(ISD::FNEG , MVT::f32, Custom);
462 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
464 // Use ANDPS and ORPS to simulate FCOPYSIGN.
465 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
466 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
468 // We don't support sin/cos/fmod
469 setOperationAction(ISD::FSIN , MVT::f32, Expand);
470 setOperationAction(ISD::FCOS , MVT::f32, Expand);
472 // Special cases we handle for FP constants.
473 addLegalFPImmediate(APFloat(+0.0f)); // xorps
474 addLegalFPImmediate(APFloat(+0.0)); // FLD0
475 addLegalFPImmediate(APFloat(+1.0)); // FLD1
476 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
477 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
480 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
481 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
483 } else if (!UseSoftFloat) {
484 // f32 and f64 in x87.
485 // Set up the FP register classes.
486 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
487 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
489 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
490 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
491 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
492 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
495 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
496 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
498 addLegalFPImmediate(APFloat(+0.0)); // FLD0
499 addLegalFPImmediate(APFloat(+1.0)); // FLD1
500 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
501 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
502 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
503 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
504 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
505 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
508 // Long double always uses X87.
510 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
511 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
512 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
515 APFloat TmpFlt(+0.0);
516 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
518 addLegalFPImmediate(TmpFlt); // FLD0
520 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
521 APFloat TmpFlt2(+1.0);
522 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
524 addLegalFPImmediate(TmpFlt2); // FLD1
525 TmpFlt2.changeSign();
526 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
530 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
531 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
535 // Always use a library call for pow.
536 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
537 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
538 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
540 setOperationAction(ISD::FLOG, MVT::f80, Expand);
541 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
542 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
543 setOperationAction(ISD::FEXP, MVT::f80, Expand);
544 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
546 // First set operation action for all vector types to either promote
547 // (for widening) or expand (for scalarization). Then we will selectively
548 // turn on ones that can be effectively codegen'd.
549 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
550 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
551 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
566 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
567 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
599 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
600 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
601 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
603 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
604 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
605 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
606 setTruncStoreAction((MVT::SimpleValueType)VT,
607 (MVT::SimpleValueType)InnerVT, Expand);
608 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
609 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
610 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
613 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
614 // with -msoft-float, disable use of MMX as well.
615 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
616 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass, false);
617 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass, false);
618 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass, false);
620 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass, false);
622 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
623 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
624 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
625 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
627 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
628 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
629 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
630 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
632 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
633 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
635 setOperationAction(ISD::AND, MVT::v8i8, Promote);
636 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
637 setOperationAction(ISD::AND, MVT::v4i16, Promote);
638 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
639 setOperationAction(ISD::AND, MVT::v2i32, Promote);
640 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
641 setOperationAction(ISD::AND, MVT::v1i64, Legal);
643 setOperationAction(ISD::OR, MVT::v8i8, Promote);
644 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
645 setOperationAction(ISD::OR, MVT::v4i16, Promote);
646 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
647 setOperationAction(ISD::OR, MVT::v2i32, Promote);
648 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
649 setOperationAction(ISD::OR, MVT::v1i64, Legal);
651 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
652 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
653 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
654 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
655 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
656 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
657 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
659 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
660 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
661 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
662 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
663 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
664 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
665 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
667 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
668 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
669 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
670 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
672 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
673 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
674 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
675 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
677 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
678 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
679 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
681 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
683 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
684 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
685 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
686 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
687 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
688 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
689 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
691 if (!X86ScalarSSEf64 && Subtarget->is64Bit()) {
692 setOperationAction(ISD::BIT_CONVERT, MVT::v8i8, Custom);
693 setOperationAction(ISD::BIT_CONVERT, MVT::v4i16, Custom);
694 setOperationAction(ISD::BIT_CONVERT, MVT::v2i32, Custom);
695 setOperationAction(ISD::BIT_CONVERT, MVT::v1i64, Custom);
699 if (!UseSoftFloat && Subtarget->hasSSE1()) {
700 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
702 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
703 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
704 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
705 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
706 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
707 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
708 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
709 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
710 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
711 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
712 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
713 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
716 if (!UseSoftFloat && Subtarget->hasSSE2()) {
717 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
719 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
720 // registers cannot be used even for integer operations.
721 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
722 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
723 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
724 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
726 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
727 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
728 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
729 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
730 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
731 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
732 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
733 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
734 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
735 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
736 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
737 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
738 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
739 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
740 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
741 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
743 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
744 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
745 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
746 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
748 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
749 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
750 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
751 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
752 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
754 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
755 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
756 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
757 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
758 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
760 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
761 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
762 EVT VT = (MVT::SimpleValueType)i;
763 // Do not attempt to custom lower non-power-of-2 vectors
764 if (!isPowerOf2_32(VT.getVectorNumElements()))
766 // Do not attempt to custom lower non-128-bit vectors
767 if (!VT.is128BitVector())
769 setOperationAction(ISD::BUILD_VECTOR,
770 VT.getSimpleVT().SimpleTy, Custom);
771 setOperationAction(ISD::VECTOR_SHUFFLE,
772 VT.getSimpleVT().SimpleTy, Custom);
773 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
774 VT.getSimpleVT().SimpleTy, Custom);
777 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
778 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
779 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
780 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
781 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
782 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
784 if (Subtarget->is64Bit()) {
785 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
786 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
789 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
790 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
791 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
794 // Do not attempt to promote non-128-bit vectors
795 if (!VT.is128BitVector())
798 setOperationAction(ISD::AND, SVT, Promote);
799 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
800 setOperationAction(ISD::OR, SVT, Promote);
801 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
802 setOperationAction(ISD::XOR, SVT, Promote);
803 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
804 setOperationAction(ISD::LOAD, SVT, Promote);
805 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
806 setOperationAction(ISD::SELECT, SVT, Promote);
807 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
810 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
812 // Custom lower v2i64 and v2f64 selects.
813 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
814 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
815 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
816 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
818 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
819 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
820 if (!DisableMMX && Subtarget->hasMMX()) {
821 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
822 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
826 if (Subtarget->hasSSE41()) {
827 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
828 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
829 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
830 setOperationAction(ISD::FRINT, MVT::f32, Legal);
831 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
832 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
833 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
834 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
835 setOperationAction(ISD::FRINT, MVT::f64, Legal);
836 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
838 // FIXME: Do we need to handle scalar-to-vector here?
839 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
841 // Can turn SHL into an integer multiply.
842 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
843 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
845 // i8 and i16 vectors are custom , because the source register and source
846 // source memory operand types are not the same width. f32 vectors are
847 // custom since the immediate controlling the insert encodes additional
849 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
850 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
851 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
852 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
854 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
855 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
856 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
857 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
859 if (Subtarget->is64Bit()) {
860 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
861 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
865 if (Subtarget->hasSSE42()) {
866 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
869 if (!UseSoftFloat && Subtarget->hasAVX()) {
870 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
871 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
872 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
873 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
874 addRegisterClass(MVT::v32i8, X86::VR256RegisterClass);
876 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
877 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
878 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
879 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
880 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
881 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
882 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
883 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
884 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
885 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
886 setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
887 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
888 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
889 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
890 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
892 // Operations to consider commented out -v16i16 v32i8
893 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
894 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
895 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
896 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
897 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
898 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
899 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
900 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
901 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
902 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
903 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
904 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
905 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
906 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
908 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
909 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
910 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
911 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
913 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
914 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
915 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
916 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
917 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
919 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
920 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
921 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
922 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
923 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
924 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
927 // Not sure we want to do this since there are no 256-bit integer
930 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
931 // This includes 256-bit vectors
932 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
933 EVT VT = (MVT::SimpleValueType)i;
935 // Do not attempt to custom lower non-power-of-2 vectors
936 if (!isPowerOf2_32(VT.getVectorNumElements()))
939 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
940 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
941 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
944 if (Subtarget->is64Bit()) {
945 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
946 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
951 // Not sure we want to do this since there are no 256-bit integer
954 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
955 // Including 256-bit vectors
956 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
957 EVT VT = (MVT::SimpleValueType)i;
959 if (!VT.is256BitVector()) {
962 setOperationAction(ISD::AND, VT, Promote);
963 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
964 setOperationAction(ISD::OR, VT, Promote);
965 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
966 setOperationAction(ISD::XOR, VT, Promote);
967 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
968 setOperationAction(ISD::LOAD, VT, Promote);
969 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
970 setOperationAction(ISD::SELECT, VT, Promote);
971 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
974 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
978 // We want to custom lower some of our intrinsics.
979 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
981 // Add/Sub/Mul with overflow operations are custom lowered.
982 setOperationAction(ISD::SADDO, MVT::i32, Custom);
983 setOperationAction(ISD::UADDO, MVT::i32, Custom);
984 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
985 setOperationAction(ISD::USUBO, MVT::i32, Custom);
986 setOperationAction(ISD::SMULO, MVT::i32, Custom);
988 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
989 // handle type legalization for these operations here.
991 // FIXME: We really should do custom legalization for addition and
992 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
993 // than generic legalization for 64-bit multiplication-with-overflow, though.
994 if (Subtarget->is64Bit()) {
995 setOperationAction(ISD::SADDO, MVT::i64, Custom);
996 setOperationAction(ISD::UADDO, MVT::i64, Custom);
997 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
998 setOperationAction(ISD::USUBO, MVT::i64, Custom);
999 setOperationAction(ISD::SMULO, MVT::i64, Custom);
1002 if (!Subtarget->is64Bit()) {
1003 // These libcalls are not available in 32-bit.
1004 setLibcallName(RTLIB::SHL_I128, 0);
1005 setLibcallName(RTLIB::SRL_I128, 0);
1006 setLibcallName(RTLIB::SRA_I128, 0);
1009 // We have target-specific dag combine patterns for the following nodes:
1010 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1011 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1012 setTargetDAGCombine(ISD::BUILD_VECTOR);
1013 setTargetDAGCombine(ISD::SELECT);
1014 setTargetDAGCombine(ISD::SHL);
1015 setTargetDAGCombine(ISD::SRA);
1016 setTargetDAGCombine(ISD::SRL);
1017 setTargetDAGCombine(ISD::OR);
1018 setTargetDAGCombine(ISD::STORE);
1019 setTargetDAGCombine(ISD::ZERO_EXTEND);
1020 if (Subtarget->is64Bit())
1021 setTargetDAGCombine(ISD::MUL);
1023 computeRegisterProperties();
1025 // FIXME: These should be based on subtarget info. Plus, the values should
1026 // be smaller when we are in optimizing for size mode.
1027 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1028 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1029 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1030 setPrefLoopAlignment(16);
1031 benefitFromCodePlacementOpt = true;
1035 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1040 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1041 /// the desired ByVal argument alignment.
1042 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1045 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1046 if (VTy->getBitWidth() == 128)
1048 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1049 unsigned EltAlign = 0;
1050 getMaxByValAlign(ATy->getElementType(), EltAlign);
1051 if (EltAlign > MaxAlign)
1052 MaxAlign = EltAlign;
1053 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1054 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1055 unsigned EltAlign = 0;
1056 getMaxByValAlign(STy->getElementType(i), EltAlign);
1057 if (EltAlign > MaxAlign)
1058 MaxAlign = EltAlign;
1066 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1067 /// function arguments in the caller parameter area. For X86, aggregates
1068 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1069 /// are at 4-byte boundaries.
1070 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1071 if (Subtarget->is64Bit()) {
1072 // Max of 8 and alignment of type.
1073 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1080 if (Subtarget->hasSSE1())
1081 getMaxByValAlign(Ty, Align);
1085 /// getOptimalMemOpType - Returns the target specific optimal type for load
1086 /// and store operations as a result of memset, memcpy, and memmove
1087 /// lowering. If DstAlign is zero that means it's safe to destination
1088 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1089 /// means there isn't a need to check it against alignment requirement,
1090 /// probably because the source does not need to be loaded. If
1091 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1092 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1093 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1094 /// constant so it does not need to be loaded.
1095 /// It returns EVT::Other if the type should be determined using generic
1096 /// target-independent logic.
1098 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1099 unsigned DstAlign, unsigned SrcAlign,
1100 bool NonScalarIntSafe,
1102 MachineFunction &MF) const {
1103 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1104 // linux. This is because the stack realignment code can't handle certain
1105 // cases like PR2962. This should be removed when PR2962 is fixed.
1106 const Function *F = MF.getFunction();
1107 if (NonScalarIntSafe &&
1108 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1110 (Subtarget->isUnalignedMemAccessFast() ||
1111 ((DstAlign == 0 || DstAlign >= 16) &&
1112 (SrcAlign == 0 || SrcAlign >= 16))) &&
1113 Subtarget->getStackAlignment() >= 16) {
1114 if (Subtarget->hasSSE2())
1116 if (Subtarget->hasSSE1())
1118 } else if (!MemcpyStrSrc && Size >= 8 &&
1119 !Subtarget->is64Bit() &&
1120 Subtarget->getStackAlignment() >= 8 &&
1121 Subtarget->hasSSE2()) {
1122 // Do not use f64 to lower memcpy if source is string constant. It's
1123 // better to use i32 to avoid the loads.
1127 if (Subtarget->is64Bit() && Size >= 8)
1132 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1133 /// current function. The returned value is a member of the
1134 /// MachineJumpTableInfo::JTEntryKind enum.
1135 unsigned X86TargetLowering::getJumpTableEncoding() const {
1136 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1138 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1139 Subtarget->isPICStyleGOT())
1140 return MachineJumpTableInfo::EK_Custom32;
1142 // Otherwise, use the normal jump table encoding heuristics.
1143 return TargetLowering::getJumpTableEncoding();
1146 /// getPICBaseSymbol - Return the X86-32 PIC base.
1148 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1149 MCContext &Ctx) const {
1150 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1151 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1152 Twine(MF->getFunctionNumber())+"$pb");
1157 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1158 const MachineBasicBlock *MBB,
1159 unsigned uid,MCContext &Ctx) const{
1160 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1161 Subtarget->isPICStyleGOT());
1162 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1164 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1165 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1168 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1170 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1171 SelectionDAG &DAG) const {
1172 if (!Subtarget->is64Bit())
1173 // This doesn't have DebugLoc associated with it, but is not really the
1174 // same as a Register.
1175 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1179 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1180 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1182 const MCExpr *X86TargetLowering::
1183 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1184 MCContext &Ctx) const {
1185 // X86-64 uses RIP relative addressing based on the jump table label.
1186 if (Subtarget->isPICStyleRIPRel())
1187 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1189 // Otherwise, the reference is relative to the PIC base.
1190 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1193 /// getFunctionAlignment - Return the Log2 alignment of this function.
1194 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1195 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1198 std::pair<const TargetRegisterClass*, uint8_t>
1199 X86TargetLowering::findRepresentativeClass(EVT VT) const{
1200 const TargetRegisterClass *RRC = 0;
1202 switch (VT.getSimpleVT().SimpleTy) {
1204 return TargetLowering::findRepresentativeClass(VT);
1205 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1206 RRC = (Subtarget->is64Bit()
1207 ? X86::GR64RegisterClass : X86::GR32RegisterClass);
1209 case MVT::v8i8: case MVT::v4i16:
1210 case MVT::v2i32: case MVT::v1i64:
1211 RRC = X86::VR64RegisterClass;
1213 case MVT::f32: case MVT::f64:
1214 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1215 case MVT::v4f32: case MVT::v2f64:
1216 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1218 RRC = X86::VR128RegisterClass;
1221 return std::make_pair(RRC, Cost);
1225 X86TargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
1226 MachineFunction &MF) const {
1227 unsigned FPDiff = RegInfo->hasFP(MF) ? 1 : 0;
1228 switch (RC->getID()) {
1231 case X86::GR32RegClassID:
1233 case X86::GR64RegClassID:
1235 case X86::VR128RegClassID:
1236 return Subtarget->is64Bit() ? 10 : 4;
1237 case X86::VR64RegClassID:
1242 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1243 unsigned &Offset) const {
1244 if (!Subtarget->isTargetLinux())
1247 if (Subtarget->is64Bit()) {
1248 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1250 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1263 //===----------------------------------------------------------------------===//
1264 // Return Value Calling Convention Implementation
1265 //===----------------------------------------------------------------------===//
1267 #include "X86GenCallingConv.inc"
1270 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1271 const SmallVectorImpl<ISD::OutputArg> &Outs,
1272 LLVMContext &Context) const {
1273 SmallVector<CCValAssign, 16> RVLocs;
1274 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1276 return CCInfo.CheckReturn(Outs, RetCC_X86);
1280 X86TargetLowering::LowerReturn(SDValue Chain,
1281 CallingConv::ID CallConv, bool isVarArg,
1282 const SmallVectorImpl<ISD::OutputArg> &Outs,
1283 const SmallVectorImpl<SDValue> &OutVals,
1284 DebugLoc dl, SelectionDAG &DAG) const {
1285 MachineFunction &MF = DAG.getMachineFunction();
1286 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1288 SmallVector<CCValAssign, 16> RVLocs;
1289 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1290 RVLocs, *DAG.getContext());
1291 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1293 // Add the regs to the liveout set for the function.
1294 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1295 for (unsigned i = 0; i != RVLocs.size(); ++i)
1296 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1297 MRI.addLiveOut(RVLocs[i].getLocReg());
1301 SmallVector<SDValue, 6> RetOps;
1302 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1303 // Operand #1 = Bytes To Pop
1304 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1307 // Copy the result values into the output registers.
1308 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1309 CCValAssign &VA = RVLocs[i];
1310 assert(VA.isRegLoc() && "Can only return in registers!");
1311 SDValue ValToCopy = OutVals[i];
1312 EVT ValVT = ValToCopy.getValueType();
1314 // If this is x86-64, and we disabled SSE, we can't return FP values
1315 if ((ValVT == MVT::f32 || ValVT == MVT::f64) &&
1316 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
1317 report_fatal_error("SSE register return with SSE disabled");
1319 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1320 // llvm-gcc has never done it right and no one has noticed, so this
1321 // should be OK for now.
1322 if (ValVT == MVT::f64 &&
1323 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
1324 report_fatal_error("SSE2 register return with SSE2 disabled");
1326 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1327 // the RET instruction and handled by the FP Stackifier.
1328 if (VA.getLocReg() == X86::ST0 ||
1329 VA.getLocReg() == X86::ST1) {
1330 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1331 // change the value to the FP stack register class.
1332 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1333 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1334 RetOps.push_back(ValToCopy);
1335 // Don't emit a copytoreg.
1339 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1340 // which is returned in RAX / RDX.
1341 if (Subtarget->is64Bit()) {
1342 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1343 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1344 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1345 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1348 // If we don't have SSE2 available, convert to v4f32 so the generated
1349 // register is legal.
1350 if (!Subtarget->hasSSE2())
1351 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32,ValToCopy);
1356 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1357 Flag = Chain.getValue(1);
1360 // The x86-64 ABI for returning structs by value requires that we copy
1361 // the sret argument into %rax for the return. We saved the argument into
1362 // a virtual register in the entry block, so now we copy the value out
1364 if (Subtarget->is64Bit() &&
1365 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1366 MachineFunction &MF = DAG.getMachineFunction();
1367 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1368 unsigned Reg = FuncInfo->getSRetReturnReg();
1370 "SRetReturnReg should have been set in LowerFormalArguments().");
1371 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1373 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1374 Flag = Chain.getValue(1);
1376 // RAX now acts like a return value.
1377 MRI.addLiveOut(X86::RAX);
1380 RetOps[0] = Chain; // Update chain.
1382 // Add the flag if we have it.
1384 RetOps.push_back(Flag);
1386 return DAG.getNode(X86ISD::RET_FLAG, dl,
1387 MVT::Other, &RetOps[0], RetOps.size());
1390 /// LowerCallResult - Lower the result values of a call into the
1391 /// appropriate copies out of appropriate physical registers.
1394 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1395 CallingConv::ID CallConv, bool isVarArg,
1396 const SmallVectorImpl<ISD::InputArg> &Ins,
1397 DebugLoc dl, SelectionDAG &DAG,
1398 SmallVectorImpl<SDValue> &InVals) const {
1400 // Assign locations to each value returned by this call.
1401 SmallVector<CCValAssign, 16> RVLocs;
1402 bool Is64Bit = Subtarget->is64Bit();
1403 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1404 RVLocs, *DAG.getContext());
1405 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1407 // Copy all of the result registers out of their specified physreg.
1408 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1409 CCValAssign &VA = RVLocs[i];
1410 EVT CopyVT = VA.getValVT();
1412 // If this is x86-64, and we disabled SSE, we can't return FP values
1413 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1414 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1415 report_fatal_error("SSE register return with SSE disabled");
1420 // If this is a call to a function that returns an fp value on the floating
1421 // point stack, we must guarantee the the value is popped from the stack, so
1422 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1423 // if the return value is not used. We use the FpGET_ST0 instructions
1425 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1426 // If we prefer to use the value in xmm registers, copy it out as f80 and
1427 // use a truncate to move it from fp stack reg to xmm reg.
1428 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1429 bool isST0 = VA.getLocReg() == X86::ST0;
1431 if (CopyVT == MVT::f32) Opc = isST0 ? X86::FpGET_ST0_32:X86::FpGET_ST1_32;
1432 if (CopyVT == MVT::f64) Opc = isST0 ? X86::FpGET_ST0_64:X86::FpGET_ST1_64;
1433 if (CopyVT == MVT::f80) Opc = isST0 ? X86::FpGET_ST0_80:X86::FpGET_ST1_80;
1434 SDValue Ops[] = { Chain, InFlag };
1435 Chain = SDValue(DAG.getMachineNode(Opc, dl, CopyVT, MVT::Other, MVT::Flag,
1437 Val = Chain.getValue(0);
1439 // Round the f80 to the right size, which also moves it to the appropriate
1441 if (CopyVT != VA.getValVT())
1442 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1443 // This truncation won't change the value.
1444 DAG.getIntPtrConstant(1));
1445 } else if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1446 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1447 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1448 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1449 MVT::v2i64, InFlag).getValue(1);
1450 Val = Chain.getValue(0);
1451 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1452 Val, DAG.getConstant(0, MVT::i64));
1454 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1455 MVT::i64, InFlag).getValue(1);
1456 Val = Chain.getValue(0);
1458 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1460 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1461 CopyVT, InFlag).getValue(1);
1462 Val = Chain.getValue(0);
1464 InFlag = Chain.getValue(2);
1465 InVals.push_back(Val);
1472 //===----------------------------------------------------------------------===//
1473 // C & StdCall & Fast Calling Convention implementation
1474 //===----------------------------------------------------------------------===//
1475 // StdCall calling convention seems to be standard for many Windows' API
1476 // routines and around. It differs from C calling convention just a little:
1477 // callee should clean up the stack, not caller. Symbols should be also
1478 // decorated in some fancy way :) It doesn't support any vector arguments.
1479 // For info on fast calling convention see Fast Calling Convention (tail call)
1480 // implementation LowerX86_32FastCCCallTo.
1482 /// CallIsStructReturn - Determines whether a call uses struct return
1484 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1488 return Outs[0].Flags.isSRet();
1491 /// ArgsAreStructReturn - Determines whether a function uses struct
1492 /// return semantics.
1494 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1498 return Ins[0].Flags.isSRet();
1501 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1502 /// given CallingConvention value.
1503 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1504 if (Subtarget->is64Bit()) {
1505 if (CC == CallingConv::GHC)
1506 return CC_X86_64_GHC;
1507 else if (Subtarget->isTargetWin64())
1508 return CC_X86_Win64_C;
1513 if (CC == CallingConv::X86_FastCall)
1514 return CC_X86_32_FastCall;
1515 else if (CC == CallingConv::X86_ThisCall)
1516 return CC_X86_32_ThisCall;
1517 else if (CC == CallingConv::Fast)
1518 return CC_X86_32_FastCC;
1519 else if (CC == CallingConv::GHC)
1520 return CC_X86_32_GHC;
1525 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1526 /// by "Src" to address "Dst" with size and alignment information specified by
1527 /// the specific parameter attribute. The copy will be passed as a byval
1528 /// function parameter.
1530 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1531 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1533 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1534 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1535 /*isVolatile*/false, /*AlwaysInline=*/true,
1539 /// IsTailCallConvention - Return true if the calling convention is one that
1540 /// supports tail call optimization.
1541 static bool IsTailCallConvention(CallingConv::ID CC) {
1542 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1545 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1546 /// a tailcall target by changing its ABI.
1547 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1548 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1552 X86TargetLowering::LowerMemArgument(SDValue Chain,
1553 CallingConv::ID CallConv,
1554 const SmallVectorImpl<ISD::InputArg> &Ins,
1555 DebugLoc dl, SelectionDAG &DAG,
1556 const CCValAssign &VA,
1557 MachineFrameInfo *MFI,
1559 // Create the nodes corresponding to a load from this parameter slot.
1560 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1561 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1562 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1565 // If value is passed by pointer we have address passed instead of the value
1567 if (VA.getLocInfo() == CCValAssign::Indirect)
1568 ValVT = VA.getLocVT();
1570 ValVT = VA.getValVT();
1572 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1573 // changed with more analysis.
1574 // In case of tail call optimization mark all arguments mutable. Since they
1575 // could be overwritten by lowering of arguments in case of a tail call.
1576 if (Flags.isByVal()) {
1577 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1578 VA.getLocMemOffset(), isImmutable);
1579 return DAG.getFrameIndex(FI, getPointerTy());
1581 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1582 VA.getLocMemOffset(), isImmutable);
1583 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1584 return DAG.getLoad(ValVT, dl, Chain, FIN,
1585 PseudoSourceValue::getFixedStack(FI), 0,
1591 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1592 CallingConv::ID CallConv,
1594 const SmallVectorImpl<ISD::InputArg> &Ins,
1597 SmallVectorImpl<SDValue> &InVals)
1599 MachineFunction &MF = DAG.getMachineFunction();
1600 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1602 const Function* Fn = MF.getFunction();
1603 if (Fn->hasExternalLinkage() &&
1604 Subtarget->isTargetCygMing() &&
1605 Fn->getName() == "main")
1606 FuncInfo->setForceFramePointer(true);
1608 MachineFrameInfo *MFI = MF.getFrameInfo();
1609 bool Is64Bit = Subtarget->is64Bit();
1610 bool IsWin64 = Subtarget->isTargetWin64();
1612 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1613 "Var args not supported with calling convention fastcc or ghc");
1615 // Assign locations to all of the incoming arguments.
1616 SmallVector<CCValAssign, 16> ArgLocs;
1617 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1618 ArgLocs, *DAG.getContext());
1619 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1621 unsigned LastVal = ~0U;
1623 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1624 CCValAssign &VA = ArgLocs[i];
1625 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1627 assert(VA.getValNo() != LastVal &&
1628 "Don't support value assigned to multiple locs yet");
1629 LastVal = VA.getValNo();
1631 if (VA.isRegLoc()) {
1632 EVT RegVT = VA.getLocVT();
1633 TargetRegisterClass *RC = NULL;
1634 if (RegVT == MVT::i32)
1635 RC = X86::GR32RegisterClass;
1636 else if (Is64Bit && RegVT == MVT::i64)
1637 RC = X86::GR64RegisterClass;
1638 else if (RegVT == MVT::f32)
1639 RC = X86::FR32RegisterClass;
1640 else if (RegVT == MVT::f64)
1641 RC = X86::FR64RegisterClass;
1642 else if (RegVT.isVector() && RegVT.getSizeInBits() == 256)
1643 RC = X86::VR256RegisterClass;
1644 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1645 RC = X86::VR128RegisterClass;
1646 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1647 RC = X86::VR64RegisterClass;
1649 llvm_unreachable("Unknown argument type!");
1651 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1652 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1654 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1655 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1657 if (VA.getLocInfo() == CCValAssign::SExt)
1658 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1659 DAG.getValueType(VA.getValVT()));
1660 else if (VA.getLocInfo() == CCValAssign::ZExt)
1661 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1662 DAG.getValueType(VA.getValVT()));
1663 else if (VA.getLocInfo() == CCValAssign::BCvt)
1664 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1666 if (VA.isExtInLoc()) {
1667 // Handle MMX values passed in XMM regs.
1668 if (RegVT.isVector()) {
1669 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1670 ArgValue, DAG.getConstant(0, MVT::i64));
1671 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1673 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1676 assert(VA.isMemLoc());
1677 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1680 // If value is passed via pointer - do a load.
1681 if (VA.getLocInfo() == CCValAssign::Indirect)
1682 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1685 InVals.push_back(ArgValue);
1688 // The x86-64 ABI for returning structs by value requires that we copy
1689 // the sret argument into %rax for the return. Save the argument into
1690 // a virtual register so that we can access it from the return points.
1691 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1692 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1693 unsigned Reg = FuncInfo->getSRetReturnReg();
1695 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1696 FuncInfo->setSRetReturnReg(Reg);
1698 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1699 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1702 unsigned StackSize = CCInfo.getNextStackOffset();
1703 // Align stack specially for tail calls.
1704 if (FuncIsMadeTailCallSafe(CallConv))
1705 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1707 // If the function takes variable number of arguments, make a frame index for
1708 // the start of the first vararg value... for expansion of llvm.va_start.
1710 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1711 CallConv != CallingConv::X86_ThisCall)) {
1712 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1715 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1717 // FIXME: We should really autogenerate these arrays
1718 static const unsigned GPR64ArgRegsWin64[] = {
1719 X86::RCX, X86::RDX, X86::R8, X86::R9
1721 static const unsigned XMMArgRegsWin64[] = {
1722 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1724 static const unsigned GPR64ArgRegs64Bit[] = {
1725 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1727 static const unsigned XMMArgRegs64Bit[] = {
1728 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1729 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1731 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1734 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1735 GPR64ArgRegs = GPR64ArgRegsWin64;
1736 XMMArgRegs = XMMArgRegsWin64;
1738 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1739 GPR64ArgRegs = GPR64ArgRegs64Bit;
1740 XMMArgRegs = XMMArgRegs64Bit;
1742 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1744 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1747 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1748 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1749 "SSE register cannot be used when SSE is disabled!");
1750 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1751 "SSE register cannot be used when SSE is disabled!");
1752 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1753 // Kernel mode asks for SSE to be disabled, so don't push them
1755 TotalNumXMMRegs = 0;
1757 // For X86-64, if there are vararg parameters that are passed via
1758 // registers, then we must store them to their spots on the stack so they
1759 // may be loaded by deferencing the result of va_next.
1760 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1761 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1762 FuncInfo->setRegSaveFrameIndex(
1763 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1766 // Store the integer parameter registers.
1767 SmallVector<SDValue, 8> MemOps;
1768 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1770 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1771 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1772 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1773 DAG.getIntPtrConstant(Offset));
1774 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1775 X86::GR64RegisterClass);
1776 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1778 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1779 PseudoSourceValue::getFixedStack(
1780 FuncInfo->getRegSaveFrameIndex()),
1781 Offset, false, false, 0);
1782 MemOps.push_back(Store);
1786 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1787 // Now store the XMM (fp + vector) parameter registers.
1788 SmallVector<SDValue, 11> SaveXMMOps;
1789 SaveXMMOps.push_back(Chain);
1791 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1792 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1793 SaveXMMOps.push_back(ALVal);
1795 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1796 FuncInfo->getRegSaveFrameIndex()));
1797 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1798 FuncInfo->getVarArgsFPOffset()));
1800 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1801 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1802 X86::VR128RegisterClass);
1803 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1804 SaveXMMOps.push_back(Val);
1806 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1808 &SaveXMMOps[0], SaveXMMOps.size()));
1811 if (!MemOps.empty())
1812 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1813 &MemOps[0], MemOps.size());
1817 // Some CCs need callee pop.
1818 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1819 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1821 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1822 // If this is an sret function, the return should pop the hidden pointer.
1823 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1824 FuncInfo->setBytesToPopOnReturn(4);
1828 // RegSaveFrameIndex is X86-64 only.
1829 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1830 if (CallConv == CallingConv::X86_FastCall ||
1831 CallConv == CallingConv::X86_ThisCall)
1832 // fastcc functions can't have varargs.
1833 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1840 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1841 SDValue StackPtr, SDValue Arg,
1842 DebugLoc dl, SelectionDAG &DAG,
1843 const CCValAssign &VA,
1844 ISD::ArgFlagsTy Flags) const {
1845 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1846 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1847 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1848 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1849 if (Flags.isByVal()) {
1850 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1852 return DAG.getStore(Chain, dl, Arg, PtrOff,
1853 PseudoSourceValue::getStack(), LocMemOffset,
1857 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1858 /// optimization is performed and it is required.
1860 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1861 SDValue &OutRetAddr, SDValue Chain,
1862 bool IsTailCall, bool Is64Bit,
1863 int FPDiff, DebugLoc dl) const {
1864 // Adjust the Return address stack slot.
1865 EVT VT = getPointerTy();
1866 OutRetAddr = getReturnAddressFrameIndex(DAG);
1868 // Load the "old" Return address.
1869 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1870 return SDValue(OutRetAddr.getNode(), 1);
1873 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1874 /// optimization is performed and it is required (FPDiff!=0).
1876 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1877 SDValue Chain, SDValue RetAddrFrIdx,
1878 bool Is64Bit, int FPDiff, DebugLoc dl) {
1879 // Store the return address to the appropriate stack slot.
1880 if (!FPDiff) return Chain;
1881 // Calculate the new stack slot for the return address.
1882 int SlotSize = Is64Bit ? 8 : 4;
1883 int NewReturnAddrFI =
1884 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1885 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1886 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1887 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1888 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1894 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1895 CallingConv::ID CallConv, bool isVarArg,
1897 const SmallVectorImpl<ISD::OutputArg> &Outs,
1898 const SmallVectorImpl<SDValue> &OutVals,
1899 const SmallVectorImpl<ISD::InputArg> &Ins,
1900 DebugLoc dl, SelectionDAG &DAG,
1901 SmallVectorImpl<SDValue> &InVals) const {
1902 MachineFunction &MF = DAG.getMachineFunction();
1903 bool Is64Bit = Subtarget->is64Bit();
1904 bool IsStructRet = CallIsStructReturn(Outs);
1905 bool IsSibcall = false;
1908 // Check if it's really possible to do a tail call.
1909 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1910 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1911 Outs, OutVals, Ins, DAG);
1913 // Sibcalls are automatically detected tailcalls which do not require
1915 if (!GuaranteedTailCallOpt && isTailCall)
1922 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1923 "Var args not supported with calling convention fastcc or ghc");
1925 // Analyze operands of the call, assigning locations to each operand.
1926 SmallVector<CCValAssign, 16> ArgLocs;
1927 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1928 ArgLocs, *DAG.getContext());
1929 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1931 // Get a count of how many bytes are to be pushed on the stack.
1932 unsigned NumBytes = CCInfo.getNextStackOffset();
1934 // This is a sibcall. The memory operands are available in caller's
1935 // own caller's stack.
1937 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1938 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1941 if (isTailCall && !IsSibcall) {
1942 // Lower arguments at fp - stackoffset + fpdiff.
1943 unsigned NumBytesCallerPushed =
1944 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1945 FPDiff = NumBytesCallerPushed - NumBytes;
1947 // Set the delta of movement of the returnaddr stackslot.
1948 // But only set if delta is greater than previous delta.
1949 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1950 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1954 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1956 SDValue RetAddrFrIdx;
1957 // Load return adress for tail calls.
1958 if (isTailCall && FPDiff)
1959 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1960 Is64Bit, FPDiff, dl);
1962 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1963 SmallVector<SDValue, 8> MemOpChains;
1966 // Walk the register/memloc assignments, inserting copies/loads. In the case
1967 // of tail call optimization arguments are handle later.
1968 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1969 CCValAssign &VA = ArgLocs[i];
1970 EVT RegVT = VA.getLocVT();
1971 SDValue Arg = OutVals[i];
1972 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1973 bool isByVal = Flags.isByVal();
1975 // Promote the value if needed.
1976 switch (VA.getLocInfo()) {
1977 default: llvm_unreachable("Unknown loc info!");
1978 case CCValAssign::Full: break;
1979 case CCValAssign::SExt:
1980 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1982 case CCValAssign::ZExt:
1983 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1985 case CCValAssign::AExt:
1986 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1987 // Special case: passing MMX values in XMM registers.
1988 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1989 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1990 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1992 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1994 case CCValAssign::BCvt:
1995 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1997 case CCValAssign::Indirect: {
1998 // Store the argument.
1999 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2000 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2001 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2002 PseudoSourceValue::getFixedStack(FI), 0,
2009 if (VA.isRegLoc()) {
2010 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2011 if (isVarArg && Subtarget->isTargetWin64()) {
2012 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2013 // shadow reg if callee is a varargs function.
2014 unsigned ShadowReg = 0;
2015 switch (VA.getLocReg()) {
2016 case X86::XMM0: ShadowReg = X86::RCX; break;
2017 case X86::XMM1: ShadowReg = X86::RDX; break;
2018 case X86::XMM2: ShadowReg = X86::R8; break;
2019 case X86::XMM3: ShadowReg = X86::R9; break;
2022 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2024 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2025 assert(VA.isMemLoc());
2026 if (StackPtr.getNode() == 0)
2027 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
2028 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2029 dl, DAG, VA, Flags));
2033 if (!MemOpChains.empty())
2034 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2035 &MemOpChains[0], MemOpChains.size());
2037 // Build a sequence of copy-to-reg nodes chained together with token chain
2038 // and flag operands which copy the outgoing args into registers.
2040 // Tail call byval lowering might overwrite argument registers so in case of
2041 // tail call optimization the copies to registers are lowered later.
2043 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2044 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2045 RegsToPass[i].second, InFlag);
2046 InFlag = Chain.getValue(1);
2049 if (Subtarget->isPICStyleGOT()) {
2050 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2053 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
2054 DAG.getNode(X86ISD::GlobalBaseReg,
2055 DebugLoc(), getPointerTy()),
2057 InFlag = Chain.getValue(1);
2059 // If we are tail calling and generating PIC/GOT style code load the
2060 // address of the callee into ECX. The value in ecx is used as target of
2061 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2062 // for tail calls on PIC/GOT architectures. Normally we would just put the
2063 // address of GOT into ebx and then call target@PLT. But for tail calls
2064 // ebx would be restored (since ebx is callee saved) before jumping to the
2067 // Note: The actual moving to ECX is done further down.
2068 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2069 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2070 !G->getGlobal()->hasProtectedVisibility())
2071 Callee = LowerGlobalAddress(Callee, DAG);
2072 else if (isa<ExternalSymbolSDNode>(Callee))
2073 Callee = LowerExternalSymbol(Callee, DAG);
2077 if (Is64Bit && isVarArg && !Subtarget->isTargetWin64()) {
2078 // From AMD64 ABI document:
2079 // For calls that may call functions that use varargs or stdargs
2080 // (prototype-less calls or calls to functions containing ellipsis (...) in
2081 // the declaration) %al is used as hidden argument to specify the number
2082 // of SSE registers used. The contents of %al do not need to match exactly
2083 // the number of registers, but must be an ubound on the number of SSE
2084 // registers used and is in the range 0 - 8 inclusive.
2086 // Count the number of XMM registers allocated.
2087 static const unsigned XMMArgRegs[] = {
2088 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2089 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2091 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2092 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2093 && "SSE registers cannot be used when SSE is disabled");
2095 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2096 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2097 InFlag = Chain.getValue(1);
2101 // For tail calls lower the arguments to the 'real' stack slot.
2103 // Force all the incoming stack arguments to be loaded from the stack
2104 // before any new outgoing arguments are stored to the stack, because the
2105 // outgoing stack slots may alias the incoming argument stack slots, and
2106 // the alias isn't otherwise explicit. This is slightly more conservative
2107 // than necessary, because it means that each store effectively depends
2108 // on every argument instead of just those arguments it would clobber.
2109 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2111 SmallVector<SDValue, 8> MemOpChains2;
2114 // Do not flag preceeding copytoreg stuff together with the following stuff.
2116 if (GuaranteedTailCallOpt) {
2117 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2118 CCValAssign &VA = ArgLocs[i];
2121 assert(VA.isMemLoc());
2122 SDValue Arg = OutVals[i];
2123 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2124 // Create frame index.
2125 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2126 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2127 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2128 FIN = DAG.getFrameIndex(FI, getPointerTy());
2130 if (Flags.isByVal()) {
2131 // Copy relative to framepointer.
2132 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2133 if (StackPtr.getNode() == 0)
2134 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2136 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2138 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2142 // Store relative to framepointer.
2143 MemOpChains2.push_back(
2144 DAG.getStore(ArgChain, dl, Arg, FIN,
2145 PseudoSourceValue::getFixedStack(FI), 0,
2151 if (!MemOpChains2.empty())
2152 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2153 &MemOpChains2[0], MemOpChains2.size());
2155 // Copy arguments to their registers.
2156 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2157 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2158 RegsToPass[i].second, InFlag);
2159 InFlag = Chain.getValue(1);
2163 // Store the return address to the appropriate stack slot.
2164 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2168 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2169 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2170 // In the 64-bit large code model, we have to make all calls
2171 // through a register, since the call instruction's 32-bit
2172 // pc-relative offset may not be large enough to hold the whole
2174 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2175 // If the callee is a GlobalAddress node (quite common, every direct call
2176 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2179 // We should use extra load for direct calls to dllimported functions in
2181 const GlobalValue *GV = G->getGlobal();
2182 if (!GV->hasDLLImportLinkage()) {
2183 unsigned char OpFlags = 0;
2185 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2186 // external symbols most go through the PLT in PIC mode. If the symbol
2187 // has hidden or protected visibility, or if it is static or local, then
2188 // we don't need to use the PLT - we can directly call it.
2189 if (Subtarget->isTargetELF() &&
2190 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2191 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2192 OpFlags = X86II::MO_PLT;
2193 } else if (Subtarget->isPICStyleStubAny() &&
2194 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2195 Subtarget->getDarwinVers() < 9) {
2196 // PC-relative references to external symbols should go through $stub,
2197 // unless we're building with the leopard linker or later, which
2198 // automatically synthesizes these stubs.
2199 OpFlags = X86II::MO_DARWIN_STUB;
2202 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2203 G->getOffset(), OpFlags);
2205 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2206 unsigned char OpFlags = 0;
2208 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2209 // symbols should go through the PLT.
2210 if (Subtarget->isTargetELF() &&
2211 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2212 OpFlags = X86II::MO_PLT;
2213 } else if (Subtarget->isPICStyleStubAny() &&
2214 Subtarget->getDarwinVers() < 9) {
2215 // PC-relative references to external symbols should go through $stub,
2216 // unless we're building with the leopard linker or later, which
2217 // automatically synthesizes these stubs.
2218 OpFlags = X86II::MO_DARWIN_STUB;
2221 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2225 // Returns a chain & a flag for retval copy to use.
2226 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2227 SmallVector<SDValue, 8> Ops;
2229 if (!IsSibcall && isTailCall) {
2230 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2231 DAG.getIntPtrConstant(0, true), InFlag);
2232 InFlag = Chain.getValue(1);
2235 Ops.push_back(Chain);
2236 Ops.push_back(Callee);
2239 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2241 // Add argument registers to the end of the list so that they are known live
2243 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2244 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2245 RegsToPass[i].second.getValueType()));
2247 // Add an implicit use GOT pointer in EBX.
2248 if (!isTailCall && Subtarget->isPICStyleGOT())
2249 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2251 // Add an implicit use of AL for non-Windows x86 64-bit vararg functions.
2252 if (Is64Bit && isVarArg && !Subtarget->isTargetWin64())
2253 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2255 if (InFlag.getNode())
2256 Ops.push_back(InFlag);
2260 //// If this is the first return lowered for this function, add the regs
2261 //// to the liveout set for the function.
2262 // This isn't right, although it's probably harmless on x86; liveouts
2263 // should be computed from returns not tail calls. Consider a void
2264 // function making a tail call to a function returning int.
2265 return DAG.getNode(X86ISD::TC_RETURN, dl,
2266 NodeTys, &Ops[0], Ops.size());
2269 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2270 InFlag = Chain.getValue(1);
2272 // Create the CALLSEQ_END node.
2273 unsigned NumBytesForCalleeToPush;
2274 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2275 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2276 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2277 // If this is a call to a struct-return function, the callee
2278 // pops the hidden struct pointer, so we have to push it back.
2279 // This is common for Darwin/X86, Linux & Mingw32 targets.
2280 NumBytesForCalleeToPush = 4;
2282 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2284 // Returns a flag for retval copy to use.
2286 Chain = DAG.getCALLSEQ_END(Chain,
2287 DAG.getIntPtrConstant(NumBytes, true),
2288 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2291 InFlag = Chain.getValue(1);
2294 // Handle result values, copying them out of physregs into vregs that we
2296 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2297 Ins, dl, DAG, InVals);
2301 //===----------------------------------------------------------------------===//
2302 // Fast Calling Convention (tail call) implementation
2303 //===----------------------------------------------------------------------===//
2305 // Like std call, callee cleans arguments, convention except that ECX is
2306 // reserved for storing the tail called function address. Only 2 registers are
2307 // free for argument passing (inreg). Tail call optimization is performed
2309 // * tailcallopt is enabled
2310 // * caller/callee are fastcc
2311 // On X86_64 architecture with GOT-style position independent code only local
2312 // (within module) calls are supported at the moment.
2313 // To keep the stack aligned according to platform abi the function
2314 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2315 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2316 // If a tail called function callee has more arguments than the caller the
2317 // caller needs to make sure that there is room to move the RETADDR to. This is
2318 // achieved by reserving an area the size of the argument delta right after the
2319 // original REtADDR, but before the saved framepointer or the spilled registers
2320 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2332 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2333 /// for a 16 byte align requirement.
2335 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2336 SelectionDAG& DAG) const {
2337 MachineFunction &MF = DAG.getMachineFunction();
2338 const TargetMachine &TM = MF.getTarget();
2339 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2340 unsigned StackAlignment = TFI.getStackAlignment();
2341 uint64_t AlignMask = StackAlignment - 1;
2342 int64_t Offset = StackSize;
2343 uint64_t SlotSize = TD->getPointerSize();
2344 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2345 // Number smaller than 12 so just add the difference.
2346 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2348 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2349 Offset = ((~AlignMask) & Offset) + StackAlignment +
2350 (StackAlignment-SlotSize);
2355 /// MatchingStackOffset - Return true if the given stack call argument is
2356 /// already available in the same position (relatively) of the caller's
2357 /// incoming argument stack.
2359 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2360 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2361 const X86InstrInfo *TII) {
2362 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2364 if (Arg.getOpcode() == ISD::CopyFromReg) {
2365 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2366 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2368 MachineInstr *Def = MRI->getVRegDef(VR);
2371 if (!Flags.isByVal()) {
2372 if (!TII->isLoadFromStackSlot(Def, FI))
2375 unsigned Opcode = Def->getOpcode();
2376 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2377 Def->getOperand(1).isFI()) {
2378 FI = Def->getOperand(1).getIndex();
2379 Bytes = Flags.getByValSize();
2383 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2384 if (Flags.isByVal())
2385 // ByVal argument is passed in as a pointer but it's now being
2386 // dereferenced. e.g.
2387 // define @foo(%struct.X* %A) {
2388 // tail call @bar(%struct.X* byval %A)
2391 SDValue Ptr = Ld->getBasePtr();
2392 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2395 FI = FINode->getIndex();
2399 assert(FI != INT_MAX);
2400 if (!MFI->isFixedObjectIndex(FI))
2402 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2405 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2406 /// for tail call optimization. Targets which want to do tail call
2407 /// optimization should implement this function.
2409 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2410 CallingConv::ID CalleeCC,
2412 bool isCalleeStructRet,
2413 bool isCallerStructRet,
2414 const SmallVectorImpl<ISD::OutputArg> &Outs,
2415 const SmallVectorImpl<SDValue> &OutVals,
2416 const SmallVectorImpl<ISD::InputArg> &Ins,
2417 SelectionDAG& DAG) const {
2418 if (!IsTailCallConvention(CalleeCC) &&
2419 CalleeCC != CallingConv::C)
2422 // If -tailcallopt is specified, make fastcc functions tail-callable.
2423 const MachineFunction &MF = DAG.getMachineFunction();
2424 const Function *CallerF = DAG.getMachineFunction().getFunction();
2425 CallingConv::ID CallerCC = CallerF->getCallingConv();
2426 bool CCMatch = CallerCC == CalleeCC;
2428 if (GuaranteedTailCallOpt) {
2429 if (IsTailCallConvention(CalleeCC) && CCMatch)
2434 // Look for obvious safe cases to perform tail call optimization that do not
2435 // require ABI changes. This is what gcc calls sibcall.
2437 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2438 // emit a special epilogue.
2439 if (RegInfo->needsStackRealignment(MF))
2442 // Do not sibcall optimize vararg calls unless the call site is not passing
2444 if (isVarArg && !Outs.empty())
2447 // Also avoid sibcall optimization if either caller or callee uses struct
2448 // return semantics.
2449 if (isCalleeStructRet || isCallerStructRet)
2452 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2453 // Therefore if it's not used by the call it is not safe to optimize this into
2455 bool Unused = false;
2456 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2463 SmallVector<CCValAssign, 16> RVLocs;
2464 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2465 RVLocs, *DAG.getContext());
2466 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2467 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2468 CCValAssign &VA = RVLocs[i];
2469 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2474 // If the calling conventions do not match, then we'd better make sure the
2475 // results are returned in the same way as what the caller expects.
2477 SmallVector<CCValAssign, 16> RVLocs1;
2478 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2479 RVLocs1, *DAG.getContext());
2480 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2482 SmallVector<CCValAssign, 16> RVLocs2;
2483 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2484 RVLocs2, *DAG.getContext());
2485 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2487 if (RVLocs1.size() != RVLocs2.size())
2489 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2490 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2492 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2494 if (RVLocs1[i].isRegLoc()) {
2495 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2498 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2504 // If the callee takes no arguments then go on to check the results of the
2506 if (!Outs.empty()) {
2507 // Check if stack adjustment is needed. For now, do not do this if any
2508 // argument is passed on the stack.
2509 SmallVector<CCValAssign, 16> ArgLocs;
2510 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2511 ArgLocs, *DAG.getContext());
2512 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2513 if (CCInfo.getNextStackOffset()) {
2514 MachineFunction &MF = DAG.getMachineFunction();
2515 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2517 if (Subtarget->isTargetWin64())
2518 // Win64 ABI has additional complications.
2521 // Check if the arguments are already laid out in the right way as
2522 // the caller's fixed stack objects.
2523 MachineFrameInfo *MFI = MF.getFrameInfo();
2524 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2525 const X86InstrInfo *TII =
2526 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2527 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2528 CCValAssign &VA = ArgLocs[i];
2529 SDValue Arg = OutVals[i];
2530 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2531 if (VA.getLocInfo() == CCValAssign::Indirect)
2533 if (!VA.isRegLoc()) {
2534 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2541 // If the tailcall address may be in a register, then make sure it's
2542 // possible to register allocate for it. In 32-bit, the call address can
2543 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2544 // callee-saved registers are restored. These happen to be the same
2545 // registers used to pass 'inreg' arguments so watch out for those.
2546 if (!Subtarget->is64Bit() &&
2547 !isa<GlobalAddressSDNode>(Callee) &&
2548 !isa<ExternalSymbolSDNode>(Callee)) {
2549 unsigned NumInRegs = 0;
2550 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2551 CCValAssign &VA = ArgLocs[i];
2554 unsigned Reg = VA.getLocReg();
2557 case X86::EAX: case X86::EDX: case X86::ECX:
2558 if (++NumInRegs == 3)
2570 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2571 return X86::createFastISel(funcInfo);
2575 //===----------------------------------------------------------------------===//
2576 // Other Lowering Hooks
2577 //===----------------------------------------------------------------------===//
2579 static bool MayFoldLoad(SDValue Op) {
2580 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
2583 static bool MayFoldIntoStore(SDValue Op) {
2584 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
2587 static bool isTargetShuffle(unsigned Opcode) {
2589 default: return false;
2590 case X86ISD::PSHUFD:
2591 case X86ISD::PSHUFHW:
2592 case X86ISD::PSHUFLW:
2593 case X86ISD::SHUFPD:
2594 case X86ISD::SHUFPS:
2595 case X86ISD::MOVLHPS:
2596 case X86ISD::MOVLHPD:
2597 case X86ISD::MOVHLPS:
2598 case X86ISD::MOVLPS:
2599 case X86ISD::MOVLPD:
2600 case X86ISD::MOVSHDUP:
2601 case X86ISD::MOVSLDUP:
2604 case X86ISD::UNPCKLPS:
2605 case X86ISD::UNPCKLPD:
2606 case X86ISD::PUNPCKLWD:
2607 case X86ISD::PUNPCKLBW:
2608 case X86ISD::PUNPCKLDQ:
2609 case X86ISD::PUNPCKLQDQ:
2610 case X86ISD::UNPCKHPS:
2611 case X86ISD::UNPCKHPD:
2612 case X86ISD::PUNPCKHWD:
2613 case X86ISD::PUNPCKHBW:
2614 case X86ISD::PUNPCKHDQ:
2615 case X86ISD::PUNPCKHQDQ:
2621 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2622 SDValue V1, SelectionDAG &DAG) {
2624 default: llvm_unreachable("Unknown x86 shuffle node");
2625 case X86ISD::MOVSHDUP:
2626 case X86ISD::MOVSLDUP:
2627 return DAG.getNode(Opc, dl, VT, V1);
2633 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2634 SDValue V1, unsigned TargetMask, SelectionDAG &DAG) {
2636 default: llvm_unreachable("Unknown x86 shuffle node");
2637 case X86ISD::PSHUFD:
2638 case X86ISD::PSHUFHW:
2639 case X86ISD::PSHUFLW:
2640 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
2646 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2647 SDValue V1, SDValue V2, unsigned TargetMask, SelectionDAG &DAG) {
2649 default: llvm_unreachable("Unknown x86 shuffle node");
2650 case X86ISD::SHUFPD:
2651 case X86ISD::SHUFPS:
2652 return DAG.getNode(Opc, dl, VT, V1, V2,
2653 DAG.getConstant(TargetMask, MVT::i8));
2658 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2659 SDValue V1, SDValue V2, SelectionDAG &DAG) {
2661 default: llvm_unreachable("Unknown x86 shuffle node");
2662 case X86ISD::MOVLHPS:
2663 case X86ISD::MOVLHPD:
2664 case X86ISD::MOVHLPS:
2665 case X86ISD::MOVLPS:
2666 case X86ISD::MOVLPD:
2669 case X86ISD::UNPCKLPS:
2670 case X86ISD::UNPCKLPD:
2671 case X86ISD::PUNPCKLWD:
2672 case X86ISD::PUNPCKLBW:
2673 case X86ISD::PUNPCKLDQ:
2674 case X86ISD::PUNPCKLQDQ:
2675 case X86ISD::UNPCKHPS:
2676 case X86ISD::UNPCKHPD:
2677 case X86ISD::PUNPCKHWD:
2678 case X86ISD::PUNPCKHBW:
2679 case X86ISD::PUNPCKHDQ:
2680 case X86ISD::PUNPCKHQDQ:
2681 return DAG.getNode(Opc, dl, VT, V1, V2);
2686 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2687 MachineFunction &MF = DAG.getMachineFunction();
2688 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2689 int ReturnAddrIndex = FuncInfo->getRAIndex();
2691 if (ReturnAddrIndex == 0) {
2692 // Set up a frame object for the return address.
2693 uint64_t SlotSize = TD->getPointerSize();
2694 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2696 FuncInfo->setRAIndex(ReturnAddrIndex);
2699 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2703 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2704 bool hasSymbolicDisplacement) {
2705 // Offset should fit into 32 bit immediate field.
2706 if (!isInt<32>(Offset))
2709 // If we don't have a symbolic displacement - we don't have any extra
2711 if (!hasSymbolicDisplacement)
2714 // FIXME: Some tweaks might be needed for medium code model.
2715 if (M != CodeModel::Small && M != CodeModel::Kernel)
2718 // For small code model we assume that latest object is 16MB before end of 31
2719 // bits boundary. We may also accept pretty large negative constants knowing
2720 // that all objects are in the positive half of address space.
2721 if (M == CodeModel::Small && Offset < 16*1024*1024)
2724 // For kernel code model we know that all object resist in the negative half
2725 // of 32bits address space. We may not accept negative offsets, since they may
2726 // be just off and we may accept pretty large positive ones.
2727 if (M == CodeModel::Kernel && Offset > 0)
2733 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2734 /// specific condition code, returning the condition code and the LHS/RHS of the
2735 /// comparison to make.
2736 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2737 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2739 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2740 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2741 // X > -1 -> X == 0, jump !sign.
2742 RHS = DAG.getConstant(0, RHS.getValueType());
2743 return X86::COND_NS;
2744 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2745 // X < 0 -> X == 0, jump on sign.
2747 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2749 RHS = DAG.getConstant(0, RHS.getValueType());
2750 return X86::COND_LE;
2754 switch (SetCCOpcode) {
2755 default: llvm_unreachable("Invalid integer condition!");
2756 case ISD::SETEQ: return X86::COND_E;
2757 case ISD::SETGT: return X86::COND_G;
2758 case ISD::SETGE: return X86::COND_GE;
2759 case ISD::SETLT: return X86::COND_L;
2760 case ISD::SETLE: return X86::COND_LE;
2761 case ISD::SETNE: return X86::COND_NE;
2762 case ISD::SETULT: return X86::COND_B;
2763 case ISD::SETUGT: return X86::COND_A;
2764 case ISD::SETULE: return X86::COND_BE;
2765 case ISD::SETUGE: return X86::COND_AE;
2769 // First determine if it is required or is profitable to flip the operands.
2771 // If LHS is a foldable load, but RHS is not, flip the condition.
2772 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2773 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2774 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2775 std::swap(LHS, RHS);
2778 switch (SetCCOpcode) {
2784 std::swap(LHS, RHS);
2788 // On a floating point condition, the flags are set as follows:
2790 // 0 | 0 | 0 | X > Y
2791 // 0 | 0 | 1 | X < Y
2792 // 1 | 0 | 0 | X == Y
2793 // 1 | 1 | 1 | unordered
2794 switch (SetCCOpcode) {
2795 default: llvm_unreachable("Condcode should be pre-legalized away");
2797 case ISD::SETEQ: return X86::COND_E;
2798 case ISD::SETOLT: // flipped
2800 case ISD::SETGT: return X86::COND_A;
2801 case ISD::SETOLE: // flipped
2803 case ISD::SETGE: return X86::COND_AE;
2804 case ISD::SETUGT: // flipped
2806 case ISD::SETLT: return X86::COND_B;
2807 case ISD::SETUGE: // flipped
2809 case ISD::SETLE: return X86::COND_BE;
2811 case ISD::SETNE: return X86::COND_NE;
2812 case ISD::SETUO: return X86::COND_P;
2813 case ISD::SETO: return X86::COND_NP;
2815 case ISD::SETUNE: return X86::COND_INVALID;
2819 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2820 /// code. Current x86 isa includes the following FP cmov instructions:
2821 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2822 static bool hasFPCMov(unsigned X86CC) {
2838 /// isFPImmLegal - Returns true if the target can instruction select the
2839 /// specified FP immediate natively. If false, the legalizer will
2840 /// materialize the FP immediate as a load from a constant pool.
2841 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2842 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2843 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2849 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2850 /// the specified range (L, H].
2851 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2852 return (Val < 0) || (Val >= Low && Val < Hi);
2855 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2856 /// specified value.
2857 static bool isUndefOrEqual(int Val, int CmpVal) {
2858 if (Val < 0 || Val == CmpVal)
2863 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2864 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2865 /// the second operand.
2866 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2867 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2868 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2869 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2870 return (Mask[0] < 2 && Mask[1] < 2);
2874 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2875 SmallVector<int, 8> M;
2877 return ::isPSHUFDMask(M, N->getValueType(0));
2880 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2881 /// is suitable for input to PSHUFHW.
2882 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2883 if (VT != MVT::v8i16)
2886 // Lower quadword copied in order or undef.
2887 for (int i = 0; i != 4; ++i)
2888 if (Mask[i] >= 0 && Mask[i] != i)
2891 // Upper quadword shuffled.
2892 for (int i = 4; i != 8; ++i)
2893 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2899 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2900 SmallVector<int, 8> M;
2902 return ::isPSHUFHWMask(M, N->getValueType(0));
2905 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2906 /// is suitable for input to PSHUFLW.
2907 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2908 if (VT != MVT::v8i16)
2911 // Upper quadword copied in order.
2912 for (int i = 4; i != 8; ++i)
2913 if (Mask[i] >= 0 && Mask[i] != i)
2916 // Lower quadword shuffled.
2917 for (int i = 0; i != 4; ++i)
2924 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2925 SmallVector<int, 8> M;
2927 return ::isPSHUFLWMask(M, N->getValueType(0));
2930 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2931 /// is suitable for input to PALIGNR.
2932 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2934 int i, e = VT.getVectorNumElements();
2936 // Do not handle v2i64 / v2f64 shuffles with palignr.
2937 if (e < 4 || !hasSSSE3)
2940 for (i = 0; i != e; ++i)
2944 // All undef, not a palignr.
2948 // Determine if it's ok to perform a palignr with only the LHS, since we
2949 // don't have access to the actual shuffle elements to see if RHS is undef.
2950 bool Unary = Mask[i] < (int)e;
2951 bool NeedsUnary = false;
2953 int s = Mask[i] - i;
2955 // Check the rest of the elements to see if they are consecutive.
2956 for (++i; i != e; ++i) {
2961 Unary = Unary && (m < (int)e);
2962 NeedsUnary = NeedsUnary || (m < s);
2964 if (NeedsUnary && !Unary)
2966 if (Unary && m != ((s+i) & (e-1)))
2968 if (!Unary && m != (s+i))
2974 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2975 SmallVector<int, 8> M;
2977 return ::isPALIGNRMask(M, N->getValueType(0), true);
2980 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2981 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2982 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2983 int NumElems = VT.getVectorNumElements();
2984 if (NumElems != 2 && NumElems != 4)
2987 int Half = NumElems / 2;
2988 for (int i = 0; i < Half; ++i)
2989 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2991 for (int i = Half; i < NumElems; ++i)
2992 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2998 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2999 SmallVector<int, 8> M;
3001 return ::isSHUFPMask(M, N->getValueType(0));
3004 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
3005 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
3006 /// half elements to come from vector 1 (which would equal the dest.) and
3007 /// the upper half to come from vector 2.
3008 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3009 int NumElems = VT.getVectorNumElements();
3011 if (NumElems != 2 && NumElems != 4)
3014 int Half = NumElems / 2;
3015 for (int i = 0; i < Half; ++i)
3016 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3018 for (int i = Half; i < NumElems; ++i)
3019 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3024 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
3025 SmallVector<int, 8> M;
3027 return isCommutedSHUFPMask(M, N->getValueType(0));
3030 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3031 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3032 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
3033 if (N->getValueType(0).getVectorNumElements() != 4)
3036 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3037 return isUndefOrEqual(N->getMaskElt(0), 6) &&
3038 isUndefOrEqual(N->getMaskElt(1), 7) &&
3039 isUndefOrEqual(N->getMaskElt(2), 2) &&
3040 isUndefOrEqual(N->getMaskElt(3), 3);
3043 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3044 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3046 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
3047 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3052 return isUndefOrEqual(N->getMaskElt(0), 2) &&
3053 isUndefOrEqual(N->getMaskElt(1), 3) &&
3054 isUndefOrEqual(N->getMaskElt(2), 2) &&
3055 isUndefOrEqual(N->getMaskElt(3), 3);
3058 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3059 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3060 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
3061 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3063 if (NumElems != 2 && NumElems != 4)
3066 for (unsigned i = 0; i < NumElems/2; ++i)
3067 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
3070 for (unsigned i = NumElems/2; i < NumElems; ++i)
3071 if (!isUndefOrEqual(N->getMaskElt(i), i))
3077 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3078 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3079 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
3080 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3082 if (NumElems != 2 && NumElems != 4)
3085 for (unsigned i = 0; i < NumElems/2; ++i)
3086 if (!isUndefOrEqual(N->getMaskElt(i), i))
3089 for (unsigned i = 0; i < NumElems/2; ++i)
3090 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
3096 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3097 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3098 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3099 bool V2IsSplat = false) {
3100 int NumElts = VT.getVectorNumElements();
3101 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3104 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3106 int BitI1 = Mask[i+1];
3107 if (!isUndefOrEqual(BitI, j))
3110 if (!isUndefOrEqual(BitI1, NumElts))
3113 if (!isUndefOrEqual(BitI1, j + NumElts))
3120 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3121 SmallVector<int, 8> M;
3123 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
3126 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3127 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3128 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3129 bool V2IsSplat = false) {
3130 int NumElts = VT.getVectorNumElements();
3131 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3134 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3136 int BitI1 = Mask[i+1];
3137 if (!isUndefOrEqual(BitI, j + NumElts/2))
3140 if (isUndefOrEqual(BitI1, NumElts))
3143 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
3150 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3151 SmallVector<int, 8> M;
3153 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3156 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3157 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3159 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3160 int NumElems = VT.getVectorNumElements();
3161 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3164 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
3166 int BitI1 = Mask[i+1];
3167 if (!isUndefOrEqual(BitI, j))
3169 if (!isUndefOrEqual(BitI1, j))
3175 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3176 SmallVector<int, 8> M;
3178 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3181 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3182 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3184 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3185 int NumElems = VT.getVectorNumElements();
3186 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3189 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3191 int BitI1 = Mask[i+1];
3192 if (!isUndefOrEqual(BitI, j))
3194 if (!isUndefOrEqual(BitI1, j))
3200 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3201 SmallVector<int, 8> M;
3203 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3206 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3207 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3208 /// MOVSD, and MOVD, i.e. setting the lowest element.
3209 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3210 if (VT.getVectorElementType().getSizeInBits() < 32)
3213 int NumElts = VT.getVectorNumElements();
3215 if (!isUndefOrEqual(Mask[0], NumElts))
3218 for (int i = 1; i < NumElts; ++i)
3219 if (!isUndefOrEqual(Mask[i], i))
3225 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3226 SmallVector<int, 8> M;
3228 return ::isMOVLMask(M, N->getValueType(0));
3231 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3232 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3233 /// element of vector 2 and the other elements to come from vector 1 in order.
3234 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3235 bool V2IsSplat = false, bool V2IsUndef = false) {
3236 int NumOps = VT.getVectorNumElements();
3237 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3240 if (!isUndefOrEqual(Mask[0], 0))
3243 for (int i = 1; i < NumOps; ++i)
3244 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3245 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3246 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3252 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3253 bool V2IsUndef = false) {
3254 SmallVector<int, 8> M;
3256 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3259 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3260 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3261 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3262 if (N->getValueType(0).getVectorNumElements() != 4)
3265 // Expect 1, 1, 3, 3
3266 for (unsigned i = 0; i < 2; ++i) {
3267 int Elt = N->getMaskElt(i);
3268 if (Elt >= 0 && Elt != 1)
3273 for (unsigned i = 2; i < 4; ++i) {
3274 int Elt = N->getMaskElt(i);
3275 if (Elt >= 0 && Elt != 3)
3280 // Don't use movshdup if it can be done with a shufps.
3281 // FIXME: verify that matching u, u, 3, 3 is what we want.
3285 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3286 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3287 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3288 if (N->getValueType(0).getVectorNumElements() != 4)
3291 // Expect 0, 0, 2, 2
3292 for (unsigned i = 0; i < 2; ++i)
3293 if (N->getMaskElt(i) > 0)
3297 for (unsigned i = 2; i < 4; ++i) {
3298 int Elt = N->getMaskElt(i);
3299 if (Elt >= 0 && Elt != 2)
3304 // Don't use movsldup if it can be done with a shufps.
3308 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3309 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3310 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3311 int e = N->getValueType(0).getVectorNumElements() / 2;
3313 for (int i = 0; i < e; ++i)
3314 if (!isUndefOrEqual(N->getMaskElt(i), i))
3316 for (int i = 0; i < e; ++i)
3317 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3322 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3323 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3324 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3325 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3326 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3328 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3330 for (int i = 0; i < NumOperands; ++i) {
3331 int Val = SVOp->getMaskElt(NumOperands-i-1);
3332 if (Val < 0) Val = 0;
3333 if (Val >= NumOperands) Val -= NumOperands;
3335 if (i != NumOperands - 1)
3341 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3342 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3343 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3344 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3346 // 8 nodes, but we only care about the last 4.
3347 for (unsigned i = 7; i >= 4; --i) {
3348 int Val = SVOp->getMaskElt(i);
3357 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3358 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3359 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3360 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3362 // 8 nodes, but we only care about the first 4.
3363 for (int i = 3; i >= 0; --i) {
3364 int Val = SVOp->getMaskElt(i);
3373 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3374 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3375 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3376 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3377 EVT VVT = N->getValueType(0);
3378 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3382 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3383 Val = SVOp->getMaskElt(i);
3387 return (Val - i) * EltSize;
3390 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3392 bool X86::isZeroNode(SDValue Elt) {
3393 return ((isa<ConstantSDNode>(Elt) &&
3394 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3395 (isa<ConstantFPSDNode>(Elt) &&
3396 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3399 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3400 /// their permute mask.
3401 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3402 SelectionDAG &DAG) {
3403 EVT VT = SVOp->getValueType(0);
3404 unsigned NumElems = VT.getVectorNumElements();
3405 SmallVector<int, 8> MaskVec;
3407 for (unsigned i = 0; i != NumElems; ++i) {
3408 int idx = SVOp->getMaskElt(i);
3410 MaskVec.push_back(idx);
3411 else if (idx < (int)NumElems)
3412 MaskVec.push_back(idx + NumElems);
3414 MaskVec.push_back(idx - NumElems);
3416 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3417 SVOp->getOperand(0), &MaskVec[0]);
3420 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3421 /// the two vector operands have swapped position.
3422 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3423 unsigned NumElems = VT.getVectorNumElements();
3424 for (unsigned i = 0; i != NumElems; ++i) {
3428 else if (idx < (int)NumElems)
3429 Mask[i] = idx + NumElems;
3431 Mask[i] = idx - NumElems;
3435 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3436 /// match movhlps. The lower half elements should come from upper half of
3437 /// V1 (and in order), and the upper half elements should come from the upper
3438 /// half of V2 (and in order).
3439 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3440 if (Op->getValueType(0).getVectorNumElements() != 4)
3442 for (unsigned i = 0, e = 2; i != e; ++i)
3443 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3445 for (unsigned i = 2; i != 4; ++i)
3446 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3451 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3452 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3454 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3455 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3457 N = N->getOperand(0).getNode();
3458 if (!ISD::isNON_EXTLoad(N))
3461 *LD = cast<LoadSDNode>(N);
3465 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3466 /// match movlp{s|d}. The lower half elements should come from lower half of
3467 /// V1 (and in order), and the upper half elements should come from the upper
3468 /// half of V2 (and in order). And since V1 will become the source of the
3469 /// MOVLP, it must be either a vector load or a scalar load to vector.
3470 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3471 ShuffleVectorSDNode *Op) {
3472 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3474 // Is V2 is a vector load, don't do this transformation. We will try to use
3475 // load folding shufps op.
3476 if (ISD::isNON_EXTLoad(V2))
3479 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3481 if (NumElems != 2 && NumElems != 4)
3483 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3484 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3486 for (unsigned i = NumElems/2; i != NumElems; ++i)
3487 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3492 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3494 static bool isSplatVector(SDNode *N) {
3495 if (N->getOpcode() != ISD::BUILD_VECTOR)
3498 SDValue SplatValue = N->getOperand(0);
3499 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3500 if (N->getOperand(i) != SplatValue)
3505 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3506 /// to an zero vector.
3507 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3508 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3509 SDValue V1 = N->getOperand(0);
3510 SDValue V2 = N->getOperand(1);
3511 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3512 for (unsigned i = 0; i != NumElems; ++i) {
3513 int Idx = N->getMaskElt(i);
3514 if (Idx >= (int)NumElems) {
3515 unsigned Opc = V2.getOpcode();
3516 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3518 if (Opc != ISD::BUILD_VECTOR ||
3519 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3521 } else if (Idx >= 0) {
3522 unsigned Opc = V1.getOpcode();
3523 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3525 if (Opc != ISD::BUILD_VECTOR ||
3526 !X86::isZeroNode(V1.getOperand(Idx)))
3533 /// getZeroVector - Returns a vector of specified type with all zero elements.
3535 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3537 assert(VT.isVector() && "Expected a vector type");
3539 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted
3540 // to their dest type. This ensures they get CSE'd.
3542 if (VT.getSizeInBits() == 64) { // MMX
3543 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3544 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3545 } else if (VT.getSizeInBits() == 128) {
3546 if (HasSSE2) { // SSE2
3547 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3548 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3550 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3551 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3553 } else if (VT.getSizeInBits() == 256) { // AVX
3554 // 256-bit logic and arithmetic instructions in AVX are
3555 // all floating-point, no support for integer ops. Default
3556 // to emitting fp zeroed vectors then.
3557 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3558 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3559 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
3561 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3564 /// getOnesVector - Returns a vector of specified type with all bits set.
3566 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3567 assert(VT.isVector() && "Expected a vector type");
3569 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3570 // type. This ensures they get CSE'd.
3571 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3573 if (VT.getSizeInBits() == 64) // MMX
3574 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3576 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3577 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3581 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3582 /// that point to V2 points to its first element.
3583 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3584 EVT VT = SVOp->getValueType(0);
3585 unsigned NumElems = VT.getVectorNumElements();
3587 bool Changed = false;
3588 SmallVector<int, 8> MaskVec;
3589 SVOp->getMask(MaskVec);
3591 for (unsigned i = 0; i != NumElems; ++i) {
3592 if (MaskVec[i] > (int)NumElems) {
3593 MaskVec[i] = NumElems;
3598 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3599 SVOp->getOperand(1), &MaskVec[0]);
3600 return SDValue(SVOp, 0);
3603 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3604 /// operation of specified width.
3605 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3607 unsigned NumElems = VT.getVectorNumElements();
3608 SmallVector<int, 8> Mask;
3609 Mask.push_back(NumElems);
3610 for (unsigned i = 1; i != NumElems; ++i)
3612 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3615 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3616 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3618 unsigned NumElems = VT.getVectorNumElements();
3619 SmallVector<int, 8> Mask;
3620 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3622 Mask.push_back(i + NumElems);
3624 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3627 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3628 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3630 unsigned NumElems = VT.getVectorNumElements();
3631 unsigned Half = NumElems/2;
3632 SmallVector<int, 8> Mask;
3633 for (unsigned i = 0; i != Half; ++i) {
3634 Mask.push_back(i + Half);
3635 Mask.push_back(i + NumElems + Half);
3637 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3640 /// PromoteSplat - Promote a splat of v4i32, v8i16 or v16i8 to v4f32.
3641 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
3642 if (SV->getValueType(0).getVectorNumElements() <= 4)
3643 return SDValue(SV, 0);
3645 EVT PVT = MVT::v4f32;
3646 EVT VT = SV->getValueType(0);
3647 DebugLoc dl = SV->getDebugLoc();
3648 SDValue V1 = SV->getOperand(0);
3649 int NumElems = VT.getVectorNumElements();
3650 int EltNo = SV->getSplatIndex();
3652 // unpack elements to the correct location
3653 while (NumElems > 4) {
3654 if (EltNo < NumElems/2) {
3655 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3657 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3658 EltNo -= NumElems/2;
3663 // Perform the splat.
3664 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3665 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3666 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3667 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3670 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3671 /// vector of zero or undef vector. This produces a shuffle where the low
3672 /// element of V2 is swizzled into the zero/undef vector, landing at element
3673 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3674 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3675 bool isZero, bool HasSSE2,
3676 SelectionDAG &DAG) {
3677 EVT VT = V2.getValueType();
3679 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3680 unsigned NumElems = VT.getVectorNumElements();
3681 SmallVector<int, 16> MaskVec;
3682 for (unsigned i = 0; i != NumElems; ++i)
3683 // If this is the insertion idx, put the low elt of V2 here.
3684 MaskVec.push_back(i == Idx ? NumElems : i);
3685 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3688 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
3689 /// element of the result of the vector shuffle.
3690 SDValue getShuffleScalarElt(SDNode *N, int Index, SelectionDAG &DAG) {
3691 SDValue V = SDValue(N, 0);
3692 EVT VT = V.getValueType();
3693 unsigned Opcode = V.getOpcode();
3695 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
3696 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
3697 Index = SV->getMaskElt(Index);
3700 return DAG.getUNDEF(VT.getVectorElementType());
3702 int NumElems = VT.getVectorNumElements();
3703 SDValue NewV = (Index < NumElems) ? SV->getOperand(0) : SV->getOperand(1);
3704 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG);
3707 // Recurse into target specific vector shuffles to find scalars.
3708 if (isTargetShuffle(Opcode)) {
3711 case X86ISD::MOVSD: {
3712 // The index 0 always comes from the first element of the second source,
3713 // this is why MOVSS and MOVSD are used in the first place. The other
3714 // elements come from the other positions of the first source vector.
3715 unsigned OpNum = (Index == 0) ? 1 : 0;
3716 return getShuffleScalarElt(V.getOperand(OpNum).getNode(), Index, DAG);
3719 assert("not implemented for target shuffle node");
3724 // Actual nodes that may contain scalar elements
3725 if (Opcode == ISD::BIT_CONVERT) {
3726 V = V.getOperand(0);
3727 EVT SrcVT = V.getValueType();
3728 unsigned NumElems = VT.getVectorNumElements();
3730 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
3734 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
3735 return (Index == 0) ? V.getOperand(0)
3736 : DAG.getUNDEF(VT.getVectorElementType());
3738 if (V.getOpcode() == ISD::BUILD_VECTOR)
3739 return V.getOperand(Index);
3744 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
3745 /// shuffle operation which come from a consecutively from a zero. The
3746 /// search can start in two diferent directions, from left or right.
3748 unsigned getNumOfConsecutiveZeros(SDNode *N, int NumElems,
3749 bool ZerosFromLeft, SelectionDAG &DAG) {
3752 while (i < NumElems) {
3753 unsigned Index = ZerosFromLeft ? i : NumElems-i-1;
3754 SDValue Elt = getShuffleScalarElt(N, Index, DAG);
3755 if (!(Elt.getNode() &&
3756 (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt))))
3764 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies from MaskI to
3765 /// MaskE correspond consecutively to elements from one of the vector operands,
3766 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
3768 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp, int MaskI, int MaskE,
3769 int OpIdx, int NumElems, unsigned &OpNum) {
3770 bool SeenV1 = false;
3771 bool SeenV2 = false;
3773 for (int i = MaskI; i <= MaskE; ++i, ++OpIdx) {
3774 int Idx = SVOp->getMaskElt(i);
3775 // Ignore undef indicies
3784 // Only accept consecutive elements from the same vector
3785 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
3789 OpNum = SeenV1 ? 0 : 1;
3793 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
3794 /// logical left shift of a vector.
3795 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3796 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3797 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3798 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
3799 false /* check zeros from right */, DAG);
3805 // Considering the elements in the mask that are not consecutive zeros,
3806 // check if they consecutively come from only one of the source vectors.
3808 // V1 = {X, A, B, C} 0
3810 // vector_shuffle V1, V2 <1, 2, 3, X>
3812 if (!isShuffleMaskConsecutive(SVOp,
3813 0, // Mask Start Index
3814 NumElems-NumZeros-1, // Mask End Index
3815 NumZeros, // Where to start looking in the src vector
3816 NumElems, // Number of elements in vector
3817 OpSrc)) // Which source operand ?
3822 ShVal = SVOp->getOperand(OpSrc);
3826 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
3827 /// logical left shift of a vector.
3828 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3829 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3830 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3831 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
3832 true /* check zeros from left */, DAG);
3838 // Considering the elements in the mask that are not consecutive zeros,
3839 // check if they consecutively come from only one of the source vectors.
3841 // 0 { A, B, X, X } = V2
3843 // vector_shuffle V1, V2 <X, X, 4, 5>
3845 if (!isShuffleMaskConsecutive(SVOp,
3846 NumZeros, // Mask Start Index
3847 NumElems-1, // Mask End Index
3848 0, // Where to start looking in the src vector
3849 NumElems, // Number of elements in vector
3850 OpSrc)) // Which source operand ?
3855 ShVal = SVOp->getOperand(OpSrc);
3859 /// isVectorShift - Returns true if the shuffle can be implemented as a
3860 /// logical left or right shift of a vector.
3861 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3862 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3863 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
3864 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
3870 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3872 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3873 unsigned NumNonZero, unsigned NumZero,
3875 const TargetLowering &TLI) {
3879 DebugLoc dl = Op.getDebugLoc();
3882 for (unsigned i = 0; i < 16; ++i) {
3883 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3884 if (ThisIsNonZero && First) {
3886 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3888 V = DAG.getUNDEF(MVT::v8i16);
3893 SDValue ThisElt(0, 0), LastElt(0, 0);
3894 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3895 if (LastIsNonZero) {
3896 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3897 MVT::i16, Op.getOperand(i-1));
3899 if (ThisIsNonZero) {
3900 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3901 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3902 ThisElt, DAG.getConstant(8, MVT::i8));
3904 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3908 if (ThisElt.getNode())
3909 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3910 DAG.getIntPtrConstant(i/2));
3914 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3917 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3919 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3920 unsigned NumNonZero, unsigned NumZero,
3922 const TargetLowering &TLI) {
3926 DebugLoc dl = Op.getDebugLoc();
3929 for (unsigned i = 0; i < 8; ++i) {
3930 bool isNonZero = (NonZeros & (1 << i)) != 0;
3934 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3936 V = DAG.getUNDEF(MVT::v8i16);
3939 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3940 MVT::v8i16, V, Op.getOperand(i),
3941 DAG.getIntPtrConstant(i));
3948 /// getVShift - Return a vector logical shift node.
3950 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3951 unsigned NumBits, SelectionDAG &DAG,
3952 const TargetLowering &TLI, DebugLoc dl) {
3953 bool isMMX = VT.getSizeInBits() == 64;
3954 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3955 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3956 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3957 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3958 DAG.getNode(Opc, dl, ShVT, SrcOp,
3959 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3963 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3964 SelectionDAG &DAG) const {
3966 // Check if the scalar load can be widened into a vector load. And if
3967 // the address is "base + cst" see if the cst can be "absorbed" into
3968 // the shuffle mask.
3969 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3970 SDValue Ptr = LD->getBasePtr();
3971 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3973 EVT PVT = LD->getValueType(0);
3974 if (PVT != MVT::i32 && PVT != MVT::f32)
3979 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3980 FI = FINode->getIndex();
3982 } else if (Ptr.getOpcode() == ISD::ADD &&
3983 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3984 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3985 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3986 Offset = Ptr.getConstantOperandVal(1);
3987 Ptr = Ptr.getOperand(0);
3992 SDValue Chain = LD->getChain();
3993 // Make sure the stack object alignment is at least 16.
3994 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3995 if (DAG.InferPtrAlignment(Ptr) < 16) {
3996 if (MFI->isFixedObjectIndex(FI)) {
3997 // Can't change the alignment. FIXME: It's possible to compute
3998 // the exact stack offset and reference FI + adjust offset instead.
3999 // If someone *really* cares about this. That's the way to implement it.
4002 MFI->setObjectAlignment(FI, 16);
4006 // (Offset % 16) must be multiple of 4. Then address is then
4007 // Ptr + (Offset & ~15).
4010 if ((Offset % 16) & 3)
4012 int64_t StartOffset = Offset & ~15;
4014 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
4015 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4017 int EltNo = (Offset - StartOffset) >> 2;
4018 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
4019 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
4020 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
4022 // Canonicalize it to a v4i32 shuffle.
4023 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
4024 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4025 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
4026 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
4032 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
4033 /// vector of type 'VT', see if the elements can be replaced by a single large
4034 /// load which has the same value as a build_vector whose operands are 'elts'.
4036 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4038 /// FIXME: we'd also like to handle the case where the last elements are zero
4039 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4040 /// There's even a handy isZeroNode for that purpose.
4041 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
4042 DebugLoc &dl, SelectionDAG &DAG) {
4043 EVT EltVT = VT.getVectorElementType();
4044 unsigned NumElems = Elts.size();
4046 LoadSDNode *LDBase = NULL;
4047 unsigned LastLoadedElt = -1U;
4049 // For each element in the initializer, see if we've found a load or an undef.
4050 // If we don't find an initial load element, or later load elements are
4051 // non-consecutive, bail out.
4052 for (unsigned i = 0; i < NumElems; ++i) {
4053 SDValue Elt = Elts[i];
4055 if (!Elt.getNode() ||
4056 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4059 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4061 LDBase = cast<LoadSDNode>(Elt.getNode());
4065 if (Elt.getOpcode() == ISD::UNDEF)
4068 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4069 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
4074 // If we have found an entire vector of loads and undefs, then return a large
4075 // load of the entire vector width starting at the base pointer. If we found
4076 // consecutive loads for the low half, generate a vzext_load node.
4077 if (LastLoadedElt == NumElems - 1) {
4078 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
4079 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
4080 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
4081 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
4082 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
4083 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
4084 LDBase->isVolatile(), LDBase->isNonTemporal(),
4085 LDBase->getAlignment());
4086 } else if (NumElems == 4 && LastLoadedElt == 1) {
4087 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4088 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4089 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
4090 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
4096 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
4097 DebugLoc dl = Op.getDebugLoc();
4098 // All zero's are handled with pxor in SSE2 and above, xorps in SSE1.
4099 // All one's are handled with pcmpeqd. In AVX, zero's are handled with
4100 // vpxor in 128-bit and xor{pd,ps} in 256-bit, but no 256 version of pcmpeqd
4101 // is present, so AllOnes is ignored.
4102 if (ISD::isBuildVectorAllZeros(Op.getNode()) ||
4103 (Op.getValueType().getSizeInBits() != 256 &&
4104 ISD::isBuildVectorAllOnes(Op.getNode()))) {
4105 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
4106 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
4107 // eliminated on x86-32 hosts.
4108 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
4111 if (ISD::isBuildVectorAllOnes(Op.getNode()))
4112 return getOnesVector(Op.getValueType(), DAG, dl);
4113 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
4116 EVT VT = Op.getValueType();
4117 EVT ExtVT = VT.getVectorElementType();
4118 unsigned EVTBits = ExtVT.getSizeInBits();
4120 unsigned NumElems = Op.getNumOperands();
4121 unsigned NumZero = 0;
4122 unsigned NumNonZero = 0;
4123 unsigned NonZeros = 0;
4124 bool IsAllConstants = true;
4125 SmallSet<SDValue, 8> Values;
4126 for (unsigned i = 0; i < NumElems; ++i) {
4127 SDValue Elt = Op.getOperand(i);
4128 if (Elt.getOpcode() == ISD::UNDEF)
4131 if (Elt.getOpcode() != ISD::Constant &&
4132 Elt.getOpcode() != ISD::ConstantFP)
4133 IsAllConstants = false;
4134 if (X86::isZeroNode(Elt))
4137 NonZeros |= (1 << i);
4142 // All undef vector. Return an UNDEF. All zero vectors were handled above.
4143 if (NumNonZero == 0)
4144 return DAG.getUNDEF(VT);
4146 // Special case for single non-zero, non-undef, element.
4147 if (NumNonZero == 1) {
4148 unsigned Idx = CountTrailingZeros_32(NonZeros);
4149 SDValue Item = Op.getOperand(Idx);
4151 // If this is an insertion of an i64 value on x86-32, and if the top bits of
4152 // the value are obviously zero, truncate the value to i32 and do the
4153 // insertion that way. Only do this if the value is non-constant or if the
4154 // value is a constant being inserted into element 0. It is cheaper to do
4155 // a constant pool load than it is to do a movd + shuffle.
4156 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
4157 (!IsAllConstants || Idx == 0)) {
4158 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
4159 // Handle MMX and SSE both.
4160 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
4161 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
4163 // Truncate the value (which may itself be a constant) to i32, and
4164 // convert it to a vector with movd (S2V+shuffle to zero extend).
4165 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
4166 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
4167 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4168 Subtarget->hasSSE2(), DAG);
4170 // Now we have our 32-bit value zero extended in the low element of
4171 // a vector. If Idx != 0, swizzle it into place.
4173 SmallVector<int, 4> Mask;
4174 Mask.push_back(Idx);
4175 for (unsigned i = 1; i != VecElts; ++i)
4177 Item = DAG.getVectorShuffle(VecVT, dl, Item,
4178 DAG.getUNDEF(Item.getValueType()),
4181 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
4185 // If we have a constant or non-constant insertion into the low element of
4186 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
4187 // the rest of the elements. This will be matched as movd/movq/movss/movsd
4188 // depending on what the source datatype is.
4191 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4192 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
4193 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
4194 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4195 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
4196 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
4198 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
4199 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
4200 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
4201 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
4202 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4203 Subtarget->hasSSE2(), DAG);
4204 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
4208 // Is it a vector logical left shift?
4209 if (NumElems == 2 && Idx == 1 &&
4210 X86::isZeroNode(Op.getOperand(0)) &&
4211 !X86::isZeroNode(Op.getOperand(1))) {
4212 unsigned NumBits = VT.getSizeInBits();
4213 return getVShift(true, VT,
4214 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4215 VT, Op.getOperand(1)),
4216 NumBits/2, DAG, *this, dl);
4219 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
4222 // Otherwise, if this is a vector with i32 or f32 elements, and the element
4223 // is a non-constant being inserted into an element other than the low one,
4224 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
4225 // movd/movss) to move this into the low element, then shuffle it into
4227 if (EVTBits == 32) {
4228 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4230 // Turn it into a shuffle of zero and zero-extended scalar to vector.
4231 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
4232 Subtarget->hasSSE2(), DAG);
4233 SmallVector<int, 8> MaskVec;
4234 for (unsigned i = 0; i < NumElems; i++)
4235 MaskVec.push_back(i == Idx ? 0 : 1);
4236 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
4240 // Splat is obviously ok. Let legalizer expand it to a shuffle.
4241 if (Values.size() == 1) {
4242 if (EVTBits == 32) {
4243 // Instead of a shuffle like this:
4244 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
4245 // Check if it's possible to issue this instead.
4246 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
4247 unsigned Idx = CountTrailingZeros_32(NonZeros);
4248 SDValue Item = Op.getOperand(Idx);
4249 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
4250 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
4255 // A vector full of immediates; various special cases are already
4256 // handled, so this is best done with a single constant-pool load.
4260 // Let legalizer expand 2-wide build_vectors.
4261 if (EVTBits == 64) {
4262 if (NumNonZero == 1) {
4263 // One half is zero or undef.
4264 unsigned Idx = CountTrailingZeros_32(NonZeros);
4265 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
4266 Op.getOperand(Idx));
4267 return getShuffleVectorZeroOrUndef(V2, Idx, true,
4268 Subtarget->hasSSE2(), DAG);
4273 // If element VT is < 32 bits, convert it to inserts into a zero vector.
4274 if (EVTBits == 8 && NumElems == 16) {
4275 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
4277 if (V.getNode()) return V;
4280 if (EVTBits == 16 && NumElems == 8) {
4281 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
4283 if (V.getNode()) return V;
4286 // If element VT is == 32 bits, turn it into a number of shuffles.
4287 SmallVector<SDValue, 8> V;
4289 if (NumElems == 4 && NumZero > 0) {
4290 for (unsigned i = 0; i < 4; ++i) {
4291 bool isZero = !(NonZeros & (1 << i));
4293 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4295 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4298 for (unsigned i = 0; i < 2; ++i) {
4299 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
4302 V[i] = V[i*2]; // Must be a zero vector.
4305 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
4308 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
4311 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
4316 SmallVector<int, 8> MaskVec;
4317 bool Reverse = (NonZeros & 0x3) == 2;
4318 for (unsigned i = 0; i < 2; ++i)
4319 MaskVec.push_back(Reverse ? 1-i : i);
4320 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4321 for (unsigned i = 0; i < 2; ++i)
4322 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4323 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4326 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4327 // Check for a build vector of consecutive loads.
4328 for (unsigned i = 0; i < NumElems; ++i)
4329 V[i] = Op.getOperand(i);
4331 // Check for elements which are consecutive loads.
4332 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4336 // For SSE 4.1, use insertps to put the high elements into the low element.
4337 if (getSubtarget()->hasSSE41()) {
4339 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
4340 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
4342 Result = DAG.getUNDEF(VT);
4344 for (unsigned i = 1; i < NumElems; ++i) {
4345 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
4346 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
4347 Op.getOperand(i), DAG.getIntPtrConstant(i));
4352 // Otherwise, expand into a number of unpckl*, start by extending each of
4353 // our (non-undef) elements to the full vector width with the element in the
4354 // bottom slot of the vector (which generates no code for SSE).
4355 for (unsigned i = 0; i < NumElems; ++i) {
4356 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4357 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4359 V[i] = DAG.getUNDEF(VT);
4362 // Next, we iteratively mix elements, e.g. for v4f32:
4363 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4364 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4365 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4366 unsigned EltStride = NumElems >> 1;
4367 while (EltStride != 0) {
4368 for (unsigned i = 0; i < EltStride; ++i) {
4369 // If V[i+EltStride] is undef and this is the first round of mixing,
4370 // then it is safe to just drop this shuffle: V[i] is already in the
4371 // right place, the one element (since it's the first round) being
4372 // inserted as undef can be dropped. This isn't safe for successive
4373 // rounds because they will permute elements within both vectors.
4374 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
4375 EltStride == NumElems/2)
4378 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
4388 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4389 // We support concatenate two MMX registers and place them in a MMX
4390 // register. This is better than doing a stack convert.
4391 DebugLoc dl = Op.getDebugLoc();
4392 EVT ResVT = Op.getValueType();
4393 assert(Op.getNumOperands() == 2);
4394 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4395 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4397 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4398 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4399 InVec = Op.getOperand(1);
4400 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4401 unsigned NumElts = ResVT.getVectorNumElements();
4402 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4403 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4404 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4406 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4407 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4408 Mask[0] = 0; Mask[1] = 2;
4409 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4411 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4414 // v8i16 shuffles - Prefer shuffles in the following order:
4415 // 1. [all] pshuflw, pshufhw, optional move
4416 // 2. [ssse3] 1 x pshufb
4417 // 3. [ssse3] 2 x pshufb + 1 x por
4418 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4420 X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
4421 SelectionDAG &DAG) const {
4422 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4423 SDValue V1 = SVOp->getOperand(0);
4424 SDValue V2 = SVOp->getOperand(1);
4425 DebugLoc dl = SVOp->getDebugLoc();
4426 SmallVector<int, 8> MaskVals;
4428 // Determine if more than 1 of the words in each of the low and high quadwords
4429 // of the result come from the same quadword of one of the two inputs. Undef
4430 // mask values count as coming from any quadword, for better codegen.
4431 SmallVector<unsigned, 4> LoQuad(4);
4432 SmallVector<unsigned, 4> HiQuad(4);
4433 BitVector InputQuads(4);
4434 for (unsigned i = 0; i < 8; ++i) {
4435 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4436 int EltIdx = SVOp->getMaskElt(i);
4437 MaskVals.push_back(EltIdx);
4446 InputQuads.set(EltIdx / 4);
4449 int BestLoQuad = -1;
4450 unsigned MaxQuad = 1;
4451 for (unsigned i = 0; i < 4; ++i) {
4452 if (LoQuad[i] > MaxQuad) {
4454 MaxQuad = LoQuad[i];
4458 int BestHiQuad = -1;
4460 for (unsigned i = 0; i < 4; ++i) {
4461 if (HiQuad[i] > MaxQuad) {
4463 MaxQuad = HiQuad[i];
4467 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4468 // of the two input vectors, shuffle them into one input vector so only a
4469 // single pshufb instruction is necessary. If There are more than 2 input
4470 // quads, disable the next transformation since it does not help SSSE3.
4471 bool V1Used = InputQuads[0] || InputQuads[1];
4472 bool V2Used = InputQuads[2] || InputQuads[3];
4473 if (Subtarget->hasSSSE3()) {
4474 if (InputQuads.count() == 2 && V1Used && V2Used) {
4475 BestLoQuad = InputQuads.find_first();
4476 BestHiQuad = InputQuads.find_next(BestLoQuad);
4478 if (InputQuads.count() > 2) {
4484 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4485 // the shuffle mask. If a quad is scored as -1, that means that it contains
4486 // words from all 4 input quadwords.
4488 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4489 SmallVector<int, 8> MaskV;
4490 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4491 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4492 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4493 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4494 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4495 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4497 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4498 // source words for the shuffle, to aid later transformations.
4499 bool AllWordsInNewV = true;
4500 bool InOrder[2] = { true, true };
4501 for (unsigned i = 0; i != 8; ++i) {
4502 int idx = MaskVals[i];
4504 InOrder[i/4] = false;
4505 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4507 AllWordsInNewV = false;
4511 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4512 if (AllWordsInNewV) {
4513 for (int i = 0; i != 8; ++i) {
4514 int idx = MaskVals[i];
4517 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4518 if ((idx != i) && idx < 4)
4520 if ((idx != i) && idx > 3)
4529 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4530 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4531 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4532 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
4533 unsigned TargetMask = 0;
4534 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4535 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4536 TargetMask = pshufhw ? X86::getShufflePSHUFHWImmediate(NewV.getNode()):
4537 X86::getShufflePSHUFLWImmediate(NewV.getNode());
4538 V1 = NewV.getOperand(0);
4539 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
4543 // If we have SSSE3, and all words of the result are from 1 input vector,
4544 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4545 // is present, fall back to case 4.
4546 if (Subtarget->hasSSSE3()) {
4547 SmallVector<SDValue,16> pshufbMask;
4549 // If we have elements from both input vectors, set the high bit of the
4550 // shuffle mask element to zero out elements that come from V2 in the V1
4551 // mask, and elements that come from V1 in the V2 mask, so that the two
4552 // results can be OR'd together.
4553 bool TwoInputs = V1Used && V2Used;
4554 for (unsigned i = 0; i != 8; ++i) {
4555 int EltIdx = MaskVals[i] * 2;
4556 if (TwoInputs && (EltIdx >= 16)) {
4557 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4558 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4561 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4562 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4564 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4565 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4566 DAG.getNode(ISD::BUILD_VECTOR, dl,
4567 MVT::v16i8, &pshufbMask[0], 16));
4569 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4571 // Calculate the shuffle mask for the second input, shuffle it, and
4572 // OR it with the first shuffled input.
4574 for (unsigned i = 0; i != 8; ++i) {
4575 int EltIdx = MaskVals[i] * 2;
4577 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4578 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4581 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4582 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4584 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4585 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4586 DAG.getNode(ISD::BUILD_VECTOR, dl,
4587 MVT::v16i8, &pshufbMask[0], 16));
4588 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4589 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4592 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4593 // and update MaskVals with new element order.
4594 BitVector InOrder(8);
4595 if (BestLoQuad >= 0) {
4596 SmallVector<int, 8> MaskV;
4597 for (int i = 0; i != 4; ++i) {
4598 int idx = MaskVals[i];
4600 MaskV.push_back(-1);
4602 } else if ((idx / 4) == BestLoQuad) {
4603 MaskV.push_back(idx & 3);
4606 MaskV.push_back(-1);
4609 for (unsigned i = 4; i != 8; ++i)
4611 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4614 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4615 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
4617 X86::getShufflePSHUFLWImmediate(NewV.getNode()),
4621 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4622 // and update MaskVals with the new element order.
4623 if (BestHiQuad >= 0) {
4624 SmallVector<int, 8> MaskV;
4625 for (unsigned i = 0; i != 4; ++i)
4627 for (unsigned i = 4; i != 8; ++i) {
4628 int idx = MaskVals[i];
4630 MaskV.push_back(-1);
4632 } else if ((idx / 4) == BestHiQuad) {
4633 MaskV.push_back((idx & 3) + 4);
4636 MaskV.push_back(-1);
4639 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4642 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4643 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
4645 X86::getShufflePSHUFHWImmediate(NewV.getNode()),
4649 // In case BestHi & BestLo were both -1, which means each quadword has a word
4650 // from each of the four input quadwords, calculate the InOrder bitvector now
4651 // before falling through to the insert/extract cleanup.
4652 if (BestLoQuad == -1 && BestHiQuad == -1) {
4654 for (int i = 0; i != 8; ++i)
4655 if (MaskVals[i] < 0 || MaskVals[i] == i)
4659 // The other elements are put in the right place using pextrw and pinsrw.
4660 for (unsigned i = 0; i != 8; ++i) {
4663 int EltIdx = MaskVals[i];
4666 SDValue ExtOp = (EltIdx < 8)
4667 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4668 DAG.getIntPtrConstant(EltIdx))
4669 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4670 DAG.getIntPtrConstant(EltIdx - 8));
4671 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4672 DAG.getIntPtrConstant(i));
4677 // v16i8 shuffles - Prefer shuffles in the following order:
4678 // 1. [ssse3] 1 x pshufb
4679 // 2. [ssse3] 2 x pshufb + 1 x por
4680 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4682 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4684 const X86TargetLowering &TLI) {
4685 SDValue V1 = SVOp->getOperand(0);
4686 SDValue V2 = SVOp->getOperand(1);
4687 DebugLoc dl = SVOp->getDebugLoc();
4688 SmallVector<int, 16> MaskVals;
4689 SVOp->getMask(MaskVals);
4691 // If we have SSSE3, case 1 is generated when all result bytes come from
4692 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4693 // present, fall back to case 3.
4694 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4697 for (unsigned i = 0; i < 16; ++i) {
4698 int EltIdx = MaskVals[i];
4707 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4708 if (TLI.getSubtarget()->hasSSSE3()) {
4709 SmallVector<SDValue,16> pshufbMask;
4711 // If all result elements are from one input vector, then only translate
4712 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4714 // Otherwise, we have elements from both input vectors, and must zero out
4715 // elements that come from V2 in the first mask, and V1 in the second mask
4716 // so that we can OR them together.
4717 bool TwoInputs = !(V1Only || V2Only);
4718 for (unsigned i = 0; i != 16; ++i) {
4719 int EltIdx = MaskVals[i];
4720 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4721 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4724 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4726 // If all the elements are from V2, assign it to V1 and return after
4727 // building the first pshufb.
4730 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4731 DAG.getNode(ISD::BUILD_VECTOR, dl,
4732 MVT::v16i8, &pshufbMask[0], 16));
4736 // Calculate the shuffle mask for the second input, shuffle it, and
4737 // OR it with the first shuffled input.
4739 for (unsigned i = 0; i != 16; ++i) {
4740 int EltIdx = MaskVals[i];
4742 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4745 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4747 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4748 DAG.getNode(ISD::BUILD_VECTOR, dl,
4749 MVT::v16i8, &pshufbMask[0], 16));
4750 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4753 // No SSSE3 - Calculate in place words and then fix all out of place words
4754 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4755 // the 16 different words that comprise the two doublequadword input vectors.
4756 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4757 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4758 SDValue NewV = V2Only ? V2 : V1;
4759 for (int i = 0; i != 8; ++i) {
4760 int Elt0 = MaskVals[i*2];
4761 int Elt1 = MaskVals[i*2+1];
4763 // This word of the result is all undef, skip it.
4764 if (Elt0 < 0 && Elt1 < 0)
4767 // This word of the result is already in the correct place, skip it.
4768 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4770 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4773 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4774 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4777 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4778 // using a single extract together, load it and store it.
4779 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4780 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4781 DAG.getIntPtrConstant(Elt1 / 2));
4782 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4783 DAG.getIntPtrConstant(i));
4787 // If Elt1 is defined, extract it from the appropriate source. If the
4788 // source byte is not also odd, shift the extracted word left 8 bits
4789 // otherwise clear the bottom 8 bits if we need to do an or.
4791 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4792 DAG.getIntPtrConstant(Elt1 / 2));
4793 if ((Elt1 & 1) == 0)
4794 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4795 DAG.getConstant(8, TLI.getShiftAmountTy()));
4797 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4798 DAG.getConstant(0xFF00, MVT::i16));
4800 // If Elt0 is defined, extract it from the appropriate source. If the
4801 // source byte is not also even, shift the extracted word right 8 bits. If
4802 // Elt1 was also defined, OR the extracted values together before
4803 // inserting them in the result.
4805 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4806 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4807 if ((Elt0 & 1) != 0)
4808 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4809 DAG.getConstant(8, TLI.getShiftAmountTy()));
4811 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4812 DAG.getConstant(0x00FF, MVT::i16));
4813 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4816 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4817 DAG.getIntPtrConstant(i));
4819 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4822 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4823 /// ones, or rewriting v4i32 / v2i32 as 2 wide ones if possible. This can be
4824 /// done when every pair / quad of shuffle mask elements point to elements in
4825 /// the right sequence. e.g.
4826 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4828 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4830 const TargetLowering &TLI, DebugLoc dl) {
4831 EVT VT = SVOp->getValueType(0);
4832 SDValue V1 = SVOp->getOperand(0);
4833 SDValue V2 = SVOp->getOperand(1);
4834 unsigned NumElems = VT.getVectorNumElements();
4835 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4836 EVT MaskVT = (NewWidth == 4) ? MVT::v4i16 : MVT::v2i32;
4838 switch (VT.getSimpleVT().SimpleTy) {
4839 default: assert(false && "Unexpected!");
4840 case MVT::v4f32: NewVT = MVT::v2f64; break;
4841 case MVT::v4i32: NewVT = MVT::v2i64; break;
4842 case MVT::v8i16: NewVT = MVT::v4i32; break;
4843 case MVT::v16i8: NewVT = MVT::v4i32; break;
4846 if (NewWidth == 2) {
4852 int Scale = NumElems / NewWidth;
4853 SmallVector<int, 8> MaskVec;
4854 for (unsigned i = 0; i < NumElems; i += Scale) {
4856 for (int j = 0; j < Scale; ++j) {
4857 int EltIdx = SVOp->getMaskElt(i+j);
4861 StartIdx = EltIdx - (EltIdx % Scale);
4862 if (EltIdx != StartIdx + j)
4866 MaskVec.push_back(-1);
4868 MaskVec.push_back(StartIdx / Scale);
4871 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4872 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4873 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4876 /// getVZextMovL - Return a zero-extending vector move low node.
4878 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4879 SDValue SrcOp, SelectionDAG &DAG,
4880 const X86Subtarget *Subtarget, DebugLoc dl) {
4881 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4882 LoadSDNode *LD = NULL;
4883 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4884 LD = dyn_cast<LoadSDNode>(SrcOp);
4886 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4888 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4889 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4890 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4891 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4892 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4894 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4895 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4896 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4897 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4905 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4906 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4907 DAG.getNode(ISD::BIT_CONVERT, dl,
4911 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4914 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4915 SDValue V1 = SVOp->getOperand(0);
4916 SDValue V2 = SVOp->getOperand(1);
4917 DebugLoc dl = SVOp->getDebugLoc();
4918 EVT VT = SVOp->getValueType(0);
4920 SmallVector<std::pair<int, int>, 8> Locs;
4922 SmallVector<int, 8> Mask1(4U, -1);
4923 SmallVector<int, 8> PermMask;
4924 SVOp->getMask(PermMask);
4928 for (unsigned i = 0; i != 4; ++i) {
4929 int Idx = PermMask[i];
4931 Locs[i] = std::make_pair(-1, -1);
4933 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4935 Locs[i] = std::make_pair(0, NumLo);
4939 Locs[i] = std::make_pair(1, NumHi);
4941 Mask1[2+NumHi] = Idx;
4947 if (NumLo <= 2 && NumHi <= 2) {
4948 // If no more than two elements come from either vector. This can be
4949 // implemented with two shuffles. First shuffle gather the elements.
4950 // The second shuffle, which takes the first shuffle as both of its
4951 // vector operands, put the elements into the right order.
4952 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4954 SmallVector<int, 8> Mask2(4U, -1);
4956 for (unsigned i = 0; i != 4; ++i) {
4957 if (Locs[i].first == -1)
4960 unsigned Idx = (i < 2) ? 0 : 4;
4961 Idx += Locs[i].first * 2 + Locs[i].second;
4966 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4967 } else if (NumLo == 3 || NumHi == 3) {
4968 // Otherwise, we must have three elements from one vector, call it X, and
4969 // one element from the other, call it Y. First, use a shufps to build an
4970 // intermediate vector with the one element from Y and the element from X
4971 // that will be in the same half in the final destination (the indexes don't
4972 // matter). Then, use a shufps to build the final vector, taking the half
4973 // containing the element from Y from the intermediate, and the other half
4976 // Normalize it so the 3 elements come from V1.
4977 CommuteVectorShuffleMask(PermMask, VT);
4981 // Find the element from V2.
4983 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4984 int Val = PermMask[HiIndex];
4991 Mask1[0] = PermMask[HiIndex];
4993 Mask1[2] = PermMask[HiIndex^1];
4995 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4998 Mask1[0] = PermMask[0];
4999 Mask1[1] = PermMask[1];
5000 Mask1[2] = HiIndex & 1 ? 6 : 4;
5001 Mask1[3] = HiIndex & 1 ? 4 : 6;
5002 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5004 Mask1[0] = HiIndex & 1 ? 2 : 0;
5005 Mask1[1] = HiIndex & 1 ? 0 : 2;
5006 Mask1[2] = PermMask[2];
5007 Mask1[3] = PermMask[3];
5012 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
5016 // Break it into (shuffle shuffle_hi, shuffle_lo).
5018 SmallVector<int,8> LoMask(4U, -1);
5019 SmallVector<int,8> HiMask(4U, -1);
5021 SmallVector<int,8> *MaskPtr = &LoMask;
5022 unsigned MaskIdx = 0;
5025 for (unsigned i = 0; i != 4; ++i) {
5032 int Idx = PermMask[i];
5034 Locs[i] = std::make_pair(-1, -1);
5035 } else if (Idx < 4) {
5036 Locs[i] = std::make_pair(MaskIdx, LoIdx);
5037 (*MaskPtr)[LoIdx] = Idx;
5040 Locs[i] = std::make_pair(MaskIdx, HiIdx);
5041 (*MaskPtr)[HiIdx] = Idx;
5046 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
5047 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
5048 SmallVector<int, 8> MaskOps;
5049 for (unsigned i = 0; i != 4; ++i) {
5050 if (Locs[i].first == -1) {
5051 MaskOps.push_back(-1);
5053 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
5054 MaskOps.push_back(Idx);
5057 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
5061 SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
5063 SDValue V1 = Op.getOperand(0);
5064 SDValue V2 = Op.getOperand(1);
5065 EVT VT = Op.getValueType();
5067 assert(VT != MVT::v2i64 && "unsupported shuffle type");
5069 if (HasSSE2 && VT == MVT::v2f64)
5070 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
5073 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V2, DAG);
5077 SDValue getMOVHighToLow(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG) {
5078 SDValue V1 = Op.getOperand(0);
5079 SDValue V2 = Op.getOperand(1);
5080 EVT VT = Op.getValueType();
5082 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
5083 "unsupported shuffle type");
5085 if (V2.getOpcode() == ISD::UNDEF)
5089 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
5093 SDValue getMOVLP(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
5094 SDValue V1 = Op.getOperand(0);
5095 SDValue V2 = Op.getOperand(1);
5096 EVT VT = Op.getValueType();
5097 unsigned NumElems = VT.getVectorNumElements();
5099 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
5100 // operand of these instructions is only memory, so check if there's a
5101 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
5103 bool CanFoldLoad = false;
5107 // Trivial case, when V2 comes from a load.
5108 if (TmpV2.hasOneUse() && TmpV2.getOpcode() == ISD::BIT_CONVERT)
5109 TmpV2 = TmpV2.getOperand(0);
5110 if (TmpV2.hasOneUse() && TmpV2.getOpcode() == ISD::SCALAR_TO_VECTOR)
5111 TmpV2 = TmpV2.getOperand(0);
5112 if (MayFoldLoad(TmpV2))
5115 // When V1 is a load, it can be folded later into a store in isel, example:
5116 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
5118 // (MOVLPSmr addr:$src1, VR128:$src2)
5119 // So, recognize this potential and also use MOVLPS or MOVLPD
5120 if (TmpV1.hasOneUse() && TmpV1.getOpcode() == ISD::BIT_CONVERT)
5121 TmpV1 = TmpV1.getOperand(0);
5122 if (MayFoldLoad(TmpV1) && MayFoldIntoStore(Op))
5126 if (HasSSE2 && NumElems == 2)
5127 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
5130 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
5133 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5134 // movl and movlp will both match v2i64, but v2i64 is never matched by
5135 // movl earlier because we make it strict to avoid messing with the movlp load
5136 // folding logic (see the code above getMOVLP call). Match it here then,
5137 // this is horrible, but will stay like this until we move all shuffle
5138 // matching to x86 specific nodes. Note that for the 1st condition all
5139 // types are matched with movsd.
5140 if ((HasSSE2 && NumElems == 2) || !X86::isMOVLMask(SVOp))
5141 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5143 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5146 assert(VT != MVT::v4i32 && "unsupported shuffle type");
5148 // Invert the operand order and use SHUFPS to match it.
5149 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V2, V1,
5150 X86::getShuffleSHUFImmediate(SVOp), DAG);
5153 static inline unsigned getUNPCKLOpcode(EVT VT) {
5154 switch(VT.getSimpleVT().SimpleTy) {
5155 case MVT::v4i32: return X86ISD::PUNPCKLDQ;
5156 case MVT::v2i64: return X86ISD::PUNPCKLQDQ;
5157 case MVT::v4f32: return X86ISD::UNPCKLPS;
5158 case MVT::v2f64: return X86ISD::UNPCKLPD;
5159 case MVT::v16i8: return X86ISD::PUNPCKLBW;
5160 case MVT::v8i16: return X86ISD::PUNPCKLWD;
5162 llvm_unreachable("Unknow type for unpckl");
5167 static inline unsigned getUNPCKHOpcode(EVT VT) {
5168 switch(VT.getSimpleVT().SimpleTy) {
5169 case MVT::v4i32: return X86ISD::PUNPCKHDQ;
5170 case MVT::v2i64: return X86ISD::PUNPCKHQDQ;
5171 case MVT::v4f32: return X86ISD::UNPCKHPS;
5172 case MVT::v2f64: return X86ISD::UNPCKHPD;
5173 case MVT::v16i8: return X86ISD::PUNPCKHBW;
5174 case MVT::v8i16: return X86ISD::PUNPCKHWD;
5176 llvm_unreachable("Unknow type for unpckh");
5182 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
5183 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5184 SDValue V1 = Op.getOperand(0);
5185 SDValue V2 = Op.getOperand(1);
5186 EVT VT = Op.getValueType();
5187 DebugLoc dl = Op.getDebugLoc();
5188 unsigned NumElems = VT.getVectorNumElements();
5189 bool isMMX = VT.getSizeInBits() == 64;
5190 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
5191 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
5192 bool V1IsSplat = false;
5193 bool V2IsSplat = false;
5194 bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
5195 bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
5196 MachineFunction &MF = DAG.getMachineFunction();
5197 bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
5199 if (isZeroShuffle(SVOp))
5200 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
5202 // Promote splats to v4f32.
5203 if (SVOp->isSplat()) {
5204 if (isMMX || NumElems < 4)
5206 return PromoteSplat(SVOp, DAG);
5209 // If the shuffle can be profitably rewritten as a narrower shuffle, then
5211 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
5212 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
5213 if (NewOp.getNode())
5214 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5215 LowerVECTOR_SHUFFLE(NewOp, DAG));
5216 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
5217 // FIXME: Figure out a cleaner way to do this.
5218 // Try to make use of movq to zero out the top part.
5219 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
5220 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
5221 if (NewOp.getNode()) {
5222 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
5223 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
5224 DAG, Subtarget, dl);
5226 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
5227 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
5228 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
5229 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
5230 DAG, Subtarget, dl);
5234 if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp)) {
5235 // NOTE: isPSHUFDMask can also match this mask, if speed is more
5236 // important than size here, this will be matched by pshufd
5237 if (VT == MVT::v4f32)
5238 return getTargetShuffleNode(X86ISD::UNPCKLPS, dl, VT, V1, V1, DAG);
5239 if (HasSSE2 && VT == MVT::v16i8)
5240 return getTargetShuffleNode(X86ISD::PUNPCKLBW, dl, VT, V1, V1, DAG);
5241 if (HasSSE2 && VT == MVT::v8i16)
5242 return getTargetShuffleNode(X86ISD::PUNPCKLWD, dl, VT, V1, V1, DAG);
5243 if (HasSSE2 && VT == MVT::v4i32)
5244 return getTargetShuffleNode(X86ISD::PUNPCKLDQ, dl, VT, V1, V1, DAG);
5247 if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp)) {
5248 // NOTE: isPSHUFDMask can also match this mask, if speed is more
5249 // important than size here, this will be matched by pshufd
5250 if (VT == MVT::v4f32)
5251 return getTargetShuffleNode(X86ISD::UNPCKHPS, dl, VT, V1, V1, DAG);
5252 if (HasSSE2 && VT == MVT::v16i8)
5253 return getTargetShuffleNode(X86ISD::PUNPCKHBW, dl, VT, V1, V1, DAG);
5254 if (HasSSE2 && VT == MVT::v8i16)
5255 return getTargetShuffleNode(X86ISD::PUNPCKHWD, dl, VT, V1, V1, DAG);
5256 if (HasSSE2 && VT == MVT::v4i32)
5257 return getTargetShuffleNode(X86ISD::PUNPCKHDQ, dl, VT, V1, V1, DAG);
5260 if (X86::isPSHUFDMask(SVOp)) {
5261 // The actual implementation will match the mask in the if above and then
5262 // during isel it can match several different instructions, not only pshufd
5263 // as its name says, sad but true, emulate the behavior for now...
5264 if (X86::isMOVDDUPMask(SVOp) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
5265 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
5267 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5269 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
5270 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
5272 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5273 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V1,
5276 if (VT == MVT::v4f32)
5277 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V1,
5281 // Check if this can be converted into a logical shift.
5282 bool isLeft = false;
5285 bool isShift = getSubtarget()->hasSSE2() &&
5286 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
5287 if (isShift && ShVal.hasOneUse()) {
5288 // If the shifted value has multiple uses, it may be cheaper to use
5289 // v_set0 + movlhps or movhlps, etc.
5290 EVT EltVT = VT.getVectorElementType();
5291 ShAmt *= EltVT.getSizeInBits();
5292 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5295 if (X86::isMOVLMask(SVOp)) {
5298 if (ISD::isBuildVectorAllZeros(V1.getNode()))
5299 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
5300 if (!isMMX && !X86::isMOVLPMask(SVOp)) {
5301 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5302 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5304 if (VT == MVT::v4i32 || VT == MVT::v4f32)
5305 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5309 // FIXME: fold these into legal mask.
5311 if (X86::isMOVLHPSMask(SVOp) && !X86::isUNPCKLMask(SVOp))
5312 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
5314 if (X86::isMOVHLPSMask(SVOp))
5315 return getMOVHighToLow(Op, dl, DAG);
5317 if (X86::isMOVSHDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5318 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
5320 if (X86::isMOVSLDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5321 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
5323 if (X86::isMOVLPMask(SVOp))
5324 return getMOVLP(Op, dl, DAG, HasSSE2);
5327 if (ShouldXformToMOVHLPS(SVOp) ||
5328 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
5329 return CommuteVectorShuffle(SVOp, DAG);
5332 // No better options. Use a vshl / vsrl.
5333 EVT EltVT = VT.getVectorElementType();
5334 ShAmt *= EltVT.getSizeInBits();
5335 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5338 bool Commuted = false;
5339 // FIXME: This should also accept a bitcast of a splat? Be careful, not
5340 // 1,1,1,1 -> v8i16 though.
5341 V1IsSplat = isSplatVector(V1.getNode());
5342 V2IsSplat = isSplatVector(V2.getNode());
5344 // Canonicalize the splat or undef, if present, to be on the RHS.
5345 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
5346 Op = CommuteVectorShuffle(SVOp, DAG);
5347 SVOp = cast<ShuffleVectorSDNode>(Op);
5348 V1 = SVOp->getOperand(0);
5349 V2 = SVOp->getOperand(1);
5350 std::swap(V1IsSplat, V2IsSplat);
5351 std::swap(V1IsUndef, V2IsUndef);
5355 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
5356 // Shuffling low element of v1 into undef, just return v1.
5359 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
5360 // the instruction selector will not match, so get a canonical MOVL with
5361 // swapped operands to undo the commute.
5362 return getMOVL(DAG, dl, VT, V2, V1);
5365 if (X86::isUNPCKLMask(SVOp) ||
5366 X86::isUNPCKHMask(SVOp))
5370 // Normalize mask so all entries that point to V2 points to its first
5371 // element then try to match unpck{h|l} again. If match, return a
5372 // new vector_shuffle with the corrected mask.
5373 SDValue NewMask = NormalizeMask(SVOp, DAG);
5374 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
5375 if (NSVOp != SVOp) {
5376 if (X86::isUNPCKLMask(NSVOp, true)) {
5378 } else if (X86::isUNPCKHMask(NSVOp, true)) {
5385 // Commute is back and try unpck* again.
5386 // FIXME: this seems wrong.
5387 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
5388 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
5389 if (X86::isUNPCKLMask(NewSVOp) ||
5390 X86::isUNPCKHMask(NewSVOp))
5394 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
5396 // Normalize the node to match x86 shuffle ops if needed
5397 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
5398 return CommuteVectorShuffle(SVOp, DAG);
5400 // Check for legal shuffle and return?
5401 SmallVector<int, 16> PermMask;
5402 SVOp->getMask(PermMask);
5403 if (isShuffleMaskLegal(PermMask, VT))
5406 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
5407 if (VT == MVT::v8i16) {
5408 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
5409 if (NewOp.getNode())
5413 if (VT == MVT::v16i8) {
5414 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
5415 if (NewOp.getNode())
5419 // Handle all 4 wide cases with a number of shuffles except for MMX.
5420 if (NumElems == 4 && !isMMX)
5421 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
5427 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
5428 SelectionDAG &DAG) const {
5429 EVT VT = Op.getValueType();
5430 DebugLoc dl = Op.getDebugLoc();
5431 if (VT.getSizeInBits() == 8) {
5432 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
5433 Op.getOperand(0), Op.getOperand(1));
5434 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5435 DAG.getValueType(VT));
5436 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5437 } else if (VT.getSizeInBits() == 16) {
5438 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5439 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
5441 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5442 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5443 DAG.getNode(ISD::BIT_CONVERT, dl,
5447 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
5448 Op.getOperand(0), Op.getOperand(1));
5449 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5450 DAG.getValueType(VT));
5451 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5452 } else if (VT == MVT::f32) {
5453 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
5454 // the result back to FR32 register. It's only worth matching if the
5455 // result has a single use which is a store or a bitcast to i32. And in
5456 // the case of a store, it's not worth it if the index is a constant 0,
5457 // because a MOVSSmr can be used instead, which is smaller and faster.
5458 if (!Op.hasOneUse())
5460 SDNode *User = *Op.getNode()->use_begin();
5461 if ((User->getOpcode() != ISD::STORE ||
5462 (isa<ConstantSDNode>(Op.getOperand(1)) &&
5463 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
5464 (User->getOpcode() != ISD::BIT_CONVERT ||
5465 User->getValueType(0) != MVT::i32))
5467 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5468 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
5471 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
5472 } else if (VT == MVT::i32) {
5473 // ExtractPS works with constant index.
5474 if (isa<ConstantSDNode>(Op.getOperand(1)))
5482 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5483 SelectionDAG &DAG) const {
5484 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5487 if (Subtarget->hasSSE41()) {
5488 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
5493 EVT VT = Op.getValueType();
5494 DebugLoc dl = Op.getDebugLoc();
5495 // TODO: handle v16i8.
5496 if (VT.getSizeInBits() == 16) {
5497 SDValue Vec = Op.getOperand(0);
5498 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5500 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5501 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5502 DAG.getNode(ISD::BIT_CONVERT, dl,
5505 // Transform it so it match pextrw which produces a 32-bit result.
5506 EVT EltVT = MVT::i32;
5507 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
5508 Op.getOperand(0), Op.getOperand(1));
5509 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
5510 DAG.getValueType(VT));
5511 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5512 } else if (VT.getSizeInBits() == 32) {
5513 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5517 // SHUFPS the element to the lowest double word, then movss.
5518 int Mask[4] = { Idx, -1, -1, -1 };
5519 EVT VVT = Op.getOperand(0).getValueType();
5520 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5521 DAG.getUNDEF(VVT), Mask);
5522 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5523 DAG.getIntPtrConstant(0));
5524 } else if (VT.getSizeInBits() == 64) {
5525 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
5526 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
5527 // to match extract_elt for f64.
5528 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5532 // UNPCKHPD the element to the lowest double word, then movsd.
5533 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
5534 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
5535 int Mask[2] = { 1, -1 };
5536 EVT VVT = Op.getOperand(0).getValueType();
5537 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5538 DAG.getUNDEF(VVT), Mask);
5539 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5540 DAG.getIntPtrConstant(0));
5547 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
5548 SelectionDAG &DAG) const {
5549 EVT VT = Op.getValueType();
5550 EVT EltVT = VT.getVectorElementType();
5551 DebugLoc dl = Op.getDebugLoc();
5553 SDValue N0 = Op.getOperand(0);
5554 SDValue N1 = Op.getOperand(1);
5555 SDValue N2 = Op.getOperand(2);
5557 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
5558 isa<ConstantSDNode>(N2)) {
5560 if (VT == MVT::v8i16)
5561 Opc = X86ISD::PINSRW;
5562 else if (VT == MVT::v4i16)
5563 Opc = X86ISD::MMX_PINSRW;
5564 else if (VT == MVT::v16i8)
5565 Opc = X86ISD::PINSRB;
5567 Opc = X86ISD::PINSRB;
5569 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5571 if (N1.getValueType() != MVT::i32)
5572 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5573 if (N2.getValueType() != MVT::i32)
5574 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5575 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5576 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5577 // Bits [7:6] of the constant are the source select. This will always be
5578 // zero here. The DAG Combiner may combine an extract_elt index into these
5579 // bits. For example (insert (extract, 3), 2) could be matched by putting
5580 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5581 // Bits [5:4] of the constant are the destination select. This is the
5582 // value of the incoming immediate.
5583 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5584 // combine either bitwise AND or insert of float 0.0 to set these bits.
5585 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5586 // Create this as a scalar to vector..
5587 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5588 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5589 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5590 // PINSR* works with constant index.
5597 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5598 EVT VT = Op.getValueType();
5599 EVT EltVT = VT.getVectorElementType();
5601 if (Subtarget->hasSSE41())
5602 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5604 if (EltVT == MVT::i8)
5607 DebugLoc dl = Op.getDebugLoc();
5608 SDValue N0 = Op.getOperand(0);
5609 SDValue N1 = Op.getOperand(1);
5610 SDValue N2 = Op.getOperand(2);
5612 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5613 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5614 // as its second argument.
5615 if (N1.getValueType() != MVT::i32)
5616 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5617 if (N2.getValueType() != MVT::i32)
5618 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5619 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5620 dl, VT, N0, N1, N2);
5626 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5627 DebugLoc dl = Op.getDebugLoc();
5629 if (Op.getValueType() == MVT::v1i64 &&
5630 Op.getOperand(0).getValueType() == MVT::i64)
5631 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5633 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5634 EVT VT = MVT::v2i32;
5635 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5642 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5643 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5646 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5647 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5648 // one of the above mentioned nodes. It has to be wrapped because otherwise
5649 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5650 // be used to form addressing mode. These wrapped nodes will be selected
5653 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5654 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5656 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5658 unsigned char OpFlag = 0;
5659 unsigned WrapperKind = X86ISD::Wrapper;
5660 CodeModel::Model M = getTargetMachine().getCodeModel();
5662 if (Subtarget->isPICStyleRIPRel() &&
5663 (M == CodeModel::Small || M == CodeModel::Kernel))
5664 WrapperKind = X86ISD::WrapperRIP;
5665 else if (Subtarget->isPICStyleGOT())
5666 OpFlag = X86II::MO_GOTOFF;
5667 else if (Subtarget->isPICStyleStubPIC())
5668 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5670 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5672 CP->getOffset(), OpFlag);
5673 DebugLoc DL = CP->getDebugLoc();
5674 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5675 // With PIC, the address is actually $g + Offset.
5677 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5678 DAG.getNode(X86ISD::GlobalBaseReg,
5679 DebugLoc(), getPointerTy()),
5686 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5687 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5689 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5691 unsigned char OpFlag = 0;
5692 unsigned WrapperKind = X86ISD::Wrapper;
5693 CodeModel::Model M = getTargetMachine().getCodeModel();
5695 if (Subtarget->isPICStyleRIPRel() &&
5696 (M == CodeModel::Small || M == CodeModel::Kernel))
5697 WrapperKind = X86ISD::WrapperRIP;
5698 else if (Subtarget->isPICStyleGOT())
5699 OpFlag = X86II::MO_GOTOFF;
5700 else if (Subtarget->isPICStyleStubPIC())
5701 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5703 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5705 DebugLoc DL = JT->getDebugLoc();
5706 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5708 // With PIC, the address is actually $g + Offset.
5710 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5711 DAG.getNode(X86ISD::GlobalBaseReg,
5712 DebugLoc(), getPointerTy()),
5720 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5721 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5723 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5725 unsigned char OpFlag = 0;
5726 unsigned WrapperKind = X86ISD::Wrapper;
5727 CodeModel::Model M = getTargetMachine().getCodeModel();
5729 if (Subtarget->isPICStyleRIPRel() &&
5730 (M == CodeModel::Small || M == CodeModel::Kernel))
5731 WrapperKind = X86ISD::WrapperRIP;
5732 else if (Subtarget->isPICStyleGOT())
5733 OpFlag = X86II::MO_GOTOFF;
5734 else if (Subtarget->isPICStyleStubPIC())
5735 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5737 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5739 DebugLoc DL = Op.getDebugLoc();
5740 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5743 // With PIC, the address is actually $g + Offset.
5744 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5745 !Subtarget->is64Bit()) {
5746 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5747 DAG.getNode(X86ISD::GlobalBaseReg,
5748 DebugLoc(), getPointerTy()),
5756 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5757 // Create the TargetBlockAddressAddress node.
5758 unsigned char OpFlags =
5759 Subtarget->ClassifyBlockAddressReference();
5760 CodeModel::Model M = getTargetMachine().getCodeModel();
5761 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5762 DebugLoc dl = Op.getDebugLoc();
5763 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5764 /*isTarget=*/true, OpFlags);
5766 if (Subtarget->isPICStyleRIPRel() &&
5767 (M == CodeModel::Small || M == CodeModel::Kernel))
5768 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5770 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5772 // With PIC, the address is actually $g + Offset.
5773 if (isGlobalRelativeToPICBase(OpFlags)) {
5774 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5775 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5783 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5785 SelectionDAG &DAG) const {
5786 // Create the TargetGlobalAddress node, folding in the constant
5787 // offset if it is legal.
5788 unsigned char OpFlags =
5789 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5790 CodeModel::Model M = getTargetMachine().getCodeModel();
5792 if (OpFlags == X86II::MO_NO_FLAG &&
5793 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5794 // A direct static reference to a global.
5795 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
5798 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
5801 if (Subtarget->isPICStyleRIPRel() &&
5802 (M == CodeModel::Small || M == CodeModel::Kernel))
5803 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5805 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5807 // With PIC, the address is actually $g + Offset.
5808 if (isGlobalRelativeToPICBase(OpFlags)) {
5809 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5810 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5814 // For globals that require a load from a stub to get the address, emit the
5816 if (isGlobalStubReference(OpFlags))
5817 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5818 PseudoSourceValue::getGOT(), 0, false, false, 0);
5820 // If there was a non-zero offset that we didn't fold, create an explicit
5823 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5824 DAG.getConstant(Offset, getPointerTy()));
5830 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
5831 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5832 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5833 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5837 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5838 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5839 unsigned char OperandFlags) {
5840 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5841 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5842 DebugLoc dl = GA->getDebugLoc();
5843 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5844 GA->getValueType(0),
5848 SDValue Ops[] = { Chain, TGA, *InFlag };
5849 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5851 SDValue Ops[] = { Chain, TGA };
5852 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5855 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5856 MFI->setAdjustsStack(true);
5858 SDValue Flag = Chain.getValue(1);
5859 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5862 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5864 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5867 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5868 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5869 DAG.getNode(X86ISD::GlobalBaseReg,
5870 DebugLoc(), PtrVT), InFlag);
5871 InFlag = Chain.getValue(1);
5873 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5876 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5878 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5880 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5881 X86::RAX, X86II::MO_TLSGD);
5884 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5885 // "local exec" model.
5886 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5887 const EVT PtrVT, TLSModel::Model model,
5889 DebugLoc dl = GA->getDebugLoc();
5890 // Get the Thread Pointer
5891 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5893 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5896 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5897 NULL, 0, false, false, 0);
5899 unsigned char OperandFlags = 0;
5900 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5902 unsigned WrapperKind = X86ISD::Wrapper;
5903 if (model == TLSModel::LocalExec) {
5904 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5905 } else if (is64Bit) {
5906 assert(model == TLSModel::InitialExec);
5907 OperandFlags = X86II::MO_GOTTPOFF;
5908 WrapperKind = X86ISD::WrapperRIP;
5910 assert(model == TLSModel::InitialExec);
5911 OperandFlags = X86II::MO_INDNTPOFF;
5914 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5916 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5917 GA->getValueType(0),
5918 GA->getOffset(), OperandFlags);
5919 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5921 if (model == TLSModel::InitialExec)
5922 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5923 PseudoSourceValue::getGOT(), 0, false, false, 0);
5925 // The address of the thread local variable is the add of the thread
5926 // pointer with the offset of the variable.
5927 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5931 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
5933 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5934 const GlobalValue *GV = GA->getGlobal();
5936 if (Subtarget->isTargetELF()) {
5937 // TODO: implement the "local dynamic" model
5938 // TODO: implement the "initial exec"model for pic executables
5940 // If GV is an alias then use the aliasee for determining
5941 // thread-localness.
5942 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5943 GV = GA->resolveAliasedGlobal(false);
5945 TLSModel::Model model
5946 = getTLSModel(GV, getTargetMachine().getRelocationModel());
5949 case TLSModel::GeneralDynamic:
5950 case TLSModel::LocalDynamic: // not implemented
5951 if (Subtarget->is64Bit())
5952 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5953 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5955 case TLSModel::InitialExec:
5956 case TLSModel::LocalExec:
5957 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5958 Subtarget->is64Bit());
5960 } else if (Subtarget->isTargetDarwin()) {
5961 // Darwin only has one model of TLS. Lower to that.
5962 unsigned char OpFlag = 0;
5963 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
5964 X86ISD::WrapperRIP : X86ISD::Wrapper;
5966 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5968 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
5969 !Subtarget->is64Bit();
5971 OpFlag = X86II::MO_TLVP_PIC_BASE;
5973 OpFlag = X86II::MO_TLVP;
5974 DebugLoc DL = Op.getDebugLoc();
5975 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
5977 GA->getOffset(), OpFlag);
5978 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5980 // With PIC32, the address is actually $g + Offset.
5982 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5983 DAG.getNode(X86ISD::GlobalBaseReg,
5984 DebugLoc(), getPointerTy()),
5987 // Lowering the machine isd will make sure everything is in the right
5989 SDValue Args[] = { Offset };
5990 SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
5992 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
5993 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5994 MFI->setAdjustsStack(true);
5996 // And our return value (tls address) is in the standard call return value
5998 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
5999 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
6003 "TLS not implemented for this target.");
6005 llvm_unreachable("Unreachable");
6010 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
6011 /// take a 2 x i32 value to shift plus a shift amount.
6012 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
6013 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
6014 EVT VT = Op.getValueType();
6015 unsigned VTBits = VT.getSizeInBits();
6016 DebugLoc dl = Op.getDebugLoc();
6017 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
6018 SDValue ShOpLo = Op.getOperand(0);
6019 SDValue ShOpHi = Op.getOperand(1);
6020 SDValue ShAmt = Op.getOperand(2);
6021 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
6022 DAG.getConstant(VTBits - 1, MVT::i8))
6023 : DAG.getConstant(0, VT);
6026 if (Op.getOpcode() == ISD::SHL_PARTS) {
6027 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
6028 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
6030 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
6031 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
6034 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
6035 DAG.getConstant(VTBits, MVT::i8));
6036 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6037 AndNode, DAG.getConstant(0, MVT::i8));
6040 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6041 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
6042 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
6044 if (Op.getOpcode() == ISD::SHL_PARTS) {
6045 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6046 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6048 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6049 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6052 SDValue Ops[2] = { Lo, Hi };
6053 return DAG.getMergeValues(Ops, 2, dl);
6056 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
6057 SelectionDAG &DAG) const {
6058 EVT SrcVT = Op.getOperand(0).getValueType();
6060 if (SrcVT.isVector()) {
6061 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
6067 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
6068 "Unknown SINT_TO_FP to lower!");
6070 // These are really Legal; return the operand so the caller accepts it as
6072 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
6074 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
6075 Subtarget->is64Bit()) {
6079 DebugLoc dl = Op.getDebugLoc();
6080 unsigned Size = SrcVT.getSizeInBits()/8;
6081 MachineFunction &MF = DAG.getMachineFunction();
6082 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
6083 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6084 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6086 PseudoSourceValue::getFixedStack(SSFI), 0,
6088 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
6091 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
6093 SelectionDAG &DAG) const {
6095 DebugLoc dl = Op.getDebugLoc();
6097 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
6099 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
6101 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
6102 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
6103 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
6104 Tys, Ops, array_lengthof(Ops));
6107 Chain = Result.getValue(1);
6108 SDValue InFlag = Result.getValue(2);
6110 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
6111 // shouldn't be necessary except that RFP cannot be live across
6112 // multiple blocks. When stackifier is fixed, they can be uncoupled.
6113 MachineFunction &MF = DAG.getMachineFunction();
6114 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
6115 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6116 Tys = DAG.getVTList(MVT::Other);
6118 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
6120 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
6121 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
6122 PseudoSourceValue::getFixedStack(SSFI), 0,
6129 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
6130 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
6131 SelectionDAG &DAG) const {
6132 // This algorithm is not obvious. Here it is in C code, more or less:
6134 double uint64_to_double( uint32_t hi, uint32_t lo ) {
6135 static const __m128i exp = { 0x4330000045300000ULL, 0 };
6136 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
6138 // Copy ints to xmm registers.
6139 __m128i xh = _mm_cvtsi32_si128( hi );
6140 __m128i xl = _mm_cvtsi32_si128( lo );
6142 // Combine into low half of a single xmm register.
6143 __m128i x = _mm_unpacklo_epi32( xh, xl );
6147 // Merge in appropriate exponents to give the integer bits the right
6149 x = _mm_unpacklo_epi32( x, exp );
6151 // Subtract away the biases to deal with the IEEE-754 double precision
6153 d = _mm_sub_pd( (__m128d) x, bias );
6155 // All conversions up to here are exact. The correctly rounded result is
6156 // calculated using the current rounding mode using the following
6158 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
6159 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
6160 // store doesn't really need to be here (except
6161 // maybe to zero the other double)
6166 DebugLoc dl = Op.getDebugLoc();
6167 LLVMContext *Context = DAG.getContext();
6169 // Build some magic constants.
6170 std::vector<Constant*> CV0;
6171 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
6172 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
6173 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6174 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6175 Constant *C0 = ConstantVector::get(CV0);
6176 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
6178 std::vector<Constant*> CV1;
6180 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
6182 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
6183 Constant *C1 = ConstantVector::get(CV1);
6184 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
6186 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6187 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6189 DAG.getIntPtrConstant(1)));
6190 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6191 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6193 DAG.getIntPtrConstant(0)));
6194 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
6195 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
6196 PseudoSourceValue::getConstantPool(), 0,
6198 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
6199 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
6200 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
6201 PseudoSourceValue::getConstantPool(), 0,
6203 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
6205 // Add the halves; easiest way is to swap them into another reg first.
6206 int ShufMask[2] = { 1, -1 };
6207 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
6208 DAG.getUNDEF(MVT::v2f64), ShufMask);
6209 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
6210 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
6211 DAG.getIntPtrConstant(0));
6214 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
6215 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
6216 SelectionDAG &DAG) const {
6217 DebugLoc dl = Op.getDebugLoc();
6218 // FP constant to bias correct the final result.
6219 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
6222 // Load the 32-bit value into an XMM register.
6223 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6224 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6226 DAG.getIntPtrConstant(0)));
6228 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6229 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
6230 DAG.getIntPtrConstant(0));
6232 // Or the load with the bias.
6233 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
6234 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6235 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6237 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6238 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6239 MVT::v2f64, Bias)));
6240 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6241 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
6242 DAG.getIntPtrConstant(0));
6244 // Subtract the bias.
6245 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
6247 // Handle final rounding.
6248 EVT DestVT = Op.getValueType();
6250 if (DestVT.bitsLT(MVT::f64)) {
6251 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
6252 DAG.getIntPtrConstant(0));
6253 } else if (DestVT.bitsGT(MVT::f64)) {
6254 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
6257 // Handle final rounding.
6261 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
6262 SelectionDAG &DAG) const {
6263 SDValue N0 = Op.getOperand(0);
6264 DebugLoc dl = Op.getDebugLoc();
6266 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
6267 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
6268 // the optimization here.
6269 if (DAG.SignBitIsZero(N0))
6270 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
6272 EVT SrcVT = N0.getValueType();
6273 EVT DstVT = Op.getValueType();
6274 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
6275 return LowerUINT_TO_FP_i64(Op, DAG);
6276 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
6277 return LowerUINT_TO_FP_i32(Op, DAG);
6279 // Make a 64-bit buffer, and use it to build an FILD.
6280 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
6281 if (SrcVT == MVT::i32) {
6282 SDValue WordOff = DAG.getConstant(4, getPointerTy());
6283 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
6284 getPointerTy(), StackSlot, WordOff);
6285 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6286 StackSlot, NULL, 0, false, false, 0);
6287 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
6288 OffsetSlot, NULL, 0, false, false, 0);
6289 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
6293 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
6294 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6295 StackSlot, NULL, 0, false, false, 0);
6296 // For i64 source, we need to add the appropriate power of 2 if the input
6297 // was negative. This is the same as the optimization in
6298 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
6299 // we must be careful to do the computation in x87 extended precision, not
6300 // in SSE. (The generic code can't know it's OK to do this, or how to.)
6301 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
6302 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
6303 SDValue Fild = DAG.getNode(X86ISD::FILD, dl, Tys, Ops, 3);
6305 APInt FF(32, 0x5F800000ULL);
6307 // Check whether the sign bit is set.
6308 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
6309 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
6312 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
6313 SDValue FudgePtr = DAG.getConstantPool(
6314 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
6317 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
6318 SDValue Zero = DAG.getIntPtrConstant(0);
6319 SDValue Four = DAG.getIntPtrConstant(4);
6320 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
6322 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
6324 // Load the value out, extending it from f32 to f80.
6325 // FIXME: Avoid the extend by constructing the right constant pool?
6326 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, MVT::f80, dl, DAG.getEntryNode(),
6327 FudgePtr, PseudoSourceValue::getConstantPool(),
6328 0, MVT::f32, false, false, 4);
6329 // Extend everything to 80 bits to force it to be done on x87.
6330 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
6331 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
6334 std::pair<SDValue,SDValue> X86TargetLowering::
6335 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
6336 DebugLoc dl = Op.getDebugLoc();
6338 EVT DstTy = Op.getValueType();
6341 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
6345 assert(DstTy.getSimpleVT() <= MVT::i64 &&
6346 DstTy.getSimpleVT() >= MVT::i16 &&
6347 "Unknown FP_TO_SINT to lower!");
6349 // These are really Legal.
6350 if (DstTy == MVT::i32 &&
6351 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
6352 return std::make_pair(SDValue(), SDValue());
6353 if (Subtarget->is64Bit() &&
6354 DstTy == MVT::i64 &&
6355 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
6356 return std::make_pair(SDValue(), SDValue());
6358 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
6360 MachineFunction &MF = DAG.getMachineFunction();
6361 unsigned MemSize = DstTy.getSizeInBits()/8;
6362 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
6363 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6366 switch (DstTy.getSimpleVT().SimpleTy) {
6367 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
6368 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
6369 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
6370 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
6373 SDValue Chain = DAG.getEntryNode();
6374 SDValue Value = Op.getOperand(0);
6375 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
6376 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
6377 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
6378 PseudoSourceValue::getFixedStack(SSFI), 0,
6380 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
6382 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
6384 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
6385 Chain = Value.getValue(1);
6386 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
6387 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6390 // Build the FP_TO_INT*_IN_MEM
6391 SDValue Ops[] = { Chain, Value, StackSlot };
6392 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
6394 return std::make_pair(FIST, StackSlot);
6397 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
6398 SelectionDAG &DAG) const {
6399 if (Op.getValueType().isVector()) {
6400 if (Op.getValueType() == MVT::v2i32 &&
6401 Op.getOperand(0).getValueType() == MVT::v2f64) {
6407 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
6408 SDValue FIST = Vals.first, StackSlot = Vals.second;
6409 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
6410 if (FIST.getNode() == 0) return Op;
6413 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
6414 FIST, StackSlot, NULL, 0, false, false, 0);
6417 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
6418 SelectionDAG &DAG) const {
6419 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
6420 SDValue FIST = Vals.first, StackSlot = Vals.second;
6421 assert(FIST.getNode() && "Unexpected failure");
6424 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
6425 FIST, StackSlot, NULL, 0, false, false, 0);
6428 SDValue X86TargetLowering::LowerFABS(SDValue Op,
6429 SelectionDAG &DAG) const {
6430 LLVMContext *Context = DAG.getContext();
6431 DebugLoc dl = Op.getDebugLoc();
6432 EVT VT = Op.getValueType();
6435 EltVT = VT.getVectorElementType();
6436 std::vector<Constant*> CV;
6437 if (EltVT == MVT::f64) {
6438 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
6442 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
6448 Constant *C = ConstantVector::get(CV);
6449 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6450 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6451 PseudoSourceValue::getConstantPool(), 0,
6453 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
6456 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
6457 LLVMContext *Context = DAG.getContext();
6458 DebugLoc dl = Op.getDebugLoc();
6459 EVT VT = Op.getValueType();
6462 EltVT = VT.getVectorElementType();
6463 std::vector<Constant*> CV;
6464 if (EltVT == MVT::f64) {
6465 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
6469 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
6475 Constant *C = ConstantVector::get(CV);
6476 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6477 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6478 PseudoSourceValue::getConstantPool(), 0,
6480 if (VT.isVector()) {
6481 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6482 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
6483 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6485 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
6487 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
6491 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
6492 LLVMContext *Context = DAG.getContext();
6493 SDValue Op0 = Op.getOperand(0);
6494 SDValue Op1 = Op.getOperand(1);
6495 DebugLoc dl = Op.getDebugLoc();
6496 EVT VT = Op.getValueType();
6497 EVT SrcVT = Op1.getValueType();
6499 // If second operand is smaller, extend it first.
6500 if (SrcVT.bitsLT(VT)) {
6501 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
6504 // And if it is bigger, shrink it first.
6505 if (SrcVT.bitsGT(VT)) {
6506 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
6510 // At this point the operands and the result should have the same
6511 // type, and that won't be f80 since that is not custom lowered.
6513 // First get the sign bit of second operand.
6514 std::vector<Constant*> CV;
6515 if (SrcVT == MVT::f64) {
6516 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
6517 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6519 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
6520 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6521 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6522 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6524 Constant *C = ConstantVector::get(CV);
6525 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6526 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
6527 PseudoSourceValue::getConstantPool(), 0,
6529 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
6531 // Shift sign bit right or left if the two operands have different types.
6532 if (SrcVT.bitsGT(VT)) {
6533 // Op0 is MVT::f32, Op1 is MVT::f64.
6534 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
6535 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
6536 DAG.getConstant(32, MVT::i32));
6537 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
6538 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
6539 DAG.getIntPtrConstant(0));
6542 // Clear first operand sign bit.
6544 if (VT == MVT::f64) {
6545 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
6546 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6548 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
6549 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6550 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6551 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6553 C = ConstantVector::get(CV);
6554 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6555 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6556 PseudoSourceValue::getConstantPool(), 0,
6558 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
6560 // Or the value with the sign bit.
6561 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
6564 /// Emit nodes that will be selected as "test Op0,Op0", or something
6566 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
6567 SelectionDAG &DAG) const {
6568 DebugLoc dl = Op.getDebugLoc();
6570 // CF and OF aren't always set the way we want. Determine which
6571 // of these we need.
6572 bool NeedCF = false;
6573 bool NeedOF = false;
6576 case X86::COND_A: case X86::COND_AE:
6577 case X86::COND_B: case X86::COND_BE:
6580 case X86::COND_G: case X86::COND_GE:
6581 case X86::COND_L: case X86::COND_LE:
6582 case X86::COND_O: case X86::COND_NO:
6587 // See if we can use the EFLAGS value from the operand instead of
6588 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
6589 // we prove that the arithmetic won't overflow, we can't use OF or CF.
6590 if (Op.getResNo() != 0 || NeedOF || NeedCF)
6591 // Emit a CMP with 0, which is the TEST pattern.
6592 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6593 DAG.getConstant(0, Op.getValueType()));
6595 unsigned Opcode = 0;
6596 unsigned NumOperands = 0;
6597 switch (Op.getNode()->getOpcode()) {
6599 // Due to an isel shortcoming, be conservative if this add is likely to be
6600 // selected as part of a load-modify-store instruction. When the root node
6601 // in a match is a store, isel doesn't know how to remap non-chain non-flag
6602 // uses of other nodes in the match, such as the ADD in this case. This
6603 // leads to the ADD being left around and reselected, with the result being
6604 // two adds in the output. Alas, even if none our users are stores, that
6605 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
6606 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
6607 // climbing the DAG back to the root, and it doesn't seem to be worth the
6609 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6610 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6611 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
6614 if (ConstantSDNode *C =
6615 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
6616 // An add of one will be selected as an INC.
6617 if (C->getAPIntValue() == 1) {
6618 Opcode = X86ISD::INC;
6623 // An add of negative one (subtract of one) will be selected as a DEC.
6624 if (C->getAPIntValue().isAllOnesValue()) {
6625 Opcode = X86ISD::DEC;
6631 // Otherwise use a regular EFLAGS-setting add.
6632 Opcode = X86ISD::ADD;
6636 // If the primary and result isn't used, don't bother using X86ISD::AND,
6637 // because a TEST instruction will be better.
6638 bool NonFlagUse = false;
6639 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6640 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6642 unsigned UOpNo = UI.getOperandNo();
6643 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6644 // Look pass truncate.
6645 UOpNo = User->use_begin().getOperandNo();
6646 User = *User->use_begin();
6649 if (User->getOpcode() != ISD::BRCOND &&
6650 User->getOpcode() != ISD::SETCC &&
6651 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6664 // Due to the ISEL shortcoming noted above, be conservative if this op is
6665 // likely to be selected as part of a load-modify-store instruction.
6666 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6667 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6668 if (UI->getOpcode() == ISD::STORE)
6671 // Otherwise use a regular EFLAGS-setting instruction.
6672 switch (Op.getNode()->getOpcode()) {
6673 default: llvm_unreachable("unexpected operator!");
6674 case ISD::SUB: Opcode = X86ISD::SUB; break;
6675 case ISD::OR: Opcode = X86ISD::OR; break;
6676 case ISD::XOR: Opcode = X86ISD::XOR; break;
6677 case ISD::AND: Opcode = X86ISD::AND; break;
6689 return SDValue(Op.getNode(), 1);
6696 // Emit a CMP with 0, which is the TEST pattern.
6697 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6698 DAG.getConstant(0, Op.getValueType()));
6700 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6701 SmallVector<SDValue, 4> Ops;
6702 for (unsigned i = 0; i != NumOperands; ++i)
6703 Ops.push_back(Op.getOperand(i));
6705 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6706 DAG.ReplaceAllUsesWith(Op, New);
6707 return SDValue(New.getNode(), 1);
6710 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6712 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6713 SelectionDAG &DAG) const {
6714 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6715 if (C->getAPIntValue() == 0)
6716 return EmitTest(Op0, X86CC, DAG);
6718 DebugLoc dl = Op0.getDebugLoc();
6719 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6722 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6723 /// if it's possible.
6724 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6725 DebugLoc dl, SelectionDAG &DAG) const {
6726 SDValue Op0 = And.getOperand(0);
6727 SDValue Op1 = And.getOperand(1);
6728 if (Op0.getOpcode() == ISD::TRUNCATE)
6729 Op0 = Op0.getOperand(0);
6730 if (Op1.getOpcode() == ISD::TRUNCATE)
6731 Op1 = Op1.getOperand(0);
6734 if (Op1.getOpcode() == ISD::SHL)
6735 std::swap(Op0, Op1);
6736 if (Op0.getOpcode() == ISD::SHL) {
6737 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6738 if (And00C->getZExtValue() == 1) {
6739 // If we looked past a truncate, check that it's only truncating away
6741 unsigned BitWidth = Op0.getValueSizeInBits();
6742 unsigned AndBitWidth = And.getValueSizeInBits();
6743 if (BitWidth > AndBitWidth) {
6744 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
6745 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
6746 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
6750 RHS = Op0.getOperand(1);
6752 } else if (Op1.getOpcode() == ISD::Constant) {
6753 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6754 SDValue AndLHS = Op0;
6755 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6756 LHS = AndLHS.getOperand(0);
6757 RHS = AndLHS.getOperand(1);
6761 if (LHS.getNode()) {
6762 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6763 // instruction. Since the shift amount is in-range-or-undefined, we know
6764 // that doing a bittest on the i32 value is ok. We extend to i32 because
6765 // the encoding for the i16 version is larger than the i32 version.
6766 // Also promote i16 to i32 for performance / code size reason.
6767 if (LHS.getValueType() == MVT::i8 ||
6768 LHS.getValueType() == MVT::i16)
6769 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6771 // If the operand types disagree, extend the shift amount to match. Since
6772 // BT ignores high bits (like shifts) we can use anyextend.
6773 if (LHS.getValueType() != RHS.getValueType())
6774 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6776 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6777 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6778 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6779 DAG.getConstant(Cond, MVT::i8), BT);
6785 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
6786 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6787 SDValue Op0 = Op.getOperand(0);
6788 SDValue Op1 = Op.getOperand(1);
6789 DebugLoc dl = Op.getDebugLoc();
6790 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6792 // Optimize to BT if possible.
6793 // Lower (X & (1 << N)) == 0 to BT(X, N).
6794 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6795 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6796 if (Op0.getOpcode() == ISD::AND &&
6798 Op1.getOpcode() == ISD::Constant &&
6799 cast<ConstantSDNode>(Op1)->isNullValue() &&
6800 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6801 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6802 if (NewSetCC.getNode())
6806 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6807 if (Op0.getOpcode() == X86ISD::SETCC &&
6808 Op1.getOpcode() == ISD::Constant &&
6809 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6810 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6811 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6812 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6813 bool Invert = (CC == ISD::SETNE) ^
6814 cast<ConstantSDNode>(Op1)->isNullValue();
6816 CCode = X86::GetOppositeBranchCondition(CCode);
6817 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6818 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6821 bool isFP = Op1.getValueType().isFloatingPoint();
6822 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6823 if (X86CC == X86::COND_INVALID)
6826 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6828 // Use sbb x, x to materialize carry bit into a GPR.
6829 if (X86CC == X86::COND_B)
6830 return DAG.getNode(ISD::AND, dl, MVT::i8,
6831 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6832 DAG.getConstant(X86CC, MVT::i8), Cond),
6833 DAG.getConstant(1, MVT::i8));
6835 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6836 DAG.getConstant(X86CC, MVT::i8), Cond);
6839 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
6841 SDValue Op0 = Op.getOperand(0);
6842 SDValue Op1 = Op.getOperand(1);
6843 SDValue CC = Op.getOperand(2);
6844 EVT VT = Op.getValueType();
6845 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6846 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6847 DebugLoc dl = Op.getDebugLoc();
6851 EVT VT0 = Op0.getValueType();
6852 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6853 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6856 switch (SetCCOpcode) {
6859 case ISD::SETEQ: SSECC = 0; break;
6861 case ISD::SETGT: Swap = true; // Fallthrough
6863 case ISD::SETOLT: SSECC = 1; break;
6865 case ISD::SETGE: Swap = true; // Fallthrough
6867 case ISD::SETOLE: SSECC = 2; break;
6868 case ISD::SETUO: SSECC = 3; break;
6870 case ISD::SETNE: SSECC = 4; break;
6871 case ISD::SETULE: Swap = true;
6872 case ISD::SETUGE: SSECC = 5; break;
6873 case ISD::SETULT: Swap = true;
6874 case ISD::SETUGT: SSECC = 6; break;
6875 case ISD::SETO: SSECC = 7; break;
6878 std::swap(Op0, Op1);
6880 // In the two special cases we can't handle, emit two comparisons.
6882 if (SetCCOpcode == ISD::SETUEQ) {
6884 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6885 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6886 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6888 else if (SetCCOpcode == ISD::SETONE) {
6890 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6891 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6892 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6894 llvm_unreachable("Illegal FP comparison");
6896 // Handle all other FP comparisons here.
6897 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6900 // We are handling one of the integer comparisons here. Since SSE only has
6901 // GT and EQ comparisons for integer, swapping operands and multiple
6902 // operations may be required for some comparisons.
6903 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6904 bool Swap = false, Invert = false, FlipSigns = false;
6906 switch (VT.getSimpleVT().SimpleTy) {
6909 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6911 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6913 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6914 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6917 switch (SetCCOpcode) {
6919 case ISD::SETNE: Invert = true;
6920 case ISD::SETEQ: Opc = EQOpc; break;
6921 case ISD::SETLT: Swap = true;
6922 case ISD::SETGT: Opc = GTOpc; break;
6923 case ISD::SETGE: Swap = true;
6924 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6925 case ISD::SETULT: Swap = true;
6926 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6927 case ISD::SETUGE: Swap = true;
6928 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6931 std::swap(Op0, Op1);
6933 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6934 // bits of the inputs before performing those operations.
6936 EVT EltVT = VT.getVectorElementType();
6937 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6939 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6940 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6942 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6943 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6946 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6948 // If the logical-not of the result is required, perform that now.
6950 Result = DAG.getNOT(dl, Result, VT);
6955 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6956 static bool isX86LogicalCmp(SDValue Op) {
6957 unsigned Opc = Op.getNode()->getOpcode();
6958 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6960 if (Op.getResNo() == 1 &&
6961 (Opc == X86ISD::ADD ||
6962 Opc == X86ISD::SUB ||
6963 Opc == X86ISD::SMUL ||
6964 Opc == X86ISD::UMUL ||
6965 Opc == X86ISD::INC ||
6966 Opc == X86ISD::DEC ||
6967 Opc == X86ISD::OR ||
6968 Opc == X86ISD::XOR ||
6969 Opc == X86ISD::AND))
6975 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
6976 bool addTest = true;
6977 SDValue Cond = Op.getOperand(0);
6978 DebugLoc dl = Op.getDebugLoc();
6981 if (Cond.getOpcode() == ISD::SETCC) {
6982 SDValue NewCond = LowerSETCC(Cond, DAG);
6983 if (NewCond.getNode())
6987 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6988 SDValue Op1 = Op.getOperand(1);
6989 SDValue Op2 = Op.getOperand(2);
6990 if (Cond.getOpcode() == X86ISD::SETCC &&
6991 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6992 SDValue Cmp = Cond.getOperand(1);
6993 if (Cmp.getOpcode() == X86ISD::CMP) {
6994 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6995 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6996 ConstantSDNode *RHSC =
6997 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6998 if (N1C && N1C->isAllOnesValue() &&
6999 N2C && N2C->isNullValue() &&
7000 RHSC && RHSC->isNullValue()) {
7001 SDValue CmpOp0 = Cmp.getOperand(0);
7002 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
7003 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
7004 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
7005 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
7010 // Look pass (and (setcc_carry (cmp ...)), 1).
7011 if (Cond.getOpcode() == ISD::AND &&
7012 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7013 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7014 if (C && C->getAPIntValue() == 1)
7015 Cond = Cond.getOperand(0);
7018 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7019 // setting operand in place of the X86ISD::SETCC.
7020 if (Cond.getOpcode() == X86ISD::SETCC ||
7021 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7022 CC = Cond.getOperand(0);
7024 SDValue Cmp = Cond.getOperand(1);
7025 unsigned Opc = Cmp.getOpcode();
7026 EVT VT = Op.getValueType();
7028 bool IllegalFPCMov = false;
7029 if (VT.isFloatingPoint() && !VT.isVector() &&
7030 !isScalarFPTypeInSSEReg(VT)) // FPStack?
7031 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
7033 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
7034 Opc == X86ISD::BT) { // FIXME
7041 // Look pass the truncate.
7042 if (Cond.getOpcode() == ISD::TRUNCATE)
7043 Cond = Cond.getOperand(0);
7045 // We know the result of AND is compared against zero. Try to match
7047 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7048 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
7049 if (NewSetCC.getNode()) {
7050 CC = NewSetCC.getOperand(0);
7051 Cond = NewSetCC.getOperand(1);
7058 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7059 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7062 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
7063 // condition is true.
7064 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
7065 SDValue Ops[] = { Op2, Op1, CC, Cond };
7066 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
7069 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
7070 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
7071 // from the AND / OR.
7072 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
7073 Opc = Op.getOpcode();
7074 if (Opc != ISD::OR && Opc != ISD::AND)
7076 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7077 Op.getOperand(0).hasOneUse() &&
7078 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
7079 Op.getOperand(1).hasOneUse());
7082 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
7083 // 1 and that the SETCC node has a single use.
7084 static bool isXor1OfSetCC(SDValue Op) {
7085 if (Op.getOpcode() != ISD::XOR)
7087 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7088 if (N1C && N1C->getAPIntValue() == 1) {
7089 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7090 Op.getOperand(0).hasOneUse();
7095 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
7096 bool addTest = true;
7097 SDValue Chain = Op.getOperand(0);
7098 SDValue Cond = Op.getOperand(1);
7099 SDValue Dest = Op.getOperand(2);
7100 DebugLoc dl = Op.getDebugLoc();
7103 if (Cond.getOpcode() == ISD::SETCC) {
7104 SDValue NewCond = LowerSETCC(Cond, DAG);
7105 if (NewCond.getNode())
7109 // FIXME: LowerXALUO doesn't handle these!!
7110 else if (Cond.getOpcode() == X86ISD::ADD ||
7111 Cond.getOpcode() == X86ISD::SUB ||
7112 Cond.getOpcode() == X86ISD::SMUL ||
7113 Cond.getOpcode() == X86ISD::UMUL)
7114 Cond = LowerXALUO(Cond, DAG);
7117 // Look pass (and (setcc_carry (cmp ...)), 1).
7118 if (Cond.getOpcode() == ISD::AND &&
7119 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7120 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7121 if (C && C->getAPIntValue() == 1)
7122 Cond = Cond.getOperand(0);
7125 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7126 // setting operand in place of the X86ISD::SETCC.
7127 if (Cond.getOpcode() == X86ISD::SETCC ||
7128 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7129 CC = Cond.getOperand(0);
7131 SDValue Cmp = Cond.getOperand(1);
7132 unsigned Opc = Cmp.getOpcode();
7133 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
7134 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
7138 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
7142 // These can only come from an arithmetic instruction with overflow,
7143 // e.g. SADDO, UADDO.
7144 Cond = Cond.getNode()->getOperand(1);
7151 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
7152 SDValue Cmp = Cond.getOperand(0).getOperand(1);
7153 if (CondOpc == ISD::OR) {
7154 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
7155 // two branches instead of an explicit OR instruction with a
7157 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7158 isX86LogicalCmp(Cmp)) {
7159 CC = Cond.getOperand(0).getOperand(0);
7160 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7161 Chain, Dest, CC, Cmp);
7162 CC = Cond.getOperand(1).getOperand(0);
7166 } else { // ISD::AND
7167 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
7168 // two branches instead of an explicit AND instruction with a
7169 // separate test. However, we only do this if this block doesn't
7170 // have a fall-through edge, because this requires an explicit
7171 // jmp when the condition is false.
7172 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7173 isX86LogicalCmp(Cmp) &&
7174 Op.getNode()->hasOneUse()) {
7175 X86::CondCode CCode =
7176 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7177 CCode = X86::GetOppositeBranchCondition(CCode);
7178 CC = DAG.getConstant(CCode, MVT::i8);
7179 SDNode *User = *Op.getNode()->use_begin();
7180 // Look for an unconditional branch following this conditional branch.
7181 // We need this because we need to reverse the successors in order
7182 // to implement FCMP_OEQ.
7183 if (User->getOpcode() == ISD::BR) {
7184 SDValue FalseBB = User->getOperand(1);
7186 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
7187 assert(NewBR == User);
7191 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7192 Chain, Dest, CC, Cmp);
7193 X86::CondCode CCode =
7194 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
7195 CCode = X86::GetOppositeBranchCondition(CCode);
7196 CC = DAG.getConstant(CCode, MVT::i8);
7202 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
7203 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
7204 // It should be transformed during dag combiner except when the condition
7205 // is set by a arithmetics with overflow node.
7206 X86::CondCode CCode =
7207 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7208 CCode = X86::GetOppositeBranchCondition(CCode);
7209 CC = DAG.getConstant(CCode, MVT::i8);
7210 Cond = Cond.getOperand(0).getOperand(1);
7216 // Look pass the truncate.
7217 if (Cond.getOpcode() == ISD::TRUNCATE)
7218 Cond = Cond.getOperand(0);
7220 // We know the result of AND is compared against zero. Try to match
7222 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7223 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
7224 if (NewSetCC.getNode()) {
7225 CC = NewSetCC.getOperand(0);
7226 Cond = NewSetCC.getOperand(1);
7233 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7234 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7236 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7237 Chain, Dest, CC, Cond);
7241 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
7242 // Calls to _alloca is needed to probe the stack when allocating more than 4k
7243 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
7244 // that the guard pages used by the OS virtual memory manager are allocated in
7245 // correct sequence.
7247 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
7248 SelectionDAG &DAG) const {
7249 assert(Subtarget->isTargetCygMing() &&
7250 "This should be used only on Cygwin/Mingw targets");
7251 DebugLoc dl = Op.getDebugLoc();
7254 SDValue Chain = Op.getOperand(0);
7255 SDValue Size = Op.getOperand(1);
7256 // FIXME: Ensure alignment here
7260 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
7262 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
7263 Flag = Chain.getValue(1);
7265 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
7267 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
7268 Flag = Chain.getValue(1);
7270 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
7272 SDValue Ops1[2] = { Chain.getValue(0), Chain };
7273 return DAG.getMergeValues(Ops1, 2, dl);
7276 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
7277 MachineFunction &MF = DAG.getMachineFunction();
7278 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
7280 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
7281 DebugLoc dl = Op.getDebugLoc();
7283 if (!Subtarget->is64Bit()) {
7284 // vastart just stores the address of the VarArgsFrameIndex slot into the
7285 // memory location argument.
7286 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
7288 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
7293 // gp_offset (0 - 6 * 8)
7294 // fp_offset (48 - 48 + 8 * 16)
7295 // overflow_arg_area (point to parameters coming in memory).
7297 SmallVector<SDValue, 8> MemOps;
7298 SDValue FIN = Op.getOperand(1);
7300 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
7301 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
7303 FIN, SV, 0, false, false, 0);
7304 MemOps.push_back(Store);
7307 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7308 FIN, DAG.getIntPtrConstant(4));
7309 Store = DAG.getStore(Op.getOperand(0), dl,
7310 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
7312 FIN, SV, 4, false, false, 0);
7313 MemOps.push_back(Store);
7315 // Store ptr to overflow_arg_area
7316 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7317 FIN, DAG.getIntPtrConstant(4));
7318 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
7320 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 8,
7322 MemOps.push_back(Store);
7324 // Store ptr to reg_save_area.
7325 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7326 FIN, DAG.getIntPtrConstant(8));
7327 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
7329 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 16,
7331 MemOps.push_back(Store);
7332 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
7333 &MemOps[0], MemOps.size());
7336 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
7337 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
7338 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
7340 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
7344 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
7345 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
7346 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
7347 SDValue Chain = Op.getOperand(0);
7348 SDValue DstPtr = Op.getOperand(1);
7349 SDValue SrcPtr = Op.getOperand(2);
7350 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
7351 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7352 DebugLoc dl = Op.getDebugLoc();
7354 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
7355 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
7356 false, DstSV, 0, SrcSV, 0);
7360 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
7361 DebugLoc dl = Op.getDebugLoc();
7362 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7364 default: return SDValue(); // Don't custom lower most intrinsics.
7365 // Comparison intrinsics.
7366 case Intrinsic::x86_sse_comieq_ss:
7367 case Intrinsic::x86_sse_comilt_ss:
7368 case Intrinsic::x86_sse_comile_ss:
7369 case Intrinsic::x86_sse_comigt_ss:
7370 case Intrinsic::x86_sse_comige_ss:
7371 case Intrinsic::x86_sse_comineq_ss:
7372 case Intrinsic::x86_sse_ucomieq_ss:
7373 case Intrinsic::x86_sse_ucomilt_ss:
7374 case Intrinsic::x86_sse_ucomile_ss:
7375 case Intrinsic::x86_sse_ucomigt_ss:
7376 case Intrinsic::x86_sse_ucomige_ss:
7377 case Intrinsic::x86_sse_ucomineq_ss:
7378 case Intrinsic::x86_sse2_comieq_sd:
7379 case Intrinsic::x86_sse2_comilt_sd:
7380 case Intrinsic::x86_sse2_comile_sd:
7381 case Intrinsic::x86_sse2_comigt_sd:
7382 case Intrinsic::x86_sse2_comige_sd:
7383 case Intrinsic::x86_sse2_comineq_sd:
7384 case Intrinsic::x86_sse2_ucomieq_sd:
7385 case Intrinsic::x86_sse2_ucomilt_sd:
7386 case Intrinsic::x86_sse2_ucomile_sd:
7387 case Intrinsic::x86_sse2_ucomigt_sd:
7388 case Intrinsic::x86_sse2_ucomige_sd:
7389 case Intrinsic::x86_sse2_ucomineq_sd: {
7391 ISD::CondCode CC = ISD::SETCC_INVALID;
7394 case Intrinsic::x86_sse_comieq_ss:
7395 case Intrinsic::x86_sse2_comieq_sd:
7399 case Intrinsic::x86_sse_comilt_ss:
7400 case Intrinsic::x86_sse2_comilt_sd:
7404 case Intrinsic::x86_sse_comile_ss:
7405 case Intrinsic::x86_sse2_comile_sd:
7409 case Intrinsic::x86_sse_comigt_ss:
7410 case Intrinsic::x86_sse2_comigt_sd:
7414 case Intrinsic::x86_sse_comige_ss:
7415 case Intrinsic::x86_sse2_comige_sd:
7419 case Intrinsic::x86_sse_comineq_ss:
7420 case Intrinsic::x86_sse2_comineq_sd:
7424 case Intrinsic::x86_sse_ucomieq_ss:
7425 case Intrinsic::x86_sse2_ucomieq_sd:
7426 Opc = X86ISD::UCOMI;
7429 case Intrinsic::x86_sse_ucomilt_ss:
7430 case Intrinsic::x86_sse2_ucomilt_sd:
7431 Opc = X86ISD::UCOMI;
7434 case Intrinsic::x86_sse_ucomile_ss:
7435 case Intrinsic::x86_sse2_ucomile_sd:
7436 Opc = X86ISD::UCOMI;
7439 case Intrinsic::x86_sse_ucomigt_ss:
7440 case Intrinsic::x86_sse2_ucomigt_sd:
7441 Opc = X86ISD::UCOMI;
7444 case Intrinsic::x86_sse_ucomige_ss:
7445 case Intrinsic::x86_sse2_ucomige_sd:
7446 Opc = X86ISD::UCOMI;
7449 case Intrinsic::x86_sse_ucomineq_ss:
7450 case Intrinsic::x86_sse2_ucomineq_sd:
7451 Opc = X86ISD::UCOMI;
7456 SDValue LHS = Op.getOperand(1);
7457 SDValue RHS = Op.getOperand(2);
7458 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
7459 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
7460 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
7461 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7462 DAG.getConstant(X86CC, MVT::i8), Cond);
7463 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7465 // ptest and testp intrinsics. The intrinsic these come from are designed to
7466 // return an integer value, not just an instruction so lower it to the ptest
7467 // or testp pattern and a setcc for the result.
7468 case Intrinsic::x86_sse41_ptestz:
7469 case Intrinsic::x86_sse41_ptestc:
7470 case Intrinsic::x86_sse41_ptestnzc:
7471 case Intrinsic::x86_avx_ptestz_256:
7472 case Intrinsic::x86_avx_ptestc_256:
7473 case Intrinsic::x86_avx_ptestnzc_256:
7474 case Intrinsic::x86_avx_vtestz_ps:
7475 case Intrinsic::x86_avx_vtestc_ps:
7476 case Intrinsic::x86_avx_vtestnzc_ps:
7477 case Intrinsic::x86_avx_vtestz_pd:
7478 case Intrinsic::x86_avx_vtestc_pd:
7479 case Intrinsic::x86_avx_vtestnzc_pd:
7480 case Intrinsic::x86_avx_vtestz_ps_256:
7481 case Intrinsic::x86_avx_vtestc_ps_256:
7482 case Intrinsic::x86_avx_vtestnzc_ps_256:
7483 case Intrinsic::x86_avx_vtestz_pd_256:
7484 case Intrinsic::x86_avx_vtestc_pd_256:
7485 case Intrinsic::x86_avx_vtestnzc_pd_256: {
7486 bool IsTestPacked = false;
7489 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
7490 case Intrinsic::x86_avx_vtestz_ps:
7491 case Intrinsic::x86_avx_vtestz_pd:
7492 case Intrinsic::x86_avx_vtestz_ps_256:
7493 case Intrinsic::x86_avx_vtestz_pd_256:
7494 IsTestPacked = true; // Fallthrough
7495 case Intrinsic::x86_sse41_ptestz:
7496 case Intrinsic::x86_avx_ptestz_256:
7498 X86CC = X86::COND_E;
7500 case Intrinsic::x86_avx_vtestc_ps:
7501 case Intrinsic::x86_avx_vtestc_pd:
7502 case Intrinsic::x86_avx_vtestc_ps_256:
7503 case Intrinsic::x86_avx_vtestc_pd_256:
7504 IsTestPacked = true; // Fallthrough
7505 case Intrinsic::x86_sse41_ptestc:
7506 case Intrinsic::x86_avx_ptestc_256:
7508 X86CC = X86::COND_B;
7510 case Intrinsic::x86_avx_vtestnzc_ps:
7511 case Intrinsic::x86_avx_vtestnzc_pd:
7512 case Intrinsic::x86_avx_vtestnzc_ps_256:
7513 case Intrinsic::x86_avx_vtestnzc_pd_256:
7514 IsTestPacked = true; // Fallthrough
7515 case Intrinsic::x86_sse41_ptestnzc:
7516 case Intrinsic::x86_avx_ptestnzc_256:
7518 X86CC = X86::COND_A;
7522 SDValue LHS = Op.getOperand(1);
7523 SDValue RHS = Op.getOperand(2);
7524 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
7525 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
7526 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
7527 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
7528 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7531 // Fix vector shift instructions where the last operand is a non-immediate
7533 case Intrinsic::x86_sse2_pslli_w:
7534 case Intrinsic::x86_sse2_pslli_d:
7535 case Intrinsic::x86_sse2_pslli_q:
7536 case Intrinsic::x86_sse2_psrli_w:
7537 case Intrinsic::x86_sse2_psrli_d:
7538 case Intrinsic::x86_sse2_psrli_q:
7539 case Intrinsic::x86_sse2_psrai_w:
7540 case Intrinsic::x86_sse2_psrai_d:
7541 case Intrinsic::x86_mmx_pslli_w:
7542 case Intrinsic::x86_mmx_pslli_d:
7543 case Intrinsic::x86_mmx_pslli_q:
7544 case Intrinsic::x86_mmx_psrli_w:
7545 case Intrinsic::x86_mmx_psrli_d:
7546 case Intrinsic::x86_mmx_psrli_q:
7547 case Intrinsic::x86_mmx_psrai_w:
7548 case Intrinsic::x86_mmx_psrai_d: {
7549 SDValue ShAmt = Op.getOperand(2);
7550 if (isa<ConstantSDNode>(ShAmt))
7553 unsigned NewIntNo = 0;
7554 EVT ShAmtVT = MVT::v4i32;
7556 case Intrinsic::x86_sse2_pslli_w:
7557 NewIntNo = Intrinsic::x86_sse2_psll_w;
7559 case Intrinsic::x86_sse2_pslli_d:
7560 NewIntNo = Intrinsic::x86_sse2_psll_d;
7562 case Intrinsic::x86_sse2_pslli_q:
7563 NewIntNo = Intrinsic::x86_sse2_psll_q;
7565 case Intrinsic::x86_sse2_psrli_w:
7566 NewIntNo = Intrinsic::x86_sse2_psrl_w;
7568 case Intrinsic::x86_sse2_psrli_d:
7569 NewIntNo = Intrinsic::x86_sse2_psrl_d;
7571 case Intrinsic::x86_sse2_psrli_q:
7572 NewIntNo = Intrinsic::x86_sse2_psrl_q;
7574 case Intrinsic::x86_sse2_psrai_w:
7575 NewIntNo = Intrinsic::x86_sse2_psra_w;
7577 case Intrinsic::x86_sse2_psrai_d:
7578 NewIntNo = Intrinsic::x86_sse2_psra_d;
7581 ShAmtVT = MVT::v2i32;
7583 case Intrinsic::x86_mmx_pslli_w:
7584 NewIntNo = Intrinsic::x86_mmx_psll_w;
7586 case Intrinsic::x86_mmx_pslli_d:
7587 NewIntNo = Intrinsic::x86_mmx_psll_d;
7589 case Intrinsic::x86_mmx_pslli_q:
7590 NewIntNo = Intrinsic::x86_mmx_psll_q;
7592 case Intrinsic::x86_mmx_psrli_w:
7593 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7595 case Intrinsic::x86_mmx_psrli_d:
7596 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7598 case Intrinsic::x86_mmx_psrli_q:
7599 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7601 case Intrinsic::x86_mmx_psrai_w:
7602 NewIntNo = Intrinsic::x86_mmx_psra_w;
7604 case Intrinsic::x86_mmx_psrai_d:
7605 NewIntNo = Intrinsic::x86_mmx_psra_d;
7607 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7613 // The vector shift intrinsics with scalars uses 32b shift amounts but
7614 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7618 ShOps[1] = DAG.getConstant(0, MVT::i32);
7619 if (ShAmtVT == MVT::v4i32) {
7620 ShOps[2] = DAG.getUNDEF(MVT::i32);
7621 ShOps[3] = DAG.getUNDEF(MVT::i32);
7622 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7624 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7627 EVT VT = Op.getValueType();
7628 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7629 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7630 DAG.getConstant(NewIntNo, MVT::i32),
7631 Op.getOperand(1), ShAmt);
7636 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7637 SelectionDAG &DAG) const {
7638 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7639 MFI->setReturnAddressIsTaken(true);
7641 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7642 DebugLoc dl = Op.getDebugLoc();
7645 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7647 DAG.getConstant(TD->getPointerSize(),
7648 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7649 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7650 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7652 NULL, 0, false, false, 0);
7655 // Just load the return address.
7656 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7657 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7658 RetAddrFI, NULL, 0, false, false, 0);
7661 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7662 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7663 MFI->setFrameAddressIsTaken(true);
7665 EVT VT = Op.getValueType();
7666 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7667 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7668 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7669 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7671 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7676 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7677 SelectionDAG &DAG) const {
7678 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7681 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7682 MachineFunction &MF = DAG.getMachineFunction();
7683 SDValue Chain = Op.getOperand(0);
7684 SDValue Offset = Op.getOperand(1);
7685 SDValue Handler = Op.getOperand(2);
7686 DebugLoc dl = Op.getDebugLoc();
7688 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
7689 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7691 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7693 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
7694 DAG.getIntPtrConstant(TD->getPointerSize()));
7695 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7696 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7697 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7698 MF.getRegInfo().addLiveOut(StoreAddrReg);
7700 return DAG.getNode(X86ISD::EH_RETURN, dl,
7702 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7705 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7706 SelectionDAG &DAG) const {
7707 SDValue Root = Op.getOperand(0);
7708 SDValue Trmp = Op.getOperand(1); // trampoline
7709 SDValue FPtr = Op.getOperand(2); // nested function
7710 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7711 DebugLoc dl = Op.getDebugLoc();
7713 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7715 if (Subtarget->is64Bit()) {
7716 SDValue OutChains[6];
7718 // Large code-model.
7719 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7720 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7722 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7723 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7725 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7727 // Load the pointer to the nested function into R11.
7728 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7729 SDValue Addr = Trmp;
7730 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7731 Addr, TrmpAddr, 0, false, false, 0);
7733 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7734 DAG.getConstant(2, MVT::i64));
7735 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7738 // Load the 'nest' parameter value into R10.
7739 // R10 is specified in X86CallingConv.td
7740 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7741 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7742 DAG.getConstant(10, MVT::i64));
7743 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7744 Addr, TrmpAddr, 10, false, false, 0);
7746 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7747 DAG.getConstant(12, MVT::i64));
7748 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7751 // Jump to the nested function.
7752 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7753 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7754 DAG.getConstant(20, MVT::i64));
7755 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7756 Addr, TrmpAddr, 20, false, false, 0);
7758 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7759 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7760 DAG.getConstant(22, MVT::i64));
7761 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7762 TrmpAddr, 22, false, false, 0);
7765 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7766 return DAG.getMergeValues(Ops, 2, dl);
7768 const Function *Func =
7769 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7770 CallingConv::ID CC = Func->getCallingConv();
7775 llvm_unreachable("Unsupported calling convention");
7776 case CallingConv::C:
7777 case CallingConv::X86_StdCall: {
7778 // Pass 'nest' parameter in ECX.
7779 // Must be kept in sync with X86CallingConv.td
7782 // Check that ECX wasn't needed by an 'inreg' parameter.
7783 const FunctionType *FTy = Func->getFunctionType();
7784 const AttrListPtr &Attrs = Func->getAttributes();
7786 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7787 unsigned InRegCount = 0;
7790 for (FunctionType::param_iterator I = FTy->param_begin(),
7791 E = FTy->param_end(); I != E; ++I, ++Idx)
7792 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7793 // FIXME: should only count parameters that are lowered to integers.
7794 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7796 if (InRegCount > 2) {
7797 report_fatal_error("Nest register in use - reduce number of inreg"
7803 case CallingConv::X86_FastCall:
7804 case CallingConv::X86_ThisCall:
7805 case CallingConv::Fast:
7806 // Pass 'nest' parameter in EAX.
7807 // Must be kept in sync with X86CallingConv.td
7812 SDValue OutChains[4];
7815 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7816 DAG.getConstant(10, MVT::i32));
7817 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7819 // This is storing the opcode for MOV32ri.
7820 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7821 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7822 OutChains[0] = DAG.getStore(Root, dl,
7823 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7824 Trmp, TrmpAddr, 0, false, false, 0);
7826 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7827 DAG.getConstant(1, MVT::i32));
7828 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7831 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7832 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7833 DAG.getConstant(5, MVT::i32));
7834 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7835 TrmpAddr, 5, false, false, 1);
7837 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7838 DAG.getConstant(6, MVT::i32));
7839 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7843 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7844 return DAG.getMergeValues(Ops, 2, dl);
7848 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
7849 SelectionDAG &DAG) const {
7851 The rounding mode is in bits 11:10 of FPSR, and has the following
7858 FLT_ROUNDS, on the other hand, expects the following:
7865 To perform the conversion, we do:
7866 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7869 MachineFunction &MF = DAG.getMachineFunction();
7870 const TargetMachine &TM = MF.getTarget();
7871 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7872 unsigned StackAlignment = TFI.getStackAlignment();
7873 EVT VT = Op.getValueType();
7874 DebugLoc dl = Op.getDebugLoc();
7876 // Save FP Control Word to stack slot
7877 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7878 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7880 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7881 DAG.getEntryNode(), StackSlot);
7883 // Load FP Control Word from stack slot
7884 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7887 // Transform as necessary
7889 DAG.getNode(ISD::SRL, dl, MVT::i16,
7890 DAG.getNode(ISD::AND, dl, MVT::i16,
7891 CWD, DAG.getConstant(0x800, MVT::i16)),
7892 DAG.getConstant(11, MVT::i8));
7894 DAG.getNode(ISD::SRL, dl, MVT::i16,
7895 DAG.getNode(ISD::AND, dl, MVT::i16,
7896 CWD, DAG.getConstant(0x400, MVT::i16)),
7897 DAG.getConstant(9, MVT::i8));
7900 DAG.getNode(ISD::AND, dl, MVT::i16,
7901 DAG.getNode(ISD::ADD, dl, MVT::i16,
7902 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7903 DAG.getConstant(1, MVT::i16)),
7904 DAG.getConstant(3, MVT::i16));
7907 return DAG.getNode((VT.getSizeInBits() < 16 ?
7908 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7911 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
7912 EVT VT = Op.getValueType();
7914 unsigned NumBits = VT.getSizeInBits();
7915 DebugLoc dl = Op.getDebugLoc();
7917 Op = Op.getOperand(0);
7918 if (VT == MVT::i8) {
7919 // Zero extend to i32 since there is not an i8 bsr.
7921 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7924 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7925 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7926 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7928 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7931 DAG.getConstant(NumBits+NumBits-1, OpVT),
7932 DAG.getConstant(X86::COND_E, MVT::i8),
7935 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7937 // Finally xor with NumBits-1.
7938 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7941 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7945 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
7946 EVT VT = Op.getValueType();
7948 unsigned NumBits = VT.getSizeInBits();
7949 DebugLoc dl = Op.getDebugLoc();
7951 Op = Op.getOperand(0);
7952 if (VT == MVT::i8) {
7954 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7957 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7958 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7959 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7961 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7964 DAG.getConstant(NumBits, OpVT),
7965 DAG.getConstant(X86::COND_E, MVT::i8),
7968 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7971 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7975 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
7976 EVT VT = Op.getValueType();
7977 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7978 DebugLoc dl = Op.getDebugLoc();
7980 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7981 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7982 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7983 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7984 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7986 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7987 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7988 // return AloBlo + AloBhi + AhiBlo;
7990 SDValue A = Op.getOperand(0);
7991 SDValue B = Op.getOperand(1);
7993 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7994 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7995 A, DAG.getConstant(32, MVT::i32));
7996 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7997 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7998 B, DAG.getConstant(32, MVT::i32));
7999 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8000 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8002 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8003 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8005 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8006 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8008 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8009 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8010 AloBhi, DAG.getConstant(32, MVT::i32));
8011 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8012 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8013 AhiBlo, DAG.getConstant(32, MVT::i32));
8014 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
8015 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
8019 SDValue X86TargetLowering::LowerSHL(SDValue Op, SelectionDAG &DAG) const {
8020 EVT VT = Op.getValueType();
8021 DebugLoc dl = Op.getDebugLoc();
8022 SDValue R = Op.getOperand(0);
8024 LLVMContext *Context = DAG.getContext();
8026 assert(Subtarget->hasSSE41() && "Cannot lower SHL without SSE4.1 or later");
8028 if (VT == MVT::v4i32) {
8029 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8030 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8031 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
8033 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
8035 std::vector<Constant*> CV(4, CI);
8036 Constant *C = ConstantVector::get(CV);
8037 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8038 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8039 PseudoSourceValue::getConstantPool(), 0,
8042 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
8043 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, Op);
8044 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
8045 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
8047 if (VT == MVT::v16i8) {
8049 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8050 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8051 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
8053 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
8054 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
8056 std::vector<Constant*> CVM1(16, CM1);
8057 std::vector<Constant*> CVM2(16, CM2);
8058 Constant *C = ConstantVector::get(CVM1);
8059 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8060 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8061 PseudoSourceValue::getConstantPool(), 0,
8064 // r = pblendv(r, psllw(r & (char16)15, 4), a);
8065 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8066 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8067 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8068 DAG.getConstant(4, MVT::i32));
8069 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8070 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8073 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8075 C = ConstantVector::get(CVM2);
8076 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8077 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8078 PseudoSourceValue::getConstantPool(), 0, false, false, 16);
8080 // r = pblendv(r, psllw(r & (char16)63, 2), a);
8081 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8082 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8083 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8084 DAG.getConstant(2, MVT::i32));
8085 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8086 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8089 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8091 // return pblendv(r, r+r, a);
8092 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8093 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
8094 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
8100 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
8101 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
8102 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
8103 // looks for this combo and may remove the "setcc" instruction if the "setcc"
8104 // has only one use.
8105 SDNode *N = Op.getNode();
8106 SDValue LHS = N->getOperand(0);
8107 SDValue RHS = N->getOperand(1);
8108 unsigned BaseOp = 0;
8110 DebugLoc dl = Op.getDebugLoc();
8112 switch (Op.getOpcode()) {
8113 default: llvm_unreachable("Unknown ovf instruction!");
8115 // A subtract of one will be selected as a INC. Note that INC doesn't
8116 // set CF, so we can't do this for UADDO.
8117 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
8118 if (C->getAPIntValue() == 1) {
8119 BaseOp = X86ISD::INC;
8123 BaseOp = X86ISD::ADD;
8127 BaseOp = X86ISD::ADD;
8131 // A subtract of one will be selected as a DEC. Note that DEC doesn't
8132 // set CF, so we can't do this for USUBO.
8133 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
8134 if (C->getAPIntValue() == 1) {
8135 BaseOp = X86ISD::DEC;
8139 BaseOp = X86ISD::SUB;
8143 BaseOp = X86ISD::SUB;
8147 BaseOp = X86ISD::SMUL;
8151 BaseOp = X86ISD::UMUL;
8156 // Also sets EFLAGS.
8157 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
8158 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
8161 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
8162 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
8164 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
8168 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
8169 DebugLoc dl = Op.getDebugLoc();
8171 if (!Subtarget->hasSSE2()) {
8172 SDValue Chain = Op.getOperand(0);
8173 SDValue Zero = DAG.getConstant(0,
8174 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
8176 DAG.getRegister(X86::ESP, MVT::i32), // Base
8177 DAG.getTargetConstant(1, MVT::i8), // Scale
8178 DAG.getRegister(0, MVT::i32), // Index
8179 DAG.getTargetConstant(0, MVT::i32), // Disp
8180 DAG.getRegister(0, MVT::i32), // Segment.
8185 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
8186 array_lengthof(Ops));
8187 return SDValue(Res, 0);
8190 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
8192 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
8194 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
8195 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
8196 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
8197 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
8199 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
8200 if (!Op1 && !Op2 && !Op3 && Op4)
8201 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
8203 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
8204 if (Op1 && !Op2 && !Op3 && !Op4)
8205 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
8207 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
8209 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
8212 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
8213 EVT T = Op.getValueType();
8214 DebugLoc dl = Op.getDebugLoc();
8217 switch(T.getSimpleVT().SimpleTy) {
8219 assert(false && "Invalid value type!");
8220 case MVT::i8: Reg = X86::AL; size = 1; break;
8221 case MVT::i16: Reg = X86::AX; size = 2; break;
8222 case MVT::i32: Reg = X86::EAX; size = 4; break;
8224 assert(Subtarget->is64Bit() && "Node not type legal!");
8225 Reg = X86::RAX; size = 8;
8228 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
8229 Op.getOperand(2), SDValue());
8230 SDValue Ops[] = { cpIn.getValue(0),
8233 DAG.getTargetConstant(size, MVT::i8),
8235 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8236 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
8238 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
8242 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
8243 SelectionDAG &DAG) const {
8244 assert(Subtarget->is64Bit() && "Result not type legalized?");
8245 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8246 SDValue TheChain = Op.getOperand(0);
8247 DebugLoc dl = Op.getDebugLoc();
8248 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
8249 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
8250 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
8252 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
8253 DAG.getConstant(32, MVT::i8));
8255 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
8258 return DAG.getMergeValues(Ops, 2, dl);
8261 SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
8262 SelectionDAG &DAG) const {
8263 EVT SrcVT = Op.getOperand(0).getValueType();
8264 EVT DstVT = Op.getValueType();
8265 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
8266 Subtarget->hasMMX() && !DisableMMX) &&
8267 "Unexpected custom BIT_CONVERT");
8268 assert((DstVT == MVT::i64 ||
8269 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
8270 "Unexpected custom BIT_CONVERT");
8271 // i64 <=> MMX conversions are Legal.
8272 if (SrcVT==MVT::i64 && DstVT.isVector())
8274 if (DstVT==MVT::i64 && SrcVT.isVector())
8276 // MMX <=> MMX conversions are Legal.
8277 if (SrcVT.isVector() && DstVT.isVector())
8279 // All other conversions need to be expanded.
8282 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
8283 SDNode *Node = Op.getNode();
8284 DebugLoc dl = Node->getDebugLoc();
8285 EVT T = Node->getValueType(0);
8286 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
8287 DAG.getConstant(0, T), Node->getOperand(2));
8288 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
8289 cast<AtomicSDNode>(Node)->getMemoryVT(),
8290 Node->getOperand(0),
8291 Node->getOperand(1), negOp,
8292 cast<AtomicSDNode>(Node)->getSrcValue(),
8293 cast<AtomicSDNode>(Node)->getAlignment());
8296 /// LowerOperation - Provide custom lowering hooks for some operations.
8298 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
8299 switch (Op.getOpcode()) {
8300 default: llvm_unreachable("Should not custom lower this!");
8301 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
8302 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
8303 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
8304 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
8305 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
8306 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
8307 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
8308 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
8309 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
8310 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
8311 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
8312 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
8313 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
8314 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
8315 case ISD::SHL_PARTS:
8316 case ISD::SRA_PARTS:
8317 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
8318 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
8319 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
8320 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
8321 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
8322 case ISD::FABS: return LowerFABS(Op, DAG);
8323 case ISD::FNEG: return LowerFNEG(Op, DAG);
8324 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
8325 case ISD::SETCC: return LowerSETCC(Op, DAG);
8326 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
8327 case ISD::SELECT: return LowerSELECT(Op, DAG);
8328 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
8329 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
8330 case ISD::VASTART: return LowerVASTART(Op, DAG);
8331 case ISD::VAARG: return LowerVAARG(Op, DAG);
8332 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
8333 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
8334 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
8335 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
8336 case ISD::FRAME_TO_ARGS_OFFSET:
8337 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
8338 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
8339 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
8340 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
8341 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
8342 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
8343 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
8344 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
8345 case ISD::SHL: return LowerSHL(Op, DAG);
8351 case ISD::UMULO: return LowerXALUO(Op, DAG);
8352 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
8353 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
8357 void X86TargetLowering::
8358 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
8359 SelectionDAG &DAG, unsigned NewOp) const {
8360 EVT T = Node->getValueType(0);
8361 DebugLoc dl = Node->getDebugLoc();
8362 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
8364 SDValue Chain = Node->getOperand(0);
8365 SDValue In1 = Node->getOperand(1);
8366 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
8367 Node->getOperand(2), DAG.getIntPtrConstant(0));
8368 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
8369 Node->getOperand(2), DAG.getIntPtrConstant(1));
8370 SDValue Ops[] = { Chain, In1, In2L, In2H };
8371 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
8373 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
8374 cast<MemSDNode>(Node)->getMemOperand());
8375 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
8376 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
8377 Results.push_back(Result.getValue(2));
8380 /// ReplaceNodeResults - Replace a node with an illegal result type
8381 /// with a new node built out of custom code.
8382 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
8383 SmallVectorImpl<SDValue>&Results,
8384 SelectionDAG &DAG) const {
8385 DebugLoc dl = N->getDebugLoc();
8386 switch (N->getOpcode()) {
8388 assert(false && "Do not know how to custom type legalize this operation!");
8390 case ISD::FP_TO_SINT: {
8391 std::pair<SDValue,SDValue> Vals =
8392 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
8393 SDValue FIST = Vals.first, StackSlot = Vals.second;
8394 if (FIST.getNode() != 0) {
8395 EVT VT = N->getValueType(0);
8396 // Return a load from the stack slot.
8397 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
8402 case ISD::READCYCLECOUNTER: {
8403 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8404 SDValue TheChain = N->getOperand(0);
8405 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
8406 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
8408 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
8410 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
8411 SDValue Ops[] = { eax, edx };
8412 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
8413 Results.push_back(edx.getValue(1));
8416 case ISD::ATOMIC_CMP_SWAP: {
8417 EVT T = N->getValueType(0);
8418 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
8419 SDValue cpInL, cpInH;
8420 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
8421 DAG.getConstant(0, MVT::i32));
8422 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
8423 DAG.getConstant(1, MVT::i32));
8424 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
8425 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
8427 SDValue swapInL, swapInH;
8428 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
8429 DAG.getConstant(0, MVT::i32));
8430 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
8431 DAG.getConstant(1, MVT::i32));
8432 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
8434 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
8435 swapInL.getValue(1));
8436 SDValue Ops[] = { swapInH.getValue(0),
8438 swapInH.getValue(1) };
8439 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
8440 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
8441 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
8442 MVT::i32, Result.getValue(1));
8443 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
8444 MVT::i32, cpOutL.getValue(2));
8445 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
8446 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
8447 Results.push_back(cpOutH.getValue(1));
8450 case ISD::ATOMIC_LOAD_ADD:
8451 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
8453 case ISD::ATOMIC_LOAD_AND:
8454 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
8456 case ISD::ATOMIC_LOAD_NAND:
8457 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
8459 case ISD::ATOMIC_LOAD_OR:
8460 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
8462 case ISD::ATOMIC_LOAD_SUB:
8463 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
8465 case ISD::ATOMIC_LOAD_XOR:
8466 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
8468 case ISD::ATOMIC_SWAP:
8469 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
8474 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
8476 default: return NULL;
8477 case X86ISD::BSF: return "X86ISD::BSF";
8478 case X86ISD::BSR: return "X86ISD::BSR";
8479 case X86ISD::SHLD: return "X86ISD::SHLD";
8480 case X86ISD::SHRD: return "X86ISD::SHRD";
8481 case X86ISD::FAND: return "X86ISD::FAND";
8482 case X86ISD::FOR: return "X86ISD::FOR";
8483 case X86ISD::FXOR: return "X86ISD::FXOR";
8484 case X86ISD::FSRL: return "X86ISD::FSRL";
8485 case X86ISD::FILD: return "X86ISD::FILD";
8486 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
8487 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
8488 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
8489 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
8490 case X86ISD::FLD: return "X86ISD::FLD";
8491 case X86ISD::FST: return "X86ISD::FST";
8492 case X86ISD::CALL: return "X86ISD::CALL";
8493 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
8494 case X86ISD::BT: return "X86ISD::BT";
8495 case X86ISD::CMP: return "X86ISD::CMP";
8496 case X86ISD::COMI: return "X86ISD::COMI";
8497 case X86ISD::UCOMI: return "X86ISD::UCOMI";
8498 case X86ISD::SETCC: return "X86ISD::SETCC";
8499 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
8500 case X86ISD::CMOV: return "X86ISD::CMOV";
8501 case X86ISD::BRCOND: return "X86ISD::BRCOND";
8502 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
8503 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
8504 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
8505 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
8506 case X86ISD::Wrapper: return "X86ISD::Wrapper";
8507 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
8508 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
8509 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
8510 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
8511 case X86ISD::PINSRB: return "X86ISD::PINSRB";
8512 case X86ISD::PINSRW: return "X86ISD::PINSRW";
8513 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
8514 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
8515 case X86ISD::FMAX: return "X86ISD::FMAX";
8516 case X86ISD::FMIN: return "X86ISD::FMIN";
8517 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
8518 case X86ISD::FRCP: return "X86ISD::FRCP";
8519 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
8520 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
8521 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
8522 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
8523 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
8524 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
8525 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
8526 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
8527 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
8528 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
8529 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
8530 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
8531 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
8532 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
8533 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
8534 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
8535 case X86ISD::VSHL: return "X86ISD::VSHL";
8536 case X86ISD::VSRL: return "X86ISD::VSRL";
8537 case X86ISD::CMPPD: return "X86ISD::CMPPD";
8538 case X86ISD::CMPPS: return "X86ISD::CMPPS";
8539 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
8540 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
8541 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
8542 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
8543 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
8544 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
8545 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
8546 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
8547 case X86ISD::ADD: return "X86ISD::ADD";
8548 case X86ISD::SUB: return "X86ISD::SUB";
8549 case X86ISD::SMUL: return "X86ISD::SMUL";
8550 case X86ISD::UMUL: return "X86ISD::UMUL";
8551 case X86ISD::INC: return "X86ISD::INC";
8552 case X86ISD::DEC: return "X86ISD::DEC";
8553 case X86ISD::OR: return "X86ISD::OR";
8554 case X86ISD::XOR: return "X86ISD::XOR";
8555 case X86ISD::AND: return "X86ISD::AND";
8556 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
8557 case X86ISD::PTEST: return "X86ISD::PTEST";
8558 case X86ISD::TESTP: return "X86ISD::TESTP";
8559 case X86ISD::PALIGN: return "X86ISD::PALIGN";
8560 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
8561 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
8562 case X86ISD::PSHUFHW_LD: return "X86ISD::PSHUFHW_LD";
8563 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
8564 case X86ISD::PSHUFLW_LD: return "X86ISD::PSHUFLW_LD";
8565 case X86ISD::SHUFPS: return "X86ISD::SHUFPS";
8566 case X86ISD::SHUFPD: return "X86ISD::SHUFPD";
8567 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
8568 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
8569 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
8570 case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
8571 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
8572 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
8573 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
8574 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
8575 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
8576 case X86ISD::MOVSHDUP_LD: return "X86ISD::MOVSHDUP_LD";
8577 case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
8578 case X86ISD::MOVSD: return "X86ISD::MOVSD";
8579 case X86ISD::MOVSS: return "X86ISD::MOVSS";
8580 case X86ISD::UNPCKLPS: return "X86ISD::UNPCKLPS";
8581 case X86ISD::UNPCKLPD: return "X86ISD::UNPCKLPD";
8582 case X86ISD::UNPCKHPS: return "X86ISD::UNPCKHPS";
8583 case X86ISD::UNPCKHPD: return "X86ISD::UNPCKHPD";
8584 case X86ISD::PUNPCKLBW: return "X86ISD::PUNPCKLBW";
8585 case X86ISD::PUNPCKLWD: return "X86ISD::PUNPCKLWD";
8586 case X86ISD::PUNPCKLDQ: return "X86ISD::PUNPCKLDQ";
8587 case X86ISD::PUNPCKLQDQ: return "X86ISD::PUNPCKLQDQ";
8588 case X86ISD::PUNPCKHBW: return "X86ISD::PUNPCKHBW";
8589 case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
8590 case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
8591 case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
8592 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
8593 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
8597 // isLegalAddressingMode - Return true if the addressing mode represented
8598 // by AM is legal for this target, for a load/store of the specified type.
8599 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
8600 const Type *Ty) const {
8601 // X86 supports extremely general addressing modes.
8602 CodeModel::Model M = getTargetMachine().getCodeModel();
8603 Reloc::Model R = getTargetMachine().getRelocationModel();
8605 // X86 allows a sign-extended 32-bit immediate field as a displacement.
8606 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
8611 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
8613 // If a reference to this global requires an extra load, we can't fold it.
8614 if (isGlobalStubReference(GVFlags))
8617 // If BaseGV requires a register for the PIC base, we cannot also have a
8618 // BaseReg specified.
8619 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
8622 // If lower 4G is not available, then we must use rip-relative addressing.
8623 if ((M != CodeModel::Small || R != Reloc::Static) &&
8624 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
8634 // These scales always work.
8639 // These scales are formed with basereg+scalereg. Only accept if there is
8644 default: // Other stuff never works.
8652 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
8653 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
8655 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
8656 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
8657 if (NumBits1 <= NumBits2)
8662 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
8663 if (!VT1.isInteger() || !VT2.isInteger())
8665 unsigned NumBits1 = VT1.getSizeInBits();
8666 unsigned NumBits2 = VT2.getSizeInBits();
8667 if (NumBits1 <= NumBits2)
8672 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
8673 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
8674 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
8677 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
8678 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
8679 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
8682 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
8683 // i16 instructions are longer (0x66 prefix) and potentially slower.
8684 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
8687 /// isShuffleMaskLegal - Targets can use this to indicate that they only
8688 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
8689 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
8690 /// are assumed to be legal.
8692 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
8694 // Very little shuffling can be done for 64-bit vectors right now.
8695 if (VT.getSizeInBits() == 64)
8696 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
8698 // FIXME: pshufb, blends, shifts.
8699 return (VT.getVectorNumElements() == 2 ||
8700 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
8701 isMOVLMask(M, VT) ||
8702 isSHUFPMask(M, VT) ||
8703 isPSHUFDMask(M, VT) ||
8704 isPSHUFHWMask(M, VT) ||
8705 isPSHUFLWMask(M, VT) ||
8706 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
8707 isUNPCKLMask(M, VT) ||
8708 isUNPCKHMask(M, VT) ||
8709 isUNPCKL_v_undef_Mask(M, VT) ||
8710 isUNPCKH_v_undef_Mask(M, VT));
8714 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
8716 unsigned NumElts = VT.getVectorNumElements();
8717 // FIXME: This collection of masks seems suspect.
8720 if (NumElts == 4 && VT.getSizeInBits() == 128) {
8721 return (isMOVLMask(Mask, VT) ||
8722 isCommutedMOVLMask(Mask, VT, true) ||
8723 isSHUFPMask(Mask, VT) ||
8724 isCommutedSHUFPMask(Mask, VT));
8729 //===----------------------------------------------------------------------===//
8730 // X86 Scheduler Hooks
8731 //===----------------------------------------------------------------------===//
8733 // private utility function
8735 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
8736 MachineBasicBlock *MBB,
8743 TargetRegisterClass *RC,
8744 bool invSrc) const {
8745 // For the atomic bitwise operator, we generate
8748 // ld t1 = [bitinstr.addr]
8749 // op t2 = t1, [bitinstr.val]
8751 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8753 // fallthrough -->nextMBB
8754 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8755 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8756 MachineFunction::iterator MBBIter = MBB;
8759 /// First build the CFG
8760 MachineFunction *F = MBB->getParent();
8761 MachineBasicBlock *thisMBB = MBB;
8762 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8763 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8764 F->insert(MBBIter, newMBB);
8765 F->insert(MBBIter, nextMBB);
8767 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8768 nextMBB->splice(nextMBB->begin(), thisMBB,
8769 llvm::next(MachineBasicBlock::iterator(bInstr)),
8771 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8773 // Update thisMBB to fall through to newMBB
8774 thisMBB->addSuccessor(newMBB);
8776 // newMBB jumps to itself and fall through to nextMBB
8777 newMBB->addSuccessor(nextMBB);
8778 newMBB->addSuccessor(newMBB);
8780 // Insert instructions into newMBB based on incoming instruction
8781 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
8782 "unexpected number of operands");
8783 DebugLoc dl = bInstr->getDebugLoc();
8784 MachineOperand& destOper = bInstr->getOperand(0);
8785 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8786 int numArgs = bInstr->getNumOperands() - 1;
8787 for (int i=0; i < numArgs; ++i)
8788 argOpers[i] = &bInstr->getOperand(i+1);
8790 // x86 address has 4 operands: base, index, scale, and displacement
8791 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8792 int valArgIndx = lastAddrIndx + 1;
8794 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8795 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8796 for (int i=0; i <= lastAddrIndx; ++i)
8797 (*MIB).addOperand(*argOpers[i]);
8799 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8801 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8806 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8807 assert((argOpers[valArgIndx]->isReg() ||
8808 argOpers[valArgIndx]->isImm()) &&
8810 if (argOpers[valArgIndx]->isReg())
8811 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8813 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8815 (*MIB).addOperand(*argOpers[valArgIndx]);
8817 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
8820 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8821 for (int i=0; i <= lastAddrIndx; ++i)
8822 (*MIB).addOperand(*argOpers[i]);
8824 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8825 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8826 bInstr->memoperands_end());
8828 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
8832 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8834 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8838 // private utility function: 64 bit atomics on 32 bit host.
8840 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8841 MachineBasicBlock *MBB,
8846 bool invSrc) const {
8847 // For the atomic bitwise operator, we generate
8848 // thisMBB (instructions are in pairs, except cmpxchg8b)
8849 // ld t1,t2 = [bitinstr.addr]
8851 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8852 // op t5, t6 <- out1, out2, [bitinstr.val]
8853 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8854 // mov ECX, EBX <- t5, t6
8855 // mov EAX, EDX <- t1, t2
8856 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8857 // mov t3, t4 <- EAX, EDX
8859 // result in out1, out2
8860 // fallthrough -->nextMBB
8862 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8863 const unsigned LoadOpc = X86::MOV32rm;
8864 const unsigned NotOpc = X86::NOT32r;
8865 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8866 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8867 MachineFunction::iterator MBBIter = MBB;
8870 /// First build the CFG
8871 MachineFunction *F = MBB->getParent();
8872 MachineBasicBlock *thisMBB = MBB;
8873 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8874 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8875 F->insert(MBBIter, newMBB);
8876 F->insert(MBBIter, nextMBB);
8878 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8879 nextMBB->splice(nextMBB->begin(), thisMBB,
8880 llvm::next(MachineBasicBlock::iterator(bInstr)),
8882 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8884 // Update thisMBB to fall through to newMBB
8885 thisMBB->addSuccessor(newMBB);
8887 // newMBB jumps to itself and fall through to nextMBB
8888 newMBB->addSuccessor(nextMBB);
8889 newMBB->addSuccessor(newMBB);
8891 DebugLoc dl = bInstr->getDebugLoc();
8892 // Insert instructions into newMBB based on incoming instruction
8893 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8894 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
8895 "unexpected number of operands");
8896 MachineOperand& dest1Oper = bInstr->getOperand(0);
8897 MachineOperand& dest2Oper = bInstr->getOperand(1);
8898 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8899 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
8900 argOpers[i] = &bInstr->getOperand(i+2);
8902 // We use some of the operands multiple times, so conservatively just
8903 // clear any kill flags that might be present.
8904 if (argOpers[i]->isReg() && argOpers[i]->isUse())
8905 argOpers[i]->setIsKill(false);
8908 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8909 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8911 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8912 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8913 for (int i=0; i <= lastAddrIndx; ++i)
8914 (*MIB).addOperand(*argOpers[i]);
8915 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8916 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8917 // add 4 to displacement.
8918 for (int i=0; i <= lastAddrIndx-2; ++i)
8919 (*MIB).addOperand(*argOpers[i]);
8920 MachineOperand newOp3 = *(argOpers[3]);
8922 newOp3.setImm(newOp3.getImm()+4);
8924 newOp3.setOffset(newOp3.getOffset()+4);
8925 (*MIB).addOperand(newOp3);
8926 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8928 // t3/4 are defined later, at the bottom of the loop
8929 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8930 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8931 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8932 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8933 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8934 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8936 // The subsequent operations should be using the destination registers of
8937 //the PHI instructions.
8939 t1 = F->getRegInfo().createVirtualRegister(RC);
8940 t2 = F->getRegInfo().createVirtualRegister(RC);
8941 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8942 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8944 t1 = dest1Oper.getReg();
8945 t2 = dest2Oper.getReg();
8948 int valArgIndx = lastAddrIndx + 1;
8949 assert((argOpers[valArgIndx]->isReg() ||
8950 argOpers[valArgIndx]->isImm()) &&
8952 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8953 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8954 if (argOpers[valArgIndx]->isReg())
8955 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8957 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8958 if (regOpcL != X86::MOV32rr)
8960 (*MIB).addOperand(*argOpers[valArgIndx]);
8961 assert(argOpers[valArgIndx + 1]->isReg() ==
8962 argOpers[valArgIndx]->isReg());
8963 assert(argOpers[valArgIndx + 1]->isImm() ==
8964 argOpers[valArgIndx]->isImm());
8965 if (argOpers[valArgIndx + 1]->isReg())
8966 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8968 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8969 if (regOpcH != X86::MOV32rr)
8971 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8973 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
8975 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
8978 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
8980 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
8983 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8984 for (int i=0; i <= lastAddrIndx; ++i)
8985 (*MIB).addOperand(*argOpers[i]);
8987 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8988 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8989 bInstr->memoperands_end());
8991 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
8992 MIB.addReg(X86::EAX);
8993 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
8994 MIB.addReg(X86::EDX);
8997 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8999 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
9003 // private utility function
9005 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
9006 MachineBasicBlock *MBB,
9007 unsigned cmovOpc) const {
9008 // For the atomic min/max operator, we generate
9011 // ld t1 = [min/max.addr]
9012 // mov t2 = [min/max.val]
9014 // cmov[cond] t2 = t1
9016 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
9018 // fallthrough -->nextMBB
9020 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9021 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9022 MachineFunction::iterator MBBIter = MBB;
9025 /// First build the CFG
9026 MachineFunction *F = MBB->getParent();
9027 MachineBasicBlock *thisMBB = MBB;
9028 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9029 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9030 F->insert(MBBIter, newMBB);
9031 F->insert(MBBIter, nextMBB);
9033 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9034 nextMBB->splice(nextMBB->begin(), thisMBB,
9035 llvm::next(MachineBasicBlock::iterator(mInstr)),
9037 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9039 // Update thisMBB to fall through to newMBB
9040 thisMBB->addSuccessor(newMBB);
9042 // newMBB jumps to newMBB and fall through to nextMBB
9043 newMBB->addSuccessor(nextMBB);
9044 newMBB->addSuccessor(newMBB);
9046 DebugLoc dl = mInstr->getDebugLoc();
9047 // Insert instructions into newMBB based on incoming instruction
9048 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
9049 "unexpected number of operands");
9050 MachineOperand& destOper = mInstr->getOperand(0);
9051 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9052 int numArgs = mInstr->getNumOperands() - 1;
9053 for (int i=0; i < numArgs; ++i)
9054 argOpers[i] = &mInstr->getOperand(i+1);
9056 // x86 address has 4 operands: base, index, scale, and displacement
9057 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9058 int valArgIndx = lastAddrIndx + 1;
9060 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9061 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
9062 for (int i=0; i <= lastAddrIndx; ++i)
9063 (*MIB).addOperand(*argOpers[i]);
9065 // We only support register and immediate values
9066 assert((argOpers[valArgIndx]->isReg() ||
9067 argOpers[valArgIndx]->isImm()) &&
9070 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9071 if (argOpers[valArgIndx]->isReg())
9072 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
9074 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
9075 (*MIB).addOperand(*argOpers[valArgIndx]);
9077 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
9080 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
9085 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9086 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
9090 // Cmp and exchange if none has modified the memory location
9091 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
9092 for (int i=0; i <= lastAddrIndx; ++i)
9093 (*MIB).addOperand(*argOpers[i]);
9095 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9096 (*MIB).setMemRefs(mInstr->memoperands_begin(),
9097 mInstr->memoperands_end());
9099 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
9100 MIB.addReg(X86::EAX);
9103 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9105 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
9109 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
9110 // or XMM0_V32I8 in AVX all of this code can be replaced with that
9113 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
9114 unsigned numArgs, bool memArg) const {
9116 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
9117 "Target must have SSE4.2 or AVX features enabled");
9119 DebugLoc dl = MI->getDebugLoc();
9120 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9124 if (!Subtarget->hasAVX()) {
9126 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
9128 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
9131 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
9133 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
9136 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
9138 for (unsigned i = 0; i < numArgs; ++i) {
9139 MachineOperand &Op = MI->getOperand(i+1);
9141 if (!(Op.isReg() && Op.isImplicit()))
9145 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
9148 MI->eraseFromParent();
9154 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
9156 MachineBasicBlock *MBB) const {
9157 // Emit code to save XMM registers to the stack. The ABI says that the
9158 // number of registers to save is given in %al, so it's theoretically
9159 // possible to do an indirect jump trick to avoid saving all of them,
9160 // however this code takes a simpler approach and just executes all
9161 // of the stores if %al is non-zero. It's less code, and it's probably
9162 // easier on the hardware branch predictor, and stores aren't all that
9163 // expensive anyway.
9165 // Create the new basic blocks. One block contains all the XMM stores,
9166 // and one block is the final destination regardless of whether any
9167 // stores were performed.
9168 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9169 MachineFunction *F = MBB->getParent();
9170 MachineFunction::iterator MBBIter = MBB;
9172 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
9173 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
9174 F->insert(MBBIter, XMMSaveMBB);
9175 F->insert(MBBIter, EndMBB);
9177 // Transfer the remainder of MBB and its successor edges to EndMBB.
9178 EndMBB->splice(EndMBB->begin(), MBB,
9179 llvm::next(MachineBasicBlock::iterator(MI)),
9181 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
9183 // The original block will now fall through to the XMM save block.
9184 MBB->addSuccessor(XMMSaveMBB);
9185 // The XMMSaveMBB will fall through to the end block.
9186 XMMSaveMBB->addSuccessor(EndMBB);
9188 // Now add the instructions.
9189 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9190 DebugLoc DL = MI->getDebugLoc();
9192 unsigned CountReg = MI->getOperand(0).getReg();
9193 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
9194 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
9196 if (!Subtarget->isTargetWin64()) {
9197 // If %al is 0, branch around the XMM save block.
9198 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
9199 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
9200 MBB->addSuccessor(EndMBB);
9203 // In the XMM save block, save all the XMM argument registers.
9204 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
9205 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
9206 MachineMemOperand *MMO =
9207 F->getMachineMemOperand(
9208 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
9209 MachineMemOperand::MOStore, Offset,
9210 /*Size=*/16, /*Align=*/16);
9211 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
9212 .addFrameIndex(RegSaveFrameIndex)
9213 .addImm(/*Scale=*/1)
9214 .addReg(/*IndexReg=*/0)
9215 .addImm(/*Disp=*/Offset)
9216 .addReg(/*Segment=*/0)
9217 .addReg(MI->getOperand(i).getReg())
9218 .addMemOperand(MMO);
9221 MI->eraseFromParent(); // The pseudo instruction is gone now.
9227 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
9228 MachineBasicBlock *BB) const {
9229 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9230 DebugLoc DL = MI->getDebugLoc();
9232 // To "insert" a SELECT_CC instruction, we actually have to insert the
9233 // diamond control-flow pattern. The incoming instruction knows the
9234 // destination vreg to set, the condition code register to branch on, the
9235 // true/false values to select between, and a branch opcode to use.
9236 const BasicBlock *LLVM_BB = BB->getBasicBlock();
9237 MachineFunction::iterator It = BB;
9243 // cmpTY ccX, r1, r2
9245 // fallthrough --> copy0MBB
9246 MachineBasicBlock *thisMBB = BB;
9247 MachineFunction *F = BB->getParent();
9248 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
9249 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
9250 F->insert(It, copy0MBB);
9251 F->insert(It, sinkMBB);
9253 // If the EFLAGS register isn't dead in the terminator, then claim that it's
9254 // live into the sink and copy blocks.
9255 const MachineFunction *MF = BB->getParent();
9256 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
9257 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
9259 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
9260 const MachineOperand &MO = MI->getOperand(I);
9261 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
9262 unsigned Reg = MO.getReg();
9263 if (Reg != X86::EFLAGS) continue;
9264 copy0MBB->addLiveIn(Reg);
9265 sinkMBB->addLiveIn(Reg);
9268 // Transfer the remainder of BB and its successor edges to sinkMBB.
9269 sinkMBB->splice(sinkMBB->begin(), BB,
9270 llvm::next(MachineBasicBlock::iterator(MI)),
9272 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
9274 // Add the true and fallthrough blocks as its successors.
9275 BB->addSuccessor(copy0MBB);
9276 BB->addSuccessor(sinkMBB);
9278 // Create the conditional branch instruction.
9280 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
9281 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
9284 // %FalseValue = ...
9285 // # fallthrough to sinkMBB
9286 copy0MBB->addSuccessor(sinkMBB);
9289 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
9291 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
9292 TII->get(X86::PHI), MI->getOperand(0).getReg())
9293 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
9294 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
9296 MI->eraseFromParent(); // The pseudo instruction is gone now.
9301 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
9302 MachineBasicBlock *BB) const {
9303 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9304 DebugLoc DL = MI->getDebugLoc();
9306 // The lowering is pretty easy: we're just emitting the call to _alloca. The
9307 // non-trivial part is impdef of ESP.
9308 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
9311 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
9312 .addExternalSymbol("_alloca")
9313 .addReg(X86::EAX, RegState::Implicit)
9314 .addReg(X86::ESP, RegState::Implicit)
9315 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
9316 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
9317 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
9319 MI->eraseFromParent(); // The pseudo instruction is gone now.
9324 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
9325 MachineBasicBlock *BB) const {
9326 // This is pretty easy. We're taking the value that we received from
9327 // our load from the relocation, sticking it in either RDI (x86-64)
9328 // or EAX and doing an indirect call. The return value will then
9329 // be in the normal return register.
9330 const X86InstrInfo *TII
9331 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
9332 DebugLoc DL = MI->getDebugLoc();
9333 MachineFunction *F = BB->getParent();
9334 bool IsWin64 = Subtarget->isTargetWin64();
9336 assert(MI->getOperand(3).isGlobal() && "This should be a global");
9338 if (Subtarget->is64Bit()) {
9339 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
9340 TII->get(X86::MOV64rm), X86::RDI)
9342 .addImm(0).addReg(0)
9343 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
9344 MI->getOperand(3).getTargetFlags())
9346 MIB = BuildMI(*BB, MI, DL, TII->get(IsWin64 ? X86::WINCALL64m : X86::CALL64m));
9347 addDirectMem(MIB, X86::RDI);
9348 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
9349 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
9350 TII->get(X86::MOV32rm), X86::EAX)
9352 .addImm(0).addReg(0)
9353 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
9354 MI->getOperand(3).getTargetFlags())
9356 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
9357 addDirectMem(MIB, X86::EAX);
9359 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
9360 TII->get(X86::MOV32rm), X86::EAX)
9361 .addReg(TII->getGlobalBaseReg(F))
9362 .addImm(0).addReg(0)
9363 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
9364 MI->getOperand(3).getTargetFlags())
9366 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
9367 addDirectMem(MIB, X86::EAX);
9370 MI->eraseFromParent(); // The pseudo instruction is gone now.
9375 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
9376 MachineBasicBlock *BB) const {
9377 switch (MI->getOpcode()) {
9378 default: assert(false && "Unexpected instr type to insert");
9379 case X86::MINGW_ALLOCA:
9380 return EmitLoweredMingwAlloca(MI, BB);
9381 case X86::TLSCall_32:
9382 case X86::TLSCall_64:
9383 return EmitLoweredTLSCall(MI, BB);
9385 case X86::CMOV_V1I64:
9386 case X86::CMOV_FR32:
9387 case X86::CMOV_FR64:
9388 case X86::CMOV_V4F32:
9389 case X86::CMOV_V2F64:
9390 case X86::CMOV_V2I64:
9391 case X86::CMOV_GR16:
9392 case X86::CMOV_GR32:
9393 case X86::CMOV_RFP32:
9394 case X86::CMOV_RFP64:
9395 case X86::CMOV_RFP80:
9396 return EmitLoweredSelect(MI, BB);
9398 case X86::FP32_TO_INT16_IN_MEM:
9399 case X86::FP32_TO_INT32_IN_MEM:
9400 case X86::FP32_TO_INT64_IN_MEM:
9401 case X86::FP64_TO_INT16_IN_MEM:
9402 case X86::FP64_TO_INT32_IN_MEM:
9403 case X86::FP64_TO_INT64_IN_MEM:
9404 case X86::FP80_TO_INT16_IN_MEM:
9405 case X86::FP80_TO_INT32_IN_MEM:
9406 case X86::FP80_TO_INT64_IN_MEM: {
9407 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9408 DebugLoc DL = MI->getDebugLoc();
9410 // Change the floating point control register to use "round towards zero"
9411 // mode when truncating to an integer value.
9412 MachineFunction *F = BB->getParent();
9413 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
9414 addFrameReference(BuildMI(*BB, MI, DL,
9415 TII->get(X86::FNSTCW16m)), CWFrameIdx);
9417 // Load the old value of the high byte of the control word...
9419 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
9420 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
9423 // Set the high part to be round to zero...
9424 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
9427 // Reload the modified control word now...
9428 addFrameReference(BuildMI(*BB, MI, DL,
9429 TII->get(X86::FLDCW16m)), CWFrameIdx);
9431 // Restore the memory image of control word to original value
9432 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
9435 // Get the X86 opcode to use.
9437 switch (MI->getOpcode()) {
9438 default: llvm_unreachable("illegal opcode!");
9439 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
9440 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
9441 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
9442 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
9443 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
9444 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
9445 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
9446 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
9447 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
9451 MachineOperand &Op = MI->getOperand(0);
9453 AM.BaseType = X86AddressMode::RegBase;
9454 AM.Base.Reg = Op.getReg();
9456 AM.BaseType = X86AddressMode::FrameIndexBase;
9457 AM.Base.FrameIndex = Op.getIndex();
9459 Op = MI->getOperand(1);
9461 AM.Scale = Op.getImm();
9462 Op = MI->getOperand(2);
9464 AM.IndexReg = Op.getImm();
9465 Op = MI->getOperand(3);
9466 if (Op.isGlobal()) {
9467 AM.GV = Op.getGlobal();
9469 AM.Disp = Op.getImm();
9471 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
9472 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
9474 // Reload the original control word now.
9475 addFrameReference(BuildMI(*BB, MI, DL,
9476 TII->get(X86::FLDCW16m)), CWFrameIdx);
9478 MI->eraseFromParent(); // The pseudo instruction is gone now.
9481 // String/text processing lowering.
9482 case X86::PCMPISTRM128REG:
9483 case X86::VPCMPISTRM128REG:
9484 return EmitPCMP(MI, BB, 3, false /* in-mem */);
9485 case X86::PCMPISTRM128MEM:
9486 case X86::VPCMPISTRM128MEM:
9487 return EmitPCMP(MI, BB, 3, true /* in-mem */);
9488 case X86::PCMPESTRM128REG:
9489 case X86::VPCMPESTRM128REG:
9490 return EmitPCMP(MI, BB, 5, false /* in mem */);
9491 case X86::PCMPESTRM128MEM:
9492 case X86::VPCMPESTRM128MEM:
9493 return EmitPCMP(MI, BB, 5, true /* in mem */);
9496 case X86::ATOMAND32:
9497 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
9498 X86::AND32ri, X86::MOV32rm,
9500 X86::NOT32r, X86::EAX,
9501 X86::GR32RegisterClass);
9503 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
9504 X86::OR32ri, X86::MOV32rm,
9506 X86::NOT32r, X86::EAX,
9507 X86::GR32RegisterClass);
9508 case X86::ATOMXOR32:
9509 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
9510 X86::XOR32ri, X86::MOV32rm,
9512 X86::NOT32r, X86::EAX,
9513 X86::GR32RegisterClass);
9514 case X86::ATOMNAND32:
9515 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
9516 X86::AND32ri, X86::MOV32rm,
9518 X86::NOT32r, X86::EAX,
9519 X86::GR32RegisterClass, true);
9520 case X86::ATOMMIN32:
9521 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
9522 case X86::ATOMMAX32:
9523 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
9524 case X86::ATOMUMIN32:
9525 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
9526 case X86::ATOMUMAX32:
9527 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
9529 case X86::ATOMAND16:
9530 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
9531 X86::AND16ri, X86::MOV16rm,
9533 X86::NOT16r, X86::AX,
9534 X86::GR16RegisterClass);
9536 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
9537 X86::OR16ri, X86::MOV16rm,
9539 X86::NOT16r, X86::AX,
9540 X86::GR16RegisterClass);
9541 case X86::ATOMXOR16:
9542 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
9543 X86::XOR16ri, X86::MOV16rm,
9545 X86::NOT16r, X86::AX,
9546 X86::GR16RegisterClass);
9547 case X86::ATOMNAND16:
9548 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
9549 X86::AND16ri, X86::MOV16rm,
9551 X86::NOT16r, X86::AX,
9552 X86::GR16RegisterClass, true);
9553 case X86::ATOMMIN16:
9554 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
9555 case X86::ATOMMAX16:
9556 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
9557 case X86::ATOMUMIN16:
9558 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
9559 case X86::ATOMUMAX16:
9560 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
9563 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
9564 X86::AND8ri, X86::MOV8rm,
9566 X86::NOT8r, X86::AL,
9567 X86::GR8RegisterClass);
9569 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
9570 X86::OR8ri, X86::MOV8rm,
9572 X86::NOT8r, X86::AL,
9573 X86::GR8RegisterClass);
9575 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
9576 X86::XOR8ri, X86::MOV8rm,
9578 X86::NOT8r, X86::AL,
9579 X86::GR8RegisterClass);
9580 case X86::ATOMNAND8:
9581 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
9582 X86::AND8ri, X86::MOV8rm,
9584 X86::NOT8r, X86::AL,
9585 X86::GR8RegisterClass, true);
9586 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
9587 // This group is for 64-bit host.
9588 case X86::ATOMAND64:
9589 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
9590 X86::AND64ri32, X86::MOV64rm,
9592 X86::NOT64r, X86::RAX,
9593 X86::GR64RegisterClass);
9595 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
9596 X86::OR64ri32, X86::MOV64rm,
9598 X86::NOT64r, X86::RAX,
9599 X86::GR64RegisterClass);
9600 case X86::ATOMXOR64:
9601 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
9602 X86::XOR64ri32, X86::MOV64rm,
9604 X86::NOT64r, X86::RAX,
9605 X86::GR64RegisterClass);
9606 case X86::ATOMNAND64:
9607 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
9608 X86::AND64ri32, X86::MOV64rm,
9610 X86::NOT64r, X86::RAX,
9611 X86::GR64RegisterClass, true);
9612 case X86::ATOMMIN64:
9613 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
9614 case X86::ATOMMAX64:
9615 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
9616 case X86::ATOMUMIN64:
9617 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
9618 case X86::ATOMUMAX64:
9619 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
9621 // This group does 64-bit operations on a 32-bit host.
9622 case X86::ATOMAND6432:
9623 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9624 X86::AND32rr, X86::AND32rr,
9625 X86::AND32ri, X86::AND32ri,
9627 case X86::ATOMOR6432:
9628 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9629 X86::OR32rr, X86::OR32rr,
9630 X86::OR32ri, X86::OR32ri,
9632 case X86::ATOMXOR6432:
9633 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9634 X86::XOR32rr, X86::XOR32rr,
9635 X86::XOR32ri, X86::XOR32ri,
9637 case X86::ATOMNAND6432:
9638 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9639 X86::AND32rr, X86::AND32rr,
9640 X86::AND32ri, X86::AND32ri,
9642 case X86::ATOMADD6432:
9643 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9644 X86::ADD32rr, X86::ADC32rr,
9645 X86::ADD32ri, X86::ADC32ri,
9647 case X86::ATOMSUB6432:
9648 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9649 X86::SUB32rr, X86::SBB32rr,
9650 X86::SUB32ri, X86::SBB32ri,
9652 case X86::ATOMSWAP6432:
9653 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9654 X86::MOV32rr, X86::MOV32rr,
9655 X86::MOV32ri, X86::MOV32ri,
9657 case X86::VASTART_SAVE_XMM_REGS:
9658 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
9662 //===----------------------------------------------------------------------===//
9663 // X86 Optimization Hooks
9664 //===----------------------------------------------------------------------===//
9666 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9670 const SelectionDAG &DAG,
9671 unsigned Depth) const {
9672 unsigned Opc = Op.getOpcode();
9673 assert((Opc >= ISD::BUILTIN_OP_END ||
9674 Opc == ISD::INTRINSIC_WO_CHAIN ||
9675 Opc == ISD::INTRINSIC_W_CHAIN ||
9676 Opc == ISD::INTRINSIC_VOID) &&
9677 "Should use MaskedValueIsZero if you don't know whether Op"
9678 " is a target node!");
9680 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
9692 // These nodes' second result is a boolean.
9693 if (Op.getResNo() == 0)
9697 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
9698 Mask.getBitWidth() - 1);
9703 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
9704 /// node is a GlobalAddress + offset.
9705 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
9706 const GlobalValue* &GA,
9707 int64_t &Offset) const {
9708 if (N->getOpcode() == X86ISD::Wrapper) {
9709 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
9710 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
9711 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
9715 return TargetLowering::isGAPlusOffset(N, GA, Offset);
9718 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
9719 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
9720 /// if the load addresses are consecutive, non-overlapping, and in the right
9722 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
9723 const TargetLowering &TLI) {
9724 DebugLoc dl = N->getDebugLoc();
9725 EVT VT = N->getValueType(0);
9727 if (VT.getSizeInBits() != 128)
9730 SmallVector<SDValue, 16> Elts;
9731 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
9732 Elts.push_back(getShuffleScalarElt(N, i, DAG));
9734 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
9737 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
9738 /// generation and convert it from being a bunch of shuffles and extracts
9739 /// to a simple store and scalar loads to extract the elements.
9740 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
9741 const TargetLowering &TLI) {
9742 SDValue InputVector = N->getOperand(0);
9744 // Only operate on vectors of 4 elements, where the alternative shuffling
9745 // gets to be more expensive.
9746 if (InputVector.getValueType() != MVT::v4i32)
9749 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
9750 // single use which is a sign-extend or zero-extend, and all elements are
9752 SmallVector<SDNode *, 4> Uses;
9753 unsigned ExtractedElements = 0;
9754 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
9755 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
9756 if (UI.getUse().getResNo() != InputVector.getResNo())
9759 SDNode *Extract = *UI;
9760 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
9763 if (Extract->getValueType(0) != MVT::i32)
9765 if (!Extract->hasOneUse())
9767 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
9768 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
9770 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
9773 // Record which element was extracted.
9774 ExtractedElements |=
9775 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
9777 Uses.push_back(Extract);
9780 // If not all the elements were used, this may not be worthwhile.
9781 if (ExtractedElements != 15)
9784 // Ok, we've now decided to do the transformation.
9785 DebugLoc dl = InputVector.getDebugLoc();
9787 // Store the value to a temporary stack slot.
9788 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
9789 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL,
9790 0, false, false, 0);
9792 // Replace each use (extract) with a load of the appropriate element.
9793 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
9794 UE = Uses.end(); UI != UE; ++UI) {
9795 SDNode *Extract = *UI;
9797 // Compute the element's address.
9798 SDValue Idx = Extract->getOperand(1);
9800 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
9801 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
9802 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
9804 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(),
9805 OffsetVal, StackPtr);
9808 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
9809 ScalarAddr, NULL, 0, false, false, 0);
9811 // Replace the exact with the load.
9812 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
9815 // The replacement was made in place; don't return anything.
9819 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
9820 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
9821 const X86Subtarget *Subtarget) {
9822 DebugLoc DL = N->getDebugLoc();
9823 SDValue Cond = N->getOperand(0);
9824 // Get the LHS/RHS of the select.
9825 SDValue LHS = N->getOperand(1);
9826 SDValue RHS = N->getOperand(2);
9828 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9829 // instructions match the semantics of the common C idiom x<y?x:y but not
9830 // x<=y?x:y, because of how they handle negative zero (which can be
9831 // ignored in unsafe-math mode).
9832 if (Subtarget->hasSSE2() &&
9833 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9834 Cond.getOpcode() == ISD::SETCC) {
9835 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9837 unsigned Opcode = 0;
9838 // Check for x CC y ? x : y.
9839 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9840 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9844 // Converting this to a min would handle NaNs incorrectly, and swapping
9845 // the operands would cause it to handle comparisons between positive
9846 // and negative zero incorrectly.
9847 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
9848 if (!UnsafeFPMath &&
9849 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9851 std::swap(LHS, RHS);
9853 Opcode = X86ISD::FMIN;
9856 // Converting this to a min would handle comparisons between positive
9857 // and negative zero incorrectly.
9858 if (!UnsafeFPMath &&
9859 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9861 Opcode = X86ISD::FMIN;
9864 // Converting this to a min would handle both negative zeros and NaNs
9865 // incorrectly, but we can swap the operands to fix both.
9866 std::swap(LHS, RHS);
9870 Opcode = X86ISD::FMIN;
9874 // Converting this to a max would handle comparisons between positive
9875 // and negative zero incorrectly.
9876 if (!UnsafeFPMath &&
9877 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9879 Opcode = X86ISD::FMAX;
9882 // Converting this to a max would handle NaNs incorrectly, and swapping
9883 // the operands would cause it to handle comparisons between positive
9884 // and negative zero incorrectly.
9885 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
9886 if (!UnsafeFPMath &&
9887 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9889 std::swap(LHS, RHS);
9891 Opcode = X86ISD::FMAX;
9894 // Converting this to a max would handle both negative zeros and NaNs
9895 // incorrectly, but we can swap the operands to fix both.
9896 std::swap(LHS, RHS);
9900 Opcode = X86ISD::FMAX;
9903 // Check for x CC y ? y : x -- a min/max with reversed arms.
9904 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9905 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9909 // Converting this to a min would handle comparisons between positive
9910 // and negative zero incorrectly, and swapping the operands would
9911 // cause it to handle NaNs incorrectly.
9912 if (!UnsafeFPMath &&
9913 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9914 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9916 std::swap(LHS, RHS);
9918 Opcode = X86ISD::FMIN;
9921 // Converting this to a min would handle NaNs incorrectly.
9922 if (!UnsafeFPMath &&
9923 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9925 Opcode = X86ISD::FMIN;
9928 // Converting this to a min would handle both negative zeros and NaNs
9929 // incorrectly, but we can swap the operands to fix both.
9930 std::swap(LHS, RHS);
9934 Opcode = X86ISD::FMIN;
9938 // Converting this to a max would handle NaNs incorrectly.
9939 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9941 Opcode = X86ISD::FMAX;
9944 // Converting this to a max would handle comparisons between positive
9945 // and negative zero incorrectly, and swapping the operands would
9946 // cause it to handle NaNs incorrectly.
9947 if (!UnsafeFPMath &&
9948 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9949 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9951 std::swap(LHS, RHS);
9953 Opcode = X86ISD::FMAX;
9956 // Converting this to a max would handle both negative zeros and NaNs
9957 // incorrectly, but we can swap the operands to fix both.
9958 std::swap(LHS, RHS);
9962 Opcode = X86ISD::FMAX;
9968 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9971 // If this is a select between two integer constants, try to do some
9973 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9974 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9975 // Don't do this for crazy integer types.
9976 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9977 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9978 // so that TrueC (the true value) is larger than FalseC.
9979 bool NeedsCondInvert = false;
9981 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9982 // Efficiently invertible.
9983 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9984 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9985 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9986 NeedsCondInvert = true;
9987 std::swap(TrueC, FalseC);
9990 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9991 if (FalseC->getAPIntValue() == 0 &&
9992 TrueC->getAPIntValue().isPowerOf2()) {
9993 if (NeedsCondInvert) // Invert the condition if needed.
9994 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9995 DAG.getConstant(1, Cond.getValueType()));
9997 // Zero extend the condition if needed.
9998 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
10000 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
10001 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
10002 DAG.getConstant(ShAmt, MVT::i8));
10005 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
10006 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
10007 if (NeedsCondInvert) // Invert the condition if needed.
10008 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
10009 DAG.getConstant(1, Cond.getValueType()));
10011 // Zero extend the condition if needed.
10012 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
10013 FalseC->getValueType(0), Cond);
10014 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10015 SDValue(FalseC, 0));
10018 // Optimize cases that will turn into an LEA instruction. This requires
10019 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
10020 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
10021 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
10022 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
10024 bool isFastMultiplier = false;
10026 switch ((unsigned char)Diff) {
10028 case 1: // result = add base, cond
10029 case 2: // result = lea base( , cond*2)
10030 case 3: // result = lea base(cond, cond*2)
10031 case 4: // result = lea base( , cond*4)
10032 case 5: // result = lea base(cond, cond*4)
10033 case 8: // result = lea base( , cond*8)
10034 case 9: // result = lea base(cond, cond*8)
10035 isFastMultiplier = true;
10040 if (isFastMultiplier) {
10041 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
10042 if (NeedsCondInvert) // Invert the condition if needed.
10043 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
10044 DAG.getConstant(1, Cond.getValueType()));
10046 // Zero extend the condition if needed.
10047 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
10049 // Scale the condition by the difference.
10051 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
10052 DAG.getConstant(Diff, Cond.getValueType()));
10054 // Add the base if non-zero.
10055 if (FalseC->getAPIntValue() != 0)
10056 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10057 SDValue(FalseC, 0));
10067 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
10068 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
10069 TargetLowering::DAGCombinerInfo &DCI) {
10070 DebugLoc DL = N->getDebugLoc();
10072 // If the flag operand isn't dead, don't touch this CMOV.
10073 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
10076 // If this is a select between two integer constants, try to do some
10077 // optimizations. Note that the operands are ordered the opposite of SELECT
10079 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
10080 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
10081 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
10082 // larger than FalseC (the false value).
10083 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
10085 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
10086 CC = X86::GetOppositeBranchCondition(CC);
10087 std::swap(TrueC, FalseC);
10090 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
10091 // This is efficient for any integer data type (including i8/i16) and
10093 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
10094 SDValue Cond = N->getOperand(3);
10095 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10096 DAG.getConstant(CC, MVT::i8), Cond);
10098 // Zero extend the condition if needed.
10099 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
10101 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
10102 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
10103 DAG.getConstant(ShAmt, MVT::i8));
10104 if (N->getNumValues() == 2) // Dead flag value?
10105 return DCI.CombineTo(N, Cond, SDValue());
10109 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
10110 // for any integer data type, including i8/i16.
10111 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
10112 SDValue Cond = N->getOperand(3);
10113 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10114 DAG.getConstant(CC, MVT::i8), Cond);
10116 // Zero extend the condition if needed.
10117 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
10118 FalseC->getValueType(0), Cond);
10119 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10120 SDValue(FalseC, 0));
10122 if (N->getNumValues() == 2) // Dead flag value?
10123 return DCI.CombineTo(N, Cond, SDValue());
10127 // Optimize cases that will turn into an LEA instruction. This requires
10128 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
10129 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
10130 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
10131 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
10133 bool isFastMultiplier = false;
10135 switch ((unsigned char)Diff) {
10137 case 1: // result = add base, cond
10138 case 2: // result = lea base( , cond*2)
10139 case 3: // result = lea base(cond, cond*2)
10140 case 4: // result = lea base( , cond*4)
10141 case 5: // result = lea base(cond, cond*4)
10142 case 8: // result = lea base( , cond*8)
10143 case 9: // result = lea base(cond, cond*8)
10144 isFastMultiplier = true;
10149 if (isFastMultiplier) {
10150 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
10151 SDValue Cond = N->getOperand(3);
10152 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10153 DAG.getConstant(CC, MVT::i8), Cond);
10154 // Zero extend the condition if needed.
10155 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
10157 // Scale the condition by the difference.
10159 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
10160 DAG.getConstant(Diff, Cond.getValueType()));
10162 // Add the base if non-zero.
10163 if (FalseC->getAPIntValue() != 0)
10164 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10165 SDValue(FalseC, 0));
10166 if (N->getNumValues() == 2) // Dead flag value?
10167 return DCI.CombineTo(N, Cond, SDValue());
10177 /// PerformMulCombine - Optimize a single multiply with constant into two
10178 /// in order to implement it with two cheaper instructions, e.g.
10179 /// LEA + SHL, LEA + LEA.
10180 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
10181 TargetLowering::DAGCombinerInfo &DCI) {
10182 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
10185 EVT VT = N->getValueType(0);
10186 if (VT != MVT::i64)
10189 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
10192 uint64_t MulAmt = C->getZExtValue();
10193 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
10196 uint64_t MulAmt1 = 0;
10197 uint64_t MulAmt2 = 0;
10198 if ((MulAmt % 9) == 0) {
10200 MulAmt2 = MulAmt / 9;
10201 } else if ((MulAmt % 5) == 0) {
10203 MulAmt2 = MulAmt / 5;
10204 } else if ((MulAmt % 3) == 0) {
10206 MulAmt2 = MulAmt / 3;
10209 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
10210 DebugLoc DL = N->getDebugLoc();
10212 if (isPowerOf2_64(MulAmt2) &&
10213 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
10214 // If second multiplifer is pow2, issue it first. We want the multiply by
10215 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
10217 std::swap(MulAmt1, MulAmt2);
10220 if (isPowerOf2_64(MulAmt1))
10221 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
10222 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
10224 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
10225 DAG.getConstant(MulAmt1, VT));
10227 if (isPowerOf2_64(MulAmt2))
10228 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
10229 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
10231 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
10232 DAG.getConstant(MulAmt2, VT));
10234 // Do not add new nodes to DAG combiner worklist.
10235 DCI.CombineTo(N, NewMul, false);
10240 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
10241 SDValue N0 = N->getOperand(0);
10242 SDValue N1 = N->getOperand(1);
10243 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
10244 EVT VT = N0.getValueType();
10246 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
10247 // since the result of setcc_c is all zero's or all ones.
10248 if (N1C && N0.getOpcode() == ISD::AND &&
10249 N0.getOperand(1).getOpcode() == ISD::Constant) {
10250 SDValue N00 = N0.getOperand(0);
10251 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
10252 ((N00.getOpcode() == ISD::ANY_EXTEND ||
10253 N00.getOpcode() == ISD::ZERO_EXTEND) &&
10254 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
10255 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
10256 APInt ShAmt = N1C->getAPIntValue();
10257 Mask = Mask.shl(ShAmt);
10259 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
10260 N00, DAG.getConstant(Mask, VT));
10267 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
10269 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
10270 const X86Subtarget *Subtarget) {
10271 EVT VT = N->getValueType(0);
10272 if (!VT.isVector() && VT.isInteger() &&
10273 N->getOpcode() == ISD::SHL)
10274 return PerformSHLCombine(N, DAG);
10276 // On X86 with SSE2 support, we can transform this to a vector shift if
10277 // all elements are shifted by the same amount. We can't do this in legalize
10278 // because the a constant vector is typically transformed to a constant pool
10279 // so we have no knowledge of the shift amount.
10280 if (!Subtarget->hasSSE2())
10283 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
10286 SDValue ShAmtOp = N->getOperand(1);
10287 EVT EltVT = VT.getVectorElementType();
10288 DebugLoc DL = N->getDebugLoc();
10289 SDValue BaseShAmt = SDValue();
10290 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
10291 unsigned NumElts = VT.getVectorNumElements();
10293 for (; i != NumElts; ++i) {
10294 SDValue Arg = ShAmtOp.getOperand(i);
10295 if (Arg.getOpcode() == ISD::UNDEF) continue;
10299 for (; i != NumElts; ++i) {
10300 SDValue Arg = ShAmtOp.getOperand(i);
10301 if (Arg.getOpcode() == ISD::UNDEF) continue;
10302 if (Arg != BaseShAmt) {
10306 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
10307 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
10308 SDValue InVec = ShAmtOp.getOperand(0);
10309 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
10310 unsigned NumElts = InVec.getValueType().getVectorNumElements();
10312 for (; i != NumElts; ++i) {
10313 SDValue Arg = InVec.getOperand(i);
10314 if (Arg.getOpcode() == ISD::UNDEF) continue;
10318 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
10319 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
10320 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
10321 if (C->getZExtValue() == SplatIdx)
10322 BaseShAmt = InVec.getOperand(1);
10325 if (BaseShAmt.getNode() == 0)
10326 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
10327 DAG.getIntPtrConstant(0));
10331 // The shift amount is an i32.
10332 if (EltVT.bitsGT(MVT::i32))
10333 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
10334 else if (EltVT.bitsLT(MVT::i32))
10335 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
10337 // The shift amount is identical so we can do a vector shift.
10338 SDValue ValOp = N->getOperand(0);
10339 switch (N->getOpcode()) {
10341 llvm_unreachable("Unknown shift opcode!");
10344 if (VT == MVT::v2i64)
10345 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10346 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
10348 if (VT == MVT::v4i32)
10349 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10350 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
10352 if (VT == MVT::v8i16)
10353 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10354 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
10358 if (VT == MVT::v4i32)
10359 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10360 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
10362 if (VT == MVT::v8i16)
10363 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10364 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
10368 if (VT == MVT::v2i64)
10369 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10370 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
10372 if (VT == MVT::v4i32)
10373 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10374 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
10376 if (VT == MVT::v8i16)
10377 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
10378 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
10385 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
10386 TargetLowering::DAGCombinerInfo &DCI,
10387 const X86Subtarget *Subtarget) {
10388 if (DCI.isBeforeLegalizeOps())
10391 EVT VT = N->getValueType(0);
10392 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
10395 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
10396 SDValue N0 = N->getOperand(0);
10397 SDValue N1 = N->getOperand(1);
10398 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
10400 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
10402 if (!N0.hasOneUse() || !N1.hasOneUse())
10405 SDValue ShAmt0 = N0.getOperand(1);
10406 if (ShAmt0.getValueType() != MVT::i8)
10408 SDValue ShAmt1 = N1.getOperand(1);
10409 if (ShAmt1.getValueType() != MVT::i8)
10411 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
10412 ShAmt0 = ShAmt0.getOperand(0);
10413 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
10414 ShAmt1 = ShAmt1.getOperand(0);
10416 DebugLoc DL = N->getDebugLoc();
10417 unsigned Opc = X86ISD::SHLD;
10418 SDValue Op0 = N0.getOperand(0);
10419 SDValue Op1 = N1.getOperand(0);
10420 if (ShAmt0.getOpcode() == ISD::SUB) {
10421 Opc = X86ISD::SHRD;
10422 std::swap(Op0, Op1);
10423 std::swap(ShAmt0, ShAmt1);
10426 unsigned Bits = VT.getSizeInBits();
10427 if (ShAmt1.getOpcode() == ISD::SUB) {
10428 SDValue Sum = ShAmt1.getOperand(0);
10429 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
10430 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
10431 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
10432 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
10433 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
10434 return DAG.getNode(Opc, DL, VT,
10436 DAG.getNode(ISD::TRUNCATE, DL,
10439 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
10440 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
10442 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
10443 return DAG.getNode(Opc, DL, VT,
10444 N0.getOperand(0), N1.getOperand(0),
10445 DAG.getNode(ISD::TRUNCATE, DL,
10452 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
10453 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
10454 const X86Subtarget *Subtarget) {
10455 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
10456 // the FP state in cases where an emms may be missing.
10457 // A preferable solution to the general problem is to figure out the right
10458 // places to insert EMMS. This qualifies as a quick hack.
10460 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
10461 StoreSDNode *St = cast<StoreSDNode>(N);
10462 EVT VT = St->getValue().getValueType();
10463 if (VT.getSizeInBits() != 64)
10466 const Function *F = DAG.getMachineFunction().getFunction();
10467 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
10468 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
10469 && Subtarget->hasSSE2();
10470 if ((VT.isVector() ||
10471 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
10472 isa<LoadSDNode>(St->getValue()) &&
10473 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
10474 St->getChain().hasOneUse() && !St->isVolatile()) {
10475 SDNode* LdVal = St->getValue().getNode();
10476 LoadSDNode *Ld = 0;
10477 int TokenFactorIndex = -1;
10478 SmallVector<SDValue, 8> Ops;
10479 SDNode* ChainVal = St->getChain().getNode();
10480 // Must be a store of a load. We currently handle two cases: the load
10481 // is a direct child, and it's under an intervening TokenFactor. It is
10482 // possible to dig deeper under nested TokenFactors.
10483 if (ChainVal == LdVal)
10484 Ld = cast<LoadSDNode>(St->getChain());
10485 else if (St->getValue().hasOneUse() &&
10486 ChainVal->getOpcode() == ISD::TokenFactor) {
10487 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
10488 if (ChainVal->getOperand(i).getNode() == LdVal) {
10489 TokenFactorIndex = i;
10490 Ld = cast<LoadSDNode>(St->getValue());
10492 Ops.push_back(ChainVal->getOperand(i));
10496 if (!Ld || !ISD::isNormalLoad(Ld))
10499 // If this is not the MMX case, i.e. we are just turning i64 load/store
10500 // into f64 load/store, avoid the transformation if there are multiple
10501 // uses of the loaded value.
10502 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
10505 DebugLoc LdDL = Ld->getDebugLoc();
10506 DebugLoc StDL = N->getDebugLoc();
10507 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
10508 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
10510 if (Subtarget->is64Bit() || F64IsLegal) {
10511 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
10512 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
10513 Ld->getBasePtr(), Ld->getSrcValue(),
10514 Ld->getSrcValueOffset(), Ld->isVolatile(),
10515 Ld->isNonTemporal(), Ld->getAlignment());
10516 SDValue NewChain = NewLd.getValue(1);
10517 if (TokenFactorIndex != -1) {
10518 Ops.push_back(NewChain);
10519 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
10522 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
10523 St->getSrcValue(), St->getSrcValueOffset(),
10524 St->isVolatile(), St->isNonTemporal(),
10525 St->getAlignment());
10528 // Otherwise, lower to two pairs of 32-bit loads / stores.
10529 SDValue LoAddr = Ld->getBasePtr();
10530 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
10531 DAG.getConstant(4, MVT::i32));
10533 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
10534 Ld->getSrcValue(), Ld->getSrcValueOffset(),
10535 Ld->isVolatile(), Ld->isNonTemporal(),
10536 Ld->getAlignment());
10537 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
10538 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
10539 Ld->isVolatile(), Ld->isNonTemporal(),
10540 MinAlign(Ld->getAlignment(), 4));
10542 SDValue NewChain = LoLd.getValue(1);
10543 if (TokenFactorIndex != -1) {
10544 Ops.push_back(LoLd);
10545 Ops.push_back(HiLd);
10546 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
10550 LoAddr = St->getBasePtr();
10551 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
10552 DAG.getConstant(4, MVT::i32));
10554 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
10555 St->getSrcValue(), St->getSrcValueOffset(),
10556 St->isVolatile(), St->isNonTemporal(),
10557 St->getAlignment());
10558 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
10560 St->getSrcValueOffset() + 4,
10562 St->isNonTemporal(),
10563 MinAlign(St->getAlignment(), 4));
10564 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
10569 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
10570 /// X86ISD::FXOR nodes.
10571 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
10572 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
10573 // F[X]OR(0.0, x) -> x
10574 // F[X]OR(x, 0.0) -> x
10575 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
10576 if (C->getValueAPF().isPosZero())
10577 return N->getOperand(1);
10578 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
10579 if (C->getValueAPF().isPosZero())
10580 return N->getOperand(0);
10584 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
10585 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
10586 // FAND(0.0, x) -> 0.0
10587 // FAND(x, 0.0) -> 0.0
10588 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
10589 if (C->getValueAPF().isPosZero())
10590 return N->getOperand(0);
10591 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
10592 if (C->getValueAPF().isPosZero())
10593 return N->getOperand(1);
10597 static SDValue PerformBTCombine(SDNode *N,
10599 TargetLowering::DAGCombinerInfo &DCI) {
10600 // BT ignores high bits in the bit index operand.
10601 SDValue Op1 = N->getOperand(1);
10602 if (Op1.hasOneUse()) {
10603 unsigned BitWidth = Op1.getValueSizeInBits();
10604 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
10605 APInt KnownZero, KnownOne;
10606 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
10607 !DCI.isBeforeLegalizeOps());
10608 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10609 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
10610 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
10611 DCI.CommitTargetLoweringOpt(TLO);
10616 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
10617 SDValue Op = N->getOperand(0);
10618 if (Op.getOpcode() == ISD::BIT_CONVERT)
10619 Op = Op.getOperand(0);
10620 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
10621 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
10622 VT.getVectorElementType().getSizeInBits() ==
10623 OpVT.getVectorElementType().getSizeInBits()) {
10624 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
10629 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
10630 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
10631 // (and (i32 x86isd::setcc_carry), 1)
10632 // This eliminates the zext. This transformation is necessary because
10633 // ISD::SETCC is always legalized to i8.
10634 DebugLoc dl = N->getDebugLoc();
10635 SDValue N0 = N->getOperand(0);
10636 EVT VT = N->getValueType(0);
10637 if (N0.getOpcode() == ISD::AND &&
10639 N0.getOperand(0).hasOneUse()) {
10640 SDValue N00 = N0.getOperand(0);
10641 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
10643 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
10644 if (!C || C->getZExtValue() != 1)
10646 return DAG.getNode(ISD::AND, dl, VT,
10647 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
10648 N00.getOperand(0), N00.getOperand(1)),
10649 DAG.getConstant(1, VT));
10655 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
10656 DAGCombinerInfo &DCI) const {
10657 SelectionDAG &DAG = DCI.DAG;
10658 switch (N->getOpcode()) {
10660 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
10661 case ISD::EXTRACT_VECTOR_ELT:
10662 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
10663 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
10664 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
10665 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
10668 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
10669 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
10670 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
10672 case X86ISD::FOR: return PerformFORCombine(N, DAG);
10673 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
10674 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
10675 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
10676 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
10682 /// isTypeDesirableForOp - Return true if the target has native support for
10683 /// the specified value type and it is 'desirable' to use the type for the
10684 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
10685 /// instruction encodings are longer and some i16 instructions are slow.
10686 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
10687 if (!isTypeLegal(VT))
10689 if (VT != MVT::i16)
10696 case ISD::SIGN_EXTEND:
10697 case ISD::ZERO_EXTEND:
10698 case ISD::ANY_EXTEND:
10711 /// IsDesirableToPromoteOp - This method query the target whether it is
10712 /// beneficial for dag combiner to promote the specified node. If true, it
10713 /// should return the desired promotion type by reference.
10714 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
10715 EVT VT = Op.getValueType();
10716 if (VT != MVT::i16)
10719 bool Promote = false;
10720 bool Commute = false;
10721 switch (Op.getOpcode()) {
10724 LoadSDNode *LD = cast<LoadSDNode>(Op);
10725 // If the non-extending load has a single use and it's not live out, then it
10726 // might be folded.
10727 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
10728 Op.hasOneUse()*/) {
10729 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
10730 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
10731 // The only case where we'd want to promote LOAD (rather then it being
10732 // promoted as an operand is when it's only use is liveout.
10733 if (UI->getOpcode() != ISD::CopyToReg)
10740 case ISD::SIGN_EXTEND:
10741 case ISD::ZERO_EXTEND:
10742 case ISD::ANY_EXTEND:
10747 SDValue N0 = Op.getOperand(0);
10748 // Look out for (store (shl (load), x)).
10749 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
10762 SDValue N0 = Op.getOperand(0);
10763 SDValue N1 = Op.getOperand(1);
10764 if (!Commute && MayFoldLoad(N1))
10766 // Avoid disabling potential load folding opportunities.
10767 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
10769 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
10779 //===----------------------------------------------------------------------===//
10780 // X86 Inline Assembly Support
10781 //===----------------------------------------------------------------------===//
10783 static bool LowerToBSwap(CallInst *CI) {
10784 // FIXME: this should verify that we are targetting a 486 or better. If not,
10785 // we will turn this bswap into something that will be lowered to logical ops
10786 // instead of emitting the bswap asm. For now, we don't support 486 or lower
10787 // so don't worry about this.
10789 // Verify this is a simple bswap.
10790 if (CI->getNumArgOperands() != 1 ||
10791 CI->getType() != CI->getArgOperand(0)->getType() ||
10792 !CI->getType()->isIntegerTy())
10795 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10796 if (!Ty || Ty->getBitWidth() % 16 != 0)
10799 // Okay, we can do this xform, do so now.
10800 const Type *Tys[] = { Ty };
10801 Module *M = CI->getParent()->getParent()->getParent();
10802 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
10804 Value *Op = CI->getArgOperand(0);
10805 Op = CallInst::Create(Int, Op, CI->getName(), CI);
10807 CI->replaceAllUsesWith(Op);
10808 CI->eraseFromParent();
10812 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
10813 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10814 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
10816 std::string AsmStr = IA->getAsmString();
10818 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
10819 SmallVector<StringRef, 4> AsmPieces;
10820 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
10822 switch (AsmPieces.size()) {
10823 default: return false;
10825 AsmStr = AsmPieces[0];
10827 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
10830 if (AsmPieces.size() == 2 &&
10831 (AsmPieces[0] == "bswap" ||
10832 AsmPieces[0] == "bswapq" ||
10833 AsmPieces[0] == "bswapl") &&
10834 (AsmPieces[1] == "$0" ||
10835 AsmPieces[1] == "${0:q}")) {
10836 // No need to check constraints, nothing other than the equivalent of
10837 // "=r,0" would be valid here.
10838 return LowerToBSwap(CI);
10840 // rorw $$8, ${0:w} --> llvm.bswap.i16
10841 if (CI->getType()->isIntegerTy(16) &&
10842 AsmPieces.size() == 3 &&
10843 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
10844 AsmPieces[1] == "$$8," &&
10845 AsmPieces[2] == "${0:w}" &&
10846 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
10848 const std::string &Constraints = IA->getConstraintString();
10849 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
10850 std::sort(AsmPieces.begin(), AsmPieces.end());
10851 if (AsmPieces.size() == 4 &&
10852 AsmPieces[0] == "~{cc}" &&
10853 AsmPieces[1] == "~{dirflag}" &&
10854 AsmPieces[2] == "~{flags}" &&
10855 AsmPieces[3] == "~{fpsr}") {
10856 return LowerToBSwap(CI);
10861 if (CI->getType()->isIntegerTy(64) &&
10862 Constraints.size() >= 2 &&
10863 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
10864 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
10865 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
10866 SmallVector<StringRef, 4> Words;
10867 SplitString(AsmPieces[0], Words, " \t");
10868 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
10870 SplitString(AsmPieces[1], Words, " \t");
10871 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
10873 SplitString(AsmPieces[2], Words, " \t,");
10874 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
10875 Words[2] == "%edx") {
10876 return LowerToBSwap(CI);
10888 /// getConstraintType - Given a constraint letter, return the type of
10889 /// constraint it is for this target.
10890 X86TargetLowering::ConstraintType
10891 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10892 if (Constraint.size() == 1) {
10893 switch (Constraint[0]) {
10905 return C_RegisterClass;
10913 return TargetLowering::getConstraintType(Constraint);
10916 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10917 /// with another that has more specific requirements based on the type of the
10918 /// corresponding operand.
10919 const char *X86TargetLowering::
10920 LowerXConstraint(EVT ConstraintVT) const {
10921 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10922 // 'f' like normal targets.
10923 if (ConstraintVT.isFloatingPoint()) {
10924 if (Subtarget->hasSSE2())
10926 if (Subtarget->hasSSE1())
10930 return TargetLowering::LowerXConstraint(ConstraintVT);
10933 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10934 /// vector. If it is invalid, don't add anything to Ops.
10935 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10937 std::vector<SDValue>&Ops,
10938 SelectionDAG &DAG) const {
10939 SDValue Result(0, 0);
10941 switch (Constraint) {
10944 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10945 if (C->getZExtValue() <= 31) {
10946 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10952 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10953 if (C->getZExtValue() <= 63) {
10954 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10960 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10961 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10962 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10968 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10969 if (C->getZExtValue() <= 255) {
10970 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10976 // 32-bit signed value
10977 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10978 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10979 C->getSExtValue())) {
10980 // Widen to 64 bits here to get it sign extended.
10981 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10984 // FIXME gcc accepts some relocatable values here too, but only in certain
10985 // memory models; it's complicated.
10990 // 32-bit unsigned value
10991 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10992 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10993 C->getZExtValue())) {
10994 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10998 // FIXME gcc accepts some relocatable values here too, but only in certain
10999 // memory models; it's complicated.
11003 // Literal immediates are always ok.
11004 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
11005 // Widen to 64 bits here to get it sign extended.
11006 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
11010 // In any sort of PIC mode addresses need to be computed at runtime by
11011 // adding in a register or some sort of table lookup. These can't
11012 // be used as immediates.
11013 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
11016 // If we are in non-pic codegen mode, we allow the address of a global (with
11017 // an optional displacement) to be used with 'i'.
11018 GlobalAddressSDNode *GA = 0;
11019 int64_t Offset = 0;
11021 // Match either (GA), (GA+C), (GA+C1+C2), etc.
11023 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
11024 Offset += GA->getOffset();
11026 } else if (Op.getOpcode() == ISD::ADD) {
11027 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
11028 Offset += C->getZExtValue();
11029 Op = Op.getOperand(0);
11032 } else if (Op.getOpcode() == ISD::SUB) {
11033 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
11034 Offset += -C->getZExtValue();
11035 Op = Op.getOperand(0);
11040 // Otherwise, this isn't something we can handle, reject it.
11044 const GlobalValue *GV = GA->getGlobal();
11045 // If we require an extra load to get this address, as in PIC mode, we
11046 // can't accept it.
11047 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
11048 getTargetMachine())))
11051 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
11052 GA->getValueType(0), Offset);
11057 if (Result.getNode()) {
11058 Ops.push_back(Result);
11061 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
11064 std::vector<unsigned> X86TargetLowering::
11065 getRegClassForInlineAsmConstraint(const std::string &Constraint,
11067 if (Constraint.size() == 1) {
11068 // FIXME: not handling fp-stack yet!
11069 switch (Constraint[0]) { // GCC X86 Constraint Letters
11070 default: break; // Unknown constraint letter
11071 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
11072 if (Subtarget->is64Bit()) {
11073 if (VT == MVT::i32)
11074 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
11075 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
11076 X86::R10D,X86::R11D,X86::R12D,
11077 X86::R13D,X86::R14D,X86::R15D,
11078 X86::EBP, X86::ESP, 0);
11079 else if (VT == MVT::i16)
11080 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
11081 X86::SI, X86::DI, X86::R8W,X86::R9W,
11082 X86::R10W,X86::R11W,X86::R12W,
11083 X86::R13W,X86::R14W,X86::R15W,
11084 X86::BP, X86::SP, 0);
11085 else if (VT == MVT::i8)
11086 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
11087 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
11088 X86::R10B,X86::R11B,X86::R12B,
11089 X86::R13B,X86::R14B,X86::R15B,
11090 X86::BPL, X86::SPL, 0);
11092 else if (VT == MVT::i64)
11093 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
11094 X86::RSI, X86::RDI, X86::R8, X86::R9,
11095 X86::R10, X86::R11, X86::R12,
11096 X86::R13, X86::R14, X86::R15,
11097 X86::RBP, X86::RSP, 0);
11101 // 32-bit fallthrough
11102 case 'Q': // Q_REGS
11103 if (VT == MVT::i32)
11104 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
11105 else if (VT == MVT::i16)
11106 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
11107 else if (VT == MVT::i8)
11108 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
11109 else if (VT == MVT::i64)
11110 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
11115 return std::vector<unsigned>();
11118 std::pair<unsigned, const TargetRegisterClass*>
11119 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
11121 // First, see if this is a constraint that directly corresponds to an LLVM
11123 if (Constraint.size() == 1) {
11124 // GCC Constraint Letters
11125 switch (Constraint[0]) {
11127 case 'r': // GENERAL_REGS
11128 case 'l': // INDEX_REGS
11130 return std::make_pair(0U, X86::GR8RegisterClass);
11131 if (VT == MVT::i16)
11132 return std::make_pair(0U, X86::GR16RegisterClass);
11133 if (VT == MVT::i32 || !Subtarget->is64Bit())
11134 return std::make_pair(0U, X86::GR32RegisterClass);
11135 return std::make_pair(0U, X86::GR64RegisterClass);
11136 case 'R': // LEGACY_REGS
11138 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
11139 if (VT == MVT::i16)
11140 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
11141 if (VT == MVT::i32 || !Subtarget->is64Bit())
11142 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
11143 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
11144 case 'f': // FP Stack registers.
11145 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
11146 // value to the correct fpstack register class.
11147 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
11148 return std::make_pair(0U, X86::RFP32RegisterClass);
11149 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
11150 return std::make_pair(0U, X86::RFP64RegisterClass);
11151 return std::make_pair(0U, X86::RFP80RegisterClass);
11152 case 'y': // MMX_REGS if MMX allowed.
11153 if (!Subtarget->hasMMX()) break;
11154 return std::make_pair(0U, X86::VR64RegisterClass);
11155 case 'Y': // SSE_REGS if SSE2 allowed
11156 if (!Subtarget->hasSSE2()) break;
11158 case 'x': // SSE_REGS if SSE1 allowed
11159 if (!Subtarget->hasSSE1()) break;
11161 switch (VT.getSimpleVT().SimpleTy) {
11163 // Scalar SSE types.
11166 return std::make_pair(0U, X86::FR32RegisterClass);
11169 return std::make_pair(0U, X86::FR64RegisterClass);
11177 return std::make_pair(0U, X86::VR128RegisterClass);
11183 // Use the default implementation in TargetLowering to convert the register
11184 // constraint into a member of a register class.
11185 std::pair<unsigned, const TargetRegisterClass*> Res;
11186 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
11188 // Not found as a standard register?
11189 if (Res.second == 0) {
11190 // Map st(0) -> st(7) -> ST0
11191 if (Constraint.size() == 7 && Constraint[0] == '{' &&
11192 tolower(Constraint[1]) == 's' &&
11193 tolower(Constraint[2]) == 't' &&
11194 Constraint[3] == '(' &&
11195 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
11196 Constraint[5] == ')' &&
11197 Constraint[6] == '}') {
11199 Res.first = X86::ST0+Constraint[4]-'0';
11200 Res.second = X86::RFP80RegisterClass;
11204 // GCC allows "st(0)" to be called just plain "st".
11205 if (StringRef("{st}").equals_lower(Constraint)) {
11206 Res.first = X86::ST0;
11207 Res.second = X86::RFP80RegisterClass;
11212 if (StringRef("{flags}").equals_lower(Constraint)) {
11213 Res.first = X86::EFLAGS;
11214 Res.second = X86::CCRRegisterClass;
11218 // 'A' means EAX + EDX.
11219 if (Constraint == "A") {
11220 Res.first = X86::EAX;
11221 Res.second = X86::GR32_ADRegisterClass;
11227 // Otherwise, check to see if this is a register class of the wrong value
11228 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
11229 // turn into {ax},{dx}.
11230 if (Res.second->hasType(VT))
11231 return Res; // Correct type already, nothing to do.
11233 // All of the single-register GCC register classes map their values onto
11234 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
11235 // really want an 8-bit or 32-bit register, map to the appropriate register
11236 // class and return the appropriate register.
11237 if (Res.second == X86::GR16RegisterClass) {
11238 if (VT == MVT::i8) {
11239 unsigned DestReg = 0;
11240 switch (Res.first) {
11242 case X86::AX: DestReg = X86::AL; break;
11243 case X86::DX: DestReg = X86::DL; break;
11244 case X86::CX: DestReg = X86::CL; break;
11245 case X86::BX: DestReg = X86::BL; break;
11248 Res.first = DestReg;
11249 Res.second = X86::GR8RegisterClass;
11251 } else if (VT == MVT::i32) {
11252 unsigned DestReg = 0;
11253 switch (Res.first) {
11255 case X86::AX: DestReg = X86::EAX; break;
11256 case X86::DX: DestReg = X86::EDX; break;
11257 case X86::CX: DestReg = X86::ECX; break;
11258 case X86::BX: DestReg = X86::EBX; break;
11259 case X86::SI: DestReg = X86::ESI; break;
11260 case X86::DI: DestReg = X86::EDI; break;
11261 case X86::BP: DestReg = X86::EBP; break;
11262 case X86::SP: DestReg = X86::ESP; break;
11265 Res.first = DestReg;
11266 Res.second = X86::GR32RegisterClass;
11268 } else if (VT == MVT::i64) {
11269 unsigned DestReg = 0;
11270 switch (Res.first) {
11272 case X86::AX: DestReg = X86::RAX; break;
11273 case X86::DX: DestReg = X86::RDX; break;
11274 case X86::CX: DestReg = X86::RCX; break;
11275 case X86::BX: DestReg = X86::RBX; break;
11276 case X86::SI: DestReg = X86::RSI; break;
11277 case X86::DI: DestReg = X86::RDI; break;
11278 case X86::BP: DestReg = X86::RBP; break;
11279 case X86::SP: DestReg = X86::RSP; break;
11282 Res.first = DestReg;
11283 Res.second = X86::GR64RegisterClass;
11286 } else if (Res.second == X86::FR32RegisterClass ||
11287 Res.second == X86::FR64RegisterClass ||
11288 Res.second == X86::VR128RegisterClass) {
11289 // Handle references to XMM physical registers that got mapped into the
11290 // wrong class. This can happen with constraints like {xmm0} where the
11291 // target independent register mapper will just pick the first match it can
11292 // find, ignoring the required type.
11293 if (VT == MVT::f32)
11294 Res.second = X86::FR32RegisterClass;
11295 else if (VT == MVT::f64)
11296 Res.second = X86::FR64RegisterClass;
11297 else if (X86::VR128RegisterClass->hasType(VT))
11298 Res.second = X86::VR128RegisterClass;