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");
1327 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1328 // the RET instruction and handled by the FP Stackifier.
1329 if (VA.getLocReg() == X86::ST0 ||
1330 VA.getLocReg() == X86::ST1) {
1331 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1332 // change the value to the FP stack register class.
1333 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1334 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1335 RetOps.push_back(ValToCopy);
1336 // Don't emit a copytoreg.
1340 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1341 // which is returned in RAX / RDX.
1342 if (Subtarget->is64Bit()) {
1343 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1344 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1345 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1346 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1351 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1352 Flag = Chain.getValue(1);
1355 // The x86-64 ABI for returning structs by value requires that we copy
1356 // the sret argument into %rax for the return. We saved the argument into
1357 // a virtual register in the entry block, so now we copy the value out
1359 if (Subtarget->is64Bit() &&
1360 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1361 MachineFunction &MF = DAG.getMachineFunction();
1362 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1363 unsigned Reg = FuncInfo->getSRetReturnReg();
1365 "SRetReturnReg should have been set in LowerFormalArguments().");
1366 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1368 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1369 Flag = Chain.getValue(1);
1371 // RAX now acts like a return value.
1372 MRI.addLiveOut(X86::RAX);
1375 RetOps[0] = Chain; // Update chain.
1377 // Add the flag if we have it.
1379 RetOps.push_back(Flag);
1381 return DAG.getNode(X86ISD::RET_FLAG, dl,
1382 MVT::Other, &RetOps[0], RetOps.size());
1385 /// LowerCallResult - Lower the result values of a call into the
1386 /// appropriate copies out of appropriate physical registers.
1389 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1390 CallingConv::ID CallConv, bool isVarArg,
1391 const SmallVectorImpl<ISD::InputArg> &Ins,
1392 DebugLoc dl, SelectionDAG &DAG,
1393 SmallVectorImpl<SDValue> &InVals) const {
1395 // Assign locations to each value returned by this call.
1396 SmallVector<CCValAssign, 16> RVLocs;
1397 bool Is64Bit = Subtarget->is64Bit();
1398 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1399 RVLocs, *DAG.getContext());
1400 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1402 // Copy all of the result registers out of their specified physreg.
1403 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1404 CCValAssign &VA = RVLocs[i];
1405 EVT CopyVT = VA.getValVT();
1407 // If this is x86-64, and we disabled SSE, we can't return FP values
1408 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1409 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1410 report_fatal_error("SSE register return with SSE disabled");
1415 // If this is a call to a function that returns an fp value on the floating
1416 // point stack, we must guarantee the the value is popped from the stack, so
1417 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1418 // if the return value is not used. We use the FpGET_ST0 instructions
1420 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1421 // If we prefer to use the value in xmm registers, copy it out as f80 and
1422 // use a truncate to move it from fp stack reg to xmm reg.
1423 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1424 bool isST0 = VA.getLocReg() == X86::ST0;
1426 if (CopyVT == MVT::f32) Opc = isST0 ? X86::FpGET_ST0_32:X86::FpGET_ST1_32;
1427 if (CopyVT == MVT::f64) Opc = isST0 ? X86::FpGET_ST0_64:X86::FpGET_ST1_64;
1428 if (CopyVT == MVT::f80) Opc = isST0 ? X86::FpGET_ST0_80:X86::FpGET_ST1_80;
1429 SDValue Ops[] = { Chain, InFlag };
1430 Chain = SDValue(DAG.getMachineNode(Opc, dl, CopyVT, MVT::Other, MVT::Flag,
1432 Val = Chain.getValue(0);
1434 // Round the f80 to the right size, which also moves it to the appropriate
1436 if (CopyVT != VA.getValVT())
1437 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1438 // This truncation won't change the value.
1439 DAG.getIntPtrConstant(1));
1440 } else if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1441 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1442 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1443 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1444 MVT::v2i64, InFlag).getValue(1);
1445 Val = Chain.getValue(0);
1446 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1447 Val, DAG.getConstant(0, MVT::i64));
1449 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1450 MVT::i64, InFlag).getValue(1);
1451 Val = Chain.getValue(0);
1453 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1455 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1456 CopyVT, InFlag).getValue(1);
1457 Val = Chain.getValue(0);
1459 InFlag = Chain.getValue(2);
1460 InVals.push_back(Val);
1467 //===----------------------------------------------------------------------===//
1468 // C & StdCall & Fast Calling Convention implementation
1469 //===----------------------------------------------------------------------===//
1470 // StdCall calling convention seems to be standard for many Windows' API
1471 // routines and around. It differs from C calling convention just a little:
1472 // callee should clean up the stack, not caller. Symbols should be also
1473 // decorated in some fancy way :) It doesn't support any vector arguments.
1474 // For info on fast calling convention see Fast Calling Convention (tail call)
1475 // implementation LowerX86_32FastCCCallTo.
1477 /// CallIsStructReturn - Determines whether a call uses struct return
1479 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1483 return Outs[0].Flags.isSRet();
1486 /// ArgsAreStructReturn - Determines whether a function uses struct
1487 /// return semantics.
1489 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1493 return Ins[0].Flags.isSRet();
1496 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1497 /// given CallingConvention value.
1498 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1499 if (Subtarget->is64Bit()) {
1500 if (CC == CallingConv::GHC)
1501 return CC_X86_64_GHC;
1502 else if (Subtarget->isTargetWin64())
1503 return CC_X86_Win64_C;
1508 if (CC == CallingConv::X86_FastCall)
1509 return CC_X86_32_FastCall;
1510 else if (CC == CallingConv::X86_ThisCall)
1511 return CC_X86_32_ThisCall;
1512 else if (CC == CallingConv::Fast)
1513 return CC_X86_32_FastCC;
1514 else if (CC == CallingConv::GHC)
1515 return CC_X86_32_GHC;
1520 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1521 /// by "Src" to address "Dst" with size and alignment information specified by
1522 /// the specific parameter attribute. The copy will be passed as a byval
1523 /// function parameter.
1525 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1526 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1528 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1529 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1530 /*isVolatile*/false, /*AlwaysInline=*/true,
1534 /// IsTailCallConvention - Return true if the calling convention is one that
1535 /// supports tail call optimization.
1536 static bool IsTailCallConvention(CallingConv::ID CC) {
1537 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1540 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1541 /// a tailcall target by changing its ABI.
1542 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1543 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1547 X86TargetLowering::LowerMemArgument(SDValue Chain,
1548 CallingConv::ID CallConv,
1549 const SmallVectorImpl<ISD::InputArg> &Ins,
1550 DebugLoc dl, SelectionDAG &DAG,
1551 const CCValAssign &VA,
1552 MachineFrameInfo *MFI,
1554 // Create the nodes corresponding to a load from this parameter slot.
1555 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1556 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1557 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1560 // If value is passed by pointer we have address passed instead of the value
1562 if (VA.getLocInfo() == CCValAssign::Indirect)
1563 ValVT = VA.getLocVT();
1565 ValVT = VA.getValVT();
1567 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1568 // changed with more analysis.
1569 // In case of tail call optimization mark all arguments mutable. Since they
1570 // could be overwritten by lowering of arguments in case of a tail call.
1571 if (Flags.isByVal()) {
1572 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1573 VA.getLocMemOffset(), isImmutable);
1574 return DAG.getFrameIndex(FI, getPointerTy());
1576 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1577 VA.getLocMemOffset(), isImmutable);
1578 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1579 return DAG.getLoad(ValVT, dl, Chain, FIN,
1580 PseudoSourceValue::getFixedStack(FI), 0,
1586 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1587 CallingConv::ID CallConv,
1589 const SmallVectorImpl<ISD::InputArg> &Ins,
1592 SmallVectorImpl<SDValue> &InVals)
1594 MachineFunction &MF = DAG.getMachineFunction();
1595 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1597 const Function* Fn = MF.getFunction();
1598 if (Fn->hasExternalLinkage() &&
1599 Subtarget->isTargetCygMing() &&
1600 Fn->getName() == "main")
1601 FuncInfo->setForceFramePointer(true);
1603 MachineFrameInfo *MFI = MF.getFrameInfo();
1604 bool Is64Bit = Subtarget->is64Bit();
1605 bool IsWin64 = Subtarget->isTargetWin64();
1607 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1608 "Var args not supported with calling convention fastcc or ghc");
1610 // Assign locations to all of the incoming arguments.
1611 SmallVector<CCValAssign, 16> ArgLocs;
1612 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1613 ArgLocs, *DAG.getContext());
1614 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1616 unsigned LastVal = ~0U;
1618 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1619 CCValAssign &VA = ArgLocs[i];
1620 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1622 assert(VA.getValNo() != LastVal &&
1623 "Don't support value assigned to multiple locs yet");
1624 LastVal = VA.getValNo();
1626 if (VA.isRegLoc()) {
1627 EVT RegVT = VA.getLocVT();
1628 TargetRegisterClass *RC = NULL;
1629 if (RegVT == MVT::i32)
1630 RC = X86::GR32RegisterClass;
1631 else if (Is64Bit && RegVT == MVT::i64)
1632 RC = X86::GR64RegisterClass;
1633 else if (RegVT == MVT::f32)
1634 RC = X86::FR32RegisterClass;
1635 else if (RegVT == MVT::f64)
1636 RC = X86::FR64RegisterClass;
1637 else if (RegVT.isVector() && RegVT.getSizeInBits() == 256)
1638 RC = X86::VR256RegisterClass;
1639 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1640 RC = X86::VR128RegisterClass;
1641 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1642 RC = X86::VR64RegisterClass;
1644 llvm_unreachable("Unknown argument type!");
1646 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1647 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1649 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1650 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1652 if (VA.getLocInfo() == CCValAssign::SExt)
1653 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1654 DAG.getValueType(VA.getValVT()));
1655 else if (VA.getLocInfo() == CCValAssign::ZExt)
1656 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1657 DAG.getValueType(VA.getValVT()));
1658 else if (VA.getLocInfo() == CCValAssign::BCvt)
1659 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1661 if (VA.isExtInLoc()) {
1662 // Handle MMX values passed in XMM regs.
1663 if (RegVT.isVector()) {
1664 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1665 ArgValue, DAG.getConstant(0, MVT::i64));
1666 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1668 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1671 assert(VA.isMemLoc());
1672 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1675 // If value is passed via pointer - do a load.
1676 if (VA.getLocInfo() == CCValAssign::Indirect)
1677 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1680 InVals.push_back(ArgValue);
1683 // The x86-64 ABI for returning structs by value requires that we copy
1684 // the sret argument into %rax for the return. Save the argument into
1685 // a virtual register so that we can access it from the return points.
1686 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1687 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1688 unsigned Reg = FuncInfo->getSRetReturnReg();
1690 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1691 FuncInfo->setSRetReturnReg(Reg);
1693 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1694 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1697 unsigned StackSize = CCInfo.getNextStackOffset();
1698 // Align stack specially for tail calls.
1699 if (FuncIsMadeTailCallSafe(CallConv))
1700 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1702 // If the function takes variable number of arguments, make a frame index for
1703 // the start of the first vararg value... for expansion of llvm.va_start.
1705 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1706 CallConv != CallingConv::X86_ThisCall)) {
1707 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1710 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1712 // FIXME: We should really autogenerate these arrays
1713 static const unsigned GPR64ArgRegsWin64[] = {
1714 X86::RCX, X86::RDX, X86::R8, X86::R9
1716 static const unsigned XMMArgRegsWin64[] = {
1717 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1719 static const unsigned GPR64ArgRegs64Bit[] = {
1720 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1722 static const unsigned XMMArgRegs64Bit[] = {
1723 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1724 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1726 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1729 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1730 GPR64ArgRegs = GPR64ArgRegsWin64;
1731 XMMArgRegs = XMMArgRegsWin64;
1733 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1734 GPR64ArgRegs = GPR64ArgRegs64Bit;
1735 XMMArgRegs = XMMArgRegs64Bit;
1737 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1739 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1742 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1743 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1744 "SSE register cannot be used when SSE is disabled!");
1745 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1746 "SSE register cannot be used when SSE is disabled!");
1747 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1748 // Kernel mode asks for SSE to be disabled, so don't push them
1750 TotalNumXMMRegs = 0;
1752 // For X86-64, if there are vararg parameters that are passed via
1753 // registers, then we must store them to their spots on the stack so they
1754 // may be loaded by deferencing the result of va_next.
1755 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1756 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1757 FuncInfo->setRegSaveFrameIndex(
1758 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1761 // Store the integer parameter registers.
1762 SmallVector<SDValue, 8> MemOps;
1763 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1765 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1766 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1767 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1768 DAG.getIntPtrConstant(Offset));
1769 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1770 X86::GR64RegisterClass);
1771 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1773 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1774 PseudoSourceValue::getFixedStack(
1775 FuncInfo->getRegSaveFrameIndex()),
1776 Offset, false, false, 0);
1777 MemOps.push_back(Store);
1781 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1782 // Now store the XMM (fp + vector) parameter registers.
1783 SmallVector<SDValue, 11> SaveXMMOps;
1784 SaveXMMOps.push_back(Chain);
1786 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1787 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1788 SaveXMMOps.push_back(ALVal);
1790 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1791 FuncInfo->getRegSaveFrameIndex()));
1792 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1793 FuncInfo->getVarArgsFPOffset()));
1795 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1796 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1797 X86::VR128RegisterClass);
1798 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1799 SaveXMMOps.push_back(Val);
1801 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1803 &SaveXMMOps[0], SaveXMMOps.size()));
1806 if (!MemOps.empty())
1807 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1808 &MemOps[0], MemOps.size());
1812 // Some CCs need callee pop.
1813 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1814 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1816 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1817 // If this is an sret function, the return should pop the hidden pointer.
1818 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1819 FuncInfo->setBytesToPopOnReturn(4);
1823 // RegSaveFrameIndex is X86-64 only.
1824 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1825 if (CallConv == CallingConv::X86_FastCall ||
1826 CallConv == CallingConv::X86_ThisCall)
1827 // fastcc functions can't have varargs.
1828 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1835 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1836 SDValue StackPtr, SDValue Arg,
1837 DebugLoc dl, SelectionDAG &DAG,
1838 const CCValAssign &VA,
1839 ISD::ArgFlagsTy Flags) const {
1840 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1841 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1842 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1843 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1844 if (Flags.isByVal()) {
1845 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1847 return DAG.getStore(Chain, dl, Arg, PtrOff,
1848 PseudoSourceValue::getStack(), LocMemOffset,
1852 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1853 /// optimization is performed and it is required.
1855 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1856 SDValue &OutRetAddr, SDValue Chain,
1857 bool IsTailCall, bool Is64Bit,
1858 int FPDiff, DebugLoc dl) const {
1859 // Adjust the Return address stack slot.
1860 EVT VT = getPointerTy();
1861 OutRetAddr = getReturnAddressFrameIndex(DAG);
1863 // Load the "old" Return address.
1864 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1865 return SDValue(OutRetAddr.getNode(), 1);
1868 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1869 /// optimization is performed and it is required (FPDiff!=0).
1871 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1872 SDValue Chain, SDValue RetAddrFrIdx,
1873 bool Is64Bit, int FPDiff, DebugLoc dl) {
1874 // Store the return address to the appropriate stack slot.
1875 if (!FPDiff) return Chain;
1876 // Calculate the new stack slot for the return address.
1877 int SlotSize = Is64Bit ? 8 : 4;
1878 int NewReturnAddrFI =
1879 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1880 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1881 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1882 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1883 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1889 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1890 CallingConv::ID CallConv, bool isVarArg,
1892 const SmallVectorImpl<ISD::OutputArg> &Outs,
1893 const SmallVectorImpl<SDValue> &OutVals,
1894 const SmallVectorImpl<ISD::InputArg> &Ins,
1895 DebugLoc dl, SelectionDAG &DAG,
1896 SmallVectorImpl<SDValue> &InVals) const {
1897 MachineFunction &MF = DAG.getMachineFunction();
1898 bool Is64Bit = Subtarget->is64Bit();
1899 bool IsStructRet = CallIsStructReturn(Outs);
1900 bool IsSibcall = false;
1903 // Check if it's really possible to do a tail call.
1904 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1905 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1906 Outs, OutVals, Ins, DAG);
1908 // Sibcalls are automatically detected tailcalls which do not require
1910 if (!GuaranteedTailCallOpt && isTailCall)
1917 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1918 "Var args not supported with calling convention fastcc or ghc");
1920 // Analyze operands of the call, assigning locations to each operand.
1921 SmallVector<CCValAssign, 16> ArgLocs;
1922 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1923 ArgLocs, *DAG.getContext());
1924 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1926 // Get a count of how many bytes are to be pushed on the stack.
1927 unsigned NumBytes = CCInfo.getNextStackOffset();
1929 // This is a sibcall. The memory operands are available in caller's
1930 // own caller's stack.
1932 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1933 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1936 if (isTailCall && !IsSibcall) {
1937 // Lower arguments at fp - stackoffset + fpdiff.
1938 unsigned NumBytesCallerPushed =
1939 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1940 FPDiff = NumBytesCallerPushed - NumBytes;
1942 // Set the delta of movement of the returnaddr stackslot.
1943 // But only set if delta is greater than previous delta.
1944 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1945 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1949 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1951 SDValue RetAddrFrIdx;
1952 // Load return adress for tail calls.
1953 if (isTailCall && FPDiff)
1954 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1955 Is64Bit, FPDiff, dl);
1957 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1958 SmallVector<SDValue, 8> MemOpChains;
1961 // Walk the register/memloc assignments, inserting copies/loads. In the case
1962 // of tail call optimization arguments are handle later.
1963 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1964 CCValAssign &VA = ArgLocs[i];
1965 EVT RegVT = VA.getLocVT();
1966 SDValue Arg = OutVals[i];
1967 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1968 bool isByVal = Flags.isByVal();
1970 // Promote the value if needed.
1971 switch (VA.getLocInfo()) {
1972 default: llvm_unreachable("Unknown loc info!");
1973 case CCValAssign::Full: break;
1974 case CCValAssign::SExt:
1975 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1977 case CCValAssign::ZExt:
1978 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1980 case CCValAssign::AExt:
1981 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1982 // Special case: passing MMX values in XMM registers.
1983 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1984 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1985 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1987 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1989 case CCValAssign::BCvt:
1990 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1992 case CCValAssign::Indirect: {
1993 // Store the argument.
1994 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1995 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1996 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1997 PseudoSourceValue::getFixedStack(FI), 0,
2004 if (VA.isRegLoc()) {
2005 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2006 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2007 assert(VA.isMemLoc());
2008 if (StackPtr.getNode() == 0)
2009 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
2010 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2011 dl, DAG, VA, Flags));
2015 if (!MemOpChains.empty())
2016 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2017 &MemOpChains[0], MemOpChains.size());
2019 // Build a sequence of copy-to-reg nodes chained together with token chain
2020 // and flag operands which copy the outgoing args into registers.
2022 // Tail call byval lowering might overwrite argument registers so in case of
2023 // tail call optimization the copies to registers are lowered later.
2025 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2026 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2027 RegsToPass[i].second, InFlag);
2028 InFlag = Chain.getValue(1);
2031 if (Subtarget->isPICStyleGOT()) {
2032 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2035 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
2036 DAG.getNode(X86ISD::GlobalBaseReg,
2037 DebugLoc(), getPointerTy()),
2039 InFlag = Chain.getValue(1);
2041 // If we are tail calling and generating PIC/GOT style code load the
2042 // address of the callee into ECX. The value in ecx is used as target of
2043 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2044 // for tail calls on PIC/GOT architectures. Normally we would just put the
2045 // address of GOT into ebx and then call target@PLT. But for tail calls
2046 // ebx would be restored (since ebx is callee saved) before jumping to the
2049 // Note: The actual moving to ECX is done further down.
2050 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2051 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2052 !G->getGlobal()->hasProtectedVisibility())
2053 Callee = LowerGlobalAddress(Callee, DAG);
2054 else if (isa<ExternalSymbolSDNode>(Callee))
2055 Callee = LowerExternalSymbol(Callee, DAG);
2059 if (Is64Bit && isVarArg && !Subtarget->isTargetWin64()) {
2060 // From AMD64 ABI document:
2061 // For calls that may call functions that use varargs or stdargs
2062 // (prototype-less calls or calls to functions containing ellipsis (...) in
2063 // the declaration) %al is used as hidden argument to specify the number
2064 // of SSE registers used. The contents of %al do not need to match exactly
2065 // the number of registers, but must be an ubound on the number of SSE
2066 // registers used and is in the range 0 - 8 inclusive.
2068 // Count the number of XMM registers allocated.
2069 static const unsigned XMMArgRegs[] = {
2070 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2071 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2073 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2074 assert((Subtarget->hasSSE1() || !NumXMMRegs)
2075 && "SSE registers cannot be used when SSE is disabled");
2077 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2078 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2079 InFlag = Chain.getValue(1);
2083 // For tail calls lower the arguments to the 'real' stack slot.
2085 // Force all the incoming stack arguments to be loaded from the stack
2086 // before any new outgoing arguments are stored to the stack, because the
2087 // outgoing stack slots may alias the incoming argument stack slots, and
2088 // the alias isn't otherwise explicit. This is slightly more conservative
2089 // than necessary, because it means that each store effectively depends
2090 // on every argument instead of just those arguments it would clobber.
2091 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2093 SmallVector<SDValue, 8> MemOpChains2;
2096 // Do not flag preceeding copytoreg stuff together with the following stuff.
2098 if (GuaranteedTailCallOpt) {
2099 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2100 CCValAssign &VA = ArgLocs[i];
2103 assert(VA.isMemLoc());
2104 SDValue Arg = OutVals[i];
2105 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2106 // Create frame index.
2107 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2108 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2109 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2110 FIN = DAG.getFrameIndex(FI, getPointerTy());
2112 if (Flags.isByVal()) {
2113 // Copy relative to framepointer.
2114 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2115 if (StackPtr.getNode() == 0)
2116 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2118 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2120 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2124 // Store relative to framepointer.
2125 MemOpChains2.push_back(
2126 DAG.getStore(ArgChain, dl, Arg, FIN,
2127 PseudoSourceValue::getFixedStack(FI), 0,
2133 if (!MemOpChains2.empty())
2134 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2135 &MemOpChains2[0], MemOpChains2.size());
2137 // Copy arguments to their registers.
2138 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2139 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2140 RegsToPass[i].second, InFlag);
2141 InFlag = Chain.getValue(1);
2145 // Store the return address to the appropriate stack slot.
2146 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2150 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2151 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2152 // In the 64-bit large code model, we have to make all calls
2153 // through a register, since the call instruction's 32-bit
2154 // pc-relative offset may not be large enough to hold the whole
2156 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2157 // If the callee is a GlobalAddress node (quite common, every direct call
2158 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2161 // We should use extra load for direct calls to dllimported functions in
2163 const GlobalValue *GV = G->getGlobal();
2164 if (!GV->hasDLLImportLinkage()) {
2165 unsigned char OpFlags = 0;
2167 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2168 // external symbols most go through the PLT in PIC mode. If the symbol
2169 // has hidden or protected visibility, or if it is static or local, then
2170 // we don't need to use the PLT - we can directly call it.
2171 if (Subtarget->isTargetELF() &&
2172 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2173 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2174 OpFlags = X86II::MO_PLT;
2175 } else if (Subtarget->isPICStyleStubAny() &&
2176 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2177 Subtarget->getDarwinVers() < 9) {
2178 // PC-relative references to external symbols should go through $stub,
2179 // unless we're building with the leopard linker or later, which
2180 // automatically synthesizes these stubs.
2181 OpFlags = X86II::MO_DARWIN_STUB;
2184 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2185 G->getOffset(), OpFlags);
2187 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2188 unsigned char OpFlags = 0;
2190 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2191 // symbols should go through the PLT.
2192 if (Subtarget->isTargetELF() &&
2193 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2194 OpFlags = X86II::MO_PLT;
2195 } else if (Subtarget->isPICStyleStubAny() &&
2196 Subtarget->getDarwinVers() < 9) {
2197 // PC-relative references to external symbols should go through $stub,
2198 // unless we're building with the leopard linker or later, which
2199 // automatically synthesizes these stubs.
2200 OpFlags = X86II::MO_DARWIN_STUB;
2203 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2207 // Returns a chain & a flag for retval copy to use.
2208 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2209 SmallVector<SDValue, 8> Ops;
2211 if (!IsSibcall && isTailCall) {
2212 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2213 DAG.getIntPtrConstant(0, true), InFlag);
2214 InFlag = Chain.getValue(1);
2217 Ops.push_back(Chain);
2218 Ops.push_back(Callee);
2221 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2223 // Add argument registers to the end of the list so that they are known live
2225 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2226 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2227 RegsToPass[i].second.getValueType()));
2229 // Add an implicit use GOT pointer in EBX.
2230 if (!isTailCall && Subtarget->isPICStyleGOT())
2231 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2233 // Add an implicit use of AL for x86 vararg functions.
2234 if (Is64Bit && isVarArg)
2235 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2237 if (InFlag.getNode())
2238 Ops.push_back(InFlag);
2242 //// If this is the first return lowered for this function, add the regs
2243 //// to the liveout set for the function.
2244 // This isn't right, although it's probably harmless on x86; liveouts
2245 // should be computed from returns not tail calls. Consider a void
2246 // function making a tail call to a function returning int.
2247 return DAG.getNode(X86ISD::TC_RETURN, dl,
2248 NodeTys, &Ops[0], Ops.size());
2251 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2252 InFlag = Chain.getValue(1);
2254 // Create the CALLSEQ_END node.
2255 unsigned NumBytesForCalleeToPush;
2256 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2257 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2258 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2259 // If this is a call to a struct-return function, the callee
2260 // pops the hidden struct pointer, so we have to push it back.
2261 // This is common for Darwin/X86, Linux & Mingw32 targets.
2262 NumBytesForCalleeToPush = 4;
2264 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2266 // Returns a flag for retval copy to use.
2268 Chain = DAG.getCALLSEQ_END(Chain,
2269 DAG.getIntPtrConstant(NumBytes, true),
2270 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2273 InFlag = Chain.getValue(1);
2276 // Handle result values, copying them out of physregs into vregs that we
2278 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2279 Ins, dl, DAG, InVals);
2283 //===----------------------------------------------------------------------===//
2284 // Fast Calling Convention (tail call) implementation
2285 //===----------------------------------------------------------------------===//
2287 // Like std call, callee cleans arguments, convention except that ECX is
2288 // reserved for storing the tail called function address. Only 2 registers are
2289 // free for argument passing (inreg). Tail call optimization is performed
2291 // * tailcallopt is enabled
2292 // * caller/callee are fastcc
2293 // On X86_64 architecture with GOT-style position independent code only local
2294 // (within module) calls are supported at the moment.
2295 // To keep the stack aligned according to platform abi the function
2296 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2297 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2298 // If a tail called function callee has more arguments than the caller the
2299 // caller needs to make sure that there is room to move the RETADDR to. This is
2300 // achieved by reserving an area the size of the argument delta right after the
2301 // original REtADDR, but before the saved framepointer or the spilled registers
2302 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2314 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2315 /// for a 16 byte align requirement.
2317 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2318 SelectionDAG& DAG) const {
2319 MachineFunction &MF = DAG.getMachineFunction();
2320 const TargetMachine &TM = MF.getTarget();
2321 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2322 unsigned StackAlignment = TFI.getStackAlignment();
2323 uint64_t AlignMask = StackAlignment - 1;
2324 int64_t Offset = StackSize;
2325 uint64_t SlotSize = TD->getPointerSize();
2326 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2327 // Number smaller than 12 so just add the difference.
2328 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2330 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2331 Offset = ((~AlignMask) & Offset) + StackAlignment +
2332 (StackAlignment-SlotSize);
2337 /// MatchingStackOffset - Return true if the given stack call argument is
2338 /// already available in the same position (relatively) of the caller's
2339 /// incoming argument stack.
2341 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2342 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2343 const X86InstrInfo *TII) {
2344 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2346 if (Arg.getOpcode() == ISD::CopyFromReg) {
2347 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2348 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2350 MachineInstr *Def = MRI->getVRegDef(VR);
2353 if (!Flags.isByVal()) {
2354 if (!TII->isLoadFromStackSlot(Def, FI))
2357 unsigned Opcode = Def->getOpcode();
2358 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2359 Def->getOperand(1).isFI()) {
2360 FI = Def->getOperand(1).getIndex();
2361 Bytes = Flags.getByValSize();
2365 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2366 if (Flags.isByVal())
2367 // ByVal argument is passed in as a pointer but it's now being
2368 // dereferenced. e.g.
2369 // define @foo(%struct.X* %A) {
2370 // tail call @bar(%struct.X* byval %A)
2373 SDValue Ptr = Ld->getBasePtr();
2374 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2377 FI = FINode->getIndex();
2381 assert(FI != INT_MAX);
2382 if (!MFI->isFixedObjectIndex(FI))
2384 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2387 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2388 /// for tail call optimization. Targets which want to do tail call
2389 /// optimization should implement this function.
2391 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2392 CallingConv::ID CalleeCC,
2394 bool isCalleeStructRet,
2395 bool isCallerStructRet,
2396 const SmallVectorImpl<ISD::OutputArg> &Outs,
2397 const SmallVectorImpl<SDValue> &OutVals,
2398 const SmallVectorImpl<ISD::InputArg> &Ins,
2399 SelectionDAG& DAG) const {
2400 if (!IsTailCallConvention(CalleeCC) &&
2401 CalleeCC != CallingConv::C)
2404 // If -tailcallopt is specified, make fastcc functions tail-callable.
2405 const MachineFunction &MF = DAG.getMachineFunction();
2406 const Function *CallerF = DAG.getMachineFunction().getFunction();
2407 CallingConv::ID CallerCC = CallerF->getCallingConv();
2408 bool CCMatch = CallerCC == CalleeCC;
2410 if (GuaranteedTailCallOpt) {
2411 if (IsTailCallConvention(CalleeCC) && CCMatch)
2416 // Look for obvious safe cases to perform tail call optimization that do not
2417 // require ABI changes. This is what gcc calls sibcall.
2419 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2420 // emit a special epilogue.
2421 if (RegInfo->needsStackRealignment(MF))
2424 // Do not sibcall optimize vararg calls unless the call site is not passing
2426 if (isVarArg && !Outs.empty())
2429 // Also avoid sibcall optimization if either caller or callee uses struct
2430 // return semantics.
2431 if (isCalleeStructRet || isCallerStructRet)
2434 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2435 // Therefore if it's not used by the call it is not safe to optimize this into
2437 bool Unused = false;
2438 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2445 SmallVector<CCValAssign, 16> RVLocs;
2446 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2447 RVLocs, *DAG.getContext());
2448 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2449 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2450 CCValAssign &VA = RVLocs[i];
2451 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2456 // If the calling conventions do not match, then we'd better make sure the
2457 // results are returned in the same way as what the caller expects.
2459 SmallVector<CCValAssign, 16> RVLocs1;
2460 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2461 RVLocs1, *DAG.getContext());
2462 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2464 SmallVector<CCValAssign, 16> RVLocs2;
2465 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2466 RVLocs2, *DAG.getContext());
2467 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2469 if (RVLocs1.size() != RVLocs2.size())
2471 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2472 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2474 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2476 if (RVLocs1[i].isRegLoc()) {
2477 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2480 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2486 // If the callee takes no arguments then go on to check the results of the
2488 if (!Outs.empty()) {
2489 // Check if stack adjustment is needed. For now, do not do this if any
2490 // argument is passed on the stack.
2491 SmallVector<CCValAssign, 16> ArgLocs;
2492 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2493 ArgLocs, *DAG.getContext());
2494 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2495 if (CCInfo.getNextStackOffset()) {
2496 MachineFunction &MF = DAG.getMachineFunction();
2497 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2499 if (Subtarget->isTargetWin64())
2500 // Win64 ABI has additional complications.
2503 // Check if the arguments are already laid out in the right way as
2504 // the caller's fixed stack objects.
2505 MachineFrameInfo *MFI = MF.getFrameInfo();
2506 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2507 const X86InstrInfo *TII =
2508 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2509 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2510 CCValAssign &VA = ArgLocs[i];
2511 SDValue Arg = OutVals[i];
2512 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2513 if (VA.getLocInfo() == CCValAssign::Indirect)
2515 if (!VA.isRegLoc()) {
2516 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2523 // If the tailcall address may be in a register, then make sure it's
2524 // possible to register allocate for it. In 32-bit, the call address can
2525 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2526 // callee-saved registers are restored. These happen to be the same
2527 // registers used to pass 'inreg' arguments so watch out for those.
2528 if (!Subtarget->is64Bit() &&
2529 !isa<GlobalAddressSDNode>(Callee) &&
2530 !isa<ExternalSymbolSDNode>(Callee)) {
2531 unsigned NumInRegs = 0;
2532 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2533 CCValAssign &VA = ArgLocs[i];
2536 unsigned Reg = VA.getLocReg();
2539 case X86::EAX: case X86::EDX: case X86::ECX:
2540 if (++NumInRegs == 3)
2552 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2553 return X86::createFastISel(funcInfo);
2557 //===----------------------------------------------------------------------===//
2558 // Other Lowering Hooks
2559 //===----------------------------------------------------------------------===//
2562 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2563 MachineFunction &MF = DAG.getMachineFunction();
2564 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2565 int ReturnAddrIndex = FuncInfo->getRAIndex();
2567 if (ReturnAddrIndex == 0) {
2568 // Set up a frame object for the return address.
2569 uint64_t SlotSize = TD->getPointerSize();
2570 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2572 FuncInfo->setRAIndex(ReturnAddrIndex);
2575 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2579 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2580 bool hasSymbolicDisplacement) {
2581 // Offset should fit into 32 bit immediate field.
2582 if (!isInt<32>(Offset))
2585 // If we don't have a symbolic displacement - we don't have any extra
2587 if (!hasSymbolicDisplacement)
2590 // FIXME: Some tweaks might be needed for medium code model.
2591 if (M != CodeModel::Small && M != CodeModel::Kernel)
2594 // For small code model we assume that latest object is 16MB before end of 31
2595 // bits boundary. We may also accept pretty large negative constants knowing
2596 // that all objects are in the positive half of address space.
2597 if (M == CodeModel::Small && Offset < 16*1024*1024)
2600 // For kernel code model we know that all object resist in the negative half
2601 // of 32bits address space. We may not accept negative offsets, since they may
2602 // be just off and we may accept pretty large positive ones.
2603 if (M == CodeModel::Kernel && Offset > 0)
2609 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2610 /// specific condition code, returning the condition code and the LHS/RHS of the
2611 /// comparison to make.
2612 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2613 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2615 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2616 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2617 // X > -1 -> X == 0, jump !sign.
2618 RHS = DAG.getConstant(0, RHS.getValueType());
2619 return X86::COND_NS;
2620 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2621 // X < 0 -> X == 0, jump on sign.
2623 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2625 RHS = DAG.getConstant(0, RHS.getValueType());
2626 return X86::COND_LE;
2630 switch (SetCCOpcode) {
2631 default: llvm_unreachable("Invalid integer condition!");
2632 case ISD::SETEQ: return X86::COND_E;
2633 case ISD::SETGT: return X86::COND_G;
2634 case ISD::SETGE: return X86::COND_GE;
2635 case ISD::SETLT: return X86::COND_L;
2636 case ISD::SETLE: return X86::COND_LE;
2637 case ISD::SETNE: return X86::COND_NE;
2638 case ISD::SETULT: return X86::COND_B;
2639 case ISD::SETUGT: return X86::COND_A;
2640 case ISD::SETULE: return X86::COND_BE;
2641 case ISD::SETUGE: return X86::COND_AE;
2645 // First determine if it is required or is profitable to flip the operands.
2647 // If LHS is a foldable load, but RHS is not, flip the condition.
2648 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2649 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2650 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2651 std::swap(LHS, RHS);
2654 switch (SetCCOpcode) {
2660 std::swap(LHS, RHS);
2664 // On a floating point condition, the flags are set as follows:
2666 // 0 | 0 | 0 | X > Y
2667 // 0 | 0 | 1 | X < Y
2668 // 1 | 0 | 0 | X == Y
2669 // 1 | 1 | 1 | unordered
2670 switch (SetCCOpcode) {
2671 default: llvm_unreachable("Condcode should be pre-legalized away");
2673 case ISD::SETEQ: return X86::COND_E;
2674 case ISD::SETOLT: // flipped
2676 case ISD::SETGT: return X86::COND_A;
2677 case ISD::SETOLE: // flipped
2679 case ISD::SETGE: return X86::COND_AE;
2680 case ISD::SETUGT: // flipped
2682 case ISD::SETLT: return X86::COND_B;
2683 case ISD::SETUGE: // flipped
2685 case ISD::SETLE: return X86::COND_BE;
2687 case ISD::SETNE: return X86::COND_NE;
2688 case ISD::SETUO: return X86::COND_P;
2689 case ISD::SETO: return X86::COND_NP;
2691 case ISD::SETUNE: return X86::COND_INVALID;
2695 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2696 /// code. Current x86 isa includes the following FP cmov instructions:
2697 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2698 static bool hasFPCMov(unsigned X86CC) {
2714 /// isFPImmLegal - Returns true if the target can instruction select the
2715 /// specified FP immediate natively. If false, the legalizer will
2716 /// materialize the FP immediate as a load from a constant pool.
2717 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2718 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2719 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2725 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2726 /// the specified range (L, H].
2727 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2728 return (Val < 0) || (Val >= Low && Val < Hi);
2731 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2732 /// specified value.
2733 static bool isUndefOrEqual(int Val, int CmpVal) {
2734 if (Val < 0 || Val == CmpVal)
2739 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2740 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2741 /// the second operand.
2742 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2743 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2744 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2745 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2746 return (Mask[0] < 2 && Mask[1] < 2);
2750 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2751 SmallVector<int, 8> M;
2753 return ::isPSHUFDMask(M, N->getValueType(0));
2756 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2757 /// is suitable for input to PSHUFHW.
2758 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2759 if (VT != MVT::v8i16)
2762 // Lower quadword copied in order or undef.
2763 for (int i = 0; i != 4; ++i)
2764 if (Mask[i] >= 0 && Mask[i] != i)
2767 // Upper quadword shuffled.
2768 for (int i = 4; i != 8; ++i)
2769 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2775 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2776 SmallVector<int, 8> M;
2778 return ::isPSHUFHWMask(M, N->getValueType(0));
2781 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2782 /// is suitable for input to PSHUFLW.
2783 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2784 if (VT != MVT::v8i16)
2787 // Upper quadword copied in order.
2788 for (int i = 4; i != 8; ++i)
2789 if (Mask[i] >= 0 && Mask[i] != i)
2792 // Lower quadword shuffled.
2793 for (int i = 0; i != 4; ++i)
2800 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2801 SmallVector<int, 8> M;
2803 return ::isPSHUFLWMask(M, N->getValueType(0));
2806 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2807 /// is suitable for input to PALIGNR.
2808 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2810 int i, e = VT.getVectorNumElements();
2812 // Do not handle v2i64 / v2f64 shuffles with palignr.
2813 if (e < 4 || !hasSSSE3)
2816 for (i = 0; i != e; ++i)
2820 // All undef, not a palignr.
2824 // Determine if it's ok to perform a palignr with only the LHS, since we
2825 // don't have access to the actual shuffle elements to see if RHS is undef.
2826 bool Unary = Mask[i] < (int)e;
2827 bool NeedsUnary = false;
2829 int s = Mask[i] - i;
2831 // Check the rest of the elements to see if they are consecutive.
2832 for (++i; i != e; ++i) {
2837 Unary = Unary && (m < (int)e);
2838 NeedsUnary = NeedsUnary || (m < s);
2840 if (NeedsUnary && !Unary)
2842 if (Unary && m != ((s+i) & (e-1)))
2844 if (!Unary && m != (s+i))
2850 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2851 SmallVector<int, 8> M;
2853 return ::isPALIGNRMask(M, N->getValueType(0), true);
2856 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2857 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2858 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2859 int NumElems = VT.getVectorNumElements();
2860 if (NumElems != 2 && NumElems != 4)
2863 int Half = NumElems / 2;
2864 for (int i = 0; i < Half; ++i)
2865 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2867 for (int i = Half; i < NumElems; ++i)
2868 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2874 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2875 SmallVector<int, 8> M;
2877 return ::isSHUFPMask(M, N->getValueType(0));
2880 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2881 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2882 /// half elements to come from vector 1 (which would equal the dest.) and
2883 /// the upper half to come from vector 2.
2884 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2885 int NumElems = VT.getVectorNumElements();
2887 if (NumElems != 2 && NumElems != 4)
2890 int Half = NumElems / 2;
2891 for (int i = 0; i < Half; ++i)
2892 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2894 for (int i = Half; i < NumElems; ++i)
2895 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2900 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2901 SmallVector<int, 8> M;
2903 return isCommutedSHUFPMask(M, N->getValueType(0));
2906 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2907 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2908 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2909 if (N->getValueType(0).getVectorNumElements() != 4)
2912 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2913 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2914 isUndefOrEqual(N->getMaskElt(1), 7) &&
2915 isUndefOrEqual(N->getMaskElt(2), 2) &&
2916 isUndefOrEqual(N->getMaskElt(3), 3);
2919 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2920 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2922 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2923 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2928 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2929 isUndefOrEqual(N->getMaskElt(1), 3) &&
2930 isUndefOrEqual(N->getMaskElt(2), 2) &&
2931 isUndefOrEqual(N->getMaskElt(3), 3);
2934 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2935 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2936 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2937 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2939 if (NumElems != 2 && NumElems != 4)
2942 for (unsigned i = 0; i < NumElems/2; ++i)
2943 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2946 for (unsigned i = NumElems/2; i < NumElems; ++i)
2947 if (!isUndefOrEqual(N->getMaskElt(i), i))
2953 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2954 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2955 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2956 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2958 if (NumElems != 2 && NumElems != 4)
2961 for (unsigned i = 0; i < NumElems/2; ++i)
2962 if (!isUndefOrEqual(N->getMaskElt(i), i))
2965 for (unsigned i = 0; i < NumElems/2; ++i)
2966 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2972 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2973 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2974 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2975 bool V2IsSplat = false) {
2976 int NumElts = VT.getVectorNumElements();
2977 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2980 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2982 int BitI1 = Mask[i+1];
2983 if (!isUndefOrEqual(BitI, j))
2986 if (!isUndefOrEqual(BitI1, NumElts))
2989 if (!isUndefOrEqual(BitI1, j + NumElts))
2996 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2997 SmallVector<int, 8> M;
2999 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
3002 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3003 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3004 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3005 bool V2IsSplat = false) {
3006 int NumElts = VT.getVectorNumElements();
3007 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3010 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3012 int BitI1 = Mask[i+1];
3013 if (!isUndefOrEqual(BitI, j + NumElts/2))
3016 if (isUndefOrEqual(BitI1, NumElts))
3019 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
3026 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3027 SmallVector<int, 8> M;
3029 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3032 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3033 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3035 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3036 int NumElems = VT.getVectorNumElements();
3037 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3040 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
3042 int BitI1 = Mask[i+1];
3043 if (!isUndefOrEqual(BitI, j))
3045 if (!isUndefOrEqual(BitI1, j))
3051 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3052 SmallVector<int, 8> M;
3054 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3057 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3058 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3060 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3061 int NumElems = VT.getVectorNumElements();
3062 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3065 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3067 int BitI1 = Mask[i+1];
3068 if (!isUndefOrEqual(BitI, j))
3070 if (!isUndefOrEqual(BitI1, j))
3076 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3077 SmallVector<int, 8> M;
3079 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3082 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3083 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3084 /// MOVSD, and MOVD, i.e. setting the lowest element.
3085 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3086 if (VT.getVectorElementType().getSizeInBits() < 32)
3089 int NumElts = VT.getVectorNumElements();
3091 if (!isUndefOrEqual(Mask[0], NumElts))
3094 for (int i = 1; i < NumElts; ++i)
3095 if (!isUndefOrEqual(Mask[i], i))
3101 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3102 SmallVector<int, 8> M;
3104 return ::isMOVLMask(M, N->getValueType(0));
3107 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3108 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3109 /// element of vector 2 and the other elements to come from vector 1 in order.
3110 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3111 bool V2IsSplat = false, bool V2IsUndef = false) {
3112 int NumOps = VT.getVectorNumElements();
3113 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3116 if (!isUndefOrEqual(Mask[0], 0))
3119 for (int i = 1; i < NumOps; ++i)
3120 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3121 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3122 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3128 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3129 bool V2IsUndef = false) {
3130 SmallVector<int, 8> M;
3132 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3135 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3136 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3137 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3138 if (N->getValueType(0).getVectorNumElements() != 4)
3141 // Expect 1, 1, 3, 3
3142 for (unsigned i = 0; i < 2; ++i) {
3143 int Elt = N->getMaskElt(i);
3144 if (Elt >= 0 && Elt != 1)
3149 for (unsigned i = 2; i < 4; ++i) {
3150 int Elt = N->getMaskElt(i);
3151 if (Elt >= 0 && Elt != 3)
3156 // Don't use movshdup if it can be done with a shufps.
3157 // FIXME: verify that matching u, u, 3, 3 is what we want.
3161 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3162 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3163 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3164 if (N->getValueType(0).getVectorNumElements() != 4)
3167 // Expect 0, 0, 2, 2
3168 for (unsigned i = 0; i < 2; ++i)
3169 if (N->getMaskElt(i) > 0)
3173 for (unsigned i = 2; i < 4; ++i) {
3174 int Elt = N->getMaskElt(i);
3175 if (Elt >= 0 && Elt != 2)
3180 // Don't use movsldup if it can be done with a shufps.
3184 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3185 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3186 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3187 int e = N->getValueType(0).getVectorNumElements() / 2;
3189 for (int i = 0; i < e; ++i)
3190 if (!isUndefOrEqual(N->getMaskElt(i), i))
3192 for (int i = 0; i < e; ++i)
3193 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3198 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3199 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3200 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3201 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3202 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3204 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3206 for (int i = 0; i < NumOperands; ++i) {
3207 int Val = SVOp->getMaskElt(NumOperands-i-1);
3208 if (Val < 0) Val = 0;
3209 if (Val >= NumOperands) Val -= NumOperands;
3211 if (i != NumOperands - 1)
3217 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3218 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3219 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3220 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3222 // 8 nodes, but we only care about the last 4.
3223 for (unsigned i = 7; i >= 4; --i) {
3224 int Val = SVOp->getMaskElt(i);
3233 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3234 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3235 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3236 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3238 // 8 nodes, but we only care about the first 4.
3239 for (int i = 3; i >= 0; --i) {
3240 int Val = SVOp->getMaskElt(i);
3249 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3250 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3251 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3252 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3253 EVT VVT = N->getValueType(0);
3254 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3258 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3259 Val = SVOp->getMaskElt(i);
3263 return (Val - i) * EltSize;
3266 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3268 bool X86::isZeroNode(SDValue Elt) {
3269 return ((isa<ConstantSDNode>(Elt) &&
3270 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3271 (isa<ConstantFPSDNode>(Elt) &&
3272 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3275 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3276 /// their permute mask.
3277 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3278 SelectionDAG &DAG) {
3279 EVT VT = SVOp->getValueType(0);
3280 unsigned NumElems = VT.getVectorNumElements();
3281 SmallVector<int, 8> MaskVec;
3283 for (unsigned i = 0; i != NumElems; ++i) {
3284 int idx = SVOp->getMaskElt(i);
3286 MaskVec.push_back(idx);
3287 else if (idx < (int)NumElems)
3288 MaskVec.push_back(idx + NumElems);
3290 MaskVec.push_back(idx - NumElems);
3292 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3293 SVOp->getOperand(0), &MaskVec[0]);
3296 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3297 /// the two vector operands have swapped position.
3298 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3299 unsigned NumElems = VT.getVectorNumElements();
3300 for (unsigned i = 0; i != NumElems; ++i) {
3304 else if (idx < (int)NumElems)
3305 Mask[i] = idx + NumElems;
3307 Mask[i] = idx - NumElems;
3311 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3312 /// match movhlps. The lower half elements should come from upper half of
3313 /// V1 (and in order), and the upper half elements should come from the upper
3314 /// half of V2 (and in order).
3315 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3316 if (Op->getValueType(0).getVectorNumElements() != 4)
3318 for (unsigned i = 0, e = 2; i != e; ++i)
3319 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3321 for (unsigned i = 2; i != 4; ++i)
3322 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3327 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3328 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3330 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3331 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3333 N = N->getOperand(0).getNode();
3334 if (!ISD::isNON_EXTLoad(N))
3337 *LD = cast<LoadSDNode>(N);
3341 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3342 /// match movlp{s|d}. The lower half elements should come from lower half of
3343 /// V1 (and in order), and the upper half elements should come from the upper
3344 /// half of V2 (and in order). And since V1 will become the source of the
3345 /// MOVLP, it must be either a vector load or a scalar load to vector.
3346 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3347 ShuffleVectorSDNode *Op) {
3348 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3350 // Is V2 is a vector load, don't do this transformation. We will try to use
3351 // load folding shufps op.
3352 if (ISD::isNON_EXTLoad(V2))
3355 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3357 if (NumElems != 2 && NumElems != 4)
3359 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3360 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3362 for (unsigned i = NumElems/2; i != NumElems; ++i)
3363 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3368 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3370 static bool isSplatVector(SDNode *N) {
3371 if (N->getOpcode() != ISD::BUILD_VECTOR)
3374 SDValue SplatValue = N->getOperand(0);
3375 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3376 if (N->getOperand(i) != SplatValue)
3381 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3382 /// to an zero vector.
3383 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3384 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3385 SDValue V1 = N->getOperand(0);
3386 SDValue V2 = N->getOperand(1);
3387 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3388 for (unsigned i = 0; i != NumElems; ++i) {
3389 int Idx = N->getMaskElt(i);
3390 if (Idx >= (int)NumElems) {
3391 unsigned Opc = V2.getOpcode();
3392 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3394 if (Opc != ISD::BUILD_VECTOR ||
3395 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3397 } else if (Idx >= 0) {
3398 unsigned Opc = V1.getOpcode();
3399 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3401 if (Opc != ISD::BUILD_VECTOR ||
3402 !X86::isZeroNode(V1.getOperand(Idx)))
3409 /// getZeroVector - Returns a vector of specified type with all zero elements.
3411 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3413 assert(VT.isVector() && "Expected a vector type");
3415 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted
3416 // to their dest type. This ensures they get CSE'd.
3418 if (VT.getSizeInBits() == 64) { // MMX
3419 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3420 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3421 } else if (VT.getSizeInBits() == 128) {
3422 if (HasSSE2) { // SSE2
3423 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3424 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3426 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3427 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3429 } else if (VT.getSizeInBits() == 256) { // AVX
3430 // 256-bit logic and arithmetic instructions in AVX are
3431 // all floating-point, no support for integer ops. Default
3432 // to emitting fp zeroed vectors then.
3433 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3434 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3435 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
3437 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3440 /// getOnesVector - Returns a vector of specified type with all bits set.
3442 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3443 assert(VT.isVector() && "Expected a vector type");
3445 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3446 // type. This ensures they get CSE'd.
3447 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3449 if (VT.getSizeInBits() == 64) // MMX
3450 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3452 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3453 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3457 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3458 /// that point to V2 points to its first element.
3459 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3460 EVT VT = SVOp->getValueType(0);
3461 unsigned NumElems = VT.getVectorNumElements();
3463 bool Changed = false;
3464 SmallVector<int, 8> MaskVec;
3465 SVOp->getMask(MaskVec);
3467 for (unsigned i = 0; i != NumElems; ++i) {
3468 if (MaskVec[i] > (int)NumElems) {
3469 MaskVec[i] = NumElems;
3474 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3475 SVOp->getOperand(1), &MaskVec[0]);
3476 return SDValue(SVOp, 0);
3479 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3480 /// operation of specified width.
3481 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3483 unsigned NumElems = VT.getVectorNumElements();
3484 SmallVector<int, 8> Mask;
3485 Mask.push_back(NumElems);
3486 for (unsigned i = 1; i != NumElems; ++i)
3488 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3491 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3492 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3494 unsigned NumElems = VT.getVectorNumElements();
3495 SmallVector<int, 8> Mask;
3496 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3498 Mask.push_back(i + NumElems);
3500 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3503 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3504 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3506 unsigned NumElems = VT.getVectorNumElements();
3507 unsigned Half = NumElems/2;
3508 SmallVector<int, 8> Mask;
3509 for (unsigned i = 0; i != Half; ++i) {
3510 Mask.push_back(i + Half);
3511 Mask.push_back(i + NumElems + Half);
3513 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3516 /// PromoteSplat - Promote a splat of v4i32, v8i16 or v16i8 to v4f32.
3517 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
3518 if (SV->getValueType(0).getVectorNumElements() <= 4)
3519 return SDValue(SV, 0);
3521 EVT PVT = MVT::v4f32;
3522 EVT VT = SV->getValueType(0);
3523 DebugLoc dl = SV->getDebugLoc();
3524 SDValue V1 = SV->getOperand(0);
3525 int NumElems = VT.getVectorNumElements();
3526 int EltNo = SV->getSplatIndex();
3528 // unpack elements to the correct location
3529 while (NumElems > 4) {
3530 if (EltNo < NumElems/2) {
3531 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3533 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3534 EltNo -= NumElems/2;
3539 // Perform the splat.
3540 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3541 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3542 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3543 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3546 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3547 /// vector of zero or undef vector. This produces a shuffle where the low
3548 /// element of V2 is swizzled into the zero/undef vector, landing at element
3549 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3550 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3551 bool isZero, bool HasSSE2,
3552 SelectionDAG &DAG) {
3553 EVT VT = V2.getValueType();
3555 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3556 unsigned NumElems = VT.getVectorNumElements();
3557 SmallVector<int, 16> MaskVec;
3558 for (unsigned i = 0; i != NumElems; ++i)
3559 // If this is the insertion idx, put the low elt of V2 here.
3560 MaskVec.push_back(i == Idx ? NumElems : i);
3561 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3564 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3565 /// a shuffle that is zero.
3567 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3568 bool Low, SelectionDAG &DAG) {
3569 unsigned NumZeros = 0;
3570 for (int i = 0; i < NumElems; ++i) {
3571 unsigned Index = Low ? i : NumElems-i-1;
3572 int Idx = SVOp->getMaskElt(Index);
3577 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3578 if (Elt.getNode() && X86::isZeroNode(Elt))
3586 /// isVectorShift - Returns true if the shuffle can be implemented as a
3587 /// logical left or right shift of a vector.
3588 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3589 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3590 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3591 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3594 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3597 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3601 bool SeenV1 = false;
3602 bool SeenV2 = false;
3603 for (unsigned i = NumZeros; i < NumElems; ++i) {
3604 unsigned Val = isLeft ? (i - NumZeros) : i;
3605 int Idx_ = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3608 unsigned Idx = (unsigned) Idx_;
3618 if (SeenV1 && SeenV2)
3621 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3627 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3629 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3630 unsigned NumNonZero, unsigned NumZero,
3632 const TargetLowering &TLI) {
3636 DebugLoc dl = Op.getDebugLoc();
3639 for (unsigned i = 0; i < 16; ++i) {
3640 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3641 if (ThisIsNonZero && First) {
3643 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3645 V = DAG.getUNDEF(MVT::v8i16);
3650 SDValue ThisElt(0, 0), LastElt(0, 0);
3651 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3652 if (LastIsNonZero) {
3653 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3654 MVT::i16, Op.getOperand(i-1));
3656 if (ThisIsNonZero) {
3657 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3658 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3659 ThisElt, DAG.getConstant(8, MVT::i8));
3661 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3665 if (ThisElt.getNode())
3666 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3667 DAG.getIntPtrConstant(i/2));
3671 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3674 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3676 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3677 unsigned NumNonZero, unsigned NumZero,
3679 const TargetLowering &TLI) {
3683 DebugLoc dl = Op.getDebugLoc();
3686 for (unsigned i = 0; i < 8; ++i) {
3687 bool isNonZero = (NonZeros & (1 << i)) != 0;
3691 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3693 V = DAG.getUNDEF(MVT::v8i16);
3696 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3697 MVT::v8i16, V, Op.getOperand(i),
3698 DAG.getIntPtrConstant(i));
3705 /// getVShift - Return a vector logical shift node.
3707 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3708 unsigned NumBits, SelectionDAG &DAG,
3709 const TargetLowering &TLI, DebugLoc dl) {
3710 bool isMMX = VT.getSizeInBits() == 64;
3711 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3712 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3713 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3714 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3715 DAG.getNode(Opc, dl, ShVT, SrcOp,
3716 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3720 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3721 SelectionDAG &DAG) const {
3723 // Check if the scalar load can be widened into a vector load. And if
3724 // the address is "base + cst" see if the cst can be "absorbed" into
3725 // the shuffle mask.
3726 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3727 SDValue Ptr = LD->getBasePtr();
3728 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3730 EVT PVT = LD->getValueType(0);
3731 if (PVT != MVT::i32 && PVT != MVT::f32)
3736 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3737 FI = FINode->getIndex();
3739 } else if (Ptr.getOpcode() == ISD::ADD &&
3740 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3741 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3742 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3743 Offset = Ptr.getConstantOperandVal(1);
3744 Ptr = Ptr.getOperand(0);
3749 SDValue Chain = LD->getChain();
3750 // Make sure the stack object alignment is at least 16.
3751 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3752 if (DAG.InferPtrAlignment(Ptr) < 16) {
3753 if (MFI->isFixedObjectIndex(FI)) {
3754 // Can't change the alignment. FIXME: It's possible to compute
3755 // the exact stack offset and reference FI + adjust offset instead.
3756 // If someone *really* cares about this. That's the way to implement it.
3759 MFI->setObjectAlignment(FI, 16);
3763 // (Offset % 16) must be multiple of 4. Then address is then
3764 // Ptr + (Offset & ~15).
3767 if ((Offset % 16) & 3)
3769 int64_t StartOffset = Offset & ~15;
3771 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3772 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3774 int EltNo = (Offset - StartOffset) >> 2;
3775 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3776 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3777 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3779 // Canonicalize it to a v4i32 shuffle.
3780 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3781 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3782 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3783 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3789 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
3790 /// vector of type 'VT', see if the elements can be replaced by a single large
3791 /// load which has the same value as a build_vector whose operands are 'elts'.
3793 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
3795 /// FIXME: we'd also like to handle the case where the last elements are zero
3796 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
3797 /// There's even a handy isZeroNode for that purpose.
3798 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
3799 DebugLoc &dl, SelectionDAG &DAG) {
3800 EVT EltVT = VT.getVectorElementType();
3801 unsigned NumElems = Elts.size();
3803 LoadSDNode *LDBase = NULL;
3804 unsigned LastLoadedElt = -1U;
3806 // For each element in the initializer, see if we've found a load or an undef.
3807 // If we don't find an initial load element, or later load elements are
3808 // non-consecutive, bail out.
3809 for (unsigned i = 0; i < NumElems; ++i) {
3810 SDValue Elt = Elts[i];
3812 if (!Elt.getNode() ||
3813 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
3816 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
3818 LDBase = cast<LoadSDNode>(Elt.getNode());
3822 if (Elt.getOpcode() == ISD::UNDEF)
3825 LoadSDNode *LD = cast<LoadSDNode>(Elt);
3826 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
3831 // If we have found an entire vector of loads and undefs, then return a large
3832 // load of the entire vector width starting at the base pointer. If we found
3833 // consecutive loads for the low half, generate a vzext_load node.
3834 if (LastLoadedElt == NumElems - 1) {
3835 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
3836 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3837 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3838 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
3839 return DAG.getLoad(VT, dl, LDBase->getChain(), LDBase->getBasePtr(),
3840 LDBase->getSrcValue(), LDBase->getSrcValueOffset(),
3841 LDBase->isVolatile(), LDBase->isNonTemporal(),
3842 LDBase->getAlignment());
3843 } else if (NumElems == 4 && LastLoadedElt == 1) {
3844 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
3845 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
3846 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
3847 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
3853 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
3854 DebugLoc dl = Op.getDebugLoc();
3855 // All zero's are handled with pxor in SSE2 and above, xorps in SSE1 and
3856 // all one's are handled with pcmpeqd. In AVX, zero's are handled with
3857 // vpxor in 128-bit and xor{pd,ps} in 256-bit, but no 256 version of pcmpeqd
3858 // is present, so AllOnes is ignored.
3859 if (ISD::isBuildVectorAllZeros(Op.getNode()) ||
3860 (Op.getValueType().getSizeInBits() != 256 &&
3861 ISD::isBuildVectorAllOnes(Op.getNode()))) {
3862 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3863 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3864 // eliminated on x86-32 hosts.
3865 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3868 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3869 return getOnesVector(Op.getValueType(), DAG, dl);
3870 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3873 EVT VT = Op.getValueType();
3874 EVT ExtVT = VT.getVectorElementType();
3875 unsigned EVTBits = ExtVT.getSizeInBits();
3877 unsigned NumElems = Op.getNumOperands();
3878 unsigned NumZero = 0;
3879 unsigned NumNonZero = 0;
3880 unsigned NonZeros = 0;
3881 bool IsAllConstants = true;
3882 SmallSet<SDValue, 8> Values;
3883 for (unsigned i = 0; i < NumElems; ++i) {
3884 SDValue Elt = Op.getOperand(i);
3885 if (Elt.getOpcode() == ISD::UNDEF)
3888 if (Elt.getOpcode() != ISD::Constant &&
3889 Elt.getOpcode() != ISD::ConstantFP)
3890 IsAllConstants = false;
3891 if (X86::isZeroNode(Elt))
3894 NonZeros |= (1 << i);
3899 if (NumNonZero == 0) {
3900 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3901 return DAG.getUNDEF(VT);
3904 // Special case for single non-zero, non-undef, element.
3905 if (NumNonZero == 1) {
3906 unsigned Idx = CountTrailingZeros_32(NonZeros);
3907 SDValue Item = Op.getOperand(Idx);
3909 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3910 // the value are obviously zero, truncate the value to i32 and do the
3911 // insertion that way. Only do this if the value is non-constant or if the
3912 // value is a constant being inserted into element 0. It is cheaper to do
3913 // a constant pool load than it is to do a movd + shuffle.
3914 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3915 (!IsAllConstants || Idx == 0)) {
3916 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3917 // Handle MMX and SSE both.
3918 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3919 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3921 // Truncate the value (which may itself be a constant) to i32, and
3922 // convert it to a vector with movd (S2V+shuffle to zero extend).
3923 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3924 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3925 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3926 Subtarget->hasSSE2(), DAG);
3928 // Now we have our 32-bit value zero extended in the low element of
3929 // a vector. If Idx != 0, swizzle it into place.
3931 SmallVector<int, 4> Mask;
3932 Mask.push_back(Idx);
3933 for (unsigned i = 1; i != VecElts; ++i)
3935 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3936 DAG.getUNDEF(Item.getValueType()),
3939 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3943 // If we have a constant or non-constant insertion into the low element of
3944 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3945 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3946 // depending on what the source datatype is.
3949 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3950 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3951 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3952 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3953 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3954 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3956 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3957 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3958 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3959 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3960 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3961 Subtarget->hasSSE2(), DAG);
3962 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3966 // Is it a vector logical left shift?
3967 if (NumElems == 2 && Idx == 1 &&
3968 X86::isZeroNode(Op.getOperand(0)) &&
3969 !X86::isZeroNode(Op.getOperand(1))) {
3970 unsigned NumBits = VT.getSizeInBits();
3971 return getVShift(true, VT,
3972 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3973 VT, Op.getOperand(1)),
3974 NumBits/2, DAG, *this, dl);
3977 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3980 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3981 // is a non-constant being inserted into an element other than the low one,
3982 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3983 // movd/movss) to move this into the low element, then shuffle it into
3985 if (EVTBits == 32) {
3986 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3988 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3989 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3990 Subtarget->hasSSE2(), DAG);
3991 SmallVector<int, 8> MaskVec;
3992 for (unsigned i = 0; i < NumElems; i++)
3993 MaskVec.push_back(i == Idx ? 0 : 1);
3994 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3998 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3999 if (Values.size() == 1) {
4000 if (EVTBits == 32) {
4001 // Instead of a shuffle like this:
4002 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
4003 // Check if it's possible to issue this instead.
4004 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
4005 unsigned Idx = CountTrailingZeros_32(NonZeros);
4006 SDValue Item = Op.getOperand(Idx);
4007 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
4008 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
4013 // A vector full of immediates; various special cases are already
4014 // handled, so this is best done with a single constant-pool load.
4018 // Let legalizer expand 2-wide build_vectors.
4019 if (EVTBits == 64) {
4020 if (NumNonZero == 1) {
4021 // One half is zero or undef.
4022 unsigned Idx = CountTrailingZeros_32(NonZeros);
4023 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
4024 Op.getOperand(Idx));
4025 return getShuffleVectorZeroOrUndef(V2, Idx, true,
4026 Subtarget->hasSSE2(), DAG);
4031 // If element VT is < 32 bits, convert it to inserts into a zero vector.
4032 if (EVTBits == 8 && NumElems == 16) {
4033 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
4035 if (V.getNode()) return V;
4038 if (EVTBits == 16 && NumElems == 8) {
4039 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
4041 if (V.getNode()) return V;
4044 // If element VT is == 32 bits, turn it into a number of shuffles.
4045 SmallVector<SDValue, 8> V;
4047 if (NumElems == 4 && NumZero > 0) {
4048 for (unsigned i = 0; i < 4; ++i) {
4049 bool isZero = !(NonZeros & (1 << i));
4051 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4053 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4056 for (unsigned i = 0; i < 2; ++i) {
4057 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
4060 V[i] = V[i*2]; // Must be a zero vector.
4063 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
4066 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
4069 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
4074 SmallVector<int, 8> MaskVec;
4075 bool Reverse = (NonZeros & 0x3) == 2;
4076 for (unsigned i = 0; i < 2; ++i)
4077 MaskVec.push_back(Reverse ? 1-i : i);
4078 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4079 for (unsigned i = 0; i < 2; ++i)
4080 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4081 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4084 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4085 // Check for a build vector of consecutive loads.
4086 for (unsigned i = 0; i < NumElems; ++i)
4087 V[i] = Op.getOperand(i);
4089 // Check for elements which are consecutive loads.
4090 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4094 // For SSE 4.1, use inserts into undef.
4095 if (getSubtarget()->hasSSE41()) {
4096 V[0] = DAG.getUNDEF(VT);
4097 for (unsigned i = 0; i < NumElems; ++i)
4098 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4099 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
4100 Op.getOperand(i), DAG.getIntPtrConstant(i));
4104 // Otherwise, expand into a number of unpckl*
4106 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4107 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4108 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4109 for (unsigned i = 0; i < NumElems; ++i)
4110 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4112 while (NumElems != 0) {
4113 for (unsigned i = 0; i < NumElems; ++i)
4114 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
4123 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4124 // We support concatenate two MMX registers and place them in a MMX
4125 // register. This is better than doing a stack convert.
4126 DebugLoc dl = Op.getDebugLoc();
4127 EVT ResVT = Op.getValueType();
4128 assert(Op.getNumOperands() == 2);
4129 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4130 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4132 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
4133 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4134 InVec = Op.getOperand(1);
4135 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4136 unsigned NumElts = ResVT.getVectorNumElements();
4137 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4138 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4139 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4141 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
4142 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4143 Mask[0] = 0; Mask[1] = 2;
4144 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4146 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
4149 // v8i16 shuffles - Prefer shuffles in the following order:
4150 // 1. [all] pshuflw, pshufhw, optional move
4151 // 2. [ssse3] 1 x pshufb
4152 // 3. [ssse3] 2 x pshufb + 1 x por
4153 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4155 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
4157 const X86TargetLowering &TLI) {
4158 SDValue V1 = SVOp->getOperand(0);
4159 SDValue V2 = SVOp->getOperand(1);
4160 DebugLoc dl = SVOp->getDebugLoc();
4161 SmallVector<int, 8> MaskVals;
4163 // Determine if more than 1 of the words in each of the low and high quadwords
4164 // of the result come from the same quadword of one of the two inputs. Undef
4165 // mask values count as coming from any quadword, for better codegen.
4166 SmallVector<unsigned, 4> LoQuad(4);
4167 SmallVector<unsigned, 4> HiQuad(4);
4168 BitVector InputQuads(4);
4169 for (unsigned i = 0; i < 8; ++i) {
4170 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4171 int EltIdx = SVOp->getMaskElt(i);
4172 MaskVals.push_back(EltIdx);
4181 InputQuads.set(EltIdx / 4);
4184 int BestLoQuad = -1;
4185 unsigned MaxQuad = 1;
4186 for (unsigned i = 0; i < 4; ++i) {
4187 if (LoQuad[i] > MaxQuad) {
4189 MaxQuad = LoQuad[i];
4193 int BestHiQuad = -1;
4195 for (unsigned i = 0; i < 4; ++i) {
4196 if (HiQuad[i] > MaxQuad) {
4198 MaxQuad = HiQuad[i];
4202 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4203 // of the two input vectors, shuffle them into one input vector so only a
4204 // single pshufb instruction is necessary. If There are more than 2 input
4205 // quads, disable the next transformation since it does not help SSSE3.
4206 bool V1Used = InputQuads[0] || InputQuads[1];
4207 bool V2Used = InputQuads[2] || InputQuads[3];
4208 if (TLI.getSubtarget()->hasSSSE3()) {
4209 if (InputQuads.count() == 2 && V1Used && V2Used) {
4210 BestLoQuad = InputQuads.find_first();
4211 BestHiQuad = InputQuads.find_next(BestLoQuad);
4213 if (InputQuads.count() > 2) {
4219 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4220 // the shuffle mask. If a quad is scored as -1, that means that it contains
4221 // words from all 4 input quadwords.
4223 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4224 SmallVector<int, 8> MaskV;
4225 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4226 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4227 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4228 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
4229 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
4230 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
4232 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4233 // source words for the shuffle, to aid later transformations.
4234 bool AllWordsInNewV = true;
4235 bool InOrder[2] = { true, true };
4236 for (unsigned i = 0; i != 8; ++i) {
4237 int idx = MaskVals[i];
4239 InOrder[i/4] = false;
4240 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4242 AllWordsInNewV = false;
4246 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4247 if (AllWordsInNewV) {
4248 for (int i = 0; i != 8; ++i) {
4249 int idx = MaskVals[i];
4252 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4253 if ((idx != i) && idx < 4)
4255 if ((idx != i) && idx > 3)
4264 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4265 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4266 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4267 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4268 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4272 // If we have SSSE3, and all words of the result are from 1 input vector,
4273 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4274 // is present, fall back to case 4.
4275 if (TLI.getSubtarget()->hasSSSE3()) {
4276 SmallVector<SDValue,16> pshufbMask;
4278 // If we have elements from both input vectors, set the high bit of the
4279 // shuffle mask element to zero out elements that come from V2 in the V1
4280 // mask, and elements that come from V1 in the V2 mask, so that the two
4281 // results can be OR'd together.
4282 bool TwoInputs = V1Used && V2Used;
4283 for (unsigned i = 0; i != 8; ++i) {
4284 int EltIdx = MaskVals[i] * 2;
4285 if (TwoInputs && (EltIdx >= 16)) {
4286 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4287 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4290 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4291 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4293 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4294 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4295 DAG.getNode(ISD::BUILD_VECTOR, dl,
4296 MVT::v16i8, &pshufbMask[0], 16));
4298 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4300 // Calculate the shuffle mask for the second input, shuffle it, and
4301 // OR it with the first shuffled input.
4303 for (unsigned i = 0; i != 8; ++i) {
4304 int EltIdx = MaskVals[i] * 2;
4306 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4307 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4310 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4311 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4313 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4314 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4315 DAG.getNode(ISD::BUILD_VECTOR, dl,
4316 MVT::v16i8, &pshufbMask[0], 16));
4317 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4318 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4321 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4322 // and update MaskVals with new element order.
4323 BitVector InOrder(8);
4324 if (BestLoQuad >= 0) {
4325 SmallVector<int, 8> MaskV;
4326 for (int i = 0; i != 4; ++i) {
4327 int idx = MaskVals[i];
4329 MaskV.push_back(-1);
4331 } else if ((idx / 4) == BestLoQuad) {
4332 MaskV.push_back(idx & 3);
4335 MaskV.push_back(-1);
4338 for (unsigned i = 4; i != 8; ++i)
4340 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4344 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4345 // and update MaskVals with the new element order.
4346 if (BestHiQuad >= 0) {
4347 SmallVector<int, 8> MaskV;
4348 for (unsigned i = 0; i != 4; ++i)
4350 for (unsigned i = 4; i != 8; ++i) {
4351 int idx = MaskVals[i];
4353 MaskV.push_back(-1);
4355 } else if ((idx / 4) == BestHiQuad) {
4356 MaskV.push_back((idx & 3) + 4);
4359 MaskV.push_back(-1);
4362 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4366 // In case BestHi & BestLo were both -1, which means each quadword has a word
4367 // from each of the four input quadwords, calculate the InOrder bitvector now
4368 // before falling through to the insert/extract cleanup.
4369 if (BestLoQuad == -1 && BestHiQuad == -1) {
4371 for (int i = 0; i != 8; ++i)
4372 if (MaskVals[i] < 0 || MaskVals[i] == i)
4376 // The other elements are put in the right place using pextrw and pinsrw.
4377 for (unsigned i = 0; i != 8; ++i) {
4380 int EltIdx = MaskVals[i];
4383 SDValue ExtOp = (EltIdx < 8)
4384 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4385 DAG.getIntPtrConstant(EltIdx))
4386 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4387 DAG.getIntPtrConstant(EltIdx - 8));
4388 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4389 DAG.getIntPtrConstant(i));
4394 // v16i8 shuffles - Prefer shuffles in the following order:
4395 // 1. [ssse3] 1 x pshufb
4396 // 2. [ssse3] 2 x pshufb + 1 x por
4397 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4399 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4401 const X86TargetLowering &TLI) {
4402 SDValue V1 = SVOp->getOperand(0);
4403 SDValue V2 = SVOp->getOperand(1);
4404 DebugLoc dl = SVOp->getDebugLoc();
4405 SmallVector<int, 16> MaskVals;
4406 SVOp->getMask(MaskVals);
4408 // If we have SSSE3, case 1 is generated when all result bytes come from
4409 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4410 // present, fall back to case 3.
4411 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4414 for (unsigned i = 0; i < 16; ++i) {
4415 int EltIdx = MaskVals[i];
4424 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4425 if (TLI.getSubtarget()->hasSSSE3()) {
4426 SmallVector<SDValue,16> pshufbMask;
4428 // If all result elements are from one input vector, then only translate
4429 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4431 // Otherwise, we have elements from both input vectors, and must zero out
4432 // elements that come from V2 in the first mask, and V1 in the second mask
4433 // so that we can OR them together.
4434 bool TwoInputs = !(V1Only || V2Only);
4435 for (unsigned i = 0; i != 16; ++i) {
4436 int EltIdx = MaskVals[i];
4437 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4438 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4441 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4443 // If all the elements are from V2, assign it to V1 and return after
4444 // building the first pshufb.
4447 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4448 DAG.getNode(ISD::BUILD_VECTOR, dl,
4449 MVT::v16i8, &pshufbMask[0], 16));
4453 // Calculate the shuffle mask for the second input, shuffle it, and
4454 // OR it with the first shuffled input.
4456 for (unsigned i = 0; i != 16; ++i) {
4457 int EltIdx = MaskVals[i];
4459 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4462 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4464 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4465 DAG.getNode(ISD::BUILD_VECTOR, dl,
4466 MVT::v16i8, &pshufbMask[0], 16));
4467 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4470 // No SSSE3 - Calculate in place words and then fix all out of place words
4471 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4472 // the 16 different words that comprise the two doublequadword input vectors.
4473 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4474 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4475 SDValue NewV = V2Only ? V2 : V1;
4476 for (int i = 0; i != 8; ++i) {
4477 int Elt0 = MaskVals[i*2];
4478 int Elt1 = MaskVals[i*2+1];
4480 // This word of the result is all undef, skip it.
4481 if (Elt0 < 0 && Elt1 < 0)
4484 // This word of the result is already in the correct place, skip it.
4485 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4487 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4490 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4491 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4494 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4495 // using a single extract together, load it and store it.
4496 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4497 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4498 DAG.getIntPtrConstant(Elt1 / 2));
4499 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4500 DAG.getIntPtrConstant(i));
4504 // If Elt1 is defined, extract it from the appropriate source. If the
4505 // source byte is not also odd, shift the extracted word left 8 bits
4506 // otherwise clear the bottom 8 bits if we need to do an or.
4508 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4509 DAG.getIntPtrConstant(Elt1 / 2));
4510 if ((Elt1 & 1) == 0)
4511 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4512 DAG.getConstant(8, TLI.getShiftAmountTy()));
4514 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4515 DAG.getConstant(0xFF00, MVT::i16));
4517 // If Elt0 is defined, extract it from the appropriate source. If the
4518 // source byte is not also even, shift the extracted word right 8 bits. If
4519 // Elt1 was also defined, OR the extracted values together before
4520 // inserting them in the result.
4522 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4523 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4524 if ((Elt0 & 1) != 0)
4525 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4526 DAG.getConstant(8, TLI.getShiftAmountTy()));
4528 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4529 DAG.getConstant(0x00FF, MVT::i16));
4530 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4533 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4534 DAG.getIntPtrConstant(i));
4536 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4539 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4540 /// ones, or rewriting v4i32 / v2i32 as 2 wide ones if possible. This can be
4541 /// done when every pair / quad of shuffle mask elements point to elements in
4542 /// the right sequence. e.g.
4543 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4545 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4547 const TargetLowering &TLI, DebugLoc dl) {
4548 EVT VT = SVOp->getValueType(0);
4549 SDValue V1 = SVOp->getOperand(0);
4550 SDValue V2 = SVOp->getOperand(1);
4551 unsigned NumElems = VT.getVectorNumElements();
4552 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4553 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4555 switch (VT.getSimpleVT().SimpleTy) {
4556 default: assert(false && "Unexpected!");
4557 case MVT::v4f32: NewVT = MVT::v2f64; break;
4558 case MVT::v4i32: NewVT = MVT::v2i64; break;
4559 case MVT::v8i16: NewVT = MVT::v4i32; break;
4560 case MVT::v16i8: NewVT = MVT::v4i32; break;
4563 if (NewWidth == 2) {
4569 int Scale = NumElems / NewWidth;
4570 SmallVector<int, 8> MaskVec;
4571 for (unsigned i = 0; i < NumElems; i += Scale) {
4573 for (int j = 0; j < Scale; ++j) {
4574 int EltIdx = SVOp->getMaskElt(i+j);
4578 StartIdx = EltIdx - (EltIdx % Scale);
4579 if (EltIdx != StartIdx + j)
4583 MaskVec.push_back(-1);
4585 MaskVec.push_back(StartIdx / Scale);
4588 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4589 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4590 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4593 /// getVZextMovL - Return a zero-extending vector move low node.
4595 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4596 SDValue SrcOp, SelectionDAG &DAG,
4597 const X86Subtarget *Subtarget, DebugLoc dl) {
4598 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4599 LoadSDNode *LD = NULL;
4600 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4601 LD = dyn_cast<LoadSDNode>(SrcOp);
4603 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4605 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4606 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4607 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4608 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4609 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4611 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4612 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4613 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4614 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4622 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4623 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4624 DAG.getNode(ISD::BIT_CONVERT, dl,
4628 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4631 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4632 SDValue V1 = SVOp->getOperand(0);
4633 SDValue V2 = SVOp->getOperand(1);
4634 DebugLoc dl = SVOp->getDebugLoc();
4635 EVT VT = SVOp->getValueType(0);
4637 SmallVector<std::pair<int, int>, 8> Locs;
4639 SmallVector<int, 8> Mask1(4U, -1);
4640 SmallVector<int, 8> PermMask;
4641 SVOp->getMask(PermMask);
4645 for (unsigned i = 0; i != 4; ++i) {
4646 int Idx = PermMask[i];
4648 Locs[i] = std::make_pair(-1, -1);
4650 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4652 Locs[i] = std::make_pair(0, NumLo);
4656 Locs[i] = std::make_pair(1, NumHi);
4658 Mask1[2+NumHi] = Idx;
4664 if (NumLo <= 2 && NumHi <= 2) {
4665 // If no more than two elements come from either vector. This can be
4666 // implemented with two shuffles. First shuffle gather the elements.
4667 // The second shuffle, which takes the first shuffle as both of its
4668 // vector operands, put the elements into the right order.
4669 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4671 SmallVector<int, 8> Mask2(4U, -1);
4673 for (unsigned i = 0; i != 4; ++i) {
4674 if (Locs[i].first == -1)
4677 unsigned Idx = (i < 2) ? 0 : 4;
4678 Idx += Locs[i].first * 2 + Locs[i].second;
4683 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4684 } else if (NumLo == 3 || NumHi == 3) {
4685 // Otherwise, we must have three elements from one vector, call it X, and
4686 // one element from the other, call it Y. First, use a shufps to build an
4687 // intermediate vector with the one element from Y and the element from X
4688 // that will be in the same half in the final destination (the indexes don't
4689 // matter). Then, use a shufps to build the final vector, taking the half
4690 // containing the element from Y from the intermediate, and the other half
4693 // Normalize it so the 3 elements come from V1.
4694 CommuteVectorShuffleMask(PermMask, VT);
4698 // Find the element from V2.
4700 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4701 int Val = PermMask[HiIndex];
4708 Mask1[0] = PermMask[HiIndex];
4710 Mask1[2] = PermMask[HiIndex^1];
4712 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4715 Mask1[0] = PermMask[0];
4716 Mask1[1] = PermMask[1];
4717 Mask1[2] = HiIndex & 1 ? 6 : 4;
4718 Mask1[3] = HiIndex & 1 ? 4 : 6;
4719 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4721 Mask1[0] = HiIndex & 1 ? 2 : 0;
4722 Mask1[1] = HiIndex & 1 ? 0 : 2;
4723 Mask1[2] = PermMask[2];
4724 Mask1[3] = PermMask[3];
4729 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4733 // Break it into (shuffle shuffle_hi, shuffle_lo).
4735 SmallVector<int,8> LoMask(4U, -1);
4736 SmallVector<int,8> HiMask(4U, -1);
4738 SmallVector<int,8> *MaskPtr = &LoMask;
4739 unsigned MaskIdx = 0;
4742 for (unsigned i = 0; i != 4; ++i) {
4749 int Idx = PermMask[i];
4751 Locs[i] = std::make_pair(-1, -1);
4752 } else if (Idx < 4) {
4753 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4754 (*MaskPtr)[LoIdx] = Idx;
4757 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4758 (*MaskPtr)[HiIdx] = Idx;
4763 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4764 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4765 SmallVector<int, 8> MaskOps;
4766 for (unsigned i = 0; i != 4; ++i) {
4767 if (Locs[i].first == -1) {
4768 MaskOps.push_back(-1);
4770 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4771 MaskOps.push_back(Idx);
4774 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4778 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
4779 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4780 SDValue V1 = Op.getOperand(0);
4781 SDValue V2 = Op.getOperand(1);
4782 EVT VT = Op.getValueType();
4783 DebugLoc dl = Op.getDebugLoc();
4784 unsigned NumElems = VT.getVectorNumElements();
4785 bool isMMX = VT.getSizeInBits() == 64;
4786 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4787 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4788 bool V1IsSplat = false;
4789 bool V2IsSplat = false;
4791 if (isZeroShuffle(SVOp))
4792 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4794 // Promote splats to v4f32.
4795 if (SVOp->isSplat()) {
4796 if (isMMX || NumElems < 4)
4798 return PromoteSplat(SVOp, DAG);
4801 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4803 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4804 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4805 if (NewOp.getNode())
4806 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4807 LowerVECTOR_SHUFFLE(NewOp, DAG));
4808 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4809 // FIXME: Figure out a cleaner way to do this.
4810 // Try to make use of movq to zero out the top part.
4811 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4812 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4813 if (NewOp.getNode()) {
4814 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4815 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4816 DAG, Subtarget, dl);
4818 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4819 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4820 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4821 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4822 DAG, Subtarget, dl);
4826 if (X86::isPSHUFDMask(SVOp))
4829 // Check if this can be converted into a logical shift.
4830 bool isLeft = false;
4833 bool isShift = getSubtarget()->hasSSE2() &&
4834 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4835 if (isShift && ShVal.hasOneUse()) {
4836 // If the shifted value has multiple uses, it may be cheaper to use
4837 // v_set0 + movlhps or movhlps, etc.
4838 EVT EltVT = VT.getVectorElementType();
4839 ShAmt *= EltVT.getSizeInBits();
4840 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4843 if (X86::isMOVLMask(SVOp)) {
4846 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4847 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4852 // FIXME: fold these into legal mask.
4853 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4854 X86::isMOVSLDUPMask(SVOp) ||
4855 X86::isMOVHLPSMask(SVOp) ||
4856 X86::isMOVLHPSMask(SVOp) ||
4857 X86::isMOVLPMask(SVOp)))
4860 if (ShouldXformToMOVHLPS(SVOp) ||
4861 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4862 return CommuteVectorShuffle(SVOp, DAG);
4865 // No better options. Use a vshl / vsrl.
4866 EVT EltVT = VT.getVectorElementType();
4867 ShAmt *= EltVT.getSizeInBits();
4868 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4871 bool Commuted = false;
4872 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4873 // 1,1,1,1 -> v8i16 though.
4874 V1IsSplat = isSplatVector(V1.getNode());
4875 V2IsSplat = isSplatVector(V2.getNode());
4877 // Canonicalize the splat or undef, if present, to be on the RHS.
4878 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4879 Op = CommuteVectorShuffle(SVOp, DAG);
4880 SVOp = cast<ShuffleVectorSDNode>(Op);
4881 V1 = SVOp->getOperand(0);
4882 V2 = SVOp->getOperand(1);
4883 std::swap(V1IsSplat, V2IsSplat);
4884 std::swap(V1IsUndef, V2IsUndef);
4888 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4889 // Shuffling low element of v1 into undef, just return v1.
4892 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4893 // the instruction selector will not match, so get a canonical MOVL with
4894 // swapped operands to undo the commute.
4895 return getMOVL(DAG, dl, VT, V2, V1);
4898 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4899 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4900 X86::isUNPCKLMask(SVOp) ||
4901 X86::isUNPCKHMask(SVOp))
4905 // Normalize mask so all entries that point to V2 points to its first
4906 // element then try to match unpck{h|l} again. If match, return a
4907 // new vector_shuffle with the corrected mask.
4908 SDValue NewMask = NormalizeMask(SVOp, DAG);
4909 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4910 if (NSVOp != SVOp) {
4911 if (X86::isUNPCKLMask(NSVOp, true)) {
4913 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4920 // Commute is back and try unpck* again.
4921 // FIXME: this seems wrong.
4922 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4923 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4924 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4925 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4926 X86::isUNPCKLMask(NewSVOp) ||
4927 X86::isUNPCKHMask(NewSVOp))
4931 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4933 // Normalize the node to match x86 shuffle ops if needed
4934 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4935 return CommuteVectorShuffle(SVOp, DAG);
4937 // Check for legal shuffle and return?
4938 SmallVector<int, 16> PermMask;
4939 SVOp->getMask(PermMask);
4940 if (isShuffleMaskLegal(PermMask, VT))
4943 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4944 if (VT == MVT::v8i16) {
4945 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4946 if (NewOp.getNode())
4950 if (VT == MVT::v16i8) {
4951 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4952 if (NewOp.getNode())
4956 // Handle all 4 wide cases with a number of shuffles except for MMX.
4957 if (NumElems == 4 && !isMMX)
4958 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4964 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4965 SelectionDAG &DAG) const {
4966 EVT VT = Op.getValueType();
4967 DebugLoc dl = Op.getDebugLoc();
4968 if (VT.getSizeInBits() == 8) {
4969 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4970 Op.getOperand(0), Op.getOperand(1));
4971 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4972 DAG.getValueType(VT));
4973 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4974 } else if (VT.getSizeInBits() == 16) {
4975 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4976 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4978 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4979 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4980 DAG.getNode(ISD::BIT_CONVERT, dl,
4984 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4985 Op.getOperand(0), Op.getOperand(1));
4986 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4987 DAG.getValueType(VT));
4988 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4989 } else if (VT == MVT::f32) {
4990 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4991 // the result back to FR32 register. It's only worth matching if the
4992 // result has a single use which is a store or a bitcast to i32. And in
4993 // the case of a store, it's not worth it if the index is a constant 0,
4994 // because a MOVSSmr can be used instead, which is smaller and faster.
4995 if (!Op.hasOneUse())
4997 SDNode *User = *Op.getNode()->use_begin();
4998 if ((User->getOpcode() != ISD::STORE ||
4999 (isa<ConstantSDNode>(Op.getOperand(1)) &&
5000 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
5001 (User->getOpcode() != ISD::BIT_CONVERT ||
5002 User->getValueType(0) != MVT::i32))
5004 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5005 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
5008 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
5009 } else if (VT == MVT::i32) {
5010 // ExtractPS works with constant index.
5011 if (isa<ConstantSDNode>(Op.getOperand(1)))
5019 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5020 SelectionDAG &DAG) const {
5021 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5024 if (Subtarget->hasSSE41()) {
5025 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
5030 EVT VT = Op.getValueType();
5031 DebugLoc dl = Op.getDebugLoc();
5032 // TODO: handle v16i8.
5033 if (VT.getSizeInBits() == 16) {
5034 SDValue Vec = Op.getOperand(0);
5035 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5037 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5038 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5039 DAG.getNode(ISD::BIT_CONVERT, dl,
5042 // Transform it so it match pextrw which produces a 32-bit result.
5043 EVT EltVT = MVT::i32;
5044 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
5045 Op.getOperand(0), Op.getOperand(1));
5046 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
5047 DAG.getValueType(VT));
5048 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5049 } else if (VT.getSizeInBits() == 32) {
5050 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5054 // SHUFPS the element to the lowest double word, then movss.
5055 int Mask[4] = { Idx, -1, -1, -1 };
5056 EVT VVT = Op.getOperand(0).getValueType();
5057 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5058 DAG.getUNDEF(VVT), Mask);
5059 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5060 DAG.getIntPtrConstant(0));
5061 } else if (VT.getSizeInBits() == 64) {
5062 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
5063 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
5064 // to match extract_elt for f64.
5065 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5069 // UNPCKHPD the element to the lowest double word, then movsd.
5070 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
5071 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
5072 int Mask[2] = { 1, -1 };
5073 EVT VVT = Op.getOperand(0).getValueType();
5074 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5075 DAG.getUNDEF(VVT), Mask);
5076 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5077 DAG.getIntPtrConstant(0));
5084 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
5085 SelectionDAG &DAG) const {
5086 EVT VT = Op.getValueType();
5087 EVT EltVT = VT.getVectorElementType();
5088 DebugLoc dl = Op.getDebugLoc();
5090 SDValue N0 = Op.getOperand(0);
5091 SDValue N1 = Op.getOperand(1);
5092 SDValue N2 = Op.getOperand(2);
5094 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
5095 isa<ConstantSDNode>(N2)) {
5097 if (VT == MVT::v8i16)
5098 Opc = X86ISD::PINSRW;
5099 else if (VT == MVT::v4i16)
5100 Opc = X86ISD::MMX_PINSRW;
5101 else if (VT == MVT::v16i8)
5102 Opc = X86ISD::PINSRB;
5104 Opc = X86ISD::PINSRB;
5106 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5108 if (N1.getValueType() != MVT::i32)
5109 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5110 if (N2.getValueType() != MVT::i32)
5111 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5112 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5113 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5114 // Bits [7:6] of the constant are the source select. This will always be
5115 // zero here. The DAG Combiner may combine an extract_elt index into these
5116 // bits. For example (insert (extract, 3), 2) could be matched by putting
5117 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5118 // Bits [5:4] of the constant are the destination select. This is the
5119 // value of the incoming immediate.
5120 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5121 // combine either bitwise AND or insert of float 0.0 to set these bits.
5122 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5123 // Create this as a scalar to vector..
5124 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5125 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5126 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5127 // PINSR* works with constant index.
5134 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5135 EVT VT = Op.getValueType();
5136 EVT EltVT = VT.getVectorElementType();
5138 if (Subtarget->hasSSE41())
5139 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5141 if (EltVT == MVT::i8)
5144 DebugLoc dl = Op.getDebugLoc();
5145 SDValue N0 = Op.getOperand(0);
5146 SDValue N1 = Op.getOperand(1);
5147 SDValue N2 = Op.getOperand(2);
5149 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5150 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5151 // as its second argument.
5152 if (N1.getValueType() != MVT::i32)
5153 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5154 if (N2.getValueType() != MVT::i32)
5155 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5156 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
5157 dl, VT, N0, N1, N2);
5163 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5164 DebugLoc dl = Op.getDebugLoc();
5166 if (Op.getValueType() == MVT::v1i64 &&
5167 Op.getOperand(0).getValueType() == MVT::i64)
5168 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
5170 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
5171 EVT VT = MVT::v2i32;
5172 switch (Op.getValueType().getSimpleVT().SimpleTy) {
5179 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
5180 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
5183 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
5184 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
5185 // one of the above mentioned nodes. It has to be wrapped because otherwise
5186 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
5187 // be used to form addressing mode. These wrapped nodes will be selected
5190 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
5191 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
5193 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5195 unsigned char OpFlag = 0;
5196 unsigned WrapperKind = X86ISD::Wrapper;
5197 CodeModel::Model M = getTargetMachine().getCodeModel();
5199 if (Subtarget->isPICStyleRIPRel() &&
5200 (M == CodeModel::Small || M == CodeModel::Kernel))
5201 WrapperKind = X86ISD::WrapperRIP;
5202 else if (Subtarget->isPICStyleGOT())
5203 OpFlag = X86II::MO_GOTOFF;
5204 else if (Subtarget->isPICStyleStubPIC())
5205 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5207 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
5209 CP->getOffset(), OpFlag);
5210 DebugLoc DL = CP->getDebugLoc();
5211 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5212 // With PIC, the address is actually $g + Offset.
5214 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5215 DAG.getNode(X86ISD::GlobalBaseReg,
5216 DebugLoc(), getPointerTy()),
5223 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
5224 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
5226 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5228 unsigned char OpFlag = 0;
5229 unsigned WrapperKind = X86ISD::Wrapper;
5230 CodeModel::Model M = getTargetMachine().getCodeModel();
5232 if (Subtarget->isPICStyleRIPRel() &&
5233 (M == CodeModel::Small || M == CodeModel::Kernel))
5234 WrapperKind = X86ISD::WrapperRIP;
5235 else if (Subtarget->isPICStyleGOT())
5236 OpFlag = X86II::MO_GOTOFF;
5237 else if (Subtarget->isPICStyleStubPIC())
5238 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5240 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
5242 DebugLoc DL = JT->getDebugLoc();
5243 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5245 // With PIC, the address is actually $g + Offset.
5247 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5248 DAG.getNode(X86ISD::GlobalBaseReg,
5249 DebugLoc(), getPointerTy()),
5257 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
5258 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
5260 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5262 unsigned char OpFlag = 0;
5263 unsigned WrapperKind = X86ISD::Wrapper;
5264 CodeModel::Model M = getTargetMachine().getCodeModel();
5266 if (Subtarget->isPICStyleRIPRel() &&
5267 (M == CodeModel::Small || M == CodeModel::Kernel))
5268 WrapperKind = X86ISD::WrapperRIP;
5269 else if (Subtarget->isPICStyleGOT())
5270 OpFlag = X86II::MO_GOTOFF;
5271 else if (Subtarget->isPICStyleStubPIC())
5272 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5274 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5276 DebugLoc DL = Op.getDebugLoc();
5277 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5280 // With PIC, the address is actually $g + Offset.
5281 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5282 !Subtarget->is64Bit()) {
5283 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5284 DAG.getNode(X86ISD::GlobalBaseReg,
5285 DebugLoc(), getPointerTy()),
5293 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
5294 // Create the TargetBlockAddressAddress node.
5295 unsigned char OpFlags =
5296 Subtarget->ClassifyBlockAddressReference();
5297 CodeModel::Model M = getTargetMachine().getCodeModel();
5298 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5299 DebugLoc dl = Op.getDebugLoc();
5300 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5301 /*isTarget=*/true, OpFlags);
5303 if (Subtarget->isPICStyleRIPRel() &&
5304 (M == CodeModel::Small || M == CodeModel::Kernel))
5305 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5307 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5309 // With PIC, the address is actually $g + Offset.
5310 if (isGlobalRelativeToPICBase(OpFlags)) {
5311 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5312 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5320 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5322 SelectionDAG &DAG) const {
5323 // Create the TargetGlobalAddress node, folding in the constant
5324 // offset if it is legal.
5325 unsigned char OpFlags =
5326 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5327 CodeModel::Model M = getTargetMachine().getCodeModel();
5329 if (OpFlags == X86II::MO_NO_FLAG &&
5330 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5331 // A direct static reference to a global.
5332 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
5335 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
5338 if (Subtarget->isPICStyleRIPRel() &&
5339 (M == CodeModel::Small || M == CodeModel::Kernel))
5340 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5342 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5344 // With PIC, the address is actually $g + Offset.
5345 if (isGlobalRelativeToPICBase(OpFlags)) {
5346 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5347 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5351 // For globals that require a load from a stub to get the address, emit the
5353 if (isGlobalStubReference(OpFlags))
5354 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5355 PseudoSourceValue::getGOT(), 0, false, false, 0);
5357 // If there was a non-zero offset that we didn't fold, create an explicit
5360 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5361 DAG.getConstant(Offset, getPointerTy()));
5367 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
5368 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5369 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5370 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5374 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5375 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5376 unsigned char OperandFlags) {
5377 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5378 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5379 DebugLoc dl = GA->getDebugLoc();
5380 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5381 GA->getValueType(0),
5385 SDValue Ops[] = { Chain, TGA, *InFlag };
5386 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5388 SDValue Ops[] = { Chain, TGA };
5389 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5392 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5393 MFI->setAdjustsStack(true);
5395 SDValue Flag = Chain.getValue(1);
5396 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5399 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5401 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5404 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5405 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5406 DAG.getNode(X86ISD::GlobalBaseReg,
5407 DebugLoc(), PtrVT), InFlag);
5408 InFlag = Chain.getValue(1);
5410 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5413 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5415 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5417 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5418 X86::RAX, X86II::MO_TLSGD);
5421 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5422 // "local exec" model.
5423 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5424 const EVT PtrVT, TLSModel::Model model,
5426 DebugLoc dl = GA->getDebugLoc();
5427 // Get the Thread Pointer
5428 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5430 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5433 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5434 NULL, 0, false, false, 0);
5436 unsigned char OperandFlags = 0;
5437 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5439 unsigned WrapperKind = X86ISD::Wrapper;
5440 if (model == TLSModel::LocalExec) {
5441 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5442 } else if (is64Bit) {
5443 assert(model == TLSModel::InitialExec);
5444 OperandFlags = X86II::MO_GOTTPOFF;
5445 WrapperKind = X86ISD::WrapperRIP;
5447 assert(model == TLSModel::InitialExec);
5448 OperandFlags = X86II::MO_INDNTPOFF;
5451 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5453 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
5454 GA->getValueType(0),
5455 GA->getOffset(), OperandFlags);
5456 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5458 if (model == TLSModel::InitialExec)
5459 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5460 PseudoSourceValue::getGOT(), 0, false, false, 0);
5462 // The address of the thread local variable is the add of the thread
5463 // pointer with the offset of the variable.
5464 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5468 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
5470 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5471 const GlobalValue *GV = GA->getGlobal();
5473 if (Subtarget->isTargetELF()) {
5474 // TODO: implement the "local dynamic" model
5475 // TODO: implement the "initial exec"model for pic executables
5477 // If GV is an alias then use the aliasee for determining
5478 // thread-localness.
5479 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5480 GV = GA->resolveAliasedGlobal(false);
5482 TLSModel::Model model
5483 = getTLSModel(GV, getTargetMachine().getRelocationModel());
5486 case TLSModel::GeneralDynamic:
5487 case TLSModel::LocalDynamic: // not implemented
5488 if (Subtarget->is64Bit())
5489 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5490 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5492 case TLSModel::InitialExec:
5493 case TLSModel::LocalExec:
5494 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5495 Subtarget->is64Bit());
5497 } else if (Subtarget->isTargetDarwin()) {
5498 // Darwin only has one model of TLS. Lower to that.
5499 unsigned char OpFlag = 0;
5500 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
5501 X86ISD::WrapperRIP : X86ISD::Wrapper;
5503 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
5505 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
5506 !Subtarget->is64Bit();
5508 OpFlag = X86II::MO_TLVP_PIC_BASE;
5510 OpFlag = X86II::MO_TLVP;
5511 DebugLoc DL = Op.getDebugLoc();
5512 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
5514 GA->getOffset(), OpFlag);
5515 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5517 // With PIC32, the address is actually $g + Offset.
5519 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5520 DAG.getNode(X86ISD::GlobalBaseReg,
5521 DebugLoc(), getPointerTy()),
5524 // Lowering the machine isd will make sure everything is in the right
5526 SDValue Args[] = { Offset };
5527 SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
5529 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
5530 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5531 MFI->setAdjustsStack(true);
5533 // And our return value (tls address) is in the standard call return value
5535 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
5536 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
5540 "TLS not implemented for this target.");
5542 llvm_unreachable("Unreachable");
5547 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5548 /// take a 2 x i32 value to shift plus a shift amount.
5549 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
5550 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5551 EVT VT = Op.getValueType();
5552 unsigned VTBits = VT.getSizeInBits();
5553 DebugLoc dl = Op.getDebugLoc();
5554 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5555 SDValue ShOpLo = Op.getOperand(0);
5556 SDValue ShOpHi = Op.getOperand(1);
5557 SDValue ShAmt = Op.getOperand(2);
5558 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5559 DAG.getConstant(VTBits - 1, MVT::i8))
5560 : DAG.getConstant(0, VT);
5563 if (Op.getOpcode() == ISD::SHL_PARTS) {
5564 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5565 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5567 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5568 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5571 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5572 DAG.getConstant(VTBits, MVT::i8));
5573 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5574 AndNode, DAG.getConstant(0, MVT::i8));
5577 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5578 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5579 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5581 if (Op.getOpcode() == ISD::SHL_PARTS) {
5582 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5583 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5585 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5586 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5589 SDValue Ops[2] = { Lo, Hi };
5590 return DAG.getMergeValues(Ops, 2, dl);
5593 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
5594 SelectionDAG &DAG) const {
5595 EVT SrcVT = Op.getOperand(0).getValueType();
5597 if (SrcVT.isVector()) {
5598 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5604 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5605 "Unknown SINT_TO_FP to lower!");
5607 // These are really Legal; return the operand so the caller accepts it as
5609 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5611 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5612 Subtarget->is64Bit()) {
5616 DebugLoc dl = Op.getDebugLoc();
5617 unsigned Size = SrcVT.getSizeInBits()/8;
5618 MachineFunction &MF = DAG.getMachineFunction();
5619 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5620 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5621 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5623 PseudoSourceValue::getFixedStack(SSFI), 0,
5625 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5628 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5630 SelectionDAG &DAG) const {
5632 DebugLoc dl = Op.getDebugLoc();
5634 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5636 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5638 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5639 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5640 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5641 Tys, Ops, array_lengthof(Ops));
5644 Chain = Result.getValue(1);
5645 SDValue InFlag = Result.getValue(2);
5647 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5648 // shouldn't be necessary except that RFP cannot be live across
5649 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5650 MachineFunction &MF = DAG.getMachineFunction();
5651 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5652 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5653 Tys = DAG.getVTList(MVT::Other);
5655 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5657 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5658 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5659 PseudoSourceValue::getFixedStack(SSFI), 0,
5666 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5667 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
5668 SelectionDAG &DAG) const {
5669 // This algorithm is not obvious. Here it is in C code, more or less:
5671 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5672 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5673 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5675 // Copy ints to xmm registers.
5676 __m128i xh = _mm_cvtsi32_si128( hi );
5677 __m128i xl = _mm_cvtsi32_si128( lo );
5679 // Combine into low half of a single xmm register.
5680 __m128i x = _mm_unpacklo_epi32( xh, xl );
5684 // Merge in appropriate exponents to give the integer bits the right
5686 x = _mm_unpacklo_epi32( x, exp );
5688 // Subtract away the biases to deal with the IEEE-754 double precision
5690 d = _mm_sub_pd( (__m128d) x, bias );
5692 // All conversions up to here are exact. The correctly rounded result is
5693 // calculated using the current rounding mode using the following
5695 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5696 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5697 // store doesn't really need to be here (except
5698 // maybe to zero the other double)
5703 DebugLoc dl = Op.getDebugLoc();
5704 LLVMContext *Context = DAG.getContext();
5706 // Build some magic constants.
5707 std::vector<Constant*> CV0;
5708 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5709 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5710 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5711 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5712 Constant *C0 = ConstantVector::get(CV0);
5713 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5715 std::vector<Constant*> CV1;
5717 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5719 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5720 Constant *C1 = ConstantVector::get(CV1);
5721 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5723 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5724 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5726 DAG.getIntPtrConstant(1)));
5727 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5728 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5730 DAG.getIntPtrConstant(0)));
5731 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5732 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5733 PseudoSourceValue::getConstantPool(), 0,
5735 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5736 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5737 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5738 PseudoSourceValue::getConstantPool(), 0,
5740 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5742 // Add the halves; easiest way is to swap them into another reg first.
5743 int ShufMask[2] = { 1, -1 };
5744 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5745 DAG.getUNDEF(MVT::v2f64), ShufMask);
5746 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5747 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5748 DAG.getIntPtrConstant(0));
5751 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5752 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
5753 SelectionDAG &DAG) const {
5754 DebugLoc dl = Op.getDebugLoc();
5755 // FP constant to bias correct the final result.
5756 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5759 // Load the 32-bit value into an XMM register.
5760 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5761 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5763 DAG.getIntPtrConstant(0)));
5765 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5766 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5767 DAG.getIntPtrConstant(0));
5769 // Or the load with the bias.
5770 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5771 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5772 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5774 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5775 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5776 MVT::v2f64, Bias)));
5777 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5778 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5779 DAG.getIntPtrConstant(0));
5781 // Subtract the bias.
5782 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5784 // Handle final rounding.
5785 EVT DestVT = Op.getValueType();
5787 if (DestVT.bitsLT(MVT::f64)) {
5788 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5789 DAG.getIntPtrConstant(0));
5790 } else if (DestVT.bitsGT(MVT::f64)) {
5791 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5794 // Handle final rounding.
5798 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
5799 SelectionDAG &DAG) const {
5800 SDValue N0 = Op.getOperand(0);
5801 DebugLoc dl = Op.getDebugLoc();
5803 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
5804 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5805 // the optimization here.
5806 if (DAG.SignBitIsZero(N0))
5807 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5809 EVT SrcVT = N0.getValueType();
5810 EVT DstVT = Op.getValueType();
5811 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
5812 return LowerUINT_TO_FP_i64(Op, DAG);
5813 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
5814 return LowerUINT_TO_FP_i32(Op, DAG);
5816 // Make a 64-bit buffer, and use it to build an FILD.
5817 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5818 if (SrcVT == MVT::i32) {
5819 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5820 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5821 getPointerTy(), StackSlot, WordOff);
5822 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5823 StackSlot, NULL, 0, false, false, 0);
5824 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5825 OffsetSlot, NULL, 0, false, false, 0);
5826 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5830 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
5831 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5832 StackSlot, NULL, 0, false, false, 0);
5833 // For i64 source, we need to add the appropriate power of 2 if the input
5834 // was negative. This is the same as the optimization in
5835 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
5836 // we must be careful to do the computation in x87 extended precision, not
5837 // in SSE. (The generic code can't know it's OK to do this, or how to.)
5838 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
5839 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
5840 SDValue Fild = DAG.getNode(X86ISD::FILD, dl, Tys, Ops, 3);
5842 APInt FF(32, 0x5F800000ULL);
5844 // Check whether the sign bit is set.
5845 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
5846 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
5849 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
5850 SDValue FudgePtr = DAG.getConstantPool(
5851 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
5854 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
5855 SDValue Zero = DAG.getIntPtrConstant(0);
5856 SDValue Four = DAG.getIntPtrConstant(4);
5857 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
5859 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
5861 // Load the value out, extending it from f32 to f80.
5862 // FIXME: Avoid the extend by constructing the right constant pool?
5863 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, MVT::f80, dl, DAG.getEntryNode(),
5864 FudgePtr, PseudoSourceValue::getConstantPool(),
5865 0, MVT::f32, false, false, 4);
5866 // Extend everything to 80 bits to force it to be done on x87.
5867 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
5868 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
5871 std::pair<SDValue,SDValue> X86TargetLowering::
5872 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
5873 DebugLoc dl = Op.getDebugLoc();
5875 EVT DstTy = Op.getValueType();
5878 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5882 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5883 DstTy.getSimpleVT() >= MVT::i16 &&
5884 "Unknown FP_TO_SINT to lower!");
5886 // These are really Legal.
5887 if (DstTy == MVT::i32 &&
5888 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5889 return std::make_pair(SDValue(), SDValue());
5890 if (Subtarget->is64Bit() &&
5891 DstTy == MVT::i64 &&
5892 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5893 return std::make_pair(SDValue(), SDValue());
5895 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5897 MachineFunction &MF = DAG.getMachineFunction();
5898 unsigned MemSize = DstTy.getSizeInBits()/8;
5899 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5900 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5903 switch (DstTy.getSimpleVT().SimpleTy) {
5904 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5905 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5906 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5907 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5910 SDValue Chain = DAG.getEntryNode();
5911 SDValue Value = Op.getOperand(0);
5912 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5913 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5914 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5915 PseudoSourceValue::getFixedStack(SSFI), 0,
5917 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5919 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5921 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5922 Chain = Value.getValue(1);
5923 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5924 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5927 // Build the FP_TO_INT*_IN_MEM
5928 SDValue Ops[] = { Chain, Value, StackSlot };
5929 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5931 return std::make_pair(FIST, StackSlot);
5934 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
5935 SelectionDAG &DAG) const {
5936 if (Op.getValueType().isVector()) {
5937 if (Op.getValueType() == MVT::v2i32 &&
5938 Op.getOperand(0).getValueType() == MVT::v2f64) {
5944 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5945 SDValue FIST = Vals.first, StackSlot = Vals.second;
5946 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5947 if (FIST.getNode() == 0) return Op;
5950 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5951 FIST, StackSlot, NULL, 0, false, false, 0);
5954 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
5955 SelectionDAG &DAG) const {
5956 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5957 SDValue FIST = Vals.first, StackSlot = Vals.second;
5958 assert(FIST.getNode() && "Unexpected failure");
5961 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5962 FIST, StackSlot, NULL, 0, false, false, 0);
5965 SDValue X86TargetLowering::LowerFABS(SDValue Op,
5966 SelectionDAG &DAG) const {
5967 LLVMContext *Context = DAG.getContext();
5968 DebugLoc dl = Op.getDebugLoc();
5969 EVT VT = Op.getValueType();
5972 EltVT = VT.getVectorElementType();
5973 std::vector<Constant*> CV;
5974 if (EltVT == MVT::f64) {
5975 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5979 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5985 Constant *C = ConstantVector::get(CV);
5986 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5987 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5988 PseudoSourceValue::getConstantPool(), 0,
5990 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5993 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
5994 LLVMContext *Context = DAG.getContext();
5995 DebugLoc dl = Op.getDebugLoc();
5996 EVT VT = Op.getValueType();
5999 EltVT = VT.getVectorElementType();
6000 std::vector<Constant*> CV;
6001 if (EltVT == MVT::f64) {
6002 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
6006 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
6012 Constant *C = ConstantVector::get(CV);
6013 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6014 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6015 PseudoSourceValue::getConstantPool(), 0,
6017 if (VT.isVector()) {
6018 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6019 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
6020 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
6022 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
6024 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
6028 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
6029 LLVMContext *Context = DAG.getContext();
6030 SDValue Op0 = Op.getOperand(0);
6031 SDValue Op1 = Op.getOperand(1);
6032 DebugLoc dl = Op.getDebugLoc();
6033 EVT VT = Op.getValueType();
6034 EVT SrcVT = Op1.getValueType();
6036 // If second operand is smaller, extend it first.
6037 if (SrcVT.bitsLT(VT)) {
6038 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
6041 // And if it is bigger, shrink it first.
6042 if (SrcVT.bitsGT(VT)) {
6043 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
6047 // At this point the operands and the result should have the same
6048 // type, and that won't be f80 since that is not custom lowered.
6050 // First get the sign bit of second operand.
6051 std::vector<Constant*> CV;
6052 if (SrcVT == MVT::f64) {
6053 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
6054 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6056 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
6057 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6058 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6059 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6061 Constant *C = ConstantVector::get(CV);
6062 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6063 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
6064 PseudoSourceValue::getConstantPool(), 0,
6066 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
6068 // Shift sign bit right or left if the two operands have different types.
6069 if (SrcVT.bitsGT(VT)) {
6070 // Op0 is MVT::f32, Op1 is MVT::f64.
6071 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
6072 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
6073 DAG.getConstant(32, MVT::i32));
6074 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
6075 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
6076 DAG.getIntPtrConstant(0));
6079 // Clear first operand sign bit.
6081 if (VT == MVT::f64) {
6082 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
6083 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6085 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
6086 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6087 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6088 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6090 C = ConstantVector::get(CV);
6091 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6092 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6093 PseudoSourceValue::getConstantPool(), 0,
6095 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
6097 // Or the value with the sign bit.
6098 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
6101 /// Emit nodes that will be selected as "test Op0,Op0", or something
6103 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
6104 SelectionDAG &DAG) const {
6105 DebugLoc dl = Op.getDebugLoc();
6107 // CF and OF aren't always set the way we want. Determine which
6108 // of these we need.
6109 bool NeedCF = false;
6110 bool NeedOF = false;
6113 case X86::COND_A: case X86::COND_AE:
6114 case X86::COND_B: case X86::COND_BE:
6117 case X86::COND_G: case X86::COND_GE:
6118 case X86::COND_L: case X86::COND_LE:
6119 case X86::COND_O: case X86::COND_NO:
6124 // See if we can use the EFLAGS value from the operand instead of
6125 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
6126 // we prove that the arithmetic won't overflow, we can't use OF or CF.
6127 if (Op.getResNo() != 0 || NeedOF || NeedCF)
6128 // Emit a CMP with 0, which is the TEST pattern.
6129 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6130 DAG.getConstant(0, Op.getValueType()));
6132 unsigned Opcode = 0;
6133 unsigned NumOperands = 0;
6134 switch (Op.getNode()->getOpcode()) {
6136 // Due to an isel shortcoming, be conservative if this add is likely to be
6137 // selected as part of a load-modify-store instruction. When the root node
6138 // in a match is a store, isel doesn't know how to remap non-chain non-flag
6139 // uses of other nodes in the match, such as the ADD in this case. This
6140 // leads to the ADD being left around and reselected, with the result being
6141 // two adds in the output. Alas, even if none our users are stores, that
6142 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
6143 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
6144 // climbing the DAG back to the root, and it doesn't seem to be worth the
6146 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6147 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6148 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
6151 if (ConstantSDNode *C =
6152 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
6153 // An add of one will be selected as an INC.
6154 if (C->getAPIntValue() == 1) {
6155 Opcode = X86ISD::INC;
6160 // An add of negative one (subtract of one) will be selected as a DEC.
6161 if (C->getAPIntValue().isAllOnesValue()) {
6162 Opcode = X86ISD::DEC;
6168 // Otherwise use a regular EFLAGS-setting add.
6169 Opcode = X86ISD::ADD;
6173 // If the primary and result isn't used, don't bother using X86ISD::AND,
6174 // because a TEST instruction will be better.
6175 bool NonFlagUse = false;
6176 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6177 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
6179 unsigned UOpNo = UI.getOperandNo();
6180 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
6181 // Look pass truncate.
6182 UOpNo = User->use_begin().getOperandNo();
6183 User = *User->use_begin();
6186 if (User->getOpcode() != ISD::BRCOND &&
6187 User->getOpcode() != ISD::SETCC &&
6188 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
6201 // Due to the ISEL shortcoming noted above, be conservative if this op is
6202 // likely to be selected as part of a load-modify-store instruction.
6203 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
6204 UE = Op.getNode()->use_end(); UI != UE; ++UI)
6205 if (UI->getOpcode() == ISD::STORE)
6208 // Otherwise use a regular EFLAGS-setting instruction.
6209 switch (Op.getNode()->getOpcode()) {
6210 default: llvm_unreachable("unexpected operator!");
6211 case ISD::SUB: Opcode = X86ISD::SUB; break;
6212 case ISD::OR: Opcode = X86ISD::OR; break;
6213 case ISD::XOR: Opcode = X86ISD::XOR; break;
6214 case ISD::AND: Opcode = X86ISD::AND; break;
6226 return SDValue(Op.getNode(), 1);
6233 // Emit a CMP with 0, which is the TEST pattern.
6234 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
6235 DAG.getConstant(0, Op.getValueType()));
6237 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
6238 SmallVector<SDValue, 4> Ops;
6239 for (unsigned i = 0; i != NumOperands; ++i)
6240 Ops.push_back(Op.getOperand(i));
6242 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
6243 DAG.ReplaceAllUsesWith(Op, New);
6244 return SDValue(New.getNode(), 1);
6247 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
6249 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
6250 SelectionDAG &DAG) const {
6251 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
6252 if (C->getAPIntValue() == 0)
6253 return EmitTest(Op0, X86CC, DAG);
6255 DebugLoc dl = Op0.getDebugLoc();
6256 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
6259 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
6260 /// if it's possible.
6261 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
6262 DebugLoc dl, SelectionDAG &DAG) const {
6263 SDValue Op0 = And.getOperand(0);
6264 SDValue Op1 = And.getOperand(1);
6265 if (Op0.getOpcode() == ISD::TRUNCATE)
6266 Op0 = Op0.getOperand(0);
6267 if (Op1.getOpcode() == ISD::TRUNCATE)
6268 Op1 = Op1.getOperand(0);
6271 if (Op1.getOpcode() == ISD::SHL)
6272 std::swap(Op0, Op1);
6273 if (Op0.getOpcode() == ISD::SHL) {
6274 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
6275 if (And00C->getZExtValue() == 1) {
6276 // If we looked past a truncate, check that it's only truncating away
6278 unsigned BitWidth = Op0.getValueSizeInBits();
6279 unsigned AndBitWidth = And.getValueSizeInBits();
6280 if (BitWidth > AndBitWidth) {
6281 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
6282 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
6283 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
6287 RHS = Op0.getOperand(1);
6289 } else if (Op1.getOpcode() == ISD::Constant) {
6290 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
6291 SDValue AndLHS = Op0;
6292 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
6293 LHS = AndLHS.getOperand(0);
6294 RHS = AndLHS.getOperand(1);
6298 if (LHS.getNode()) {
6299 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
6300 // instruction. Since the shift amount is in-range-or-undefined, we know
6301 // that doing a bittest on the i32 value is ok. We extend to i32 because
6302 // the encoding for the i16 version is larger than the i32 version.
6303 // Also promote i16 to i32 for performance / code size reason.
6304 if (LHS.getValueType() == MVT::i8 ||
6305 LHS.getValueType() == MVT::i16)
6306 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
6308 // If the operand types disagree, extend the shift amount to match. Since
6309 // BT ignores high bits (like shifts) we can use anyextend.
6310 if (LHS.getValueType() != RHS.getValueType())
6311 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
6313 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
6314 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
6315 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6316 DAG.getConstant(Cond, MVT::i8), BT);
6322 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
6323 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
6324 SDValue Op0 = Op.getOperand(0);
6325 SDValue Op1 = Op.getOperand(1);
6326 DebugLoc dl = Op.getDebugLoc();
6327 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
6329 // Optimize to BT if possible.
6330 // Lower (X & (1 << N)) == 0 to BT(X, N).
6331 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
6332 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
6333 if (Op0.getOpcode() == ISD::AND &&
6335 Op1.getOpcode() == ISD::Constant &&
6336 cast<ConstantSDNode>(Op1)->isNullValue() &&
6337 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6338 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
6339 if (NewSetCC.getNode())
6343 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
6344 if (Op0.getOpcode() == X86ISD::SETCC &&
6345 Op1.getOpcode() == ISD::Constant &&
6346 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
6347 cast<ConstantSDNode>(Op1)->isNullValue()) &&
6348 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
6349 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
6350 bool Invert = (CC == ISD::SETNE) ^
6351 cast<ConstantSDNode>(Op1)->isNullValue();
6353 CCode = X86::GetOppositeBranchCondition(CCode);
6354 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6355 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
6358 bool isFP = Op1.getValueType().isFloatingPoint();
6359 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
6360 if (X86CC == X86::COND_INVALID)
6363 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
6365 // Use sbb x, x to materialize carry bit into a GPR.
6366 if (X86CC == X86::COND_B)
6367 return DAG.getNode(ISD::AND, dl, MVT::i8,
6368 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
6369 DAG.getConstant(X86CC, MVT::i8), Cond),
6370 DAG.getConstant(1, MVT::i8));
6372 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6373 DAG.getConstant(X86CC, MVT::i8), Cond);
6376 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
6378 SDValue Op0 = Op.getOperand(0);
6379 SDValue Op1 = Op.getOperand(1);
6380 SDValue CC = Op.getOperand(2);
6381 EVT VT = Op.getValueType();
6382 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6383 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6384 DebugLoc dl = Op.getDebugLoc();
6388 EVT VT0 = Op0.getValueType();
6389 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6390 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6393 switch (SetCCOpcode) {
6396 case ISD::SETEQ: SSECC = 0; break;
6398 case ISD::SETGT: Swap = true; // Fallthrough
6400 case ISD::SETOLT: SSECC = 1; break;
6402 case ISD::SETGE: Swap = true; // Fallthrough
6404 case ISD::SETOLE: SSECC = 2; break;
6405 case ISD::SETUO: SSECC = 3; break;
6407 case ISD::SETNE: SSECC = 4; break;
6408 case ISD::SETULE: Swap = true;
6409 case ISD::SETUGE: SSECC = 5; break;
6410 case ISD::SETULT: Swap = true;
6411 case ISD::SETUGT: SSECC = 6; break;
6412 case ISD::SETO: SSECC = 7; break;
6415 std::swap(Op0, Op1);
6417 // In the two special cases we can't handle, emit two comparisons.
6419 if (SetCCOpcode == ISD::SETUEQ) {
6421 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6422 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6423 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6425 else if (SetCCOpcode == ISD::SETONE) {
6427 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6428 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6429 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6431 llvm_unreachable("Illegal FP comparison");
6433 // Handle all other FP comparisons here.
6434 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6437 // We are handling one of the integer comparisons here. Since SSE only has
6438 // GT and EQ comparisons for integer, swapping operands and multiple
6439 // operations may be required for some comparisons.
6440 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6441 bool Swap = false, Invert = false, FlipSigns = false;
6443 switch (VT.getSimpleVT().SimpleTy) {
6446 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6448 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6450 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6451 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6454 switch (SetCCOpcode) {
6456 case ISD::SETNE: Invert = true;
6457 case ISD::SETEQ: Opc = EQOpc; break;
6458 case ISD::SETLT: Swap = true;
6459 case ISD::SETGT: Opc = GTOpc; break;
6460 case ISD::SETGE: Swap = true;
6461 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6462 case ISD::SETULT: Swap = true;
6463 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6464 case ISD::SETUGE: Swap = true;
6465 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6468 std::swap(Op0, Op1);
6470 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6471 // bits of the inputs before performing those operations.
6473 EVT EltVT = VT.getVectorElementType();
6474 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6476 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6477 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6479 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6480 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6483 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6485 // If the logical-not of the result is required, perform that now.
6487 Result = DAG.getNOT(dl, Result, VT);
6492 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6493 static bool isX86LogicalCmp(SDValue Op) {
6494 unsigned Opc = Op.getNode()->getOpcode();
6495 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6497 if (Op.getResNo() == 1 &&
6498 (Opc == X86ISD::ADD ||
6499 Opc == X86ISD::SUB ||
6500 Opc == X86ISD::SMUL ||
6501 Opc == X86ISD::UMUL ||
6502 Opc == X86ISD::INC ||
6503 Opc == X86ISD::DEC ||
6504 Opc == X86ISD::OR ||
6505 Opc == X86ISD::XOR ||
6506 Opc == X86ISD::AND))
6512 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
6513 bool addTest = true;
6514 SDValue Cond = Op.getOperand(0);
6515 DebugLoc dl = Op.getDebugLoc();
6518 if (Cond.getOpcode() == ISD::SETCC) {
6519 SDValue NewCond = LowerSETCC(Cond, DAG);
6520 if (NewCond.getNode())
6524 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6525 SDValue Op1 = Op.getOperand(1);
6526 SDValue Op2 = Op.getOperand(2);
6527 if (Cond.getOpcode() == X86ISD::SETCC &&
6528 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6529 SDValue Cmp = Cond.getOperand(1);
6530 if (Cmp.getOpcode() == X86ISD::CMP) {
6531 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6532 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6533 ConstantSDNode *RHSC =
6534 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6535 if (N1C && N1C->isAllOnesValue() &&
6536 N2C && N2C->isNullValue() &&
6537 RHSC && RHSC->isNullValue()) {
6538 SDValue CmpOp0 = Cmp.getOperand(0);
6539 Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6540 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6541 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6542 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6547 // Look pass (and (setcc_carry (cmp ...)), 1).
6548 if (Cond.getOpcode() == ISD::AND &&
6549 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6550 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6551 if (C && C->getAPIntValue() == 1)
6552 Cond = Cond.getOperand(0);
6555 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6556 // setting operand in place of the X86ISD::SETCC.
6557 if (Cond.getOpcode() == X86ISD::SETCC ||
6558 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6559 CC = Cond.getOperand(0);
6561 SDValue Cmp = Cond.getOperand(1);
6562 unsigned Opc = Cmp.getOpcode();
6563 EVT VT = Op.getValueType();
6565 bool IllegalFPCMov = false;
6566 if (VT.isFloatingPoint() && !VT.isVector() &&
6567 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6568 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6570 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6571 Opc == X86ISD::BT) { // FIXME
6578 // Look pass the truncate.
6579 if (Cond.getOpcode() == ISD::TRUNCATE)
6580 Cond = Cond.getOperand(0);
6582 // We know the result of AND is compared against zero. Try to match
6584 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6585 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6586 if (NewSetCC.getNode()) {
6587 CC = NewSetCC.getOperand(0);
6588 Cond = NewSetCC.getOperand(1);
6595 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6596 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6599 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6600 // condition is true.
6601 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6602 SDValue Ops[] = { Op2, Op1, CC, Cond };
6603 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6606 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6607 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6608 // from the AND / OR.
6609 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6610 Opc = Op.getOpcode();
6611 if (Opc != ISD::OR && Opc != ISD::AND)
6613 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6614 Op.getOperand(0).hasOneUse() &&
6615 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6616 Op.getOperand(1).hasOneUse());
6619 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6620 // 1 and that the SETCC node has a single use.
6621 static bool isXor1OfSetCC(SDValue Op) {
6622 if (Op.getOpcode() != ISD::XOR)
6624 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6625 if (N1C && N1C->getAPIntValue() == 1) {
6626 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6627 Op.getOperand(0).hasOneUse();
6632 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
6633 bool addTest = true;
6634 SDValue Chain = Op.getOperand(0);
6635 SDValue Cond = Op.getOperand(1);
6636 SDValue Dest = Op.getOperand(2);
6637 DebugLoc dl = Op.getDebugLoc();
6640 if (Cond.getOpcode() == ISD::SETCC) {
6641 SDValue NewCond = LowerSETCC(Cond, DAG);
6642 if (NewCond.getNode())
6646 // FIXME: LowerXALUO doesn't handle these!!
6647 else if (Cond.getOpcode() == X86ISD::ADD ||
6648 Cond.getOpcode() == X86ISD::SUB ||
6649 Cond.getOpcode() == X86ISD::SMUL ||
6650 Cond.getOpcode() == X86ISD::UMUL)
6651 Cond = LowerXALUO(Cond, DAG);
6654 // Look pass (and (setcc_carry (cmp ...)), 1).
6655 if (Cond.getOpcode() == ISD::AND &&
6656 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6657 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6658 if (C && C->getAPIntValue() == 1)
6659 Cond = Cond.getOperand(0);
6662 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6663 // setting operand in place of the X86ISD::SETCC.
6664 if (Cond.getOpcode() == X86ISD::SETCC ||
6665 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6666 CC = Cond.getOperand(0);
6668 SDValue Cmp = Cond.getOperand(1);
6669 unsigned Opc = Cmp.getOpcode();
6670 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6671 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6675 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6679 // These can only come from an arithmetic instruction with overflow,
6680 // e.g. SADDO, UADDO.
6681 Cond = Cond.getNode()->getOperand(1);
6688 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6689 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6690 if (CondOpc == ISD::OR) {
6691 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6692 // two branches instead of an explicit OR instruction with a
6694 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6695 isX86LogicalCmp(Cmp)) {
6696 CC = Cond.getOperand(0).getOperand(0);
6697 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6698 Chain, Dest, CC, Cmp);
6699 CC = Cond.getOperand(1).getOperand(0);
6703 } else { // ISD::AND
6704 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6705 // two branches instead of an explicit AND instruction with a
6706 // separate test. However, we only do this if this block doesn't
6707 // have a fall-through edge, because this requires an explicit
6708 // jmp when the condition is false.
6709 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6710 isX86LogicalCmp(Cmp) &&
6711 Op.getNode()->hasOneUse()) {
6712 X86::CondCode CCode =
6713 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6714 CCode = X86::GetOppositeBranchCondition(CCode);
6715 CC = DAG.getConstant(CCode, MVT::i8);
6716 SDNode *User = *Op.getNode()->use_begin();
6717 // Look for an unconditional branch following this conditional branch.
6718 // We need this because we need to reverse the successors in order
6719 // to implement FCMP_OEQ.
6720 if (User->getOpcode() == ISD::BR) {
6721 SDValue FalseBB = User->getOperand(1);
6723 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
6724 assert(NewBR == User);
6728 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6729 Chain, Dest, CC, Cmp);
6730 X86::CondCode CCode =
6731 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6732 CCode = X86::GetOppositeBranchCondition(CCode);
6733 CC = DAG.getConstant(CCode, MVT::i8);
6739 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6740 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6741 // It should be transformed during dag combiner except when the condition
6742 // is set by a arithmetics with overflow node.
6743 X86::CondCode CCode =
6744 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6745 CCode = X86::GetOppositeBranchCondition(CCode);
6746 CC = DAG.getConstant(CCode, MVT::i8);
6747 Cond = Cond.getOperand(0).getOperand(1);
6753 // Look pass the truncate.
6754 if (Cond.getOpcode() == ISD::TRUNCATE)
6755 Cond = Cond.getOperand(0);
6757 // We know the result of AND is compared against zero. Try to match
6759 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6760 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6761 if (NewSetCC.getNode()) {
6762 CC = NewSetCC.getOperand(0);
6763 Cond = NewSetCC.getOperand(1);
6770 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6771 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6773 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6774 Chain, Dest, CC, Cond);
6778 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6779 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6780 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6781 // that the guard pages used by the OS virtual memory manager are allocated in
6782 // correct sequence.
6784 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6785 SelectionDAG &DAG) const {
6786 assert(Subtarget->isTargetCygMing() &&
6787 "This should be used only on Cygwin/Mingw targets");
6788 DebugLoc dl = Op.getDebugLoc();
6791 SDValue Chain = Op.getOperand(0);
6792 SDValue Size = Op.getOperand(1);
6793 // FIXME: Ensure alignment here
6797 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6799 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6800 Flag = Chain.getValue(1);
6802 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6804 Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
6805 Flag = Chain.getValue(1);
6807 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6809 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6810 return DAG.getMergeValues(Ops1, 2, dl);
6813 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
6814 MachineFunction &MF = DAG.getMachineFunction();
6815 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
6817 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6818 DebugLoc dl = Op.getDebugLoc();
6820 if (!Subtarget->is64Bit()) {
6821 // vastart just stores the address of the VarArgsFrameIndex slot into the
6822 // memory location argument.
6823 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6825 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6830 // gp_offset (0 - 6 * 8)
6831 // fp_offset (48 - 48 + 8 * 16)
6832 // overflow_arg_area (point to parameters coming in memory).
6834 SmallVector<SDValue, 8> MemOps;
6835 SDValue FIN = Op.getOperand(1);
6837 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6838 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
6840 FIN, SV, 0, false, false, 0);
6841 MemOps.push_back(Store);
6844 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6845 FIN, DAG.getIntPtrConstant(4));
6846 Store = DAG.getStore(Op.getOperand(0), dl,
6847 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
6849 FIN, SV, 4, false, false, 0);
6850 MemOps.push_back(Store);
6852 // Store ptr to overflow_arg_area
6853 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6854 FIN, DAG.getIntPtrConstant(4));
6855 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
6857 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 8,
6859 MemOps.push_back(Store);
6861 // Store ptr to reg_save_area.
6862 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6863 FIN, DAG.getIntPtrConstant(8));
6864 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
6866 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 16,
6868 MemOps.push_back(Store);
6869 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6870 &MemOps[0], MemOps.size());
6873 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
6874 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6875 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6877 report_fatal_error("VAArgInst is not yet implemented for x86-64!");
6881 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
6882 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6883 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6884 SDValue Chain = Op.getOperand(0);
6885 SDValue DstPtr = Op.getOperand(1);
6886 SDValue SrcPtr = Op.getOperand(2);
6887 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6888 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6889 DebugLoc dl = Op.getDebugLoc();
6891 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6892 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
6893 false, DstSV, 0, SrcSV, 0);
6897 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
6898 DebugLoc dl = Op.getDebugLoc();
6899 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6901 default: return SDValue(); // Don't custom lower most intrinsics.
6902 // Comparison intrinsics.
6903 case Intrinsic::x86_sse_comieq_ss:
6904 case Intrinsic::x86_sse_comilt_ss:
6905 case Intrinsic::x86_sse_comile_ss:
6906 case Intrinsic::x86_sse_comigt_ss:
6907 case Intrinsic::x86_sse_comige_ss:
6908 case Intrinsic::x86_sse_comineq_ss:
6909 case Intrinsic::x86_sse_ucomieq_ss:
6910 case Intrinsic::x86_sse_ucomilt_ss:
6911 case Intrinsic::x86_sse_ucomile_ss:
6912 case Intrinsic::x86_sse_ucomigt_ss:
6913 case Intrinsic::x86_sse_ucomige_ss:
6914 case Intrinsic::x86_sse_ucomineq_ss:
6915 case Intrinsic::x86_sse2_comieq_sd:
6916 case Intrinsic::x86_sse2_comilt_sd:
6917 case Intrinsic::x86_sse2_comile_sd:
6918 case Intrinsic::x86_sse2_comigt_sd:
6919 case Intrinsic::x86_sse2_comige_sd:
6920 case Intrinsic::x86_sse2_comineq_sd:
6921 case Intrinsic::x86_sse2_ucomieq_sd:
6922 case Intrinsic::x86_sse2_ucomilt_sd:
6923 case Intrinsic::x86_sse2_ucomile_sd:
6924 case Intrinsic::x86_sse2_ucomigt_sd:
6925 case Intrinsic::x86_sse2_ucomige_sd:
6926 case Intrinsic::x86_sse2_ucomineq_sd: {
6928 ISD::CondCode CC = ISD::SETCC_INVALID;
6931 case Intrinsic::x86_sse_comieq_ss:
6932 case Intrinsic::x86_sse2_comieq_sd:
6936 case Intrinsic::x86_sse_comilt_ss:
6937 case Intrinsic::x86_sse2_comilt_sd:
6941 case Intrinsic::x86_sse_comile_ss:
6942 case Intrinsic::x86_sse2_comile_sd:
6946 case Intrinsic::x86_sse_comigt_ss:
6947 case Intrinsic::x86_sse2_comigt_sd:
6951 case Intrinsic::x86_sse_comige_ss:
6952 case Intrinsic::x86_sse2_comige_sd:
6956 case Intrinsic::x86_sse_comineq_ss:
6957 case Intrinsic::x86_sse2_comineq_sd:
6961 case Intrinsic::x86_sse_ucomieq_ss:
6962 case Intrinsic::x86_sse2_ucomieq_sd:
6963 Opc = X86ISD::UCOMI;
6966 case Intrinsic::x86_sse_ucomilt_ss:
6967 case Intrinsic::x86_sse2_ucomilt_sd:
6968 Opc = X86ISD::UCOMI;
6971 case Intrinsic::x86_sse_ucomile_ss:
6972 case Intrinsic::x86_sse2_ucomile_sd:
6973 Opc = X86ISD::UCOMI;
6976 case Intrinsic::x86_sse_ucomigt_ss:
6977 case Intrinsic::x86_sse2_ucomigt_sd:
6978 Opc = X86ISD::UCOMI;
6981 case Intrinsic::x86_sse_ucomige_ss:
6982 case Intrinsic::x86_sse2_ucomige_sd:
6983 Opc = X86ISD::UCOMI;
6986 case Intrinsic::x86_sse_ucomineq_ss:
6987 case Intrinsic::x86_sse2_ucomineq_sd:
6988 Opc = X86ISD::UCOMI;
6993 SDValue LHS = Op.getOperand(1);
6994 SDValue RHS = Op.getOperand(2);
6995 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6996 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6997 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6998 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6999 DAG.getConstant(X86CC, MVT::i8), Cond);
7000 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7002 // ptest and testp intrinsics. The intrinsic these come from are designed to
7003 // return an integer value, not just an instruction so lower it to the ptest
7004 // or testp pattern and a setcc for the result.
7005 case Intrinsic::x86_sse41_ptestz:
7006 case Intrinsic::x86_sse41_ptestc:
7007 case Intrinsic::x86_sse41_ptestnzc:
7008 case Intrinsic::x86_avx_ptestz_256:
7009 case Intrinsic::x86_avx_ptestc_256:
7010 case Intrinsic::x86_avx_ptestnzc_256:
7011 case Intrinsic::x86_avx_vtestz_ps:
7012 case Intrinsic::x86_avx_vtestc_ps:
7013 case Intrinsic::x86_avx_vtestnzc_ps:
7014 case Intrinsic::x86_avx_vtestz_pd:
7015 case Intrinsic::x86_avx_vtestc_pd:
7016 case Intrinsic::x86_avx_vtestnzc_pd:
7017 case Intrinsic::x86_avx_vtestz_ps_256:
7018 case Intrinsic::x86_avx_vtestc_ps_256:
7019 case Intrinsic::x86_avx_vtestnzc_ps_256:
7020 case Intrinsic::x86_avx_vtestz_pd_256:
7021 case Intrinsic::x86_avx_vtestc_pd_256:
7022 case Intrinsic::x86_avx_vtestnzc_pd_256: {
7023 bool IsTestPacked = false;
7026 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
7027 case Intrinsic::x86_avx_vtestz_ps:
7028 case Intrinsic::x86_avx_vtestz_pd:
7029 case Intrinsic::x86_avx_vtestz_ps_256:
7030 case Intrinsic::x86_avx_vtestz_pd_256:
7031 IsTestPacked = true; // Fallthrough
7032 case Intrinsic::x86_sse41_ptestz:
7033 case Intrinsic::x86_avx_ptestz_256:
7035 X86CC = X86::COND_E;
7037 case Intrinsic::x86_avx_vtestc_ps:
7038 case Intrinsic::x86_avx_vtestc_pd:
7039 case Intrinsic::x86_avx_vtestc_ps_256:
7040 case Intrinsic::x86_avx_vtestc_pd_256:
7041 IsTestPacked = true; // Fallthrough
7042 case Intrinsic::x86_sse41_ptestc:
7043 case Intrinsic::x86_avx_ptestc_256:
7045 X86CC = X86::COND_B;
7047 case Intrinsic::x86_avx_vtestnzc_ps:
7048 case Intrinsic::x86_avx_vtestnzc_pd:
7049 case Intrinsic::x86_avx_vtestnzc_ps_256:
7050 case Intrinsic::x86_avx_vtestnzc_pd_256:
7051 IsTestPacked = true; // Fallthrough
7052 case Intrinsic::x86_sse41_ptestnzc:
7053 case Intrinsic::x86_avx_ptestnzc_256:
7055 X86CC = X86::COND_A;
7059 SDValue LHS = Op.getOperand(1);
7060 SDValue RHS = Op.getOperand(2);
7061 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
7062 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
7063 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
7064 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
7065 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
7068 // Fix vector shift instructions where the last operand is a non-immediate
7070 case Intrinsic::x86_sse2_pslli_w:
7071 case Intrinsic::x86_sse2_pslli_d:
7072 case Intrinsic::x86_sse2_pslli_q:
7073 case Intrinsic::x86_sse2_psrli_w:
7074 case Intrinsic::x86_sse2_psrli_d:
7075 case Intrinsic::x86_sse2_psrli_q:
7076 case Intrinsic::x86_sse2_psrai_w:
7077 case Intrinsic::x86_sse2_psrai_d:
7078 case Intrinsic::x86_mmx_pslli_w:
7079 case Intrinsic::x86_mmx_pslli_d:
7080 case Intrinsic::x86_mmx_pslli_q:
7081 case Intrinsic::x86_mmx_psrli_w:
7082 case Intrinsic::x86_mmx_psrli_d:
7083 case Intrinsic::x86_mmx_psrli_q:
7084 case Intrinsic::x86_mmx_psrai_w:
7085 case Intrinsic::x86_mmx_psrai_d: {
7086 SDValue ShAmt = Op.getOperand(2);
7087 if (isa<ConstantSDNode>(ShAmt))
7090 unsigned NewIntNo = 0;
7091 EVT ShAmtVT = MVT::v4i32;
7093 case Intrinsic::x86_sse2_pslli_w:
7094 NewIntNo = Intrinsic::x86_sse2_psll_w;
7096 case Intrinsic::x86_sse2_pslli_d:
7097 NewIntNo = Intrinsic::x86_sse2_psll_d;
7099 case Intrinsic::x86_sse2_pslli_q:
7100 NewIntNo = Intrinsic::x86_sse2_psll_q;
7102 case Intrinsic::x86_sse2_psrli_w:
7103 NewIntNo = Intrinsic::x86_sse2_psrl_w;
7105 case Intrinsic::x86_sse2_psrli_d:
7106 NewIntNo = Intrinsic::x86_sse2_psrl_d;
7108 case Intrinsic::x86_sse2_psrli_q:
7109 NewIntNo = Intrinsic::x86_sse2_psrl_q;
7111 case Intrinsic::x86_sse2_psrai_w:
7112 NewIntNo = Intrinsic::x86_sse2_psra_w;
7114 case Intrinsic::x86_sse2_psrai_d:
7115 NewIntNo = Intrinsic::x86_sse2_psra_d;
7118 ShAmtVT = MVT::v2i32;
7120 case Intrinsic::x86_mmx_pslli_w:
7121 NewIntNo = Intrinsic::x86_mmx_psll_w;
7123 case Intrinsic::x86_mmx_pslli_d:
7124 NewIntNo = Intrinsic::x86_mmx_psll_d;
7126 case Intrinsic::x86_mmx_pslli_q:
7127 NewIntNo = Intrinsic::x86_mmx_psll_q;
7129 case Intrinsic::x86_mmx_psrli_w:
7130 NewIntNo = Intrinsic::x86_mmx_psrl_w;
7132 case Intrinsic::x86_mmx_psrli_d:
7133 NewIntNo = Intrinsic::x86_mmx_psrl_d;
7135 case Intrinsic::x86_mmx_psrli_q:
7136 NewIntNo = Intrinsic::x86_mmx_psrl_q;
7138 case Intrinsic::x86_mmx_psrai_w:
7139 NewIntNo = Intrinsic::x86_mmx_psra_w;
7141 case Intrinsic::x86_mmx_psrai_d:
7142 NewIntNo = Intrinsic::x86_mmx_psra_d;
7144 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
7150 // The vector shift intrinsics with scalars uses 32b shift amounts but
7151 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
7155 ShOps[1] = DAG.getConstant(0, MVT::i32);
7156 if (ShAmtVT == MVT::v4i32) {
7157 ShOps[2] = DAG.getUNDEF(MVT::i32);
7158 ShOps[3] = DAG.getUNDEF(MVT::i32);
7159 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
7161 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
7164 EVT VT = Op.getValueType();
7165 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
7166 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7167 DAG.getConstant(NewIntNo, MVT::i32),
7168 Op.getOperand(1), ShAmt);
7173 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
7174 SelectionDAG &DAG) const {
7175 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7176 MFI->setReturnAddressIsTaken(true);
7178 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7179 DebugLoc dl = Op.getDebugLoc();
7182 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
7184 DAG.getConstant(TD->getPointerSize(),
7185 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7186 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7187 DAG.getNode(ISD::ADD, dl, getPointerTy(),
7189 NULL, 0, false, false, 0);
7192 // Just load the return address.
7193 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
7194 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
7195 RetAddrFI, NULL, 0, false, false, 0);
7198 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
7199 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7200 MFI->setFrameAddressIsTaken(true);
7202 EVT VT = Op.getValueType();
7203 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7204 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7205 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7206 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7208 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7213 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7214 SelectionDAG &DAG) const {
7215 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7218 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
7219 MachineFunction &MF = DAG.getMachineFunction();
7220 SDValue Chain = Op.getOperand(0);
7221 SDValue Offset = Op.getOperand(1);
7222 SDValue Handler = Op.getOperand(2);
7223 DebugLoc dl = Op.getDebugLoc();
7225 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
7226 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7228 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7230 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
7231 DAG.getIntPtrConstant(TD->getPointerSize()));
7232 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7233 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7234 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7235 MF.getRegInfo().addLiveOut(StoreAddrReg);
7237 return DAG.getNode(X86ISD::EH_RETURN, dl,
7239 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7242 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7243 SelectionDAG &DAG) const {
7244 SDValue Root = Op.getOperand(0);
7245 SDValue Trmp = Op.getOperand(1); // trampoline
7246 SDValue FPtr = Op.getOperand(2); // nested function
7247 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7248 DebugLoc dl = Op.getDebugLoc();
7250 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7252 if (Subtarget->is64Bit()) {
7253 SDValue OutChains[6];
7255 // Large code-model.
7256 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7257 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7259 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7260 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7262 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7264 // Load the pointer to the nested function into R11.
7265 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7266 SDValue Addr = Trmp;
7267 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7268 Addr, TrmpAddr, 0, false, false, 0);
7270 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7271 DAG.getConstant(2, MVT::i64));
7272 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7275 // Load the 'nest' parameter value into R10.
7276 // R10 is specified in X86CallingConv.td
7277 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7278 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7279 DAG.getConstant(10, MVT::i64));
7280 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7281 Addr, TrmpAddr, 10, false, false, 0);
7283 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7284 DAG.getConstant(12, MVT::i64));
7285 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7288 // Jump to the nested function.
7289 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7290 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7291 DAG.getConstant(20, MVT::i64));
7292 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7293 Addr, TrmpAddr, 20, false, false, 0);
7295 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7296 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7297 DAG.getConstant(22, MVT::i64));
7298 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7299 TrmpAddr, 22, false, false, 0);
7302 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7303 return DAG.getMergeValues(Ops, 2, dl);
7305 const Function *Func =
7306 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7307 CallingConv::ID CC = Func->getCallingConv();
7312 llvm_unreachable("Unsupported calling convention");
7313 case CallingConv::C:
7314 case CallingConv::X86_StdCall: {
7315 // Pass 'nest' parameter in ECX.
7316 // Must be kept in sync with X86CallingConv.td
7319 // Check that ECX wasn't needed by an 'inreg' parameter.
7320 const FunctionType *FTy = Func->getFunctionType();
7321 const AttrListPtr &Attrs = Func->getAttributes();
7323 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7324 unsigned InRegCount = 0;
7327 for (FunctionType::param_iterator I = FTy->param_begin(),
7328 E = FTy->param_end(); I != E; ++I, ++Idx)
7329 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7330 // FIXME: should only count parameters that are lowered to integers.
7331 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7333 if (InRegCount > 2) {
7334 report_fatal_error("Nest register in use - reduce number of inreg"
7340 case CallingConv::X86_FastCall:
7341 case CallingConv::X86_ThisCall:
7342 case CallingConv::Fast:
7343 // Pass 'nest' parameter in EAX.
7344 // Must be kept in sync with X86CallingConv.td
7349 SDValue OutChains[4];
7352 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7353 DAG.getConstant(10, MVT::i32));
7354 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7356 // This is storing the opcode for MOV32ri.
7357 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7358 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7359 OutChains[0] = DAG.getStore(Root, dl,
7360 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7361 Trmp, TrmpAddr, 0, false, false, 0);
7363 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7364 DAG.getConstant(1, MVT::i32));
7365 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7368 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7369 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7370 DAG.getConstant(5, MVT::i32));
7371 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7372 TrmpAddr, 5, false, false, 1);
7374 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7375 DAG.getConstant(6, MVT::i32));
7376 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7380 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7381 return DAG.getMergeValues(Ops, 2, dl);
7385 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
7386 SelectionDAG &DAG) const {
7388 The rounding mode is in bits 11:10 of FPSR, and has the following
7395 FLT_ROUNDS, on the other hand, expects the following:
7402 To perform the conversion, we do:
7403 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7406 MachineFunction &MF = DAG.getMachineFunction();
7407 const TargetMachine &TM = MF.getTarget();
7408 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7409 unsigned StackAlignment = TFI.getStackAlignment();
7410 EVT VT = Op.getValueType();
7411 DebugLoc dl = Op.getDebugLoc();
7413 // Save FP Control Word to stack slot
7414 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7415 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7417 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7418 DAG.getEntryNode(), StackSlot);
7420 // Load FP Control Word from stack slot
7421 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7424 // Transform as necessary
7426 DAG.getNode(ISD::SRL, dl, MVT::i16,
7427 DAG.getNode(ISD::AND, dl, MVT::i16,
7428 CWD, DAG.getConstant(0x800, MVT::i16)),
7429 DAG.getConstant(11, MVT::i8));
7431 DAG.getNode(ISD::SRL, dl, MVT::i16,
7432 DAG.getNode(ISD::AND, dl, MVT::i16,
7433 CWD, DAG.getConstant(0x400, MVT::i16)),
7434 DAG.getConstant(9, MVT::i8));
7437 DAG.getNode(ISD::AND, dl, MVT::i16,
7438 DAG.getNode(ISD::ADD, dl, MVT::i16,
7439 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7440 DAG.getConstant(1, MVT::i16)),
7441 DAG.getConstant(3, MVT::i16));
7444 return DAG.getNode((VT.getSizeInBits() < 16 ?
7445 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7448 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
7449 EVT VT = Op.getValueType();
7451 unsigned NumBits = VT.getSizeInBits();
7452 DebugLoc dl = Op.getDebugLoc();
7454 Op = Op.getOperand(0);
7455 if (VT == MVT::i8) {
7456 // Zero extend to i32 since there is not an i8 bsr.
7458 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7461 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7462 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7463 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7465 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7468 DAG.getConstant(NumBits+NumBits-1, OpVT),
7469 DAG.getConstant(X86::COND_E, MVT::i8),
7472 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7474 // Finally xor with NumBits-1.
7475 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7478 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7482 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
7483 EVT VT = Op.getValueType();
7485 unsigned NumBits = VT.getSizeInBits();
7486 DebugLoc dl = Op.getDebugLoc();
7488 Op = Op.getOperand(0);
7489 if (VT == MVT::i8) {
7491 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7494 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7495 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7496 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7498 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7501 DAG.getConstant(NumBits, OpVT),
7502 DAG.getConstant(X86::COND_E, MVT::i8),
7505 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7508 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7512 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
7513 EVT VT = Op.getValueType();
7514 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7515 DebugLoc dl = Op.getDebugLoc();
7517 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7518 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7519 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7520 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7521 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7523 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7524 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7525 // return AloBlo + AloBhi + AhiBlo;
7527 SDValue A = Op.getOperand(0);
7528 SDValue B = Op.getOperand(1);
7530 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7531 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7532 A, DAG.getConstant(32, MVT::i32));
7533 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7534 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7535 B, DAG.getConstant(32, MVT::i32));
7536 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7537 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7539 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7540 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7542 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7543 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7545 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7546 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7547 AloBhi, DAG.getConstant(32, MVT::i32));
7548 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7549 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7550 AhiBlo, DAG.getConstant(32, MVT::i32));
7551 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7552 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7556 SDValue X86TargetLowering::LowerSHL(SDValue Op, SelectionDAG &DAG) const {
7557 EVT VT = Op.getValueType();
7558 DebugLoc dl = Op.getDebugLoc();
7559 SDValue R = Op.getOperand(0);
7561 LLVMContext *Context = DAG.getContext();
7563 assert(Subtarget->hasSSE41() && "Cannot lower SHL without SSE4.1 or later");
7565 if (VT == MVT::v4i32) {
7566 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7567 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
7568 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
7570 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
7572 std::vector<Constant*> CV(4, CI);
7573 Constant *C = ConstantVector::get(CV);
7574 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7575 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7576 PseudoSourceValue::getConstantPool(), 0,
7579 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
7580 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, Op);
7581 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
7582 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
7584 if (VT == MVT::v16i8) {
7586 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7587 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
7588 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
7590 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
7591 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
7593 std::vector<Constant*> CVM1(16, CM1);
7594 std::vector<Constant*> CVM2(16, CM2);
7595 Constant *C = ConstantVector::get(CVM1);
7596 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7597 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7598 PseudoSourceValue::getConstantPool(), 0,
7601 // r = pblendv(r, psllw(r & (char16)15, 4), a);
7602 M = DAG.getNode(ISD::AND, dl, VT, R, M);
7603 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7604 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
7605 DAG.getConstant(4, MVT::i32));
7606 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7607 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
7610 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
7612 C = ConstantVector::get(CVM2);
7613 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7614 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7615 PseudoSourceValue::getConstantPool(), 0, false, false, 16);
7617 // r = pblendv(r, psllw(r & (char16)63, 2), a);
7618 M = DAG.getNode(ISD::AND, dl, VT, R, M);
7619 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7620 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
7621 DAG.getConstant(2, MVT::i32));
7622 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7623 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
7626 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
7628 // return pblendv(r, r+r, a);
7629 R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7630 DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
7631 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
7637 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
7638 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7639 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7640 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7641 // has only one use.
7642 SDNode *N = Op.getNode();
7643 SDValue LHS = N->getOperand(0);
7644 SDValue RHS = N->getOperand(1);
7645 unsigned BaseOp = 0;
7647 DebugLoc dl = Op.getDebugLoc();
7649 switch (Op.getOpcode()) {
7650 default: llvm_unreachable("Unknown ovf instruction!");
7652 // A subtract of one will be selected as a INC. Note that INC doesn't
7653 // set CF, so we can't do this for UADDO.
7654 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7655 if (C->getAPIntValue() == 1) {
7656 BaseOp = X86ISD::INC;
7660 BaseOp = X86ISD::ADD;
7664 BaseOp = X86ISD::ADD;
7668 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7669 // set CF, so we can't do this for USUBO.
7670 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7671 if (C->getAPIntValue() == 1) {
7672 BaseOp = X86ISD::DEC;
7676 BaseOp = X86ISD::SUB;
7680 BaseOp = X86ISD::SUB;
7684 BaseOp = X86ISD::SMUL;
7688 BaseOp = X86ISD::UMUL;
7693 // Also sets EFLAGS.
7694 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7695 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7698 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7699 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7701 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7705 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
7706 DebugLoc dl = Op.getDebugLoc();
7708 if (!Subtarget->hasSSE2()) {
7709 SDValue Zero = DAG.getConstant(0,
7710 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
7711 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
7715 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
7717 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
7719 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
7720 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
7721 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
7722 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
7724 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
7725 if (!Op1 && !Op2 && !Op3 && Op4)
7726 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
7728 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
7729 if (Op1 && !Op2 && !Op3 && !Op4)
7730 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
7732 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
7734 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
7737 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
7738 EVT T = Op.getValueType();
7739 DebugLoc dl = Op.getDebugLoc();
7742 switch(T.getSimpleVT().SimpleTy) {
7744 assert(false && "Invalid value type!");
7745 case MVT::i8: Reg = X86::AL; size = 1; break;
7746 case MVT::i16: Reg = X86::AX; size = 2; break;
7747 case MVT::i32: Reg = X86::EAX; size = 4; break;
7749 assert(Subtarget->is64Bit() && "Node not type legal!");
7750 Reg = X86::RAX; size = 8;
7753 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7754 Op.getOperand(2), SDValue());
7755 SDValue Ops[] = { cpIn.getValue(0),
7758 DAG.getTargetConstant(size, MVT::i8),
7760 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7761 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7763 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7767 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7768 SelectionDAG &DAG) const {
7769 assert(Subtarget->is64Bit() && "Result not type legalized?");
7770 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7771 SDValue TheChain = Op.getOperand(0);
7772 DebugLoc dl = Op.getDebugLoc();
7773 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7774 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7775 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7777 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7778 DAG.getConstant(32, MVT::i8));
7780 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7783 return DAG.getMergeValues(Ops, 2, dl);
7786 SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
7787 SelectionDAG &DAG) const {
7788 EVT SrcVT = Op.getOperand(0).getValueType();
7789 EVT DstVT = Op.getValueType();
7790 assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
7791 Subtarget->hasMMX() && !DisableMMX) &&
7792 "Unexpected custom BIT_CONVERT");
7793 assert((DstVT == MVT::i64 ||
7794 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
7795 "Unexpected custom BIT_CONVERT");
7796 // i64 <=> MMX conversions are Legal.
7797 if (SrcVT==MVT::i64 && DstVT.isVector())
7799 if (DstVT==MVT::i64 && SrcVT.isVector())
7801 // MMX <=> MMX conversions are Legal.
7802 if (SrcVT.isVector() && DstVT.isVector())
7804 // All other conversions need to be expanded.
7807 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
7808 SDNode *Node = Op.getNode();
7809 DebugLoc dl = Node->getDebugLoc();
7810 EVT T = Node->getValueType(0);
7811 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7812 DAG.getConstant(0, T), Node->getOperand(2));
7813 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7814 cast<AtomicSDNode>(Node)->getMemoryVT(),
7815 Node->getOperand(0),
7816 Node->getOperand(1), negOp,
7817 cast<AtomicSDNode>(Node)->getSrcValue(),
7818 cast<AtomicSDNode>(Node)->getAlignment());
7821 /// LowerOperation - Provide custom lowering hooks for some operations.
7823 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
7824 switch (Op.getOpcode()) {
7825 default: llvm_unreachable("Should not custom lower this!");
7826 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
7827 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7828 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7829 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7830 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7831 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7832 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7833 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7834 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7835 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7836 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7837 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7838 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7839 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7840 case ISD::SHL_PARTS:
7841 case ISD::SRA_PARTS:
7842 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7843 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7844 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7845 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7846 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7847 case ISD::FABS: return LowerFABS(Op, DAG);
7848 case ISD::FNEG: return LowerFNEG(Op, DAG);
7849 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7850 case ISD::SETCC: return LowerSETCC(Op, DAG);
7851 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7852 case ISD::SELECT: return LowerSELECT(Op, DAG);
7853 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7854 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7855 case ISD::VASTART: return LowerVASTART(Op, DAG);
7856 case ISD::VAARG: return LowerVAARG(Op, DAG);
7857 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7858 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7859 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7860 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7861 case ISD::FRAME_TO_ARGS_OFFSET:
7862 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7863 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7864 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7865 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7866 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7867 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7868 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7869 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7870 case ISD::SHL: return LowerSHL(Op, DAG);
7876 case ISD::UMULO: return LowerXALUO(Op, DAG);
7877 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7878 case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
7882 void X86TargetLowering::
7883 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7884 SelectionDAG &DAG, unsigned NewOp) const {
7885 EVT T = Node->getValueType(0);
7886 DebugLoc dl = Node->getDebugLoc();
7887 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7889 SDValue Chain = Node->getOperand(0);
7890 SDValue In1 = Node->getOperand(1);
7891 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7892 Node->getOperand(2), DAG.getIntPtrConstant(0));
7893 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7894 Node->getOperand(2), DAG.getIntPtrConstant(1));
7895 SDValue Ops[] = { Chain, In1, In2L, In2H };
7896 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7898 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7899 cast<MemSDNode>(Node)->getMemOperand());
7900 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7901 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7902 Results.push_back(Result.getValue(2));
7905 /// ReplaceNodeResults - Replace a node with an illegal result type
7906 /// with a new node built out of custom code.
7907 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7908 SmallVectorImpl<SDValue>&Results,
7909 SelectionDAG &DAG) const {
7910 DebugLoc dl = N->getDebugLoc();
7911 switch (N->getOpcode()) {
7913 assert(false && "Do not know how to custom type legalize this operation!");
7915 case ISD::FP_TO_SINT: {
7916 std::pair<SDValue,SDValue> Vals =
7917 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7918 SDValue FIST = Vals.first, StackSlot = Vals.second;
7919 if (FIST.getNode() != 0) {
7920 EVT VT = N->getValueType(0);
7921 // Return a load from the stack slot.
7922 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7927 case ISD::READCYCLECOUNTER: {
7928 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7929 SDValue TheChain = N->getOperand(0);
7930 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7931 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7933 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7935 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7936 SDValue Ops[] = { eax, edx };
7937 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7938 Results.push_back(edx.getValue(1));
7941 case ISD::ATOMIC_CMP_SWAP: {
7942 EVT T = N->getValueType(0);
7943 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7944 SDValue cpInL, cpInH;
7945 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7946 DAG.getConstant(0, MVT::i32));
7947 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7948 DAG.getConstant(1, MVT::i32));
7949 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7950 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7952 SDValue swapInL, swapInH;
7953 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7954 DAG.getConstant(0, MVT::i32));
7955 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7956 DAG.getConstant(1, MVT::i32));
7957 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7959 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7960 swapInL.getValue(1));
7961 SDValue Ops[] = { swapInH.getValue(0),
7963 swapInH.getValue(1) };
7964 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7965 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7966 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7967 MVT::i32, Result.getValue(1));
7968 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7969 MVT::i32, cpOutL.getValue(2));
7970 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7971 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7972 Results.push_back(cpOutH.getValue(1));
7975 case ISD::ATOMIC_LOAD_ADD:
7976 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7978 case ISD::ATOMIC_LOAD_AND:
7979 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7981 case ISD::ATOMIC_LOAD_NAND:
7982 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7984 case ISD::ATOMIC_LOAD_OR:
7985 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7987 case ISD::ATOMIC_LOAD_SUB:
7988 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7990 case ISD::ATOMIC_LOAD_XOR:
7991 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7993 case ISD::ATOMIC_SWAP:
7994 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7999 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
8001 default: return NULL;
8002 case X86ISD::BSF: return "X86ISD::BSF";
8003 case X86ISD::BSR: return "X86ISD::BSR";
8004 case X86ISD::SHLD: return "X86ISD::SHLD";
8005 case X86ISD::SHRD: return "X86ISD::SHRD";
8006 case X86ISD::FAND: return "X86ISD::FAND";
8007 case X86ISD::FOR: return "X86ISD::FOR";
8008 case X86ISD::FXOR: return "X86ISD::FXOR";
8009 case X86ISD::FSRL: return "X86ISD::FSRL";
8010 case X86ISD::FILD: return "X86ISD::FILD";
8011 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
8012 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
8013 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
8014 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
8015 case X86ISD::FLD: return "X86ISD::FLD";
8016 case X86ISD::FST: return "X86ISD::FST";
8017 case X86ISD::CALL: return "X86ISD::CALL";
8018 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
8019 case X86ISD::BT: return "X86ISD::BT";
8020 case X86ISD::CMP: return "X86ISD::CMP";
8021 case X86ISD::COMI: return "X86ISD::COMI";
8022 case X86ISD::UCOMI: return "X86ISD::UCOMI";
8023 case X86ISD::SETCC: return "X86ISD::SETCC";
8024 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
8025 case X86ISD::CMOV: return "X86ISD::CMOV";
8026 case X86ISD::BRCOND: return "X86ISD::BRCOND";
8027 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
8028 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
8029 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
8030 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
8031 case X86ISD::Wrapper: return "X86ISD::Wrapper";
8032 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
8033 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
8034 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
8035 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
8036 case X86ISD::PINSRB: return "X86ISD::PINSRB";
8037 case X86ISD::PINSRW: return "X86ISD::PINSRW";
8038 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
8039 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
8040 case X86ISD::FMAX: return "X86ISD::FMAX";
8041 case X86ISD::FMIN: return "X86ISD::FMIN";
8042 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
8043 case X86ISD::FRCP: return "X86ISD::FRCP";
8044 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
8045 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
8046 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
8047 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
8048 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
8049 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
8050 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
8051 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
8052 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
8053 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
8054 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
8055 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
8056 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
8057 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
8058 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
8059 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
8060 case X86ISD::VSHL: return "X86ISD::VSHL";
8061 case X86ISD::VSRL: return "X86ISD::VSRL";
8062 case X86ISD::CMPPD: return "X86ISD::CMPPD";
8063 case X86ISD::CMPPS: return "X86ISD::CMPPS";
8064 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
8065 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
8066 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
8067 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
8068 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
8069 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
8070 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
8071 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
8072 case X86ISD::ADD: return "X86ISD::ADD";
8073 case X86ISD::SUB: return "X86ISD::SUB";
8074 case X86ISD::SMUL: return "X86ISD::SMUL";
8075 case X86ISD::UMUL: return "X86ISD::UMUL";
8076 case X86ISD::INC: return "X86ISD::INC";
8077 case X86ISD::DEC: return "X86ISD::DEC";
8078 case X86ISD::OR: return "X86ISD::OR";
8079 case X86ISD::XOR: return "X86ISD::XOR";
8080 case X86ISD::AND: return "X86ISD::AND";
8081 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
8082 case X86ISD::PTEST: return "X86ISD::PTEST";
8083 case X86ISD::TESTP: return "X86ISD::TESTP";
8084 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
8085 case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
8089 // isLegalAddressingMode - Return true if the addressing mode represented
8090 // by AM is legal for this target, for a load/store of the specified type.
8091 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
8092 const Type *Ty) const {
8093 // X86 supports extremely general addressing modes.
8094 CodeModel::Model M = getTargetMachine().getCodeModel();
8096 // X86 allows a sign-extended 32-bit immediate field as a displacement.
8097 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
8102 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
8104 // If a reference to this global requires an extra load, we can't fold it.
8105 if (isGlobalStubReference(GVFlags))
8108 // If BaseGV requires a register for the PIC base, we cannot also have a
8109 // BaseReg specified.
8110 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
8113 // If lower 4G is not available, then we must use rip-relative addressing.
8114 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
8124 // These scales always work.
8129 // These scales are formed with basereg+scalereg. Only accept if there is
8134 default: // Other stuff never works.
8142 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
8143 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
8145 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
8146 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
8147 if (NumBits1 <= NumBits2)
8152 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
8153 if (!VT1.isInteger() || !VT2.isInteger())
8155 unsigned NumBits1 = VT1.getSizeInBits();
8156 unsigned NumBits2 = VT2.getSizeInBits();
8157 if (NumBits1 <= NumBits2)
8162 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
8163 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
8164 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
8167 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
8168 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
8169 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
8172 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
8173 // i16 instructions are longer (0x66 prefix) and potentially slower.
8174 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
8177 /// isShuffleMaskLegal - Targets can use this to indicate that they only
8178 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
8179 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
8180 /// are assumed to be legal.
8182 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
8184 // Very little shuffling can be done for 64-bit vectors right now.
8185 if (VT.getSizeInBits() == 64)
8186 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
8188 // FIXME: pshufb, blends, shifts.
8189 return (VT.getVectorNumElements() == 2 ||
8190 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
8191 isMOVLMask(M, VT) ||
8192 isSHUFPMask(M, VT) ||
8193 isPSHUFDMask(M, VT) ||
8194 isPSHUFHWMask(M, VT) ||
8195 isPSHUFLWMask(M, VT) ||
8196 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
8197 isUNPCKLMask(M, VT) ||
8198 isUNPCKHMask(M, VT) ||
8199 isUNPCKL_v_undef_Mask(M, VT) ||
8200 isUNPCKH_v_undef_Mask(M, VT));
8204 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
8206 unsigned NumElts = VT.getVectorNumElements();
8207 // FIXME: This collection of masks seems suspect.
8210 if (NumElts == 4 && VT.getSizeInBits() == 128) {
8211 return (isMOVLMask(Mask, VT) ||
8212 isCommutedMOVLMask(Mask, VT, true) ||
8213 isSHUFPMask(Mask, VT) ||
8214 isCommutedSHUFPMask(Mask, VT));
8219 //===----------------------------------------------------------------------===//
8220 // X86 Scheduler Hooks
8221 //===----------------------------------------------------------------------===//
8223 // private utility function
8225 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
8226 MachineBasicBlock *MBB,
8233 TargetRegisterClass *RC,
8234 bool invSrc) const {
8235 // For the atomic bitwise operator, we generate
8238 // ld t1 = [bitinstr.addr]
8239 // op t2 = t1, [bitinstr.val]
8241 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8243 // fallthrough -->nextMBB
8244 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8245 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8246 MachineFunction::iterator MBBIter = MBB;
8249 /// First build the CFG
8250 MachineFunction *F = MBB->getParent();
8251 MachineBasicBlock *thisMBB = MBB;
8252 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8253 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8254 F->insert(MBBIter, newMBB);
8255 F->insert(MBBIter, nextMBB);
8257 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8258 nextMBB->splice(nextMBB->begin(), thisMBB,
8259 llvm::next(MachineBasicBlock::iterator(bInstr)),
8261 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8263 // Update thisMBB to fall through to newMBB
8264 thisMBB->addSuccessor(newMBB);
8266 // newMBB jumps to itself and fall through to nextMBB
8267 newMBB->addSuccessor(nextMBB);
8268 newMBB->addSuccessor(newMBB);
8270 // Insert instructions into newMBB based on incoming instruction
8271 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
8272 "unexpected number of operands");
8273 DebugLoc dl = bInstr->getDebugLoc();
8274 MachineOperand& destOper = bInstr->getOperand(0);
8275 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8276 int numArgs = bInstr->getNumOperands() - 1;
8277 for (int i=0; i < numArgs; ++i)
8278 argOpers[i] = &bInstr->getOperand(i+1);
8280 // x86 address has 4 operands: base, index, scale, and displacement
8281 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8282 int valArgIndx = lastAddrIndx + 1;
8284 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8285 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
8286 for (int i=0; i <= lastAddrIndx; ++i)
8287 (*MIB).addOperand(*argOpers[i]);
8289 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
8291 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
8296 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8297 assert((argOpers[valArgIndx]->isReg() ||
8298 argOpers[valArgIndx]->isImm()) &&
8300 if (argOpers[valArgIndx]->isReg())
8301 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
8303 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
8305 (*MIB).addOperand(*argOpers[valArgIndx]);
8307 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
8310 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
8311 for (int i=0; i <= lastAddrIndx; ++i)
8312 (*MIB).addOperand(*argOpers[i]);
8314 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8315 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8316 bInstr->memoperands_end());
8318 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
8322 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8324 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8328 // private utility function: 64 bit atomics on 32 bit host.
8330 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
8331 MachineBasicBlock *MBB,
8336 bool invSrc) const {
8337 // For the atomic bitwise operator, we generate
8338 // thisMBB (instructions are in pairs, except cmpxchg8b)
8339 // ld t1,t2 = [bitinstr.addr]
8341 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
8342 // op t5, t6 <- out1, out2, [bitinstr.val]
8343 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
8344 // mov ECX, EBX <- t5, t6
8345 // mov EAX, EDX <- t1, t2
8346 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
8347 // mov t3, t4 <- EAX, EDX
8349 // result in out1, out2
8350 // fallthrough -->nextMBB
8352 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8353 const unsigned LoadOpc = X86::MOV32rm;
8354 const unsigned NotOpc = X86::NOT32r;
8355 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8356 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8357 MachineFunction::iterator MBBIter = MBB;
8360 /// First build the CFG
8361 MachineFunction *F = MBB->getParent();
8362 MachineBasicBlock *thisMBB = MBB;
8363 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8364 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8365 F->insert(MBBIter, newMBB);
8366 F->insert(MBBIter, nextMBB);
8368 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8369 nextMBB->splice(nextMBB->begin(), thisMBB,
8370 llvm::next(MachineBasicBlock::iterator(bInstr)),
8372 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8374 // Update thisMBB to fall through to newMBB
8375 thisMBB->addSuccessor(newMBB);
8377 // newMBB jumps to itself and fall through to nextMBB
8378 newMBB->addSuccessor(nextMBB);
8379 newMBB->addSuccessor(newMBB);
8381 DebugLoc dl = bInstr->getDebugLoc();
8382 // Insert instructions into newMBB based on incoming instruction
8383 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8384 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
8385 "unexpected number of operands");
8386 MachineOperand& dest1Oper = bInstr->getOperand(0);
8387 MachineOperand& dest2Oper = bInstr->getOperand(1);
8388 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8389 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
8390 argOpers[i] = &bInstr->getOperand(i+2);
8392 // We use some of the operands multiple times, so conservatively just
8393 // clear any kill flags that might be present.
8394 if (argOpers[i]->isReg() && argOpers[i]->isUse())
8395 argOpers[i]->setIsKill(false);
8398 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8399 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8401 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8402 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8403 for (int i=0; i <= lastAddrIndx; ++i)
8404 (*MIB).addOperand(*argOpers[i]);
8405 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8406 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8407 // add 4 to displacement.
8408 for (int i=0; i <= lastAddrIndx-2; ++i)
8409 (*MIB).addOperand(*argOpers[i]);
8410 MachineOperand newOp3 = *(argOpers[3]);
8412 newOp3.setImm(newOp3.getImm()+4);
8414 newOp3.setOffset(newOp3.getOffset()+4);
8415 (*MIB).addOperand(newOp3);
8416 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8418 // t3/4 are defined later, at the bottom of the loop
8419 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8420 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8421 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8422 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8423 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8424 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8426 // The subsequent operations should be using the destination registers of
8427 //the PHI instructions.
8429 t1 = F->getRegInfo().createVirtualRegister(RC);
8430 t2 = F->getRegInfo().createVirtualRegister(RC);
8431 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8432 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8434 t1 = dest1Oper.getReg();
8435 t2 = dest2Oper.getReg();
8438 int valArgIndx = lastAddrIndx + 1;
8439 assert((argOpers[valArgIndx]->isReg() ||
8440 argOpers[valArgIndx]->isImm()) &&
8442 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8443 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8444 if (argOpers[valArgIndx]->isReg())
8445 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8447 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8448 if (regOpcL != X86::MOV32rr)
8450 (*MIB).addOperand(*argOpers[valArgIndx]);
8451 assert(argOpers[valArgIndx + 1]->isReg() ==
8452 argOpers[valArgIndx]->isReg());
8453 assert(argOpers[valArgIndx + 1]->isImm() ==
8454 argOpers[valArgIndx]->isImm());
8455 if (argOpers[valArgIndx + 1]->isReg())
8456 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8458 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8459 if (regOpcH != X86::MOV32rr)
8461 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8463 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
8465 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
8468 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
8470 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
8473 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8474 for (int i=0; i <= lastAddrIndx; ++i)
8475 (*MIB).addOperand(*argOpers[i]);
8477 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8478 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8479 bInstr->memoperands_end());
8481 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
8482 MIB.addReg(X86::EAX);
8483 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
8484 MIB.addReg(X86::EDX);
8487 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8489 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
8493 // private utility function
8495 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8496 MachineBasicBlock *MBB,
8497 unsigned cmovOpc) const {
8498 // For the atomic min/max operator, we generate
8501 // ld t1 = [min/max.addr]
8502 // mov t2 = [min/max.val]
8504 // cmov[cond] t2 = t1
8506 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8508 // fallthrough -->nextMBB
8510 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8511 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8512 MachineFunction::iterator MBBIter = MBB;
8515 /// First build the CFG
8516 MachineFunction *F = MBB->getParent();
8517 MachineBasicBlock *thisMBB = MBB;
8518 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8519 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8520 F->insert(MBBIter, newMBB);
8521 F->insert(MBBIter, nextMBB);
8523 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
8524 nextMBB->splice(nextMBB->begin(), thisMBB,
8525 llvm::next(MachineBasicBlock::iterator(mInstr)),
8527 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
8529 // Update thisMBB to fall through to newMBB
8530 thisMBB->addSuccessor(newMBB);
8532 // newMBB jumps to newMBB and fall through to nextMBB
8533 newMBB->addSuccessor(nextMBB);
8534 newMBB->addSuccessor(newMBB);
8536 DebugLoc dl = mInstr->getDebugLoc();
8537 // Insert instructions into newMBB based on incoming instruction
8538 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
8539 "unexpected number of operands");
8540 MachineOperand& destOper = mInstr->getOperand(0);
8541 MachineOperand* argOpers[2 + X86::AddrNumOperands];
8542 int numArgs = mInstr->getNumOperands() - 1;
8543 for (int i=0; i < numArgs; ++i)
8544 argOpers[i] = &mInstr->getOperand(i+1);
8546 // x86 address has 4 operands: base, index, scale, and displacement
8547 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
8548 int valArgIndx = lastAddrIndx + 1;
8550 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8551 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8552 for (int i=0; i <= lastAddrIndx; ++i)
8553 (*MIB).addOperand(*argOpers[i]);
8555 // We only support register and immediate values
8556 assert((argOpers[valArgIndx]->isReg() ||
8557 argOpers[valArgIndx]->isImm()) &&
8560 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8561 if (argOpers[valArgIndx]->isReg())
8562 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
8564 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8565 (*MIB).addOperand(*argOpers[valArgIndx]);
8567 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
8570 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8575 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8576 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8580 // Cmp and exchange if none has modified the memory location
8581 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8582 for (int i=0; i <= lastAddrIndx; ++i)
8583 (*MIB).addOperand(*argOpers[i]);
8585 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8586 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8587 mInstr->memoperands_end());
8589 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
8590 MIB.addReg(X86::EAX);
8593 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8595 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
8599 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8600 // or XMM0_V32I8 in AVX all of this code can be replaced with that
8603 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8604 unsigned numArgs, bool memArg) const {
8606 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
8607 "Target must have SSE4.2 or AVX features enabled");
8609 DebugLoc dl = MI->getDebugLoc();
8610 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8614 if (!Subtarget->hasAVX()) {
8616 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8618 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8621 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
8623 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
8626 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8628 for (unsigned i = 0; i < numArgs; ++i) {
8629 MachineOperand &Op = MI->getOperand(i+1);
8631 if (!(Op.isReg() && Op.isImplicit()))
8635 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8638 MI->eraseFromParent();
8644 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8646 MachineBasicBlock *MBB) const {
8647 // Emit code to save XMM registers to the stack. The ABI says that the
8648 // number of registers to save is given in %al, so it's theoretically
8649 // possible to do an indirect jump trick to avoid saving all of them,
8650 // however this code takes a simpler approach and just executes all
8651 // of the stores if %al is non-zero. It's less code, and it's probably
8652 // easier on the hardware branch predictor, and stores aren't all that
8653 // expensive anyway.
8655 // Create the new basic blocks. One block contains all the XMM stores,
8656 // and one block is the final destination regardless of whether any
8657 // stores were performed.
8658 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8659 MachineFunction *F = MBB->getParent();
8660 MachineFunction::iterator MBBIter = MBB;
8662 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8663 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8664 F->insert(MBBIter, XMMSaveMBB);
8665 F->insert(MBBIter, EndMBB);
8667 // Transfer the remainder of MBB and its successor edges to EndMBB.
8668 EndMBB->splice(EndMBB->begin(), MBB,
8669 llvm::next(MachineBasicBlock::iterator(MI)),
8671 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
8673 // The original block will now fall through to the XMM save block.
8674 MBB->addSuccessor(XMMSaveMBB);
8675 // The XMMSaveMBB will fall through to the end block.
8676 XMMSaveMBB->addSuccessor(EndMBB);
8678 // Now add the instructions.
8679 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8680 DebugLoc DL = MI->getDebugLoc();
8682 unsigned CountReg = MI->getOperand(0).getReg();
8683 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8684 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8686 if (!Subtarget->isTargetWin64()) {
8687 // If %al is 0, branch around the XMM save block.
8688 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8689 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8690 MBB->addSuccessor(EndMBB);
8693 // In the XMM save block, save all the XMM argument registers.
8694 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8695 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8696 MachineMemOperand *MMO =
8697 F->getMachineMemOperand(
8698 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8699 MachineMemOperand::MOStore, Offset,
8700 /*Size=*/16, /*Align=*/16);
8701 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8702 .addFrameIndex(RegSaveFrameIndex)
8703 .addImm(/*Scale=*/1)
8704 .addReg(/*IndexReg=*/0)
8705 .addImm(/*Disp=*/Offset)
8706 .addReg(/*Segment=*/0)
8707 .addReg(MI->getOperand(i).getReg())
8708 .addMemOperand(MMO);
8711 MI->eraseFromParent(); // The pseudo instruction is gone now.
8717 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8718 MachineBasicBlock *BB) const {
8719 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8720 DebugLoc DL = MI->getDebugLoc();
8722 // To "insert" a SELECT_CC instruction, we actually have to insert the
8723 // diamond control-flow pattern. The incoming instruction knows the
8724 // destination vreg to set, the condition code register to branch on, the
8725 // true/false values to select between, and a branch opcode to use.
8726 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8727 MachineFunction::iterator It = BB;
8733 // cmpTY ccX, r1, r2
8735 // fallthrough --> copy0MBB
8736 MachineBasicBlock *thisMBB = BB;
8737 MachineFunction *F = BB->getParent();
8738 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8739 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8740 F->insert(It, copy0MBB);
8741 F->insert(It, sinkMBB);
8743 // If the EFLAGS register isn't dead in the terminator, then claim that it's
8744 // live into the sink and copy blocks.
8745 const MachineFunction *MF = BB->getParent();
8746 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
8747 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
8749 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
8750 const MachineOperand &MO = MI->getOperand(I);
8751 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
8752 unsigned Reg = MO.getReg();
8753 if (Reg != X86::EFLAGS) continue;
8754 copy0MBB->addLiveIn(Reg);
8755 sinkMBB->addLiveIn(Reg);
8758 // Transfer the remainder of BB and its successor edges to sinkMBB.
8759 sinkMBB->splice(sinkMBB->begin(), BB,
8760 llvm::next(MachineBasicBlock::iterator(MI)),
8762 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
8764 // Add the true and fallthrough blocks as its successors.
8765 BB->addSuccessor(copy0MBB);
8766 BB->addSuccessor(sinkMBB);
8768 // Create the conditional branch instruction.
8770 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8771 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8774 // %FalseValue = ...
8775 // # fallthrough to sinkMBB
8776 copy0MBB->addSuccessor(sinkMBB);
8779 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8781 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
8782 TII->get(X86::PHI), MI->getOperand(0).getReg())
8783 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8784 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8786 MI->eraseFromParent(); // The pseudo instruction is gone now.
8791 X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
8792 MachineBasicBlock *BB) const {
8793 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8794 DebugLoc DL = MI->getDebugLoc();
8796 // The lowering is pretty easy: we're just emitting the call to _alloca. The
8797 // non-trivial part is impdef of ESP.
8798 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
8801 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
8802 .addExternalSymbol("_alloca")
8803 .addReg(X86::EAX, RegState::Implicit)
8804 .addReg(X86::ESP, RegState::Implicit)
8805 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
8806 .addReg(X86::ESP, RegState::Define | RegState::Implicit);
8808 MI->eraseFromParent(); // The pseudo instruction is gone now.
8813 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
8814 MachineBasicBlock *BB) const {
8815 // This is pretty easy. We're taking the value that we received from
8816 // our load from the relocation, sticking it in either RDI (x86-64)
8817 // or EAX and doing an indirect call. The return value will then
8818 // be in the normal return register.
8819 const X86InstrInfo *TII
8820 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
8821 DebugLoc DL = MI->getDebugLoc();
8822 MachineFunction *F = BB->getParent();
8824 assert(MI->getOperand(3).isGlobal() && "This should be a global");
8826 if (Subtarget->is64Bit()) {
8827 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8828 TII->get(X86::MOV64rm), X86::RDI)
8830 .addImm(0).addReg(0)
8831 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8832 MI->getOperand(3).getTargetFlags())
8834 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
8835 addDirectMem(MIB, X86::RDI);
8836 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
8837 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8838 TII->get(X86::MOV32rm), X86::EAX)
8840 .addImm(0).addReg(0)
8841 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8842 MI->getOperand(3).getTargetFlags())
8844 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
8845 addDirectMem(MIB, X86::EAX);
8847 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
8848 TII->get(X86::MOV32rm), X86::EAX)
8849 .addReg(TII->getGlobalBaseReg(F))
8850 .addImm(0).addReg(0)
8851 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
8852 MI->getOperand(3).getTargetFlags())
8854 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
8855 addDirectMem(MIB, X86::EAX);
8858 MI->eraseFromParent(); // The pseudo instruction is gone now.
8863 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8864 MachineBasicBlock *BB) const {
8865 switch (MI->getOpcode()) {
8866 default: assert(false && "Unexpected instr type to insert");
8867 case X86::MINGW_ALLOCA:
8868 return EmitLoweredMingwAlloca(MI, BB);
8869 case X86::TLSCall_32:
8870 case X86::TLSCall_64:
8871 return EmitLoweredTLSCall(MI, BB);
8873 case X86::CMOV_V1I64:
8874 case X86::CMOV_FR32:
8875 case X86::CMOV_FR64:
8876 case X86::CMOV_V4F32:
8877 case X86::CMOV_V2F64:
8878 case X86::CMOV_V2I64:
8879 case X86::CMOV_GR16:
8880 case X86::CMOV_GR32:
8881 case X86::CMOV_RFP32:
8882 case X86::CMOV_RFP64:
8883 case X86::CMOV_RFP80:
8884 return EmitLoweredSelect(MI, BB);
8886 case X86::FP32_TO_INT16_IN_MEM:
8887 case X86::FP32_TO_INT32_IN_MEM:
8888 case X86::FP32_TO_INT64_IN_MEM:
8889 case X86::FP64_TO_INT16_IN_MEM:
8890 case X86::FP64_TO_INT32_IN_MEM:
8891 case X86::FP64_TO_INT64_IN_MEM:
8892 case X86::FP80_TO_INT16_IN_MEM:
8893 case X86::FP80_TO_INT32_IN_MEM:
8894 case X86::FP80_TO_INT64_IN_MEM: {
8895 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8896 DebugLoc DL = MI->getDebugLoc();
8898 // Change the floating point control register to use "round towards zero"
8899 // mode when truncating to an integer value.
8900 MachineFunction *F = BB->getParent();
8901 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8902 addFrameReference(BuildMI(*BB, MI, DL,
8903 TII->get(X86::FNSTCW16m)), CWFrameIdx);
8905 // Load the old value of the high byte of the control word...
8907 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8908 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
8911 // Set the high part to be round to zero...
8912 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8915 // Reload the modified control word now...
8916 addFrameReference(BuildMI(*BB, MI, DL,
8917 TII->get(X86::FLDCW16m)), CWFrameIdx);
8919 // Restore the memory image of control word to original value
8920 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8923 // Get the X86 opcode to use.
8925 switch (MI->getOpcode()) {
8926 default: llvm_unreachable("illegal opcode!");
8927 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8928 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8929 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8930 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8931 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8932 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8933 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8934 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8935 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8939 MachineOperand &Op = MI->getOperand(0);
8941 AM.BaseType = X86AddressMode::RegBase;
8942 AM.Base.Reg = Op.getReg();
8944 AM.BaseType = X86AddressMode::FrameIndexBase;
8945 AM.Base.FrameIndex = Op.getIndex();
8947 Op = MI->getOperand(1);
8949 AM.Scale = Op.getImm();
8950 Op = MI->getOperand(2);
8952 AM.IndexReg = Op.getImm();
8953 Op = MI->getOperand(3);
8954 if (Op.isGlobal()) {
8955 AM.GV = Op.getGlobal();
8957 AM.Disp = Op.getImm();
8959 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
8960 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
8962 // Reload the original control word now.
8963 addFrameReference(BuildMI(*BB, MI, DL,
8964 TII->get(X86::FLDCW16m)), CWFrameIdx);
8966 MI->eraseFromParent(); // The pseudo instruction is gone now.
8969 // String/text processing lowering.
8970 case X86::PCMPISTRM128REG:
8971 case X86::VPCMPISTRM128REG:
8972 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8973 case X86::PCMPISTRM128MEM:
8974 case X86::VPCMPISTRM128MEM:
8975 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8976 case X86::PCMPESTRM128REG:
8977 case X86::VPCMPESTRM128REG:
8978 return EmitPCMP(MI, BB, 5, false /* in mem */);
8979 case X86::PCMPESTRM128MEM:
8980 case X86::VPCMPESTRM128MEM:
8981 return EmitPCMP(MI, BB, 5, true /* in mem */);
8984 case X86::ATOMAND32:
8985 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8986 X86::AND32ri, X86::MOV32rm,
8988 X86::NOT32r, X86::EAX,
8989 X86::GR32RegisterClass);
8991 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8992 X86::OR32ri, X86::MOV32rm,
8994 X86::NOT32r, X86::EAX,
8995 X86::GR32RegisterClass);
8996 case X86::ATOMXOR32:
8997 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8998 X86::XOR32ri, X86::MOV32rm,
9000 X86::NOT32r, X86::EAX,
9001 X86::GR32RegisterClass);
9002 case X86::ATOMNAND32:
9003 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
9004 X86::AND32ri, X86::MOV32rm,
9006 X86::NOT32r, X86::EAX,
9007 X86::GR32RegisterClass, true);
9008 case X86::ATOMMIN32:
9009 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
9010 case X86::ATOMMAX32:
9011 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
9012 case X86::ATOMUMIN32:
9013 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
9014 case X86::ATOMUMAX32:
9015 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
9017 case X86::ATOMAND16:
9018 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
9019 X86::AND16ri, X86::MOV16rm,
9021 X86::NOT16r, X86::AX,
9022 X86::GR16RegisterClass);
9024 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
9025 X86::OR16ri, X86::MOV16rm,
9027 X86::NOT16r, X86::AX,
9028 X86::GR16RegisterClass);
9029 case X86::ATOMXOR16:
9030 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
9031 X86::XOR16ri, X86::MOV16rm,
9033 X86::NOT16r, X86::AX,
9034 X86::GR16RegisterClass);
9035 case X86::ATOMNAND16:
9036 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
9037 X86::AND16ri, X86::MOV16rm,
9039 X86::NOT16r, X86::AX,
9040 X86::GR16RegisterClass, true);
9041 case X86::ATOMMIN16:
9042 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
9043 case X86::ATOMMAX16:
9044 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
9045 case X86::ATOMUMIN16:
9046 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
9047 case X86::ATOMUMAX16:
9048 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
9051 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
9052 X86::AND8ri, X86::MOV8rm,
9054 X86::NOT8r, X86::AL,
9055 X86::GR8RegisterClass);
9057 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
9058 X86::OR8ri, X86::MOV8rm,
9060 X86::NOT8r, X86::AL,
9061 X86::GR8RegisterClass);
9063 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
9064 X86::XOR8ri, X86::MOV8rm,
9066 X86::NOT8r, X86::AL,
9067 X86::GR8RegisterClass);
9068 case X86::ATOMNAND8:
9069 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
9070 X86::AND8ri, X86::MOV8rm,
9072 X86::NOT8r, X86::AL,
9073 X86::GR8RegisterClass, true);
9074 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
9075 // This group is for 64-bit host.
9076 case X86::ATOMAND64:
9077 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
9078 X86::AND64ri32, X86::MOV64rm,
9080 X86::NOT64r, X86::RAX,
9081 X86::GR64RegisterClass);
9083 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
9084 X86::OR64ri32, X86::MOV64rm,
9086 X86::NOT64r, X86::RAX,
9087 X86::GR64RegisterClass);
9088 case X86::ATOMXOR64:
9089 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
9090 X86::XOR64ri32, X86::MOV64rm,
9092 X86::NOT64r, X86::RAX,
9093 X86::GR64RegisterClass);
9094 case X86::ATOMNAND64:
9095 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
9096 X86::AND64ri32, X86::MOV64rm,
9098 X86::NOT64r, X86::RAX,
9099 X86::GR64RegisterClass, true);
9100 case X86::ATOMMIN64:
9101 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
9102 case X86::ATOMMAX64:
9103 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
9104 case X86::ATOMUMIN64:
9105 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
9106 case X86::ATOMUMAX64:
9107 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
9109 // This group does 64-bit operations on a 32-bit host.
9110 case X86::ATOMAND6432:
9111 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9112 X86::AND32rr, X86::AND32rr,
9113 X86::AND32ri, X86::AND32ri,
9115 case X86::ATOMOR6432:
9116 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9117 X86::OR32rr, X86::OR32rr,
9118 X86::OR32ri, X86::OR32ri,
9120 case X86::ATOMXOR6432:
9121 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9122 X86::XOR32rr, X86::XOR32rr,
9123 X86::XOR32ri, X86::XOR32ri,
9125 case X86::ATOMNAND6432:
9126 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9127 X86::AND32rr, X86::AND32rr,
9128 X86::AND32ri, X86::AND32ri,
9130 case X86::ATOMADD6432:
9131 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9132 X86::ADD32rr, X86::ADC32rr,
9133 X86::ADD32ri, X86::ADC32ri,
9135 case X86::ATOMSUB6432:
9136 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9137 X86::SUB32rr, X86::SBB32rr,
9138 X86::SUB32ri, X86::SBB32ri,
9140 case X86::ATOMSWAP6432:
9141 return EmitAtomicBit6432WithCustomInserter(MI, BB,
9142 X86::MOV32rr, X86::MOV32rr,
9143 X86::MOV32ri, X86::MOV32ri,
9145 case X86::VASTART_SAVE_XMM_REGS:
9146 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
9150 //===----------------------------------------------------------------------===//
9151 // X86 Optimization Hooks
9152 //===----------------------------------------------------------------------===//
9154 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
9158 const SelectionDAG &DAG,
9159 unsigned Depth) const {
9160 unsigned Opc = Op.getOpcode();
9161 assert((Opc >= ISD::BUILTIN_OP_END ||
9162 Opc == ISD::INTRINSIC_WO_CHAIN ||
9163 Opc == ISD::INTRINSIC_W_CHAIN ||
9164 Opc == ISD::INTRINSIC_VOID) &&
9165 "Should use MaskedValueIsZero if you don't know whether Op"
9166 " is a target node!");
9168 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
9180 // These nodes' second result is a boolean.
9181 if (Op.getResNo() == 0)
9185 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
9186 Mask.getBitWidth() - 1);
9191 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
9192 /// node is a GlobalAddress + offset.
9193 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
9194 const GlobalValue* &GA,
9195 int64_t &Offset) const {
9196 if (N->getOpcode() == X86ISD::Wrapper) {
9197 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
9198 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
9199 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
9203 return TargetLowering::isGAPlusOffset(N, GA, Offset);
9206 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
9207 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
9208 /// if the load addresses are consecutive, non-overlapping, and in the right
9210 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
9211 const TargetLowering &TLI) {
9212 DebugLoc dl = N->getDebugLoc();
9213 EVT VT = N->getValueType(0);
9214 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
9216 if (VT.getSizeInBits() != 128)
9219 SmallVector<SDValue, 16> Elts;
9220 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
9221 Elts.push_back(DAG.getShuffleScalarElt(SVN, i));
9223 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
9226 /// PerformShuffleCombine - Detect vector gather/scatter index generation
9227 /// and convert it from being a bunch of shuffles and extracts to a simple
9228 /// store and scalar loads to extract the elements.
9229 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
9230 const TargetLowering &TLI) {
9231 SDValue InputVector = N->getOperand(0);
9233 // Only operate on vectors of 4 elements, where the alternative shuffling
9234 // gets to be more expensive.
9235 if (InputVector.getValueType() != MVT::v4i32)
9238 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
9239 // single use which is a sign-extend or zero-extend, and all elements are
9241 SmallVector<SDNode *, 4> Uses;
9242 unsigned ExtractedElements = 0;
9243 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
9244 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
9245 if (UI.getUse().getResNo() != InputVector.getResNo())
9248 SDNode *Extract = *UI;
9249 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
9252 if (Extract->getValueType(0) != MVT::i32)
9254 if (!Extract->hasOneUse())
9256 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
9257 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
9259 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
9262 // Record which element was extracted.
9263 ExtractedElements |=
9264 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
9266 Uses.push_back(Extract);
9269 // If not all the elements were used, this may not be worthwhile.
9270 if (ExtractedElements != 15)
9273 // Ok, we've now decided to do the transformation.
9274 DebugLoc dl = InputVector.getDebugLoc();
9276 // Store the value to a temporary stack slot.
9277 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
9278 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr, NULL,
9279 0, false, false, 0);
9281 // Replace each use (extract) with a load of the appropriate element.
9282 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
9283 UE = Uses.end(); UI != UE; ++UI) {
9284 SDNode *Extract = *UI;
9286 // Compute the element's address.
9287 SDValue Idx = Extract->getOperand(1);
9289 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
9290 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
9291 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
9293 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(),
9294 OffsetVal, StackPtr);
9297 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
9298 ScalarAddr, NULL, 0, false, false, 0);
9300 // Replace the exact with the load.
9301 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
9304 // The replacement was made in place; don't return anything.
9308 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
9309 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
9310 const X86Subtarget *Subtarget) {
9311 DebugLoc DL = N->getDebugLoc();
9312 SDValue Cond = N->getOperand(0);
9313 // Get the LHS/RHS of the select.
9314 SDValue LHS = N->getOperand(1);
9315 SDValue RHS = N->getOperand(2);
9317 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
9318 // instructions match the semantics of the common C idiom x<y?x:y but not
9319 // x<=y?x:y, because of how they handle negative zero (which can be
9320 // ignored in unsafe-math mode).
9321 if (Subtarget->hasSSE2() &&
9322 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
9323 Cond.getOpcode() == ISD::SETCC) {
9324 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
9326 unsigned Opcode = 0;
9327 // Check for x CC y ? x : y.
9328 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
9329 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
9333 // Converting this to a min would handle NaNs incorrectly, and swapping
9334 // the operands would cause it to handle comparisons between positive
9335 // and negative zero incorrectly.
9336 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
9337 if (!UnsafeFPMath &&
9338 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9340 std::swap(LHS, RHS);
9342 Opcode = X86ISD::FMIN;
9345 // Converting this to a min would handle comparisons between positive
9346 // and negative zero incorrectly.
9347 if (!UnsafeFPMath &&
9348 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
9350 Opcode = X86ISD::FMIN;
9353 // Converting this to a min would handle both negative zeros and NaNs
9354 // incorrectly, but we can swap the operands to fix both.
9355 std::swap(LHS, RHS);
9359 Opcode = X86ISD::FMIN;
9363 // Converting this to a max would handle comparisons between positive
9364 // and negative zero incorrectly.
9365 if (!UnsafeFPMath &&
9366 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
9368 Opcode = X86ISD::FMAX;
9371 // Converting this to a max would handle NaNs incorrectly, and swapping
9372 // the operands would cause it to handle comparisons between positive
9373 // and negative zero incorrectly.
9374 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
9375 if (!UnsafeFPMath &&
9376 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
9378 std::swap(LHS, RHS);
9380 Opcode = X86ISD::FMAX;
9383 // Converting this to a max would handle both negative zeros and NaNs
9384 // incorrectly, but we can swap the operands to fix both.
9385 std::swap(LHS, RHS);
9389 Opcode = X86ISD::FMAX;
9392 // Check for x CC y ? y : x -- a min/max with reversed arms.
9393 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
9394 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
9398 // Converting this to a min would handle comparisons between positive
9399 // and negative zero incorrectly, and swapping the operands would
9400 // cause it to handle NaNs incorrectly.
9401 if (!UnsafeFPMath &&
9402 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
9403 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9405 std::swap(LHS, RHS);
9407 Opcode = X86ISD::FMIN;
9410 // Converting this to a min would handle NaNs incorrectly.
9411 if (!UnsafeFPMath &&
9412 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
9414 Opcode = X86ISD::FMIN;
9417 // Converting this to a min would handle both negative zeros and NaNs
9418 // incorrectly, but we can swap the operands to fix both.
9419 std::swap(LHS, RHS);
9423 Opcode = X86ISD::FMIN;
9427 // Converting this to a max would handle NaNs incorrectly.
9428 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9430 Opcode = X86ISD::FMAX;
9433 // Converting this to a max would handle comparisons between positive
9434 // and negative zero incorrectly, and swapping the operands would
9435 // cause it to handle NaNs incorrectly.
9436 if (!UnsafeFPMath &&
9437 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
9438 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
9440 std::swap(LHS, RHS);
9442 Opcode = X86ISD::FMAX;
9445 // Converting this to a max would handle both negative zeros and NaNs
9446 // incorrectly, but we can swap the operands to fix both.
9447 std::swap(LHS, RHS);
9451 Opcode = X86ISD::FMAX;
9457 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
9460 // If this is a select between two integer constants, try to do some
9462 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
9463 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
9464 // Don't do this for crazy integer types.
9465 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
9466 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
9467 // so that TrueC (the true value) is larger than FalseC.
9468 bool NeedsCondInvert = false;
9470 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
9471 // Efficiently invertible.
9472 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
9473 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
9474 isa<ConstantSDNode>(Cond.getOperand(1))))) {
9475 NeedsCondInvert = true;
9476 std::swap(TrueC, FalseC);
9479 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
9480 if (FalseC->getAPIntValue() == 0 &&
9481 TrueC->getAPIntValue().isPowerOf2()) {
9482 if (NeedsCondInvert) // Invert the condition if needed.
9483 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9484 DAG.getConstant(1, Cond.getValueType()));
9486 // Zero extend the condition if needed.
9487 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
9489 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9490 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
9491 DAG.getConstant(ShAmt, MVT::i8));
9494 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
9495 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9496 if (NeedsCondInvert) // Invert the condition if needed.
9497 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9498 DAG.getConstant(1, Cond.getValueType()));
9500 // Zero extend the condition if needed.
9501 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9502 FalseC->getValueType(0), Cond);
9503 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9504 SDValue(FalseC, 0));
9507 // Optimize cases that will turn into an LEA instruction. This requires
9508 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9509 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9510 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9511 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9513 bool isFastMultiplier = false;
9515 switch ((unsigned char)Diff) {
9517 case 1: // result = add base, cond
9518 case 2: // result = lea base( , cond*2)
9519 case 3: // result = lea base(cond, cond*2)
9520 case 4: // result = lea base( , cond*4)
9521 case 5: // result = lea base(cond, cond*4)
9522 case 8: // result = lea base( , cond*8)
9523 case 9: // result = lea base(cond, cond*8)
9524 isFastMultiplier = true;
9529 if (isFastMultiplier) {
9530 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9531 if (NeedsCondInvert) // Invert the condition if needed.
9532 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9533 DAG.getConstant(1, Cond.getValueType()));
9535 // Zero extend the condition if needed.
9536 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9538 // Scale the condition by the difference.
9540 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9541 DAG.getConstant(Diff, Cond.getValueType()));
9543 // Add the base if non-zero.
9544 if (FalseC->getAPIntValue() != 0)
9545 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9546 SDValue(FalseC, 0));
9556 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9557 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9558 TargetLowering::DAGCombinerInfo &DCI) {
9559 DebugLoc DL = N->getDebugLoc();
9561 // If the flag operand isn't dead, don't touch this CMOV.
9562 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9565 // If this is a select between two integer constants, try to do some
9566 // optimizations. Note that the operands are ordered the opposite of SELECT
9568 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9569 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9570 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9571 // larger than FalseC (the false value).
9572 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9574 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9575 CC = X86::GetOppositeBranchCondition(CC);
9576 std::swap(TrueC, FalseC);
9579 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9580 // This is efficient for any integer data type (including i8/i16) and
9582 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9583 SDValue Cond = N->getOperand(3);
9584 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9585 DAG.getConstant(CC, MVT::i8), Cond);
9587 // Zero extend the condition if needed.
9588 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9590 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9591 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9592 DAG.getConstant(ShAmt, MVT::i8));
9593 if (N->getNumValues() == 2) // Dead flag value?
9594 return DCI.CombineTo(N, Cond, SDValue());
9598 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9599 // for any integer data type, including i8/i16.
9600 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9601 SDValue Cond = N->getOperand(3);
9602 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9603 DAG.getConstant(CC, MVT::i8), Cond);
9605 // Zero extend the condition if needed.
9606 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9607 FalseC->getValueType(0), Cond);
9608 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9609 SDValue(FalseC, 0));
9611 if (N->getNumValues() == 2) // Dead flag value?
9612 return DCI.CombineTo(N, Cond, SDValue());
9616 // Optimize cases that will turn into an LEA instruction. This requires
9617 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9618 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9619 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9620 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9622 bool isFastMultiplier = false;
9624 switch ((unsigned char)Diff) {
9626 case 1: // result = add base, cond
9627 case 2: // result = lea base( , cond*2)
9628 case 3: // result = lea base(cond, cond*2)
9629 case 4: // result = lea base( , cond*4)
9630 case 5: // result = lea base(cond, cond*4)
9631 case 8: // result = lea base( , cond*8)
9632 case 9: // result = lea base(cond, cond*8)
9633 isFastMultiplier = true;
9638 if (isFastMultiplier) {
9639 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9640 SDValue Cond = N->getOperand(3);
9641 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9642 DAG.getConstant(CC, MVT::i8), Cond);
9643 // Zero extend the condition if needed.
9644 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9646 // Scale the condition by the difference.
9648 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9649 DAG.getConstant(Diff, Cond.getValueType()));
9651 // Add the base if non-zero.
9652 if (FalseC->getAPIntValue() != 0)
9653 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9654 SDValue(FalseC, 0));
9655 if (N->getNumValues() == 2) // Dead flag value?
9656 return DCI.CombineTo(N, Cond, SDValue());
9666 /// PerformMulCombine - Optimize a single multiply with constant into two
9667 /// in order to implement it with two cheaper instructions, e.g.
9668 /// LEA + SHL, LEA + LEA.
9669 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9670 TargetLowering::DAGCombinerInfo &DCI) {
9671 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9674 EVT VT = N->getValueType(0);
9678 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9681 uint64_t MulAmt = C->getZExtValue();
9682 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9685 uint64_t MulAmt1 = 0;
9686 uint64_t MulAmt2 = 0;
9687 if ((MulAmt % 9) == 0) {
9689 MulAmt2 = MulAmt / 9;
9690 } else if ((MulAmt % 5) == 0) {
9692 MulAmt2 = MulAmt / 5;
9693 } else if ((MulAmt % 3) == 0) {
9695 MulAmt2 = MulAmt / 3;
9698 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9699 DebugLoc DL = N->getDebugLoc();
9701 if (isPowerOf2_64(MulAmt2) &&
9702 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9703 // If second multiplifer is pow2, issue it first. We want the multiply by
9704 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9706 std::swap(MulAmt1, MulAmt2);
9709 if (isPowerOf2_64(MulAmt1))
9710 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9711 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9713 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9714 DAG.getConstant(MulAmt1, VT));
9716 if (isPowerOf2_64(MulAmt2))
9717 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9718 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9720 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9721 DAG.getConstant(MulAmt2, VT));
9723 // Do not add new nodes to DAG combiner worklist.
9724 DCI.CombineTo(N, NewMul, false);
9729 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9730 SDValue N0 = N->getOperand(0);
9731 SDValue N1 = N->getOperand(1);
9732 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9733 EVT VT = N0.getValueType();
9735 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9736 // since the result of setcc_c is all zero's or all ones.
9737 if (N1C && N0.getOpcode() == ISD::AND &&
9738 N0.getOperand(1).getOpcode() == ISD::Constant) {
9739 SDValue N00 = N0.getOperand(0);
9740 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9741 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9742 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9743 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9744 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9745 APInt ShAmt = N1C->getAPIntValue();
9746 Mask = Mask.shl(ShAmt);
9748 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9749 N00, DAG.getConstant(Mask, VT));
9756 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9758 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9759 const X86Subtarget *Subtarget) {
9760 EVT VT = N->getValueType(0);
9761 if (!VT.isVector() && VT.isInteger() &&
9762 N->getOpcode() == ISD::SHL)
9763 return PerformSHLCombine(N, DAG);
9765 // On X86 with SSE2 support, we can transform this to a vector shift if
9766 // all elements are shifted by the same amount. We can't do this in legalize
9767 // because the a constant vector is typically transformed to a constant pool
9768 // so we have no knowledge of the shift amount.
9769 if (!Subtarget->hasSSE2())
9772 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9775 SDValue ShAmtOp = N->getOperand(1);
9776 EVT EltVT = VT.getVectorElementType();
9777 DebugLoc DL = N->getDebugLoc();
9778 SDValue BaseShAmt = SDValue();
9779 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9780 unsigned NumElts = VT.getVectorNumElements();
9782 for (; i != NumElts; ++i) {
9783 SDValue Arg = ShAmtOp.getOperand(i);
9784 if (Arg.getOpcode() == ISD::UNDEF) continue;
9788 for (; i != NumElts; ++i) {
9789 SDValue Arg = ShAmtOp.getOperand(i);
9790 if (Arg.getOpcode() == ISD::UNDEF) continue;
9791 if (Arg != BaseShAmt) {
9795 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9796 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9797 SDValue InVec = ShAmtOp.getOperand(0);
9798 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9799 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9801 for (; i != NumElts; ++i) {
9802 SDValue Arg = InVec.getOperand(i);
9803 if (Arg.getOpcode() == ISD::UNDEF) continue;
9807 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9808 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9809 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9810 if (C->getZExtValue() == SplatIdx)
9811 BaseShAmt = InVec.getOperand(1);
9814 if (BaseShAmt.getNode() == 0)
9815 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9816 DAG.getIntPtrConstant(0));
9820 // The shift amount is an i32.
9821 if (EltVT.bitsGT(MVT::i32))
9822 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9823 else if (EltVT.bitsLT(MVT::i32))
9824 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9826 // The shift amount is identical so we can do a vector shift.
9827 SDValue ValOp = N->getOperand(0);
9828 switch (N->getOpcode()) {
9830 llvm_unreachable("Unknown shift opcode!");
9833 if (VT == MVT::v2i64)
9834 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9835 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9837 if (VT == MVT::v4i32)
9838 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9839 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9841 if (VT == MVT::v8i16)
9842 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9843 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9847 if (VT == MVT::v4i32)
9848 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9849 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9851 if (VT == MVT::v8i16)
9852 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9853 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9857 if (VT == MVT::v2i64)
9858 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9859 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9861 if (VT == MVT::v4i32)
9862 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9863 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9865 if (VT == MVT::v8i16)
9866 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9867 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9874 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9875 TargetLowering::DAGCombinerInfo &DCI,
9876 const X86Subtarget *Subtarget) {
9877 if (DCI.isBeforeLegalizeOps())
9880 EVT VT = N->getValueType(0);
9881 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
9884 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9885 SDValue N0 = N->getOperand(0);
9886 SDValue N1 = N->getOperand(1);
9887 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9889 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9891 if (!N0.hasOneUse() || !N1.hasOneUse())
9894 SDValue ShAmt0 = N0.getOperand(1);
9895 if (ShAmt0.getValueType() != MVT::i8)
9897 SDValue ShAmt1 = N1.getOperand(1);
9898 if (ShAmt1.getValueType() != MVT::i8)
9900 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9901 ShAmt0 = ShAmt0.getOperand(0);
9902 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9903 ShAmt1 = ShAmt1.getOperand(0);
9905 DebugLoc DL = N->getDebugLoc();
9906 unsigned Opc = X86ISD::SHLD;
9907 SDValue Op0 = N0.getOperand(0);
9908 SDValue Op1 = N1.getOperand(0);
9909 if (ShAmt0.getOpcode() == ISD::SUB) {
9911 std::swap(Op0, Op1);
9912 std::swap(ShAmt0, ShAmt1);
9915 unsigned Bits = VT.getSizeInBits();
9916 if (ShAmt1.getOpcode() == ISD::SUB) {
9917 SDValue Sum = ShAmt1.getOperand(0);
9918 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9919 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
9920 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
9921 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
9922 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
9923 return DAG.getNode(Opc, DL, VT,
9925 DAG.getNode(ISD::TRUNCATE, DL,
9928 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9929 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9931 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
9932 return DAG.getNode(Opc, DL, VT,
9933 N0.getOperand(0), N1.getOperand(0),
9934 DAG.getNode(ISD::TRUNCATE, DL,
9941 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9942 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9943 const X86Subtarget *Subtarget) {
9944 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9945 // the FP state in cases where an emms may be missing.
9946 // A preferable solution to the general problem is to figure out the right
9947 // places to insert EMMS. This qualifies as a quick hack.
9949 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9950 StoreSDNode *St = cast<StoreSDNode>(N);
9951 EVT VT = St->getValue().getValueType();
9952 if (VT.getSizeInBits() != 64)
9955 const Function *F = DAG.getMachineFunction().getFunction();
9956 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9957 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9958 && Subtarget->hasSSE2();
9959 if ((VT.isVector() ||
9960 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9961 isa<LoadSDNode>(St->getValue()) &&
9962 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9963 St->getChain().hasOneUse() && !St->isVolatile()) {
9964 SDNode* LdVal = St->getValue().getNode();
9966 int TokenFactorIndex = -1;
9967 SmallVector<SDValue, 8> Ops;
9968 SDNode* ChainVal = St->getChain().getNode();
9969 // Must be a store of a load. We currently handle two cases: the load
9970 // is a direct child, and it's under an intervening TokenFactor. It is
9971 // possible to dig deeper under nested TokenFactors.
9972 if (ChainVal == LdVal)
9973 Ld = cast<LoadSDNode>(St->getChain());
9974 else if (St->getValue().hasOneUse() &&
9975 ChainVal->getOpcode() == ISD::TokenFactor) {
9976 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9977 if (ChainVal->getOperand(i).getNode() == LdVal) {
9978 TokenFactorIndex = i;
9979 Ld = cast<LoadSDNode>(St->getValue());
9981 Ops.push_back(ChainVal->getOperand(i));
9985 if (!Ld || !ISD::isNormalLoad(Ld))
9988 // If this is not the MMX case, i.e. we are just turning i64 load/store
9989 // into f64 load/store, avoid the transformation if there are multiple
9990 // uses of the loaded value.
9991 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9994 DebugLoc LdDL = Ld->getDebugLoc();
9995 DebugLoc StDL = N->getDebugLoc();
9996 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9997 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9999 if (Subtarget->is64Bit() || F64IsLegal) {
10000 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
10001 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
10002 Ld->getBasePtr(), Ld->getSrcValue(),
10003 Ld->getSrcValueOffset(), Ld->isVolatile(),
10004 Ld->isNonTemporal(), Ld->getAlignment());
10005 SDValue NewChain = NewLd.getValue(1);
10006 if (TokenFactorIndex != -1) {
10007 Ops.push_back(NewChain);
10008 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
10011 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
10012 St->getSrcValue(), St->getSrcValueOffset(),
10013 St->isVolatile(), St->isNonTemporal(),
10014 St->getAlignment());
10017 // Otherwise, lower to two pairs of 32-bit loads / stores.
10018 SDValue LoAddr = Ld->getBasePtr();
10019 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
10020 DAG.getConstant(4, MVT::i32));
10022 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
10023 Ld->getSrcValue(), Ld->getSrcValueOffset(),
10024 Ld->isVolatile(), Ld->isNonTemporal(),
10025 Ld->getAlignment());
10026 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
10027 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
10028 Ld->isVolatile(), Ld->isNonTemporal(),
10029 MinAlign(Ld->getAlignment(), 4));
10031 SDValue NewChain = LoLd.getValue(1);
10032 if (TokenFactorIndex != -1) {
10033 Ops.push_back(LoLd);
10034 Ops.push_back(HiLd);
10035 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
10039 LoAddr = St->getBasePtr();
10040 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
10041 DAG.getConstant(4, MVT::i32));
10043 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
10044 St->getSrcValue(), St->getSrcValueOffset(),
10045 St->isVolatile(), St->isNonTemporal(),
10046 St->getAlignment());
10047 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
10049 St->getSrcValueOffset() + 4,
10051 St->isNonTemporal(),
10052 MinAlign(St->getAlignment(), 4));
10053 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
10058 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
10059 /// X86ISD::FXOR nodes.
10060 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
10061 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
10062 // F[X]OR(0.0, x) -> x
10063 // F[X]OR(x, 0.0) -> x
10064 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
10065 if (C->getValueAPF().isPosZero())
10066 return N->getOperand(1);
10067 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
10068 if (C->getValueAPF().isPosZero())
10069 return N->getOperand(0);
10073 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
10074 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
10075 // FAND(0.0, x) -> 0.0
10076 // FAND(x, 0.0) -> 0.0
10077 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
10078 if (C->getValueAPF().isPosZero())
10079 return N->getOperand(0);
10080 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
10081 if (C->getValueAPF().isPosZero())
10082 return N->getOperand(1);
10086 static SDValue PerformBTCombine(SDNode *N,
10088 TargetLowering::DAGCombinerInfo &DCI) {
10089 // BT ignores high bits in the bit index operand.
10090 SDValue Op1 = N->getOperand(1);
10091 if (Op1.hasOneUse()) {
10092 unsigned BitWidth = Op1.getValueSizeInBits();
10093 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
10094 APInt KnownZero, KnownOne;
10095 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
10096 !DCI.isBeforeLegalizeOps());
10097 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10098 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
10099 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
10100 DCI.CommitTargetLoweringOpt(TLO);
10105 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
10106 SDValue Op = N->getOperand(0);
10107 if (Op.getOpcode() == ISD::BIT_CONVERT)
10108 Op = Op.getOperand(0);
10109 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
10110 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
10111 VT.getVectorElementType().getSizeInBits() ==
10112 OpVT.getVectorElementType().getSizeInBits()) {
10113 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
10118 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
10119 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
10120 // (and (i32 x86isd::setcc_carry), 1)
10121 // This eliminates the zext. This transformation is necessary because
10122 // ISD::SETCC is always legalized to i8.
10123 DebugLoc dl = N->getDebugLoc();
10124 SDValue N0 = N->getOperand(0);
10125 EVT VT = N->getValueType(0);
10126 if (N0.getOpcode() == ISD::AND &&
10128 N0.getOperand(0).hasOneUse()) {
10129 SDValue N00 = N0.getOperand(0);
10130 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
10132 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
10133 if (!C || C->getZExtValue() != 1)
10135 return DAG.getNode(ISD::AND, dl, VT,
10136 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
10137 N00.getOperand(0), N00.getOperand(1)),
10138 DAG.getConstant(1, VT));
10144 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
10145 DAGCombinerInfo &DCI) const {
10146 SelectionDAG &DAG = DCI.DAG;
10147 switch (N->getOpcode()) {
10149 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
10150 case ISD::EXTRACT_VECTOR_ELT:
10151 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
10152 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
10153 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
10154 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
10157 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
10158 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
10159 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
10161 case X86ISD::FOR: return PerformFORCombine(N, DAG);
10162 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
10163 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
10164 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
10165 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
10171 /// isTypeDesirableForOp - Return true if the target has native support for
10172 /// the specified value type and it is 'desirable' to use the type for the
10173 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
10174 /// instruction encodings are longer and some i16 instructions are slow.
10175 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
10176 if (!isTypeLegal(VT))
10178 if (VT != MVT::i16)
10185 case ISD::SIGN_EXTEND:
10186 case ISD::ZERO_EXTEND:
10187 case ISD::ANY_EXTEND:
10200 static bool MayFoldLoad(SDValue Op) {
10201 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
10204 static bool MayFoldIntoStore(SDValue Op) {
10205 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
10208 /// IsDesirableToPromoteOp - This method query the target whether it is
10209 /// beneficial for dag combiner to promote the specified node. If true, it
10210 /// should return the desired promotion type by reference.
10211 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
10212 EVT VT = Op.getValueType();
10213 if (VT != MVT::i16)
10216 bool Promote = false;
10217 bool Commute = false;
10218 switch (Op.getOpcode()) {
10221 LoadSDNode *LD = cast<LoadSDNode>(Op);
10222 // If the non-extending load has a single use and it's not live out, then it
10223 // might be folded.
10224 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
10225 Op.hasOneUse()*/) {
10226 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
10227 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
10228 // The only case where we'd want to promote LOAD (rather then it being
10229 // promoted as an operand is when it's only use is liveout.
10230 if (UI->getOpcode() != ISD::CopyToReg)
10237 case ISD::SIGN_EXTEND:
10238 case ISD::ZERO_EXTEND:
10239 case ISD::ANY_EXTEND:
10244 SDValue N0 = Op.getOperand(0);
10245 // Look out for (store (shl (load), x)).
10246 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
10259 SDValue N0 = Op.getOperand(0);
10260 SDValue N1 = Op.getOperand(1);
10261 if (!Commute && MayFoldLoad(N1))
10263 // Avoid disabling potential load folding opportunities.
10264 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
10266 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
10276 //===----------------------------------------------------------------------===//
10277 // X86 Inline Assembly Support
10278 //===----------------------------------------------------------------------===//
10280 static bool LowerToBSwap(CallInst *CI) {
10281 // FIXME: this should verify that we are targetting a 486 or better. If not,
10282 // we will turn this bswap into something that will be lowered to logical ops
10283 // instead of emitting the bswap asm. For now, we don't support 486 or lower
10284 // so don't worry about this.
10286 // Verify this is a simple bswap.
10287 if (CI->getNumArgOperands() != 1 ||
10288 CI->getType() != CI->getArgOperand(0)->getType() ||
10289 !CI->getType()->isIntegerTy())
10292 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
10293 if (!Ty || Ty->getBitWidth() % 16 != 0)
10296 // Okay, we can do this xform, do so now.
10297 const Type *Tys[] = { Ty };
10298 Module *M = CI->getParent()->getParent()->getParent();
10299 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
10301 Value *Op = CI->getArgOperand(0);
10302 Op = CallInst::Create(Int, Op, CI->getName(), CI);
10304 CI->replaceAllUsesWith(Op);
10305 CI->eraseFromParent();
10309 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
10310 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
10311 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
10313 std::string AsmStr = IA->getAsmString();
10315 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
10316 SmallVector<StringRef, 4> AsmPieces;
10317 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
10319 switch (AsmPieces.size()) {
10320 default: return false;
10322 AsmStr = AsmPieces[0];
10324 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
10327 if (AsmPieces.size() == 2 &&
10328 (AsmPieces[0] == "bswap" ||
10329 AsmPieces[0] == "bswapq" ||
10330 AsmPieces[0] == "bswapl") &&
10331 (AsmPieces[1] == "$0" ||
10332 AsmPieces[1] == "${0:q}")) {
10333 // No need to check constraints, nothing other than the equivalent of
10334 // "=r,0" would be valid here.
10335 return LowerToBSwap(CI);
10337 // rorw $$8, ${0:w} --> llvm.bswap.i16
10338 if (CI->getType()->isIntegerTy(16) &&
10339 AsmPieces.size() == 3 &&
10340 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
10341 AsmPieces[1] == "$$8," &&
10342 AsmPieces[2] == "${0:w}" &&
10343 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
10345 const std::string &Constraints = IA->getConstraintString();
10346 SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
10347 std::sort(AsmPieces.begin(), AsmPieces.end());
10348 if (AsmPieces.size() == 4 &&
10349 AsmPieces[0] == "~{cc}" &&
10350 AsmPieces[1] == "~{dirflag}" &&
10351 AsmPieces[2] == "~{flags}" &&
10352 AsmPieces[3] == "~{fpsr}") {
10353 return LowerToBSwap(CI);
10358 if (CI->getType()->isIntegerTy(64) &&
10359 Constraints.size() >= 2 &&
10360 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
10361 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
10362 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
10363 SmallVector<StringRef, 4> Words;
10364 SplitString(AsmPieces[0], Words, " \t");
10365 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
10367 SplitString(AsmPieces[1], Words, " \t");
10368 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
10370 SplitString(AsmPieces[2], Words, " \t,");
10371 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
10372 Words[2] == "%edx") {
10373 return LowerToBSwap(CI);
10385 /// getConstraintType - Given a constraint letter, return the type of
10386 /// constraint it is for this target.
10387 X86TargetLowering::ConstraintType
10388 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
10389 if (Constraint.size() == 1) {
10390 switch (Constraint[0]) {
10402 return C_RegisterClass;
10410 return TargetLowering::getConstraintType(Constraint);
10413 /// LowerXConstraint - try to replace an X constraint, which matches anything,
10414 /// with another that has more specific requirements based on the type of the
10415 /// corresponding operand.
10416 const char *X86TargetLowering::
10417 LowerXConstraint(EVT ConstraintVT) const {
10418 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
10419 // 'f' like normal targets.
10420 if (ConstraintVT.isFloatingPoint()) {
10421 if (Subtarget->hasSSE2())
10423 if (Subtarget->hasSSE1())
10427 return TargetLowering::LowerXConstraint(ConstraintVT);
10430 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
10431 /// vector. If it is invalid, don't add anything to Ops.
10432 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
10434 std::vector<SDValue>&Ops,
10435 SelectionDAG &DAG) const {
10436 SDValue Result(0, 0);
10438 switch (Constraint) {
10441 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10442 if (C->getZExtValue() <= 31) {
10443 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10449 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10450 if (C->getZExtValue() <= 63) {
10451 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10457 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10458 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
10459 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10465 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10466 if (C->getZExtValue() <= 255) {
10467 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10473 // 32-bit signed value
10474 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10475 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10476 C->getSExtValue())) {
10477 // Widen to 64 bits here to get it sign extended.
10478 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
10481 // FIXME gcc accepts some relocatable values here too, but only in certain
10482 // memory models; it's complicated.
10487 // 32-bit unsigned value
10488 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
10489 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
10490 C->getZExtValue())) {
10491 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
10495 // FIXME gcc accepts some relocatable values here too, but only in certain
10496 // memory models; it's complicated.
10500 // Literal immediates are always ok.
10501 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
10502 // Widen to 64 bits here to get it sign extended.
10503 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
10507 // In any sort of PIC mode addresses need to be computed at runtime by
10508 // adding in a register or some sort of table lookup. These can't
10509 // be used as immediates.
10510 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
10513 // If we are in non-pic codegen mode, we allow the address of a global (with
10514 // an optional displacement) to be used with 'i'.
10515 GlobalAddressSDNode *GA = 0;
10516 int64_t Offset = 0;
10518 // Match either (GA), (GA+C), (GA+C1+C2), etc.
10520 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
10521 Offset += GA->getOffset();
10523 } else if (Op.getOpcode() == ISD::ADD) {
10524 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10525 Offset += C->getZExtValue();
10526 Op = Op.getOperand(0);
10529 } else if (Op.getOpcode() == ISD::SUB) {
10530 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
10531 Offset += -C->getZExtValue();
10532 Op = Op.getOperand(0);
10537 // Otherwise, this isn't something we can handle, reject it.
10541 const GlobalValue *GV = GA->getGlobal();
10542 // If we require an extra load to get this address, as in PIC mode, we
10543 // can't accept it.
10544 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
10545 getTargetMachine())))
10548 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
10549 GA->getValueType(0), Offset);
10554 if (Result.getNode()) {
10555 Ops.push_back(Result);
10558 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
10561 std::vector<unsigned> X86TargetLowering::
10562 getRegClassForInlineAsmConstraint(const std::string &Constraint,
10564 if (Constraint.size() == 1) {
10565 // FIXME: not handling fp-stack yet!
10566 switch (Constraint[0]) { // GCC X86 Constraint Letters
10567 default: break; // Unknown constraint letter
10568 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10569 if (Subtarget->is64Bit()) {
10570 if (VT == MVT::i32)
10571 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10572 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10573 X86::R10D,X86::R11D,X86::R12D,
10574 X86::R13D,X86::R14D,X86::R15D,
10575 X86::EBP, X86::ESP, 0);
10576 else if (VT == MVT::i16)
10577 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10578 X86::SI, X86::DI, X86::R8W,X86::R9W,
10579 X86::R10W,X86::R11W,X86::R12W,
10580 X86::R13W,X86::R14W,X86::R15W,
10581 X86::BP, X86::SP, 0);
10582 else if (VT == MVT::i8)
10583 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10584 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10585 X86::R10B,X86::R11B,X86::R12B,
10586 X86::R13B,X86::R14B,X86::R15B,
10587 X86::BPL, X86::SPL, 0);
10589 else if (VT == MVT::i64)
10590 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10591 X86::RSI, X86::RDI, X86::R8, X86::R9,
10592 X86::R10, X86::R11, X86::R12,
10593 X86::R13, X86::R14, X86::R15,
10594 X86::RBP, X86::RSP, 0);
10598 // 32-bit fallthrough
10599 case 'Q': // Q_REGS
10600 if (VT == MVT::i32)
10601 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10602 else if (VT == MVT::i16)
10603 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10604 else if (VT == MVT::i8)
10605 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10606 else if (VT == MVT::i64)
10607 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10612 return std::vector<unsigned>();
10615 std::pair<unsigned, const TargetRegisterClass*>
10616 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10618 // First, see if this is a constraint that directly corresponds to an LLVM
10620 if (Constraint.size() == 1) {
10621 // GCC Constraint Letters
10622 switch (Constraint[0]) {
10624 case 'r': // GENERAL_REGS
10625 case 'l': // INDEX_REGS
10627 return std::make_pair(0U, X86::GR8RegisterClass);
10628 if (VT == MVT::i16)
10629 return std::make_pair(0U, X86::GR16RegisterClass);
10630 if (VT == MVT::i32 || !Subtarget->is64Bit())
10631 return std::make_pair(0U, X86::GR32RegisterClass);
10632 return std::make_pair(0U, X86::GR64RegisterClass);
10633 case 'R': // LEGACY_REGS
10635 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10636 if (VT == MVT::i16)
10637 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10638 if (VT == MVT::i32 || !Subtarget->is64Bit())
10639 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10640 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10641 case 'f': // FP Stack registers.
10642 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10643 // value to the correct fpstack register class.
10644 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10645 return std::make_pair(0U, X86::RFP32RegisterClass);
10646 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10647 return std::make_pair(0U, X86::RFP64RegisterClass);
10648 return std::make_pair(0U, X86::RFP80RegisterClass);
10649 case 'y': // MMX_REGS if MMX allowed.
10650 if (!Subtarget->hasMMX()) break;
10651 return std::make_pair(0U, X86::VR64RegisterClass);
10652 case 'Y': // SSE_REGS if SSE2 allowed
10653 if (!Subtarget->hasSSE2()) break;
10655 case 'x': // SSE_REGS if SSE1 allowed
10656 if (!Subtarget->hasSSE1()) break;
10658 switch (VT.getSimpleVT().SimpleTy) {
10660 // Scalar SSE types.
10663 return std::make_pair(0U, X86::FR32RegisterClass);
10666 return std::make_pair(0U, X86::FR64RegisterClass);
10674 return std::make_pair(0U, X86::VR128RegisterClass);
10680 // Use the default implementation in TargetLowering to convert the register
10681 // constraint into a member of a register class.
10682 std::pair<unsigned, const TargetRegisterClass*> Res;
10683 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10685 // Not found as a standard register?
10686 if (Res.second == 0) {
10687 // Map st(0) -> st(7) -> ST0
10688 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10689 tolower(Constraint[1]) == 's' &&
10690 tolower(Constraint[2]) == 't' &&
10691 Constraint[3] == '(' &&
10692 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10693 Constraint[5] == ')' &&
10694 Constraint[6] == '}') {
10696 Res.first = X86::ST0+Constraint[4]-'0';
10697 Res.second = X86::RFP80RegisterClass;
10701 // GCC allows "st(0)" to be called just plain "st".
10702 if (StringRef("{st}").equals_lower(Constraint)) {
10703 Res.first = X86::ST0;
10704 Res.second = X86::RFP80RegisterClass;
10709 if (StringRef("{flags}").equals_lower(Constraint)) {
10710 Res.first = X86::EFLAGS;
10711 Res.second = X86::CCRRegisterClass;
10715 // 'A' means EAX + EDX.
10716 if (Constraint == "A") {
10717 Res.first = X86::EAX;
10718 Res.second = X86::GR32_ADRegisterClass;
10724 // Otherwise, check to see if this is a register class of the wrong value
10725 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10726 // turn into {ax},{dx}.
10727 if (Res.second->hasType(VT))
10728 return Res; // Correct type already, nothing to do.
10730 // All of the single-register GCC register classes map their values onto
10731 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10732 // really want an 8-bit or 32-bit register, map to the appropriate register
10733 // class and return the appropriate register.
10734 if (Res.second == X86::GR16RegisterClass) {
10735 if (VT == MVT::i8) {
10736 unsigned DestReg = 0;
10737 switch (Res.first) {
10739 case X86::AX: DestReg = X86::AL; break;
10740 case X86::DX: DestReg = X86::DL; break;
10741 case X86::CX: DestReg = X86::CL; break;
10742 case X86::BX: DestReg = X86::BL; break;
10745 Res.first = DestReg;
10746 Res.second = X86::GR8RegisterClass;
10748 } else if (VT == MVT::i32) {
10749 unsigned DestReg = 0;
10750 switch (Res.first) {
10752 case X86::AX: DestReg = X86::EAX; break;
10753 case X86::DX: DestReg = X86::EDX; break;
10754 case X86::CX: DestReg = X86::ECX; break;
10755 case X86::BX: DestReg = X86::EBX; break;
10756 case X86::SI: DestReg = X86::ESI; break;
10757 case X86::DI: DestReg = X86::EDI; break;
10758 case X86::BP: DestReg = X86::EBP; break;
10759 case X86::SP: DestReg = X86::ESP; break;
10762 Res.first = DestReg;
10763 Res.second = X86::GR32RegisterClass;
10765 } else if (VT == MVT::i64) {
10766 unsigned DestReg = 0;
10767 switch (Res.first) {
10769 case X86::AX: DestReg = X86::RAX; break;
10770 case X86::DX: DestReg = X86::RDX; break;
10771 case X86::CX: DestReg = X86::RCX; break;
10772 case X86::BX: DestReg = X86::RBX; break;
10773 case X86::SI: DestReg = X86::RSI; break;
10774 case X86::DI: DestReg = X86::RDI; break;
10775 case X86::BP: DestReg = X86::RBP; break;
10776 case X86::SP: DestReg = X86::RSP; break;
10779 Res.first = DestReg;
10780 Res.second = X86::GR64RegisterClass;
10783 } else if (Res.second == X86::FR32RegisterClass ||
10784 Res.second == X86::FR64RegisterClass ||
10785 Res.second == X86::VR128RegisterClass) {
10786 // Handle references to XMM physical registers that got mapped into the
10787 // wrong class. This can happen with constraints like {xmm0} where the
10788 // target independent register mapper will just pick the first match it can
10789 // find, ignoring the required type.
10790 if (VT == MVT::f32)
10791 Res.second = X86::FR32RegisterClass;
10792 else if (VT == MVT::f64)
10793 Res.second = X86::FR64RegisterClass;
10794 else if (X86::VR128RegisterClass->hasType(VT))
10795 Res.second = X86::VR128RegisterClass;